xref: /illumos-gate/usr/src/uts/common/io/mac/mac_sched.c (revision bde334a8dbd66dfa70ce4d7fc9dcad6e1ae45fe4)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  * Copyright 2017 Joyent, Inc.
25  * Copyright 2013 Nexenta Systems, Inc. All rights reserved.
26  */
27 
28 /*
29  * MAC data path
30  *
31  * The MAC data path is concerned with the flow of traffic from mac clients --
32  * DLS, IP, etc. -- to various GLDv3 device drivers -- e1000g, vnic, aggr,
33  * ixgbe, etc. -- and from the GLDv3 device drivers back to clients.
34  *
35  * -----------
36  * Terminology
37  * -----------
38  *
39  * MAC uses a lot of different, but related terms that are associated with the
40  * design and structure of the data path. Before we cover other aspects, first
41  * let's review the terminology that MAC uses.
42  *
43  * MAC
44  *
45  *	This driver. It interfaces with device drivers and provides abstractions
46  *	that the rest of the system consumes. All data links -- things managed
47  *	with dladm(1M), are accessed through MAC.
48  *
49  * GLDv3 DEVICE DRIVER
50  *
51  *	A GLDv3 device driver refers to a driver, both for pseudo-devices and
52  *	real devices, which implement the GLDv3 driver API. Common examples of
53  *	these are igb and ixgbe, which are drivers for various Intel networking
54  *	cards. These devices may or may not have various features, such as
55  *	hardware rings and checksum offloading. For MAC, a GLDv3 device is the
56  *	final point for the transmission of a packet and the starting point for
57  *	the receipt of a packet.
58  *
59  * FLOWS
60  *
61  *	At a high level, a flow refers to a series of packets that are related.
62  *	Often times the term is used in the context of TCP to indicate a unique
63  *	TCP connection and the traffic over it. However, a flow can exist at
64  *	other levels of the system as well. MAC has a notion of a default flow
65  *	which is used for all unicast traffic addressed to the address of a MAC
66  *	device. For example, when a VNIC is created, a default flow is created
67  *	for the VNIC's MAC address. In addition, flows are created for broadcast
68  *	groups and a user may create a flow with flowadm(1M).
69  *
70  * CLASSIFICATION
71  *
72  *	Classification refers to the notion of identifying an incoming frame
73  *	based on its destination address and optionally its source addresses and
74  *	doing different processing based on that information. Classification can
75  *	be done in both hardware and software. In general, we usually only
76  *	classify based on the layer two destination, eg. for Ethernet, the
77  *	destination MAC address.
78  *
79  *	The system also will do classification based on layer three and layer
80  *	four properties. This is used to support things like flowadm(1M), which
81  *	allows setting QoS and other properties on a per-flow basis.
82  *
83  * RING
84  *
85  *	Conceptually, a ring represents a series of framed messages, often in a
86  *	contiguous chunk of memory that acts as a circular buffer. Rings come in
87  *	a couple of forms. Generally they are either a hardware construct (hw
88  *	ring) or they are a software construct (sw ring) maintained by MAC.
89  *
90  * HW RING
91  *
92  *	A hardware ring is a set of resources provided by a GLDv3 device driver
93  *	(even if it is a pseudo-device). A hardware ring comes in two different
94  *	forms: receive (rx) rings and transmit (tx) rings. An rx hw ring is
95  *	something that has a unique DMA (direct memory access) region and
96  *	generally supports some form of classification (though it isn't always
97  *	used), as well as a means of generating an interrupt specific to that
98  *	ring. For example, the device may generate a specific MSI-X for a PCI
99  *	express device. A tx ring is similar, except that it is dedicated to
100  *	transmission. It may also be a vector for enabling features such as VLAN
101  *	tagging and large transmit offloading. It usually has its own dedicated
102  *	interrupts for transmit being completed.
103  *
104  * SW RING
105  *
106  *	A software ring is a construction of MAC. It represents the same thing
107  *	that a hardware ring generally does, a collection of frames. However,
108  *	instead of being in a contiguous ring of memory, they're instead linked
109  *	by using the mblk_t's b_next pointer. Each frame may itself be multiple
110  *	mblk_t's linked together by the b_cont pointer. A software ring always
111  *	represents a collection of classified packets; however, it varies as to
112  *	whether it uses only layer two information, or a combination of that and
113  *	additional layer three and layer four data.
114  *
115  * FANOUT
116  *
117  *	Fanout is the idea of spreading out the load of processing frames based
118  *	on the source and destination information contained in the layer two,
119  *	three, and four headers, such that the data can then be processed in
120  *	parallel using multiple hardware threads.
121  *
122  *	A fanout algorithm hashes the headers and uses that to place different
123  *	flows into a bucket. The most important thing is that packets that are
124  *	in the same flow end up in the same bucket. If they do not, performance
125  *	can be adversely affected. Consider the case of TCP.  TCP severely
126  *	penalizes a connection if the data arrives out of order. If a given flow
127  *	is processed on different CPUs, then the data will appear out of order,
128  *	hence the invariant that fanout always hash a given flow to the same
129  *	bucket and thus get processed on the same CPU.
130  *
131  * RECEIVE SIDE SCALING (RSS)
132  *
133  *
134  *	Receive side scaling is a term that isn't common in illumos, but is used
135  *	by vendors and was popularized by Microsoft. It refers to the idea of
136  *	spreading the incoming receive load out across multiple interrupts which
137  *	can be directed to different CPUs. This allows a device to leverage
138  *	hardware rings even when it doesn't support hardware classification. The
139  *	hardware uses an algorithm to perform fanout that ensures the flow
140  *	invariant is maintained.
141  *
142  * SOFT RING SET
143  *
144  *	A soft ring set, commonly abbreviated SRS, is a collection of rings and
145  *	is used for both transmitting and receiving. It is maintained in the
146  *	structure mac_soft_ring_set_t. A soft ring set is usually associated
147  *	with flows, and coordinates both the use of hardware and software rings.
148  *	Because the use of hardware rings can change as devices such as VNICs
149  *	come and go, we always ensure that the set has software classification
150  *	rules that correspond to the hardware classification rules from rings.
151  *
152  *	Soft ring sets are also used for the enforcement of various QoS
153  *	properties. For example, if a bandwidth limit has been placed on a
154  *	specific flow or device, then that will be enforced by the soft ring
155  *	set.
156  *
157  * SERVICE ATTACHMENT POINT (SAP)
158  *
159  *	The service attachment point is a DLPI (Data Link Provider Interface)
160  *	concept; however, it comes up quite often in MAC. Most MAC devices speak
161  *	a protocol that has some notion of different channels or message type
162  *	identifiers. For example, Ethernet defines an EtherType which is a part
163  *	of the Ethernet header and defines the particular protocol of the data
164  *	payload. If the EtherType is set to 0x0800, then it defines that the
165  *	contents of that Ethernet frame is IPv4 traffic. For Ethernet, the
166  *	EtherType is the SAP.
167  *
168  *	In DLPI, a given consumer attaches to a specific SAP. In illumos, the ip
169  *	and arp drivers attach to the EtherTypes for IPv4, IPv6, and ARP. Using
170  *	libdlpi(3LIB) user software can attach to arbitrary SAPs. With the
171  *	exception of 802.1Q VLAN tagged traffic, MAC itself does not directly
172  *	consume the SAP; however, it uses that information as part of hashing
173  *	and it may be used as part of the construction of flows.
174  *
175  * PRIMARY MAC CLIENT
176  *
177  *	The primary mac client refers to a mac client whose unicast address
178  *	matches the address of the device itself. For example, if the system has
179  *	instance of the e1000g driver such as e1000g0, e1000g1, etc., the
180  *	primary mac client is the one named after the device itself. VNICs that
181  *	are created on top of such devices are not the primary client.
182  *
183  * TRANSMIT DESCRIPTORS
184  *
185  *	Transmit descriptors are a resource that most GLDv3 device drivers have.
186  *	Generally, a GLDv3 device driver takes a frame that's meant to be output
187  *	and puts a copy of it into a region of memory. Each region of memory
188  *	usually has an associated descriptor that the device uses to manage
189  *	properties of the frames. Devices have a limited number of such
190  *	descriptors. They get reclaimed once the device finishes putting the
191  *	frame on the wire.
192  *
193  *	If the driver runs out of transmit descriptors, for example, the OS is
194  *	generating more frames than it can put on the wire, then it will return
195  *	them back to the MAC layer.
196  *
197  * ---------------------------------
198  * Rings, Classification, and Fanout
199  * ---------------------------------
200  *
201  * The heart of MAC is made up of rings, and not those that Elven-kings wear.
202  * When receiving a packet, MAC breaks the work into two different, though
203  * interrelated phases. The first phase is generally classification and then the
204  * second phase is generally fanout. When a frame comes in from a GLDv3 Device,
205  * MAC needs to determine where that frame should be delivered. If it's a
206  * unicast frame (say a normal TCP/IP packet), then it will be delivered to a
207  * single MAC client; however, if it's a broadcast or multicast frame, then MAC
208  * may need to deliver it to multiple MAC clients.
209  *
210  * On transmit, classification isn't quite as important, but may still be used.
211  * Unlike with the receive path, the classification is not used to determine
212  * devices that should transmit something, but rather is used for special
213  * properties of a flow, eg. bandwidth limits for a given IP address, device, or
214  * connection.
215  *
216  * MAC employs a software classifier and leverages hardware classification as
217  * well. The software classifier can leverage the full layer two information,
218  * source, destination, VLAN, and SAP. If the SAP indicates that IP traffic is
219  * being sent, it can classify based on the IP header, and finally, it also
220  * knows how to classify based on the local and remote ports of TCP, UDP, and
221  * SCTP.
222  *
223  * Hardware classifiers vary in capability. Generally all hardware classifiers
224  * provide the capability to classify based on the destination MAC address. Some
225  * hardware has additional filters built in for performing more in-depth
226  * classification; however, it often has much more limited resources for these
227  * activities as compared to the layer two destination address classification.
228  *
229  * The modus operandi in MAC is to always ensure that we have software-based
230  * capabilities and rules in place and then to supplement that with hardware
231  * resources when available. In general, simple layer two classification is
232  * sufficient and nothing else is used, unless a specific flow is created with
233  * tools such as flowadm(1M) or bandwidth limits are set on a device with
234  * dladm(1M).
235  *
236  * RINGS AND GROUPS
237  *
238  * To get into how rings and classification play together, it's first important
239  * to understand how hardware devices commonly associate rings and allow them to
240  * be programmed. Recall that a hardware ring should be thought of as a DMA
241  * buffer and an interrupt resource. Rings are then collected into groups. A
242  * group itself has a series of classification rules. One or more MAC addresses
243  * are assigned to a group.
244  *
245  * Hardware devices vary in terms of what capabilities they provide. Sometimes
246  * they allow for a dynamic assignment of rings to a group and sometimes they
247  * have a static assignment of rings to a group. For example, the ixgbe driver
248  * has a static assignment of rings to groups such that every group has exactly
249  * one ring and the number of groups is equal to the number of rings.
250  *
251  * Classification and receive side scaling both come into play with how a device
252  * advertises itself to MAC and how MAC uses it. If a device supports layer two
253  * classification of frames, then MAC will assign MAC addresses to a group as a
254  * form of primary classification. If a single MAC address is assigned to a
255  * group, a common case, then MAC will consider packets that come in from rings
256  * on that group to be fully classified and will not need to do any software
257  * classification unless a specific flow has been created.
258  *
259  * If a device supports receive side scaling, then it may advertise or support
260  * groups with multiple rings. In those cases, then receive side scaling will
261  * come into play and MAC will use that as a means of fanning out received
262  * frames across multiple CPUs. This can also be combined with groups that
263  * support layer two classification.
264  *
265  * If a device supports dynamic assignments of rings to groups, then MAC will
266  * change around the way that rings are assigned to various groups as devices
267  * come and go from the system. For example, when a VNIC is created, a new flow
268  * will be created for the VNIC's MAC address. If a hardware ring is available,
269  * MAC may opt to reassign it from one group to another.
270  *
271  * ASSIGNMENT OF HARDWARE RINGS
272  *
273  * This is a bit of a complicated subject that varies depending on the device,
274  * the use of aggregations, the special nature of the primary mac client. This
275  * section deserves being fleshed out.
276  *
277  * FANOUT
278  *
279  * illumos uses fanout to help spread out the incoming processing load of chains
280  * of frames away from a single CPU. If a device supports receive side scaling,
281  * then that provides an initial form of fanout; however, what we're concerned
282  * with all happens after the context of a given set of frames being classified
283  * to a soft ring set.
284  *
285  * After frames reach a soft ring set and account for any potential bandwidth
286  * related accounting, they may be fanned out based on one of the following
287  * three modes:
288  *
289  *     o No Fanout
290  *     o Protocol level fanout
291  *     o Full software ring protocol fanout
292  *
293  * MAC makes the determination as to which of these modes a given soft ring set
294  * obtains based on parameters such as whether or not it's the primary mac
295  * client, whether it's on a 10 GbE or faster device, user controlled dladm(1M)
296  * properties, and the nature of the hardware and the resources that it has.
297  *
298  * When there is no fanout, MAC does not create any soft rings for a device and
299  * the device has frames delivered directly to the MAC client.
300  *
301  * Otherwise, all fanout is performed by software. MAC divides incoming frames
302  * into one of three buckets -- IPv4 TCP traffic, IPv4 UDP traffic, and
303  * everything else. Note, VLAN tagged traffic is considered other, regardless of
304  * the interior EtherType. Regardless of the type of fanout, these three
305  * categories or buckets are always used.
306  *
307  * The difference between protocol level fanout and full software ring protocol
308  * fanout is the number of software rings that end up getting created. The
309  * system always uses the same number of software rings per protocol bucket. So
310  * in the first case when we're just doing protocol level fanout, we just create
311  * one software ring each for IPv4 TCP traffic, IPv4 UDP traffic, and everything
312  * else.
313  *
314  * In the case where we do full software ring protocol fanout, we generally use
315  * mac_compute_soft_ring_count() to determine the number of rings. There are
316  * other combinations of properties and devices that may send us down other
317  * paths, but this is a common starting point. If it's a non-bandwidth enforced
318  * device and we're on at least a 10 GbE link, then we'll use eight soft rings
319  * per protocol bucket as a starting point. See mac_compute_soft_ring_count()
320  * for more information on the total number.
321  *
322  * For each of these rings, we create a mac_soft_ring_t and an associated worker
323  * thread. Particularly when doing full software ring protocol fanout, we bind
324  * each of the worker threads to individual CPUs.
325  *
326  * The other advantage of these software rings is that it allows upper layers to
327  * optionally poll on them. For example, TCP can leverage an squeue to poll on
328  * the software ring, see squeue.c for more information.
329  *
330  * DLS BYPASS
331  *
332  * DLS is the data link services module. It interfaces with DLPI, which is the
333  * primary way that other parts of the system such as IP interface with the MAC
334  * layer. While DLS is traditionally a STREAMS-based interface, it allows for
335  * certain modules such as IP to negotiate various more modern interfaces to be
336  * used, which are useful for higher performance and allow it to use direct
337  * function calls to DLS instead of using STREAMS.
338  *
339  * When we have IPv4 TCP or UDP software rings, then traffic on those rings is
340  * eligible for what we call the dls bypass. In those cases, rather than going
341  * out mac_rx_deliver() to DLS, DLS instead registers them to go directly via
342  * the direct callback registered with DLS, generally ip_input().
343  *
344  * HARDWARE RING POLLING
345  *
346  * GLDv3 devices with hardware rings generally deliver chains of messages
347  * (mblk_t chain) during the context of a single interrupt. However, interrupts
348  * are not the only way that these devices may be used. As part of implementing
349  * ring support, a GLDv3 device driver must have a way to disable the generation
350  * of that interrupt and allow for the operating system to poll on that ring.
351  *
352  * To implement this, every soft ring set has a worker thread and a polling
353  * thread. If a sufficient packet rate comes into the system, MAC will 'blank'
354  * (disable) interrupts on that specific ring and the polling thread will start
355  * consuming packets from the hardware device and deliver them to the soft ring
356  * set, where the worker thread will take over.
357  *
358  * Once the rate of packet intake drops down below a certain threshold, then
359  * polling on the hardware ring will be quiesced and interrupts will be
360  * re-enabled for the given ring. This effectively allows the system to shift
361  * how it handles a ring based on its load. At high packet rates, polling on the
362  * device as opposed to relying on interrupts can actually reduce overall system
363  * load due to the minimization of interrupt activity.
364  *
365  * Note the importance of each ring having its own interrupt source. The whole
366  * idea here is that we do not disable interrupts on the device as a whole, but
367  * rather each ring can be independently toggled.
368  *
369  * USE OF WORKER THREADS
370  *
371  * Both the soft ring set and individual soft rings have a worker thread
372  * associated with them that may be bound to a specific CPU in the system. Any
373  * such assignment will get reassessed as part of dynamic reconfiguration events
374  * in the system such as the onlining and offlining of CPUs and the creation of
375  * CPU partitions.
376  *
377  * In many cases, while in an interrupt, we try to deliver a frame all the way
378  * through the stack in the context of the interrupt itself. However, if the
379  * amount of queued frames has exceeded a threshold, then we instead defer to
380  * the worker thread to do this work and signal it. This is particularly useful
381  * when you have the soft ring set delivering frames into multiple software
382  * rings. If it was only delivering frames into a single software ring then
383  * there'd be no need to have another thread take over. However, if it's
384  * delivering chains of frames to multiple rings, then it's worthwhile to have
385  * the worker for the software ring take over so that the different software
386  * rings can be processed in parallel.
387  *
388  * In a similar fashion to the hardware polling thread, if we don't have a
389  * backlog or there's nothing to do, then the worker thread will go back to
390  * sleep and frames can be delivered all the way from an interrupt. This
391  * behavior is useful as it's designed to minimize latency and the default
392  * disposition of MAC is to optimize for latency.
393  *
394  * MAINTAINING CHAINS
395  *
396  * Another useful idea that MAC uses is to try and maintain frames in chains for
397  * as long as possible. The idea is that all of MAC can handle chains of frames
398  * structured as a series of mblk_t structures linked with the b_next pointer.
399  * When performing software classification and software fanout, MAC does not
400  * simply determine the destination and send the frame along. Instead, in the
401  * case of classification, it tries to maintain a chain for as long as possible
402  * before passing it along and performing additional processing.
403  *
404  * In the case of fanout, MAC first determines what the target software ring is
405  * for every frame in the original chain and constructs a new chain for each
406  * target. MAC then delivers the new chain to each software ring in succession.
407  *
408  * The whole rationale for doing this is that we want to try and maintain the
409  * pipe as much as possible and deliver as many frames through the stack at once
410  * that we can, rather than just pushing a single frame through. This can often
411  * help bring down latency and allows MAC to get a better sense of the overall
412  * activity in the system and properly engage worker threads.
413  *
414  * --------------------
415  * Bandwidth Management
416  * --------------------
417  *
418  * Bandwidth management is something that's built into the soft ring set itself.
419  * When bandwidth limits are placed on a flow, a corresponding soft ring set is
420  * toggled into bandwidth mode. This changes how we transmit and receive the
421  * frames in question.
422  *
423  * Bandwidth management is done on a per-tick basis. We translate the user's
424  * requested bandwidth from a quantity per-second into a quantity per-tick. MAC
425  * cannot process a frame across more than one tick, thus it sets a lower bound
426  * for the bandwidth cap to be a single MTU. This also means that when
427  * hires ticks are enabled (hz is set to 1000), that the minimum amount of
428  * bandwidth is higher, because the number of ticks has increased and MAC has to
429  * go from accepting 100 packets / sec to 1000 / sec.
430  *
431  * The bandwidth counter is reset by either the soft ring set's worker thread or
432  * a thread that is doing an inline transmit or receive if they discover that
433  * the current tick is in the future from the recorded tick.
434  *
435  * Whenever we're receiving or transmitting data, we end up leaving most of the
436  * work to the soft ring set's worker thread. This forces data inserted into the
437  * soft ring set to be effectively serialized and allows us to exhume bandwidth
438  * at a reasonable rate. If there is nothing in the soft ring set at the moment
439  * and the set has available bandwidth, then it may processed inline.
440  * Otherwise, the worker is responsible for taking care of the soft ring set.
441  *
442  * ---------------------
443  * The Receive Data Path
444  * ---------------------
445  *
446  * The following series of ASCII art images breaks apart the way that a frame
447  * comes in and is processed in MAC.
448  *
449  * Part 1 -- Initial frame receipt, SRS classification
450  *
451  * Here, a frame is received by a GLDv3 driver, generally in the context of an
452  * interrupt, and it ends up in mac_rx_common(). A driver calls either mac_rx or
453  * mac_rx_ring, depending on whether or not it supports rings and can identify
454  * the interrupt as having come from a specific ring. Here we determine whether
455  * or not it's fully classified and perform software classification as
456  * appropriate. From here, everything always ends up going to either entry [A]
457  * or entry [B] based on whether or not they have subflow processing needed. We
458  * leave via fanout or delivery.
459  *
460  *           +===========+
461  *           v hardware  v
462  *           v interrupt v
463  *           +===========+
464  *                 |
465  *                 * . . appropriate
466  *                 |     upcall made
467  *                 |     by GLDv3 driver  . . always
468  *                 |                      .
469  *  +--------+     |     +----------+     .    +---------------+
470  *  | GLDv3  |     +---->| mac_rx   |-----*--->| mac_rx_common |
471  *  | Driver |-->--+     +----------+          +---------------+
472  *  +--------+     |        ^                         |
473  *      |          |        ^                         v
474  *      ^          |        * . . always   +----------------------+
475  *      |          |        |              | mac_promisc_dispatch |
476  *      |          |    +-------------+    +----------------------+
477  *      |          +--->| mac_rx_ring |               |
478  *      |               +-------------+               * . . hw classified
479  *      |                                             v     or single flow?
480  *      |                                             |
481  *      |                                   +--------++--------------+
482  *      |                                   |        |               * hw class,
483  *      |                                   |        * hw classified | subflows
484  *      |                 no hw class and . *        | or single     | exist
485  *      |                 subflows          |        | flow          |
486  *      |                                   |        v               v
487  *      |                                   |   +-----------+   +-----------+
488  *      |                                   |   |   goto    |   |  goto     |
489  *      |                                   |   | entry [A] |   | entry [B] |
490  *      |                                   |   +-----------+   +-----------+
491  *      |                                   v          ^
492  *      |                            +-------------+   |
493  *      |                            | mac_rx_flow |   * SRS and flow found,
494  *      |                            +-------------+   | call flow cb
495  *      |                                   |          +------+
496  *      |                                   v                 |
497  *      v                             +==========+    +-----------------+
498  *      |                             v For each v--->| mac_rx_classify |
499  * +----------+                       v  mblk_t  v    +-----------------+
500  * |   srs    |                       +==========+
501  * | pollling |
502  * |  thread  |->------------------------------------------+
503  * +----------+                                            |
504  *                                                         v       . inline
505  *            +--------------------+   +----------+   +---------+  .
506  *    [A]---->| mac_rx_srs_process |-->| check bw |-->| enqueue |--*---------+
507  *            +--------------------+   |  limits  |   | frames  |            |
508  *               ^                     +----------+   | to SRS  |            |
509  *               |                                    +---------+            |
510  *               |  send chain              +--------+    |                  |
511  *               *  when clasified          | signal |    * BW limits,       |
512  *               |  flow changes            |  srs   |<---+ loopback,        |
513  *               |                          | worker |      stack too        |
514  *               |                          +--------+      deep             |
515  *      +-----------------+        +--------+                                |
516  *      | mac_flow_lookup |        |  srs   |     +---------------------+    |
517  *      +-----------------+        | worker |---->| mac_rx_srs_drain    |<---+
518  *               ^                 | thread |     | mac_rx_srs_drain_bw |
519  *               |                 +--------+     +---------------------+
520  *               |                                          |
521  *         +----------------------------+                   * software rings
522  *   [B]-->| mac_rx_srs_subflow_process |                   | for fanout?
523  *         +----------------------------+                   |
524  *                                               +----------+-----------+
525  *                                               |                      |
526  *                                               v                      v
527  *                                          +--------+             +--------+
528  *                                          |  goto  |             |  goto  |
529  *                                          | Part 2 |             | Part 3 |
530  *                                          +--------+             +--------+
531  *
532  * Part 2 -- Fanout
533  *
534  * This part is concerned with using software fanout to assign frames to
535  * software rings and then deliver them to MAC clients or allow those rings to
536  * be polled upon. While there are two different primary fanout entry points,
537  * mac_rx_fanout and mac_rx_proto_fanout, they behave in similar ways, and aside
538  * from some of the individual hashing techniques used, most of the general
539  * flow is the same.
540  *
541  *  +--------+              +-------------------+
542  *  |  From  |---+--------->| mac_rx_srs_fanout |----+
543  *  | Part 1 |   |          +-------------------+    |    +=================+
544  *  +--------+   |                                   |    v for each mblk_t v
545  *               * . . protocol only                 +--->v assign to new   v
546  *               |     fanout                        |    v chain based on  v
547  *               |                                   |    v hash % nrings   v
548  *               |    +-------------------------+    |    +=================+
549  *               +--->| mac_rx_srs_proto_fanout |----+             |
550  *                    +-------------------------+                  |
551  *                                                                 v
552  *    +------------+    +--------------------------+       +================+
553  *    | enqueue in |<---| mac_rx_soft_ring_process |<------v for each chain v
554  *    | soft ring  |    +--------------------------+       +================+
555  *    +------------+
556  *         |                                    +-----------+
557  *         * soft ring set                      | soft ring |
558  *         | empty and no                       |  worker   |
559  *         | worker?                            |  thread   |
560  *         |                                    +-----------+
561  *         +------*----------------+                  |
562  *         |      .                |                  v
563  *    No . *      . Yes            |       +------------------------+
564  *         |                       +----<--| mac_rx_soft_ring_drain |
565  *         |                       |       +------------------------+
566  *         v                       |
567  *   +-----------+                 v
568  *   |   signal  |         +---------------+
569  *   | soft ring |         | Deliver chain |
570  *   |   worker  |         | goto Part 3   |
571  *   +-----------+         +---------------+
572  *
573  *
574  * Part 3 -- Packet Delivery
575  *
576  * Here, we go through and deliver the mblk_t chain directly to a given
577  * processing function. In a lot of cases this is mac_rx_deliver(). In the case
578  * of DLS bypass being used, then instead we end up going ahead and deliver it
579  * to the direct callback registered with DLS, generally ip_input.
580  *
581  *
582  *   +---------+            +----------------+    +------------------+
583  *   |  From   |---+------->| mac_rx_deliver |--->| Off to DLS, or   |
584  *   | Parts 1 |   |        +----------------+    | other MAC client |
585  *   |  and 2  |   * DLS bypass                   +------------------+
586  *   +---------+   | enabled   +----------+    +-------------+
587  *                 +---------->| ip_input |--->|    To IP    |
588  *                             +----------+    | and beyond! |
589  *                                             +-------------+
590  *
591  * ----------------------
592  * The Transmit Data Path
593  * ----------------------
594  *
595  * Before we go into the images, it's worth talking about a problem that is a
596  * bit different from the receive data path. GLDv3 device drivers have a finite
597  * amount of transmit descriptors. When they run out, they return unused frames
598  * back to MAC. MAC, at this point has several options about what it will do,
599  * which vary based upon the settings that the client uses.
600  *
601  * When a device runs out of descriptors, the next thing that MAC does is
602  * enqueue them off of the soft ring set or a software ring, depending on the
603  * configuration of the soft ring set. MAC will enqueue up to a high watermark
604  * of mblk_t chains, at which point it will indicate flow control back to the
605  * client. Once this condition is reached, any mblk_t chains that were not
606  * enqueued will be returned to the caller and they will have to decide what to
607  * do with them. There are various flags that control this behavior that a
608  * client may pass, which are discussed below.
609  *
610  * When this condition is hit, MAC also returns a cookie to the client in
611  * addition to unconsumed frames. Clients can poll on that cookie and register a
612  * callback with MAC to be notified when they are no longer subject to flow
613  * control, at which point they may continue to call mac_tx(). This flow control
614  * actually manages to work itself all the way up the stack, back through dls,
615  * to ip, through the various protocols, and to sockfs.
616  *
617  * While the behavior described above is the default, this behavior can be
618  * modified. There are two alternate modes, described below, which are
619  * controlled with flags.
620  *
621  * DROP MODE
622  *
623  * This mode is controlled by having the client pass the MAC_DROP_ON_NO_DESC
624  * flag. When this is passed, if a device driver runs out of transmit
625  * descriptors, then the MAC layer will drop any unsent traffic. The client in
626  * this case will never have any frames returned to it.
627  *
628  * DON'T ENQUEUE
629  *
630  * This mode is controlled by having the client pass the MAC_TX_NO_ENQUEUE flag.
631  * If the MAC_DROP_ON_NO_DESC flag is also passed, it takes precedence. In this
632  * mode, when we hit a case where a driver runs out of transmit descriptors,
633  * then instead of enqueuing packets in a soft ring set or software ring, we
634  * instead return the mblk_t chain back to the caller and immediately put the
635  * soft ring set into flow control mode.
636  *
637  * The following series of ASCII art images describe the transmit data path that
638  * MAC clients enter into based on calling into mac_tx(). A soft ring set has a
639  * transmission function associated with it. There are seven possible
640  * transmission modes, some of which share function entry points. The one that a
641  * soft ring set gets depends on properties such as whether there are
642  * transmission rings for fanout, whether the device involves aggregations,
643  * whether any bandwidth limits exist, etc.
644  *
645  *
646  * Part 1 -- Initial checks
647  *
648  *      * . called by
649  *      |   MAC clients
650  *      v                     . . No
651  *  +--------+  +-----------+ .   +-------------------+  +====================+
652  *  | mac_tx |->| device    |-*-->| mac_protect_check |->v Is this the simple v
653  *  +--------+  | quiesced? |     +-------------------+  v case? See [1]      v
654  *              +-----------+            |               +====================+
655  *                  * . Yes              * failed                 |
656  *                  v                    | frames                 |
657  *             +--------------+          |                +-------+---------+
658  *             | freemsgchain |<---------+          Yes . *            No . *
659  *             +--------------+                           v                 v
660  *                                                  +-----------+     +--------+
661  *                                                  |   goto    |     |  goto  |
662  *                                                  |  Part 2   |     | SRS TX |
663  *                                                  | Entry [A] |     |  func  |
664  *                                                  +-----------+     +--------+
665  *                                                        |                 |
666  *                                                        |                 v
667  *                                                        |           +--------+
668  *                                                        +---------->| return |
669  *                                                                    | cookie |
670  *                                                                    +--------+
671  *
672  * [1] The simple case refers to the SRS being configured with the
673  * SRS_TX_DEFAULT transmission mode, having a single mblk_t (not a chain), their
674  * being only a single active client, and not having a backlog in the srs.
675  *
676  *
677  * Part 2 -- The SRS transmission functions
678  *
679  * This part is a bit more complicated. The different transmission paths often
680  * leverage one another. In this case, we'll draw out the more common ones
681  * before the parts that depend upon them. Here, we're going to start with the
682  * workings of mac_tx_send() a common function that most of the others end up
683  * calling.
684  *
685  *      +-------------+
686  *      | mac_tx_send |
687  *      +-------------+
688  *            |
689  *            v
690  *      +=============+    +==============+
691  *      v  more than  v--->v    check     v
692  *      v one client? v    v VLAN and add v
693  *      +=============+    v  VLAN tags   v
694  *            |            +==============+
695  *            |                  |
696  *            +------------------+
697  *            |
698  *            |                 [A]
699  *            v                  |
700  *       +============+ . No     v
701  *       v more than  v .     +==========+     +--------------------------+
702  *       v one active v-*---->v for each v---->| mac_promisc_dispatch_one |---+
703  *       v  client?   v       v mblk_t   v     +--------------------------+   |
704  *       +============+       +==========+        ^                           |
705  *            |                                   |       +==========+        |
706  *            * . Yes                             |       v hardware v<-------+
707  *            v                      +------------+       v  rings?  v
708  *       +==========+                |                    +==========+
709  *       v for each v       No . . . *                         |
710  *       v mblk_t   v       specific |                         |
711  *       +==========+       flow     |                   +-----+-----+
712  *            |                      |                   |           |
713  *            v                      |                   v           v
714  *    +-----------------+            |               +-------+  +---------+
715  *    | mac_tx_classify |------------+               | GLDv3 |  |  GLDv3  |
716  *    +-----------------+                            |TX func|  | ring tx |
717  *            |                                      +-------+  |  func   |
718  *            * Specific flow, generally                 |      +---------+
719  *            | bcast, mcast, loopback                   |           |
720  *            v                                          +-----+-----+
721  *      +==========+       +---------+                         |
722  *      v valid L2 v--*--->| freemsg |                         v
723  *      v  header  v  . No +---------+               +-------------------+
724  *      +==========+                                 | return unconsumed |
725  *            * . Yes                                |   frames to the   |
726  *            v                                      |      caller       |
727  *      +===========+                                +-------------------+
728  *      v braodcast v      +----------------+                  ^
729  *      v   flow?   v--*-->| mac_bcast_send |------------------+
730  *      +===========+  .   +----------------+                  |
731  *            |        . . Yes                                 |
732  *       No . *                                                v
733  *            |  +---------------------+  +---------------+  +----------+
734  *            +->|mac_promisc_dispatch |->| mac_fix_cksum |->|   flow   |
735  *               +---------------------+  +---------------+  | callback |
736  *                                                           +----------+
737  *
738  *
739  * In addition, many but not all of the routines, all rely on
740  * mac_tx_softring_process as an entry point.
741  *
742  *
743  *                                           . No             . No
744  * +--------------------------+   +========+ .  +===========+ .  +-------------+
745  * | mac_tx_soft_ring_process |-->v worker v-*->v out of tx v-*->|    goto     |
746  * +--------------------------+   v only?  v    v  descr.?  v    | mac_tx_send |
747  *                                +========+    +===========+    +-------------+
748  *                              Yes . *               * . Yes           |
749  *                   . No             v               |                 v
750  *     v=========+   .          +===========+ . Yes   |     Yes .  +==========+
751  *     v apppend v<--*----------v out of tx v-*-------+---------*--v returned v
752  *     v mblk_t  v              v  descr.?  v         |            v frames?  v
753  *     v chain   v              +===========+         |            +==========+
754  *     +=========+                                    |                 *. No
755  *         |                                          |                 v
756  *         v                                          v           +------------+
757  * +===================+           +----------------------+       |   done     |
758  * v worker scheduled? v           | mac_tx_sring_enqueue |       | processing |
759  * v Out of tx descr?  v           +----------------------+       +------------+
760  * +===================+                      |
761  *    |           |           . Yes           v
762  *    * Yes       * No        .         +============+
763  *    |           v         +-*---------v drop on no v
764  *    |      +========+     v           v  TX desc?  v
765  *    |      v  wake  v  +----------+   +============+
766  *    |      v worker v  | mac_pkt_ |         * . No
767  *    |      +========+  | drop     |         |         . Yes         . No
768  *    |           |      +----------+         v         .             .
769  *    |           |         v   ^     +===============+ .  +========+ .
770  *    +--+--------+---------+   |     v Don't enqueue v-*->v ring   v-*----+
771  *       |                      |     v     Set?      v    v empty? v      |
772  *       |      +---------------+     +===============+    +========+      |
773  *       |      |                            |                |            |
774  *       |      |        +-------------------+                |            |
775  *       |      *. Yes   |                          +---------+            |
776  *       |      |        v                          v                      v
777  *       |      |  +===========+               +========+      +--------------+
778  *       |      +<-v At hiwat? v               v append v      |    return    |
779  *       |         +===========+               v mblk_t v      | mblk_t chain |
780  *       |                  * No               v chain  v      |   and flow   |
781  *       |                  v                  +========+      |    control   |
782  *       |               +=========+                |          |    cookie    |
783  *       |               v  append v                v          +--------------+
784  *       |               v  mblk_t v           +========+
785  *       |               v  chain  v           v  wake  v   +------------+
786  *       |               +=========+           v worker v-->|    done    |
787  *       |                    |                +========+   | processing |
788  *       |                    v       .. Yes                +------------+
789  *       |               +=========+  .   +========+
790  *       |               v  first  v--*-->v  wake  v
791  *       |               v append? v      v worker v
792  *       |               +=========+      +========+
793  *       |                   |                |
794  *       |              No . *                |
795  *       |                   v                |
796  *       |       +--------------+             |
797  *       +------>|   Return     |             |
798  *               | flow control |<------------+
799  *               |   cookie     |
800  *               +--------------+
801  *
802  *
803  * The remaining images are all specific to each of the different transmission
804  * modes.
805  *
806  * SRS TX DEFAULT
807  *
808  *      [ From Part 1 ]
809  *             |
810  *             v
811  * +-------------------------+
812  * | mac_tx_single_ring_mode |
813  * +-------------------------+
814  *            |
815  *            |       . Yes
816  *            v       .
817  *       +==========+ .  +============+
818  *       v   SRS    v-*->v   Try to   v---->---------------------+
819  *       v backlog? v    v enqueue in v                          |
820  *       +==========+    v     SRS    v-->------+                * . . Queue too
821  *            |          +============+         * don't enqueue  |     deep or
822  *            * . No         ^     |            | flag or at     |     drop flag
823  *            |              |     v            | hiwat,         |
824  *            v              |     |            | return    +---------+
825  *     +-------------+       |     |            | cookie    | freemsg |
826  *     |    goto     |-*-----+     |            |           +---------+
827  *     | mac_tx_send | . returned  |            |                |
828  *     +-------------+   mblk_t    |            |                |
829  *            |                    |            |                |
830  *            |                    |            |                |
831  *            * . . all mblk_t     * queued,    |                |
832  *            v     consumed       | may return |                |
833  *     +-------------+             | tx cookie  |                |
834  *     | SRS TX func |<------------+------------+----------------+
835  *     |  completed  |
836  *     +-------------+
837  *
838  * SRS_TX_SERIALIZE
839  *
840  *   +------------------------+
841  *   | mac_tx_serializer_mode |
842  *   +------------------------+
843  *               |
844  *               |        . No
845  *               v        .
846  *         +============+ .  +============+    +-------------+   +============+
847  *         v srs being  v-*->v  set SRS   v--->|    goto     |-->v remove SRS v
848  *         v processed? v    v proc flags v    | mac_tx_send |   v proc flag  v
849  *         +============+    +============+    +-------------+   +============+
850  *               |                                                     |
851  *               * Yes                                                 |
852  *               v                                       . No          v
853  *      +--------------------+                           .        +==========+
854  *      | mac_tx_srs_enqueue |  +------------------------*-----<--v returned v
855  *      +--------------------+  |                                 v frames?  v
856  *               |              |   . Yes                         +==========+
857  *               |              |   .                                  |
858  *               |              |   . +=========+                      v
859  *               v              +-<-*-v queued  v     +--------------------+
860  *        +-------------+       |     v frames? v<----| mac_tx_srs_enqueue |
861  *        | SRS TX func |       |     +=========+     +--------------------+
862  *        | completed,  |<------+         * . Yes
863  *        | may return  |       |         v
864  *        |   cookie    |       |     +========+
865  *        +-------------+       +-<---v  wake  v
866  *                                    v worker v
867  *                                    +========+
868  *
869  *
870  * SRS_TX_FANOUT
871  *
872  *                                             . Yes
873  *   +--------------------+    +=============+ .   +--------------------------+
874  *   | mac_tx_fanout_mode |--->v Have fanout v-*-->|           goto           |
875  *   +--------------------+    v   hint?     v     | mac_rx_soft_ring_process |
876  *                             +=============+     +--------------------------+
877  *                                   * . No                    |
878  *                                   v                         ^
879  *                             +===========+                   |
880  *                        +--->v for each  v           +===============+
881  *                        |    v   mblk_t  v           v pick softring v
882  *                 same   *    +===========+           v   from hash   v
883  *                 hash   |          |                 +===============+
884  *                        |          v                         |
885  *                        |   +--------------+                 |
886  *                        +---| mac_pkt_hash |--->*------------+
887  *                            +--------------+    . different
888  *                                                  hash or
889  *                                                  done proc.
890  * SRS_TX_AGGR                                      chain
891  *
892  *   +------------------+    +================================+
893  *   | mac_tx_aggr_mode |--->v Use aggr capab function to     v
894  *   +------------------+    v find appropriate tx ring.      v
895  *                           v Applies hash based on aggr     v
896  *                           v policy, see mac_tx_aggr_mode() v
897  *                           +================================+
898  *                                          |
899  *                                          v
900  *                           +-------------------------------+
901  *                           |            goto               |
902  *                           |  mac_rx_srs_soft_ring_process |
903  *                           +-------------------------------+
904  *
905  *
906  * SRS_TX_BW, SRS_TX_BW_FANOUT, SRS_TX_BW_AGGR
907  *
908  * Note, all three of these tx functions start from the same place --
909  * mac_tx_bw_mode().
910  *
911  *  +----------------+
912  *  | mac_tx_bw_mode |
913  *  +----------------+
914  *         |
915  *         v          . No               . No               . Yes
916  *  +==============+  .  +============+  .  +=============+ .  +=========+
917  *  v  Out of BW?  v--*->v SRS empty? v--*->v  reset BW   v-*->v Bump BW v
918  *  +==============+     +============+     v tick count? v    v Usage   v
919  *         |                   |            +=============+    +=========+
920  *         |         +---------+                   |                |
921  *         |         |        +--------------------+                |
922  *         |         |        |              +----------------------+
923  *         v         |        v              v
924  * +===============+ |  +==========+   +==========+      +------------------+
925  * v Don't enqueue v |  v  set bw  v   v Is aggr? v--*-->|       goto       |
926  * v   flag set?   v |  v enforced v   +==========+  .   | mac_tx_aggr_mode |-+
927  * +===============+ |  +==========+         |       .   +------------------+ |
928  *   |    Yes .*     |        |         No . *       .                        |
929  *   |         |     |        |              |       . Yes                    |
930  *   * . No    |     |        v              |                                |
931  *   |  +---------+  |   +========+          v              +======+          |
932  *   |  | freemsg |  |   v append v   +============+  . Yes v pick v          |
933  *   |  +---------+  |   v mblk_t v   v Is fanout? v--*---->v ring v          |
934  *   |      |        |   v chain  v   +============+        +======+          |
935  *   +------+        |   +========+          |                  |             |
936  *          v        |        |              v                  v             |
937  *    +---------+    |        v       +-------------+ +--------------------+  |
938  *    | return  |    |   +========+   |    goto     | |       goto         |  |
939  *    |  flow   |    |   v wakeup v   | mac_tx_send | | mac_tx_fanout_mode |  |
940  *    | control |    |   v worker v   +-------------+ +--------------------+  |
941  *    | cookie  |    |   +========+          |                  |             |
942  *    +---------+    |        |              |                  +------+------+
943  *                   |        v              |                         |
944  *                   |   +---------+         |                         v
945  *                   |   | return  |   +============+           +------------+
946  *                   |   |  flow   |   v unconsumed v-------+   |   done     |
947  *                   |   | control |   v   frames?  v       |   | processing |
948  *                   |   | cookie  |   +============+       |   +------------+
949  *                   |   +---------+         |              |
950  *                   |                  Yes  *              |
951  *                   |                       |              |
952  *                   |                 +===========+        |
953  *                   |                 v subtract  v        |
954  *                   |                 v unused bw v        |
955  *                   |                 +===========+        |
956  *                   |                       |              |
957  *                   |                       v              |
958  *                   |              +--------------------+  |
959  *                   +------------->| mac_tx_srs_enqueue |  |
960  *                                  +--------------------+  |
961  *                                           |              |
962  *                                           |              |
963  *                                     +------------+       |
964  *                                     |  return fc |       |
965  *                                     | cookie and |<------+
966  *                                     |    mblk_t  |
967  *                                     +------------+
968  */
969 
970 #include <sys/types.h>
971 #include <sys/callb.h>
972 #include <sys/sdt.h>
973 #include <sys/strsubr.h>
974 #include <sys/strsun.h>
975 #include <sys/vlan.h>
976 #include <sys/stack.h>
977 #include <sys/archsystm.h>
978 #include <inet/ipsec_impl.h>
979 #include <inet/ip_impl.h>
980 #include <inet/sadb.h>
981 #include <inet/ipsecesp.h>
982 #include <inet/ipsecah.h>
983 #include <inet/ip6.h>
984 
985 #include <sys/mac_impl.h>
986 #include <sys/mac_client_impl.h>
987 #include <sys/mac_client_priv.h>
988 #include <sys/mac_soft_ring.h>
989 #include <sys/mac_flow_impl.h>
990 
991 static mac_tx_cookie_t mac_tx_single_ring_mode(mac_soft_ring_set_t *, mblk_t *,
992     uintptr_t, uint16_t, mblk_t **);
993 static mac_tx_cookie_t mac_tx_serializer_mode(mac_soft_ring_set_t *, mblk_t *,
994     uintptr_t, uint16_t, mblk_t **);
995 static mac_tx_cookie_t mac_tx_fanout_mode(mac_soft_ring_set_t *, mblk_t *,
996     uintptr_t, uint16_t, mblk_t **);
997 static mac_tx_cookie_t mac_tx_bw_mode(mac_soft_ring_set_t *, mblk_t *,
998     uintptr_t, uint16_t, mblk_t **);
999 static mac_tx_cookie_t mac_tx_aggr_mode(mac_soft_ring_set_t *, mblk_t *,
1000     uintptr_t, uint16_t, mblk_t **);
1001 
1002 typedef struct mac_tx_mode_s {
1003 	mac_tx_srs_mode_t	mac_tx_mode;
1004 	mac_tx_func_t		mac_tx_func;
1005 } mac_tx_mode_t;
1006 
1007 /*
1008  * There are seven modes of operation on the Tx side. These modes get set
1009  * in mac_tx_srs_setup(). Except for the experimental TX_SERIALIZE mode,
1010  * none of the other modes are user configurable. They get selected by
1011  * the system depending upon whether the link (or flow) has multiple Tx
1012  * rings or a bandwidth configured, or if the link is an aggr, etc.
1013  *
1014  * When the Tx SRS is operating in aggr mode (st_mode) or if there are
1015  * multiple Tx rings owned by Tx SRS, then each Tx ring (pseudo or
1016  * otherwise) will have a soft ring associated with it. These soft rings
1017  * are stored in srs_tx_soft_rings[] array.
1018  *
1019  * Additionally in the case of aggr, there is the st_soft_rings[] array
1020  * in the mac_srs_tx_t structure. This array is used to store the same
1021  * set of soft rings that are present in srs_tx_soft_rings[] array but
1022  * in a different manner. The soft ring associated with the pseudo Tx
1023  * ring is saved at mr_index (of the pseudo ring) in st_soft_rings[]
1024  * array. This helps in quickly getting the soft ring associated with the
1025  * Tx ring when aggr_find_tx_ring() returns the pseudo Tx ring that is to
1026  * be used for transmit.
1027  */
1028 mac_tx_mode_t mac_tx_mode_list[] = {
1029 	{SRS_TX_DEFAULT,	mac_tx_single_ring_mode},
1030 	{SRS_TX_SERIALIZE,	mac_tx_serializer_mode},
1031 	{SRS_TX_FANOUT,		mac_tx_fanout_mode},
1032 	{SRS_TX_BW,		mac_tx_bw_mode},
1033 	{SRS_TX_BW_FANOUT,	mac_tx_bw_mode},
1034 	{SRS_TX_AGGR,		mac_tx_aggr_mode},
1035 	{SRS_TX_BW_AGGR,	mac_tx_bw_mode}
1036 };
1037 
1038 /*
1039  * Soft Ring Set (SRS) - The Run time code that deals with
1040  * dynamic polling from the hardware, bandwidth enforcement,
1041  * fanout etc.
1042  *
1043  * We try to use H/W classification on NIC and assign traffic for
1044  * a MAC address to a particular Rx ring or ring group. There is a
1045  * 1-1 mapping between a SRS and a Rx ring. The SRS dynamically
1046  * switches the underlying Rx ring between interrupt and
1047  * polling mode and enforces any specified B/W control.
1048  *
1049  * There is always a SRS created and tied to each H/W and S/W rule.
1050  * Whenever we create a H/W rule, we always add the the same rule to
1051  * S/W classifier and tie a SRS to it.
1052  *
1053  * In case a B/W control is specified, it is broken into bytes
1054  * per ticks and as soon as the quota for a tick is exhausted,
1055  * the underlying Rx ring is forced into poll mode for remainder of
1056  * the tick. The SRS poll thread only polls for bytes that are
1057  * allowed to come in the SRS. We typically let 4x the configured
1058  * B/W worth of packets to come in the SRS (to prevent unnecessary
1059  * drops due to bursts) but only process the specified amount.
1060  *
1061  * A MAC client (e.g. a VNIC or aggr) can have 1 or more
1062  * Rx rings (and corresponding SRSs) assigned to it. The SRS
1063  * in turn can have softrings to do protocol level fanout or
1064  * softrings to do S/W based fanout or both. In case the NIC
1065  * has no Rx rings, we do S/W classification to respective SRS.
1066  * The S/W classification rule is always setup and ready. This
1067  * allows the MAC layer to reassign Rx rings whenever needed
1068  * but packets still continue to flow via the default path and
1069  * getting S/W classified to correct SRS.
1070  *
1071  * The SRS's are used on both Tx and Rx side. They use the same
1072  * data structure but the processing routines have slightly different
1073  * semantics due to the fact that Rx side needs to do dynamic
1074  * polling etc.
1075  *
1076  * Dynamic Polling Notes
1077  * =====================
1078  *
1079  * Each Soft ring set is capable of switching its Rx ring between
1080  * interrupt and poll mode and actively 'polls' for packets in
1081  * poll mode. If the SRS is implementing a B/W limit, it makes
1082  * sure that only Max allowed packets are pulled in poll mode
1083  * and goes to poll mode as soon as B/W limit is exceeded. As
1084  * such, there are no overheads to implement B/W limits.
1085  *
1086  * In poll mode, its better to keep the pipeline going where the
1087  * SRS worker thread keeps processing packets and poll thread
1088  * keeps bringing more packets (specially if they get to run
1089  * on different CPUs). This also prevents the overheads associated
1090  * by excessive signalling (on NUMA machines, this can be
1091  * pretty devastating). The exception is latency optimized case
1092  * where worker thread does no work and interrupt and poll thread
1093  * are allowed to do their own drain.
1094  *
1095  * We use the following policy to control Dynamic Polling:
1096  * 1) We switch to poll mode anytime the processing
1097  *    thread causes a backlog to build up in SRS and
1098  *    its associated Soft Rings (sr_poll_pkt_cnt > 0).
1099  * 2) As long as the backlog stays under the low water
1100  *    mark (sr_lowat), we poll the H/W for more packets.
1101  * 3) If the backlog (sr_poll_pkt_cnt) exceeds low
1102  *    water mark, we stay in poll mode but don't poll
1103  *    the H/W for more packets.
1104  * 4) Anytime in polling mode, if we poll the H/W for
1105  *    packets and find nothing plus we have an existing
1106  *    backlog (sr_poll_pkt_cnt > 0), we stay in polling
1107  *    mode but don't poll the H/W for packets anymore
1108  *    (let the polling thread go to sleep).
1109  * 5) Once the backlog is relived (packets are processed)
1110  *    we reenable polling (by signalling the poll thread)
1111  *    only when the backlog dips below sr_poll_thres.
1112  * 6) sr_hiwat is used exclusively when we are not
1113  *    polling capable and is used to decide when to
1114  *    drop packets so the SRS queue length doesn't grow
1115  *    infinitely.
1116  *
1117  * NOTE: Also see the block level comment on top of mac_soft_ring.c
1118  */
1119 
1120 /*
1121  * mac_latency_optimize
1122  *
1123  * Controls whether the poll thread can process the packets inline
1124  * or let the SRS worker thread do the processing. This applies if
1125  * the SRS was not being processed. For latency sensitive traffic,
1126  * this needs to be true to allow inline processing. For throughput
1127  * under load, this should be false.
1128  *
1129  * This (and other similar) tunable should be rolled into a link
1130  * or flow specific workload hint that can be set using dladm
1131  * linkprop (instead of multiple such tunables).
1132  */
1133 boolean_t mac_latency_optimize = B_TRUE;
1134 
1135 /*
1136  * MAC_RX_SRS_ENQUEUE_CHAIN and MAC_TX_SRS_ENQUEUE_CHAIN
1137  *
1138  * queue a mp or chain in soft ring set and increment the
1139  * local count (srs_count) for the SRS and the shared counter
1140  * (srs_poll_pkt_cnt - shared between SRS and its soft rings
1141  * to track the total unprocessed packets for polling to work
1142  * correctly).
1143  *
1144  * The size (total bytes queued) counters are incremented only
1145  * if we are doing B/W control.
1146  */
1147 #define	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {		\
1148 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1149 	if ((mac_srs)->srs_last != NULL)				\
1150 		(mac_srs)->srs_last->b_next = (head);			\
1151 	else								\
1152 		(mac_srs)->srs_first = (head);				\
1153 	(mac_srs)->srs_last = (tail);					\
1154 	(mac_srs)->srs_count += count;					\
1155 }
1156 
1157 #define	MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {	\
1158 	mac_srs_rx_t	*srs_rx = &(mac_srs)->srs_rx;			\
1159 									\
1160 	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz);		\
1161 	srs_rx->sr_poll_pkt_cnt += count;				\
1162 	ASSERT(srs_rx->sr_poll_pkt_cnt > 0);				\
1163 	if ((mac_srs)->srs_type & SRST_BW_CONTROL) {			\
1164 		(mac_srs)->srs_size += (sz);				\
1165 		mutex_enter(&(mac_srs)->srs_bw->mac_bw_lock);		\
1166 		(mac_srs)->srs_bw->mac_bw_sz += (sz);			\
1167 		mutex_exit(&(mac_srs)->srs_bw->mac_bw_lock);		\
1168 	}								\
1169 }
1170 
1171 #define	MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {	\
1172 	mac_srs->srs_state |= SRS_ENQUEUED;				\
1173 	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz);		\
1174 	if ((mac_srs)->srs_type & SRST_BW_CONTROL) {			\
1175 		(mac_srs)->srs_size += (sz);				\
1176 		(mac_srs)->srs_bw->mac_bw_sz += (sz);			\
1177 	}								\
1178 }
1179 
1180 /*
1181  * Turn polling on routines
1182  */
1183 #define	MAC_SRS_POLLING_ON(mac_srs) {					\
1184 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1185 	if (((mac_srs)->srs_state &					\
1186 	    (SRS_POLLING_CAPAB|SRS_POLLING)) == SRS_POLLING_CAPAB) {	\
1187 		(mac_srs)->srs_state |= SRS_POLLING;			\
1188 		(void) mac_hwring_disable_intr((mac_ring_handle_t)	\
1189 		    (mac_srs)->srs_ring);				\
1190 		(mac_srs)->srs_rx.sr_poll_on++;				\
1191 	}								\
1192 }
1193 
1194 #define	MAC_SRS_WORKER_POLLING_ON(mac_srs) {				\
1195 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1196 	if (((mac_srs)->srs_state &					\
1197 	    (SRS_POLLING_CAPAB|SRS_WORKER|SRS_POLLING)) ==		\
1198 	    (SRS_POLLING_CAPAB|SRS_WORKER)) {				\
1199 		(mac_srs)->srs_state |= SRS_POLLING;			\
1200 		(void) mac_hwring_disable_intr((mac_ring_handle_t)	\
1201 		    (mac_srs)->srs_ring);				\
1202 		(mac_srs)->srs_rx.sr_worker_poll_on++;			\
1203 	}								\
1204 }
1205 
1206 /*
1207  * MAC_SRS_POLL_RING
1208  *
1209  * Signal the SRS poll thread to poll the underlying H/W ring
1210  * provided it wasn't already polling (SRS_GET_PKTS was set).
1211  *
1212  * Poll thread gets to run only from mac_rx_srs_drain() and only
1213  * if the drain was being done by the worker thread.
1214  */
1215 #define	MAC_SRS_POLL_RING(mac_srs) {					\
1216 	mac_srs_rx_t	*srs_rx = &(mac_srs)->srs_rx;			\
1217 									\
1218 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1219 	srs_rx->sr_poll_thr_sig++;					\
1220 	if (((mac_srs)->srs_state &					\
1221 	    (SRS_POLLING_CAPAB|SRS_WORKER|SRS_GET_PKTS)) ==		\
1222 		(SRS_WORKER|SRS_POLLING_CAPAB)) {			\
1223 		(mac_srs)->srs_state |= SRS_GET_PKTS;			\
1224 		cv_signal(&(mac_srs)->srs_cv);				\
1225 	} else {							\
1226 		srs_rx->sr_poll_thr_busy++;				\
1227 	}								\
1228 }
1229 
1230 /*
1231  * MAC_SRS_CHECK_BW_CONTROL
1232  *
1233  * Check to see if next tick has started so we can reset the
1234  * SRS_BW_ENFORCED flag and allow more packets to come in the
1235  * system.
1236  */
1237 #define	MAC_SRS_CHECK_BW_CONTROL(mac_srs) {				\
1238 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1239 	ASSERT(((mac_srs)->srs_type & SRST_TX) ||			\
1240 	    MUTEX_HELD(&(mac_srs)->srs_bw->mac_bw_lock));		\
1241 	clock_t now = ddi_get_lbolt();					\
1242 	if ((mac_srs)->srs_bw->mac_bw_curr_time != now) {		\
1243 		(mac_srs)->srs_bw->mac_bw_curr_time = now;		\
1244 		(mac_srs)->srs_bw->mac_bw_used = 0;			\
1245 		if ((mac_srs)->srs_bw->mac_bw_state & SRS_BW_ENFORCED)	\
1246 			(mac_srs)->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED; \
1247 	}								\
1248 }
1249 
1250 /*
1251  * MAC_SRS_WORKER_WAKEUP
1252  *
1253  * Wake up the SRS worker thread to process the queue as long as
1254  * no one else is processing the queue. If we are optimizing for
1255  * latency, we wake up the worker thread immediately or else we
1256  * wait mac_srs_worker_wakeup_ticks before worker thread gets
1257  * woken up.
1258  */
1259 int mac_srs_worker_wakeup_ticks = 0;
1260 #define	MAC_SRS_WORKER_WAKEUP(mac_srs) {				\
1261 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1262 	if (!((mac_srs)->srs_state & SRS_PROC) &&			\
1263 		(mac_srs)->srs_tid == NULL) {				\
1264 		if (((mac_srs)->srs_state & SRS_LATENCY_OPT) ||		\
1265 			(mac_srs_worker_wakeup_ticks == 0))		\
1266 			cv_signal(&(mac_srs)->srs_async);		\
1267 		else							\
1268 			(mac_srs)->srs_tid =				\
1269 				timeout(mac_srs_fire, (mac_srs),	\
1270 					mac_srs_worker_wakeup_ticks);	\
1271 	}								\
1272 }
1273 
1274 #define	TX_BANDWIDTH_MODE(mac_srs)				\
1275 	((mac_srs)->srs_tx.st_mode == SRS_TX_BW ||		\
1276 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW_FANOUT ||	\
1277 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW_AGGR)
1278 
1279 #define	TX_SRS_TO_SOFT_RING(mac_srs, head, hint) {			\
1280 	if (tx_mode == SRS_TX_BW_FANOUT)				\
1281 		(void) mac_tx_fanout_mode(mac_srs, head, hint, 0, NULL);\
1282 	else								\
1283 		(void) mac_tx_aggr_mode(mac_srs, head, hint, 0, NULL);	\
1284 }
1285 
1286 /*
1287  * MAC_TX_SRS_BLOCK
1288  *
1289  * Always called from mac_tx_srs_drain() function. SRS_TX_BLOCKED
1290  * will be set only if srs_tx_woken_up is FALSE. If
1291  * srs_tx_woken_up is TRUE, it indicates that the wakeup arrived
1292  * before we grabbed srs_lock to set SRS_TX_BLOCKED. We need to
1293  * attempt to transmit again and not setting SRS_TX_BLOCKED does
1294  * that.
1295  */
1296 #define	MAC_TX_SRS_BLOCK(srs, mp)	{			\
1297 	ASSERT(MUTEX_HELD(&(srs)->srs_lock));			\
1298 	if ((srs)->srs_tx.st_woken_up) {			\
1299 		(srs)->srs_tx.st_woken_up = B_FALSE;		\
1300 	} else {						\
1301 		ASSERT(!((srs)->srs_state & SRS_TX_BLOCKED));	\
1302 		(srs)->srs_state |= SRS_TX_BLOCKED;		\
1303 		(srs)->srs_tx.st_stat.mts_blockcnt++;		\
1304 	}							\
1305 }
1306 
1307 /*
1308  * MAC_TX_SRS_TEST_HIWAT
1309  *
1310  * Called before queueing a packet onto Tx SRS to test and set
1311  * SRS_TX_HIWAT if srs_count exceeds srs_tx_hiwat.
1312  */
1313 #define	MAC_TX_SRS_TEST_HIWAT(srs, mp, tail, cnt, sz, cookie) {		\
1314 	boolean_t enqueue = 1;						\
1315 									\
1316 	if ((srs)->srs_count > (srs)->srs_tx.st_hiwat) {		\
1317 		/*							\
1318 		 * flow-controlled. Store srs in cookie so that it	\
1319 		 * can be returned as mac_tx_cookie_t to client		\
1320 		 */							\
1321 		(srs)->srs_state |= SRS_TX_HIWAT;			\
1322 		cookie = (mac_tx_cookie_t)srs;				\
1323 		(srs)->srs_tx.st_hiwat_cnt++;				\
1324 		if ((srs)->srs_count > (srs)->srs_tx.st_max_q_cnt) {	\
1325 			/* increment freed stats */			\
1326 			(srs)->srs_tx.st_stat.mts_sdrops += cnt;	\
1327 			/*						\
1328 			 * b_prev may be set to the fanout hint		\
1329 			 * hence can't use freemsg directly		\
1330 			 */						\
1331 			mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);	\
1332 			DTRACE_PROBE1(tx_queued_hiwat,			\
1333 			    mac_soft_ring_set_t *, srs);		\
1334 			enqueue = 0;					\
1335 		}							\
1336 	}								\
1337 	if (enqueue)							\
1338 		MAC_TX_SRS_ENQUEUE_CHAIN(srs, mp, tail, cnt, sz);	\
1339 }
1340 
1341 /* Some utility macros */
1342 #define	MAC_SRS_BW_LOCK(srs)						\
1343 	if (!(srs->srs_type & SRST_TX))					\
1344 		mutex_enter(&srs->srs_bw->mac_bw_lock);
1345 
1346 #define	MAC_SRS_BW_UNLOCK(srs)						\
1347 	if (!(srs->srs_type & SRST_TX))					\
1348 		mutex_exit(&srs->srs_bw->mac_bw_lock);
1349 
1350 #define	MAC_TX_SRS_DROP_MESSAGE(srs, mp, cookie) {		\
1351 	mac_pkt_drop(NULL, NULL, mp, B_FALSE);			\
1352 	/* increment freed stats */				\
1353 	mac_srs->srs_tx.st_stat.mts_sdrops++;			\
1354 	cookie = (mac_tx_cookie_t)srs;				\
1355 }
1356 
1357 #define	MAC_TX_SET_NO_ENQUEUE(srs, mp_chain, ret_mp, cookie) {		\
1358 	mac_srs->srs_state |= SRS_TX_WAKEUP_CLIENT;			\
1359 	cookie = (mac_tx_cookie_t)srs;					\
1360 	*ret_mp = mp_chain;						\
1361 }
1362 
1363 /*
1364  * MAC_RX_SRS_TOODEEP
1365  *
1366  * Macro called as part of receive-side processing to determine if handling
1367  * can occur in situ (in the interrupt thread) or if it should be left to a
1368  * worker thread.  Note that the constant used to make this determination is
1369  * not entirely made-up, and is a result of some emprical validation. That
1370  * said, the constant is left as a static variable to allow it to be
1371  * dynamically tuned in the field if and as needed.
1372  */
1373 static uintptr_t mac_rx_srs_stack_needed = 10240;
1374 static uint_t mac_rx_srs_stack_toodeep;
1375 
1376 #ifndef STACK_GROWTH_DOWN
1377 #error Downward stack growth assumed.
1378 #endif
1379 
1380 #define	MAC_RX_SRS_TOODEEP() (STACK_BIAS + (uintptr_t)getfp() - \
1381 	(uintptr_t)curthread->t_stkbase < mac_rx_srs_stack_needed && \
1382 	++mac_rx_srs_stack_toodeep)
1383 
1384 
1385 /*
1386  * Drop the rx packet and advance to the next one in the chain.
1387  */
1388 static void
1389 mac_rx_drop_pkt(mac_soft_ring_set_t *srs, mblk_t *mp)
1390 {
1391 	mac_srs_rx_t	*srs_rx = &srs->srs_rx;
1392 
1393 	ASSERT(mp->b_next == NULL);
1394 	mutex_enter(&srs->srs_lock);
1395 	MAC_UPDATE_SRS_COUNT_LOCKED(srs, 1);
1396 	MAC_UPDATE_SRS_SIZE_LOCKED(srs, msgdsize(mp));
1397 	mutex_exit(&srs->srs_lock);
1398 
1399 	srs_rx->sr_stat.mrs_sdrops++;
1400 	freemsg(mp);
1401 }
1402 
1403 /* DATAPATH RUNTIME ROUTINES */
1404 
1405 /*
1406  * mac_srs_fire
1407  *
1408  * Timer callback routine for waking up the SRS worker thread.
1409  */
1410 static void
1411 mac_srs_fire(void *arg)
1412 {
1413 	mac_soft_ring_set_t *mac_srs = (mac_soft_ring_set_t *)arg;
1414 
1415 	mutex_enter(&mac_srs->srs_lock);
1416 	if (mac_srs->srs_tid == NULL) {
1417 		mutex_exit(&mac_srs->srs_lock);
1418 		return;
1419 	}
1420 
1421 	mac_srs->srs_tid = NULL;
1422 	if (!(mac_srs->srs_state & SRS_PROC))
1423 		cv_signal(&mac_srs->srs_async);
1424 
1425 	mutex_exit(&mac_srs->srs_lock);
1426 }
1427 
1428 /*
1429  * 'hint' is fanout_hint (type of uint64_t) which is given by the TCP/IP stack,
1430  * and it is used on the TX path.
1431  */
1432 #define	HASH_HINT(hint)	\
1433 	((hint) ^ ((hint) >> 24) ^ ((hint) >> 16) ^ ((hint) >> 8))
1434 
1435 
1436 /*
1437  * hash based on the src address, dst address and the port information.
1438  */
1439 #define	HASH_ADDR(src, dst, ports)					\
1440 	(ntohl((src) + (dst)) ^ ((ports) >> 24) ^ ((ports) >> 16) ^	\
1441 	((ports) >> 8) ^ (ports))
1442 
1443 #define	COMPUTE_INDEX(key, sz)	(key % sz)
1444 
1445 #define	FANOUT_ENQUEUE_MP(head, tail, cnt, bw_ctl, sz, sz0, mp) {	\
1446 	if ((tail) != NULL) {						\
1447 		ASSERT((tail)->b_next == NULL);				\
1448 		(tail)->b_next = (mp);					\
1449 	} else {							\
1450 		ASSERT((head) == NULL);					\
1451 		(head) = (mp);						\
1452 	}								\
1453 	(tail) = (mp);							\
1454 	(cnt)++;							\
1455 	if ((bw_ctl))							\
1456 		(sz) += (sz0);						\
1457 }
1458 
1459 #define	MAC_FANOUT_DEFAULT	0
1460 #define	MAC_FANOUT_RND_ROBIN	1
1461 int mac_fanout_type = MAC_FANOUT_DEFAULT;
1462 
1463 #define	MAX_SR_TYPES	3
1464 /* fanout types for port based hashing */
1465 enum pkt_type {
1466 	V4_TCP = 0,
1467 	V4_UDP,
1468 	OTH,
1469 	UNDEF
1470 };
1471 
1472 /*
1473  * Pair of local and remote ports in the transport header
1474  */
1475 #define	PORTS_SIZE 4
1476 
1477 /*
1478  * mac_rx_srs_proto_fanout
1479  *
1480  * This routine delivers packets destined to an SRS into one of the
1481  * protocol soft rings.
1482  *
1483  * Given a chain of packets we need to split it up into multiple sub chains
1484  * destined into TCP, UDP or OTH soft ring. Instead of entering
1485  * the soft ring one packet at a time, we want to enter it in the form of a
1486  * chain otherwise we get this start/stop behaviour where the worker thread
1487  * goes to sleep and then next packets comes in forcing it to wake up etc.
1488  */
1489 static void
1490 mac_rx_srs_proto_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
1491 {
1492 	struct ether_header		*ehp;
1493 	struct ether_vlan_header	*evhp;
1494 	uint32_t			sap;
1495 	ipha_t				*ipha;
1496 	uint8_t				*dstaddr;
1497 	size_t				hdrsize;
1498 	mblk_t				*mp;
1499 	mblk_t				*headmp[MAX_SR_TYPES];
1500 	mblk_t				*tailmp[MAX_SR_TYPES];
1501 	int				cnt[MAX_SR_TYPES];
1502 	size_t				sz[MAX_SR_TYPES];
1503 	size_t				sz1;
1504 	boolean_t			bw_ctl;
1505 	boolean_t			hw_classified;
1506 	boolean_t			dls_bypass;
1507 	boolean_t			is_ether;
1508 	boolean_t			is_unicast;
1509 	enum pkt_type			type;
1510 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
1511 
1512 	is_ether = (mcip->mci_mip->mi_info.mi_nativemedia == DL_ETHER);
1513 	bw_ctl = ((mac_srs->srs_type & SRST_BW_CONTROL) != 0);
1514 
1515 	/*
1516 	 * If we don't have a Rx ring, S/W classification would have done
1517 	 * its job and its a packet meant for us. If we were polling on
1518 	 * the default ring (i.e. there was a ring assigned to this SRS),
1519 	 * then we need to make sure that the mac address really belongs
1520 	 * to us.
1521 	 */
1522 	hw_classified = mac_srs->srs_ring != NULL &&
1523 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
1524 
1525 	/*
1526 	 * Special clients (eg. VLAN, non ether, etc) need DLS
1527 	 * processing in the Rx path. SRST_DLS_BYPASS will be clear for
1528 	 * such SRSs. Another way of disabling bypass is to set the
1529 	 * MCIS_RX_BYPASS_DISABLE flag.
1530 	 */
1531 	dls_bypass = ((mac_srs->srs_type & SRST_DLS_BYPASS) != 0) &&
1532 	    ((mcip->mci_state_flags & MCIS_RX_BYPASS_DISABLE) == 0);
1533 
1534 	bzero(headmp, MAX_SR_TYPES * sizeof (mblk_t *));
1535 	bzero(tailmp, MAX_SR_TYPES * sizeof (mblk_t *));
1536 	bzero(cnt, MAX_SR_TYPES * sizeof (int));
1537 	bzero(sz, MAX_SR_TYPES * sizeof (size_t));
1538 
1539 	/*
1540 	 * We got a chain from SRS that we need to send to the soft rings.
1541 	 * Since squeues for TCP & IPv4 sap poll their soft rings (for
1542 	 * performance reasons), we need to separate out v4_tcp, v4_udp
1543 	 * and the rest goes in other.
1544 	 */
1545 	while (head != NULL) {
1546 		mp = head;
1547 		head = head->b_next;
1548 		mp->b_next = NULL;
1549 
1550 		type = OTH;
1551 		sz1 = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
1552 
1553 		if (is_ether) {
1554 			/*
1555 			 * At this point we can be sure the packet at least
1556 			 * has an ether header.
1557 			 */
1558 			if (sz1 < sizeof (struct ether_header)) {
1559 				mac_rx_drop_pkt(mac_srs, mp);
1560 				continue;
1561 			}
1562 			ehp = (struct ether_header *)mp->b_rptr;
1563 
1564 			/*
1565 			 * Determine if this is a VLAN or non-VLAN packet.
1566 			 */
1567 			if ((sap = ntohs(ehp->ether_type)) == VLAN_TPID) {
1568 				evhp = (struct ether_vlan_header *)mp->b_rptr;
1569 				sap = ntohs(evhp->ether_type);
1570 				hdrsize = sizeof (struct ether_vlan_header);
1571 				/*
1572 				 * Check if the VID of the packet, if any,
1573 				 * belongs to this client.
1574 				 */
1575 				if (!mac_client_check_flow_vid(mcip,
1576 				    VLAN_ID(ntohs(evhp->ether_tci)))) {
1577 					mac_rx_drop_pkt(mac_srs, mp);
1578 					continue;
1579 				}
1580 			} else {
1581 				hdrsize = sizeof (struct ether_header);
1582 			}
1583 			is_unicast =
1584 			    ((((uint8_t *)&ehp->ether_dhost)[0] & 0x01) == 0);
1585 			dstaddr = (uint8_t *)&ehp->ether_dhost;
1586 		} else {
1587 			mac_header_info_t		mhi;
1588 
1589 			if (mac_header_info((mac_handle_t)mcip->mci_mip,
1590 			    mp, &mhi) != 0) {
1591 				mac_rx_drop_pkt(mac_srs, mp);
1592 				continue;
1593 			}
1594 			hdrsize = mhi.mhi_hdrsize;
1595 			sap = mhi.mhi_bindsap;
1596 			is_unicast = (mhi.mhi_dsttype == MAC_ADDRTYPE_UNICAST);
1597 			dstaddr = (uint8_t *)mhi.mhi_daddr;
1598 		}
1599 
1600 		if (!dls_bypass) {
1601 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
1602 			    cnt[type], bw_ctl, sz[type], sz1, mp);
1603 			continue;
1604 		}
1605 
1606 		if (sap == ETHERTYPE_IP) {
1607 			/*
1608 			 * If we are H/W classified, but we have promisc
1609 			 * on, then we need to check for the unicast address.
1610 			 */
1611 			if (hw_classified && mcip->mci_promisc_list != NULL) {
1612 				mac_address_t		*map;
1613 
1614 				rw_enter(&mcip->mci_rw_lock, RW_READER);
1615 				map = mcip->mci_unicast;
1616 				if (bcmp(dstaddr, map->ma_addr,
1617 				    map->ma_len) == 0)
1618 					type = UNDEF;
1619 				rw_exit(&mcip->mci_rw_lock);
1620 			} else if (is_unicast) {
1621 				type = UNDEF;
1622 			}
1623 		}
1624 
1625 		/*
1626 		 * This needs to become a contract with the driver for
1627 		 * the fast path.
1628 		 *
1629 		 * In the normal case the packet will have at least the L2
1630 		 * header and the IP + Transport header in the same mblk.
1631 		 * This is usually the case when the NIC driver sends up
1632 		 * the packet. This is also true when the stack generates
1633 		 * a packet that is looped back and when the stack uses the
1634 		 * fastpath mechanism. The normal case is optimized for
1635 		 * performance and may bypass DLS. All other cases go through
1636 		 * the 'OTH' type path without DLS bypass.
1637 		 */
1638 
1639 		ipha = (ipha_t *)(mp->b_rptr + hdrsize);
1640 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha))
1641 			type = OTH;
1642 
1643 		if (type == OTH) {
1644 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
1645 			    cnt[type], bw_ctl, sz[type], sz1, mp);
1646 			continue;
1647 		}
1648 
1649 		ASSERT(type == UNDEF);
1650 		/*
1651 		 * We look for at least 4 bytes past the IP header to get
1652 		 * the port information. If we get an IP fragment, we don't
1653 		 * have the port information, and we use just the protocol
1654 		 * information.
1655 		 */
1656 		switch (ipha->ipha_protocol) {
1657 		case IPPROTO_TCP:
1658 			type = V4_TCP;
1659 			mp->b_rptr += hdrsize;
1660 			break;
1661 		case IPPROTO_UDP:
1662 			type = V4_UDP;
1663 			mp->b_rptr += hdrsize;
1664 			break;
1665 		default:
1666 			type = OTH;
1667 			break;
1668 		}
1669 
1670 		FANOUT_ENQUEUE_MP(headmp[type], tailmp[type], cnt[type],
1671 		    bw_ctl, sz[type], sz1, mp);
1672 	}
1673 
1674 	for (type = V4_TCP; type < UNDEF; type++) {
1675 		if (headmp[type] != NULL) {
1676 			mac_soft_ring_t			*softring;
1677 
1678 			ASSERT(tailmp[type]->b_next == NULL);
1679 			switch (type) {
1680 			case V4_TCP:
1681 				softring = mac_srs->srs_tcp_soft_rings[0];
1682 				break;
1683 			case V4_UDP:
1684 				softring = mac_srs->srs_udp_soft_rings[0];
1685 				break;
1686 			case OTH:
1687 				softring = mac_srs->srs_oth_soft_rings[0];
1688 			}
1689 			mac_rx_soft_ring_process(mcip, softring,
1690 			    headmp[type], tailmp[type], cnt[type], sz[type]);
1691 		}
1692 	}
1693 }
1694 
1695 int	fanout_unaligned = 0;
1696 
1697 /*
1698  * mac_rx_srs_long_fanout
1699  *
1700  * The fanout routine for VLANs, and for anything else that isn't performing
1701  * explicit dls bypass.  Returns -1 on an error (drop the packet due to a
1702  * malformed packet), 0 on success, with values written in *indx and *type.
1703  */
1704 static int
1705 mac_rx_srs_long_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *mp,
1706     uint32_t sap, size_t hdrsize, enum pkt_type *type, uint_t *indx)
1707 {
1708 	ip6_t		*ip6h;
1709 	ipha_t		*ipha;
1710 	uint8_t		*whereptr;
1711 	uint_t		hash;
1712 	uint16_t	remlen;
1713 	uint8_t		nexthdr;
1714 	uint16_t	hdr_len;
1715 	uint32_t	src_val, dst_val;
1716 	boolean_t	modifiable = B_TRUE;
1717 	boolean_t	v6;
1718 
1719 	ASSERT(MBLKL(mp) >= hdrsize);
1720 
1721 	if (sap == ETHERTYPE_IPV6) {
1722 		v6 = B_TRUE;
1723 		hdr_len = IPV6_HDR_LEN;
1724 	} else if (sap == ETHERTYPE_IP) {
1725 		v6 = B_FALSE;
1726 		hdr_len = IP_SIMPLE_HDR_LENGTH;
1727 	} else {
1728 		*indx = 0;
1729 		*type = OTH;
1730 		return (0);
1731 	}
1732 
1733 	ip6h = (ip6_t *)(mp->b_rptr + hdrsize);
1734 	ipha = (ipha_t *)ip6h;
1735 
1736 	if ((uint8_t *)ip6h == mp->b_wptr) {
1737 		/*
1738 		 * The first mblk_t only includes the mac header.
1739 		 * Note that it is safe to change the mp pointer here,
1740 		 * as the subsequent operation does not assume mp
1741 		 * points to the start of the mac header.
1742 		 */
1743 		mp = mp->b_cont;
1744 
1745 		/*
1746 		 * Make sure the IP header points to an entire one.
1747 		 */
1748 		if (mp == NULL)
1749 			return (-1);
1750 
1751 		if (MBLKL(mp) < hdr_len) {
1752 			modifiable = (DB_REF(mp) == 1);
1753 
1754 			if (modifiable && !pullupmsg(mp, hdr_len))
1755 				return (-1);
1756 		}
1757 
1758 		ip6h = (ip6_t *)mp->b_rptr;
1759 		ipha = (ipha_t *)ip6h;
1760 	}
1761 
1762 	if (!modifiable || !(OK_32PTR((char *)ip6h)) ||
1763 	    ((uint8_t *)ip6h + hdr_len > mp->b_wptr)) {
1764 		/*
1765 		 * If either the IP header is not aligned, or it does not hold
1766 		 * the complete simple structure (a pullupmsg() is not an
1767 		 * option since it would result in an unaligned IP header),
1768 		 * fanout to the default ring.
1769 		 *
1770 		 * Note that this may cause packet reordering.
1771 		 */
1772 		*indx = 0;
1773 		*type = OTH;
1774 		fanout_unaligned++;
1775 		return (0);
1776 	}
1777 
1778 	/*
1779 	 * Extract next-header, full header length, and source-hash value
1780 	 * using v4/v6 specific fields.
1781 	 */
1782 	if (v6) {
1783 		remlen = ntohs(ip6h->ip6_plen);
1784 		nexthdr = ip6h->ip6_nxt;
1785 		src_val = V4_PART_OF_V6(ip6h->ip6_src);
1786 		dst_val = V4_PART_OF_V6(ip6h->ip6_dst);
1787 		/*
1788 		 * Do src based fanout if below tunable is set to B_TRUE or
1789 		 * when mac_ip_hdr_length_v6() fails because of malformed
1790 		 * packets or because mblks need to be concatenated using
1791 		 * pullupmsg().
1792 		 *
1793 		 * Perform a version check to prevent parsing weirdness...
1794 		 */
1795 		if (IPH_HDR_VERSION(ip6h) != IPV6_VERSION ||
1796 		    !mac_ip_hdr_length_v6(ip6h, mp->b_wptr, &hdr_len, &nexthdr,
1797 		    NULL)) {
1798 			goto src_dst_based_fanout;
1799 		}
1800 	} else {
1801 		hdr_len = IPH_HDR_LENGTH(ipha);
1802 		remlen = ntohs(ipha->ipha_length) - hdr_len;
1803 		nexthdr = ipha->ipha_protocol;
1804 		src_val = (uint32_t)ipha->ipha_src;
1805 		dst_val = (uint32_t)ipha->ipha_dst;
1806 		/*
1807 		 * Catch IPv4 fragment case here.  IPv6 has nexthdr == FRAG
1808 		 * for its equivalent case.
1809 		 */
1810 		if ((ntohs(ipha->ipha_fragment_offset_and_flags) &
1811 		    (IPH_MF | IPH_OFFSET)) != 0) {
1812 			goto src_dst_based_fanout;
1813 		}
1814 	}
1815 	if (remlen < MIN_EHDR_LEN)
1816 		return (-1);
1817 	whereptr = (uint8_t *)ip6h + hdr_len;
1818 
1819 	/* If the transport is one of below, we do port/SPI based fanout */
1820 	switch (nexthdr) {
1821 	case IPPROTO_TCP:
1822 	case IPPROTO_UDP:
1823 	case IPPROTO_SCTP:
1824 	case IPPROTO_ESP:
1825 		/*
1826 		 * If the ports or SPI in the transport header is not part of
1827 		 * the mblk, do src_based_fanout, instead of calling
1828 		 * pullupmsg().
1829 		 */
1830 		if (mp->b_cont == NULL || whereptr + PORTS_SIZE <= mp->b_wptr)
1831 			break;	/* out of switch... */
1832 		/* FALLTHRU */
1833 	default:
1834 		goto src_dst_based_fanout;
1835 	}
1836 
1837 	switch (nexthdr) {
1838 	case IPPROTO_TCP:
1839 		hash = HASH_ADDR(src_val, dst_val, *(uint32_t *)whereptr);
1840 		*indx = COMPUTE_INDEX(hash, mac_srs->srs_tcp_ring_count);
1841 		*type = OTH;
1842 		break;
1843 	case IPPROTO_UDP:
1844 	case IPPROTO_SCTP:
1845 	case IPPROTO_ESP:
1846 		if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
1847 			hash = HASH_ADDR(src_val, dst_val,
1848 			    *(uint32_t *)whereptr);
1849 			*indx = COMPUTE_INDEX(hash,
1850 			    mac_srs->srs_udp_ring_count);
1851 		} else {
1852 			*indx = mac_srs->srs_ind % mac_srs->srs_udp_ring_count;
1853 			mac_srs->srs_ind++;
1854 		}
1855 		*type = OTH;
1856 		break;
1857 	}
1858 	return (0);
1859 
1860 src_dst_based_fanout:
1861 	hash = HASH_ADDR(src_val, dst_val, (uint32_t)0);
1862 	*indx = COMPUTE_INDEX(hash, mac_srs->srs_oth_ring_count);
1863 	*type = OTH;
1864 	return (0);
1865 }
1866 
1867 /*
1868  * mac_rx_srs_fanout
1869  *
1870  * This routine delivers packets destined to an SRS into a soft ring member
1871  * of the set.
1872  *
1873  * Given a chain of packets we need to split it up into multiple sub chains
1874  * destined for one of the TCP, UDP or OTH soft rings. Instead of entering
1875  * the soft ring one packet at a time, we want to enter it in the form of a
1876  * chain otherwise we get this start/stop behaviour where the worker thread
1877  * goes to sleep and then next packets comes in forcing it to wake up etc.
1878  *
1879  * Note:
1880  * Since we know what is the maximum fanout possible, we create a 2D array
1881  * of 'softring types * MAX_SR_FANOUT' for the head, tail, cnt and sz
1882  * variables so that we can enter the softrings with chain. We need the
1883  * MAX_SR_FANOUT so we can allocate the arrays on the stack (a kmem_alloc
1884  * for each packet would be expensive). If we ever want to have the
1885  * ability to have unlimited fanout, we should probably declare a head,
1886  * tail, cnt, sz with each soft ring (a data struct which contains a softring
1887  * along with these members) and create an array of this uber struct so we
1888  * don't have to do kmem_alloc.
1889  */
1890 int	fanout_oth1 = 0;
1891 int	fanout_oth2 = 0;
1892 int	fanout_oth3 = 0;
1893 int	fanout_oth4 = 0;
1894 int	fanout_oth5 = 0;
1895 
1896 static void
1897 mac_rx_srs_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
1898 {
1899 	struct ether_header		*ehp;
1900 	struct ether_vlan_header	*evhp;
1901 	uint32_t			sap;
1902 	ipha_t				*ipha;
1903 	uint8_t				*dstaddr;
1904 	uint_t				indx;
1905 	size_t				ports_offset;
1906 	size_t				ipha_len;
1907 	size_t				hdrsize;
1908 	uint_t				hash;
1909 	mblk_t				*mp;
1910 	mblk_t				*headmp[MAX_SR_TYPES][MAX_SR_FANOUT];
1911 	mblk_t				*tailmp[MAX_SR_TYPES][MAX_SR_FANOUT];
1912 	int				cnt[MAX_SR_TYPES][MAX_SR_FANOUT];
1913 	size_t				sz[MAX_SR_TYPES][MAX_SR_FANOUT];
1914 	size_t				sz1;
1915 	boolean_t			bw_ctl;
1916 	boolean_t			hw_classified;
1917 	boolean_t			dls_bypass;
1918 	boolean_t			is_ether;
1919 	boolean_t			is_unicast;
1920 	int				fanout_cnt;
1921 	enum pkt_type			type;
1922 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
1923 
1924 	is_ether = (mcip->mci_mip->mi_info.mi_nativemedia == DL_ETHER);
1925 	bw_ctl = ((mac_srs->srs_type & SRST_BW_CONTROL) != 0);
1926 
1927 	/*
1928 	 * If we don't have a Rx ring, S/W classification would have done
1929 	 * its job and its a packet meant for us. If we were polling on
1930 	 * the default ring (i.e. there was a ring assigned to this SRS),
1931 	 * then we need to make sure that the mac address really belongs
1932 	 * to us.
1933 	 */
1934 	hw_classified = mac_srs->srs_ring != NULL &&
1935 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
1936 
1937 	/*
1938 	 * Special clients (eg. VLAN, non ether, etc) need DLS
1939 	 * processing in the Rx path. SRST_DLS_BYPASS will be clear for
1940 	 * such SRSs. Another way of disabling bypass is to set the
1941 	 * MCIS_RX_BYPASS_DISABLE flag.
1942 	 */
1943 	dls_bypass = ((mac_srs->srs_type & SRST_DLS_BYPASS) != 0) &&
1944 	    ((mcip->mci_state_flags & MCIS_RX_BYPASS_DISABLE) == 0);
1945 
1946 	/*
1947 	 * Since the softrings are never destroyed and we always
1948 	 * create equal number of softrings for TCP, UDP and rest,
1949 	 * its OK to check one of them for count and use it without
1950 	 * any lock. In future, if soft rings get destroyed because
1951 	 * of reduction in fanout, we will need to ensure that happens
1952 	 * behind the SRS_PROC.
1953 	 */
1954 	fanout_cnt = mac_srs->srs_tcp_ring_count;
1955 
1956 	bzero(headmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
1957 	bzero(tailmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
1958 	bzero(cnt, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (int));
1959 	bzero(sz, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (size_t));
1960 
1961 	/*
1962 	 * We got a chain from SRS that we need to send to the soft rings.
1963 	 * Since squeues for TCP & IPv4 sap poll their soft rings (for
1964 	 * performance reasons), we need to separate out v4_tcp, v4_udp
1965 	 * and the rest goes in other.
1966 	 */
1967 	while (head != NULL) {
1968 		mp = head;
1969 		head = head->b_next;
1970 		mp->b_next = NULL;
1971 
1972 		type = OTH;
1973 		sz1 = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
1974 
1975 		if (is_ether) {
1976 			/*
1977 			 * At this point we can be sure the packet at least
1978 			 * has an ether header.
1979 			 */
1980 			if (sz1 < sizeof (struct ether_header)) {
1981 				mac_rx_drop_pkt(mac_srs, mp);
1982 				continue;
1983 			}
1984 			ehp = (struct ether_header *)mp->b_rptr;
1985 
1986 			/*
1987 			 * Determine if this is a VLAN or non-VLAN packet.
1988 			 */
1989 			if ((sap = ntohs(ehp->ether_type)) == VLAN_TPID) {
1990 				evhp = (struct ether_vlan_header *)mp->b_rptr;
1991 				sap = ntohs(evhp->ether_type);
1992 				hdrsize = sizeof (struct ether_vlan_header);
1993 				/*
1994 				 * Check if the VID of the packet, if any,
1995 				 * belongs to this client.
1996 				 */
1997 				if (!mac_client_check_flow_vid(mcip,
1998 				    VLAN_ID(ntohs(evhp->ether_tci)))) {
1999 					mac_rx_drop_pkt(mac_srs, mp);
2000 					continue;
2001 				}
2002 			} else {
2003 				hdrsize = sizeof (struct ether_header);
2004 			}
2005 			is_unicast =
2006 			    ((((uint8_t *)&ehp->ether_dhost)[0] & 0x01) == 0);
2007 			dstaddr = (uint8_t *)&ehp->ether_dhost;
2008 		} else {
2009 			mac_header_info_t		mhi;
2010 
2011 			if (mac_header_info((mac_handle_t)mcip->mci_mip,
2012 			    mp, &mhi) != 0) {
2013 				mac_rx_drop_pkt(mac_srs, mp);
2014 				continue;
2015 			}
2016 			hdrsize = mhi.mhi_hdrsize;
2017 			sap = mhi.mhi_bindsap;
2018 			is_unicast = (mhi.mhi_dsttype == MAC_ADDRTYPE_UNICAST);
2019 			dstaddr = (uint8_t *)mhi.mhi_daddr;
2020 		}
2021 
2022 		if (!dls_bypass) {
2023 			if (mac_rx_srs_long_fanout(mac_srs, mp, sap,
2024 			    hdrsize, &type, &indx) == -1) {
2025 				mac_rx_drop_pkt(mac_srs, mp);
2026 				continue;
2027 			}
2028 
2029 			FANOUT_ENQUEUE_MP(headmp[type][indx],
2030 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
2031 			    sz[type][indx], sz1, mp);
2032 			continue;
2033 		}
2034 
2035 
2036 		/*
2037 		 * If we are using the default Rx ring where H/W or S/W
2038 		 * classification has not happened, we need to verify if
2039 		 * this unicast packet really belongs to us.
2040 		 */
2041 		if (sap == ETHERTYPE_IP) {
2042 			/*
2043 			 * If we are H/W classified, but we have promisc
2044 			 * on, then we need to check for the unicast address.
2045 			 */
2046 			if (hw_classified && mcip->mci_promisc_list != NULL) {
2047 				mac_address_t		*map;
2048 
2049 				rw_enter(&mcip->mci_rw_lock, RW_READER);
2050 				map = mcip->mci_unicast;
2051 				if (bcmp(dstaddr, map->ma_addr,
2052 				    map->ma_len) == 0)
2053 					type = UNDEF;
2054 				rw_exit(&mcip->mci_rw_lock);
2055 			} else if (is_unicast) {
2056 				type = UNDEF;
2057 			}
2058 		}
2059 
2060 		/*
2061 		 * This needs to become a contract with the driver for
2062 		 * the fast path.
2063 		 */
2064 
2065 		ipha = (ipha_t *)(mp->b_rptr + hdrsize);
2066 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha)) {
2067 			type = OTH;
2068 			fanout_oth1++;
2069 		}
2070 
2071 		if (type != OTH) {
2072 			uint16_t	frag_offset_flags;
2073 
2074 			switch (ipha->ipha_protocol) {
2075 			case IPPROTO_TCP:
2076 			case IPPROTO_UDP:
2077 			case IPPROTO_SCTP:
2078 			case IPPROTO_ESP:
2079 				ipha_len = IPH_HDR_LENGTH(ipha);
2080 				if ((uchar_t *)ipha + ipha_len + PORTS_SIZE >
2081 				    mp->b_wptr) {
2082 					type = OTH;
2083 					break;
2084 				}
2085 				frag_offset_flags =
2086 				    ntohs(ipha->ipha_fragment_offset_and_flags);
2087 				if ((frag_offset_flags &
2088 				    (IPH_MF | IPH_OFFSET)) != 0) {
2089 					type = OTH;
2090 					fanout_oth3++;
2091 					break;
2092 				}
2093 				ports_offset = hdrsize + ipha_len;
2094 				break;
2095 			default:
2096 				type = OTH;
2097 				fanout_oth4++;
2098 				break;
2099 			}
2100 		}
2101 
2102 		if (type == OTH) {
2103 			if (mac_rx_srs_long_fanout(mac_srs, mp, sap,
2104 			    hdrsize, &type, &indx) == -1) {
2105 				mac_rx_drop_pkt(mac_srs, mp);
2106 				continue;
2107 			}
2108 
2109 			FANOUT_ENQUEUE_MP(headmp[type][indx],
2110 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
2111 			    sz[type][indx], sz1, mp);
2112 			continue;
2113 		}
2114 
2115 		ASSERT(type == UNDEF);
2116 
2117 		/*
2118 		 * XXX-Sunay: We should hold srs_lock since ring_count
2119 		 * below can change. But if we are always called from
2120 		 * mac_rx_srs_drain and SRS_PROC is set, then we can
2121 		 * enforce that ring_count can't be changed i.e.
2122 		 * to change fanout type or ring count, the calling
2123 		 * thread needs to be behind SRS_PROC.
2124 		 */
2125 		switch (ipha->ipha_protocol) {
2126 		case IPPROTO_TCP:
2127 			/*
2128 			 * Note that for ESP, we fanout on SPI and it is at the
2129 			 * same offset as the 2x16-bit ports. So it is clumped
2130 			 * along with TCP, UDP and SCTP.
2131 			 */
2132 			hash = HASH_ADDR(ipha->ipha_src, ipha->ipha_dst,
2133 			    *(uint32_t *)(mp->b_rptr + ports_offset));
2134 			indx = COMPUTE_INDEX(hash, mac_srs->srs_tcp_ring_count);
2135 			type = V4_TCP;
2136 			mp->b_rptr += hdrsize;
2137 			break;
2138 		case IPPROTO_UDP:
2139 		case IPPROTO_SCTP:
2140 		case IPPROTO_ESP:
2141 			if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
2142 				hash = HASH_ADDR(ipha->ipha_src, ipha->ipha_dst,
2143 				    *(uint32_t *)(mp->b_rptr + ports_offset));
2144 				indx = COMPUTE_INDEX(hash,
2145 				    mac_srs->srs_udp_ring_count);
2146 			} else {
2147 				indx = mac_srs->srs_ind %
2148 				    mac_srs->srs_udp_ring_count;
2149 				mac_srs->srs_ind++;
2150 			}
2151 			type = V4_UDP;
2152 			mp->b_rptr += hdrsize;
2153 			break;
2154 		default:
2155 			indx = 0;
2156 			type = OTH;
2157 		}
2158 
2159 		FANOUT_ENQUEUE_MP(headmp[type][indx], tailmp[type][indx],
2160 		    cnt[type][indx], bw_ctl, sz[type][indx], sz1, mp);
2161 	}
2162 
2163 	for (type = V4_TCP; type < UNDEF; type++) {
2164 		int	i;
2165 
2166 		for (i = 0; i < fanout_cnt; i++) {
2167 			if (headmp[type][i] != NULL) {
2168 				mac_soft_ring_t	*softring;
2169 
2170 				ASSERT(tailmp[type][i]->b_next == NULL);
2171 				switch (type) {
2172 				case V4_TCP:
2173 					softring =
2174 					    mac_srs->srs_tcp_soft_rings[i];
2175 					break;
2176 				case V4_UDP:
2177 					softring =
2178 					    mac_srs->srs_udp_soft_rings[i];
2179 					break;
2180 				case OTH:
2181 					softring =
2182 					    mac_srs->srs_oth_soft_rings[i];
2183 					break;
2184 				}
2185 				mac_rx_soft_ring_process(mcip,
2186 				    softring, headmp[type][i], tailmp[type][i],
2187 				    cnt[type][i], sz[type][i]);
2188 			}
2189 		}
2190 	}
2191 }
2192 
2193 #define	SRS_BYTES_TO_PICKUP	150000
2194 ssize_t	max_bytes_to_pickup = SRS_BYTES_TO_PICKUP;
2195 
2196 /*
2197  * mac_rx_srs_poll_ring
2198  *
2199  * This SRS Poll thread uses this routine to poll the underlying hardware
2200  * Rx ring to get a chain of packets. It can inline process that chain
2201  * if mac_latency_optimize is set (default) or signal the SRS worker thread
2202  * to do the remaining processing.
2203  *
2204  * Since packets come in the system via interrupt or poll path, we also
2205  * update the stats and deal with promiscous clients here.
2206  */
2207 void
2208 mac_rx_srs_poll_ring(mac_soft_ring_set_t *mac_srs)
2209 {
2210 	kmutex_t		*lock = &mac_srs->srs_lock;
2211 	kcondvar_t		*async = &mac_srs->srs_cv;
2212 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
2213 	mblk_t			*head, *tail, *mp;
2214 	callb_cpr_t		cprinfo;
2215 	ssize_t			bytes_to_pickup;
2216 	size_t			sz;
2217 	int			count;
2218 	mac_client_impl_t	*smcip;
2219 
2220 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "mac_srs_poll");
2221 	mutex_enter(lock);
2222 
2223 start:
2224 	for (;;) {
2225 		if (mac_srs->srs_state & SRS_PAUSE)
2226 			goto done;
2227 
2228 		CALLB_CPR_SAFE_BEGIN(&cprinfo);
2229 		cv_wait(async, lock);
2230 		CALLB_CPR_SAFE_END(&cprinfo, lock);
2231 
2232 		if (mac_srs->srs_state & SRS_PAUSE)
2233 			goto done;
2234 
2235 check_again:
2236 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
2237 			/*
2238 			 * We pick as many bytes as we are allowed to queue.
2239 			 * Its possible that we will exceed the total
2240 			 * packets queued in case this SRS is part of the
2241 			 * Rx ring group since > 1 poll thread can be pulling
2242 			 * upto the max allowed packets at the same time
2243 			 * but that should be OK.
2244 			 */
2245 			mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2246 			bytes_to_pickup =
2247 			    mac_srs->srs_bw->mac_bw_drop_threshold -
2248 			    mac_srs->srs_bw->mac_bw_sz;
2249 			/*
2250 			 * We shouldn't have been signalled if we
2251 			 * have 0 or less bytes to pick but since
2252 			 * some of the bytes accounting is driver
2253 			 * dependant, we do the safety check.
2254 			 */
2255 			if (bytes_to_pickup < 0)
2256 				bytes_to_pickup = 0;
2257 			mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2258 		} else {
2259 			/*
2260 			 * ToDO: Need to change the polling API
2261 			 * to add a packet count and a flag which
2262 			 * tells the driver whether we want packets
2263 			 * based on a count, or bytes, or all the
2264 			 * packets queued in the driver/HW. This
2265 			 * way, we never have to check the limits
2266 			 * on poll path. We truly let only as many
2267 			 * packets enter the system as we are willing
2268 			 * to process or queue.
2269 			 *
2270 			 * Something along the lines of
2271 			 * pkts_to_pickup = mac_soft_ring_max_q_cnt -
2272 			 *	mac_srs->srs_poll_pkt_cnt
2273 			 */
2274 
2275 			/*
2276 			 * Since we are not doing B/W control, pick
2277 			 * as many packets as allowed.
2278 			 */
2279 			bytes_to_pickup = max_bytes_to_pickup;
2280 		}
2281 
2282 		/* Poll the underlying Hardware */
2283 		mutex_exit(lock);
2284 		head = MAC_HWRING_POLL(mac_srs->srs_ring, (int)bytes_to_pickup);
2285 		mutex_enter(lock);
2286 
2287 		ASSERT((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
2288 		    SRS_POLL_THR_OWNER);
2289 
2290 		mp = tail = head;
2291 		count = 0;
2292 		sz = 0;
2293 		while (mp != NULL) {
2294 			tail = mp;
2295 			sz += msgdsize(mp);
2296 			mp = mp->b_next;
2297 			count++;
2298 		}
2299 
2300 		if (head != NULL) {
2301 			tail->b_next = NULL;
2302 			smcip = mac_srs->srs_mcip;
2303 
2304 			SRS_RX_STAT_UPDATE(mac_srs, pollbytes, sz);
2305 			SRS_RX_STAT_UPDATE(mac_srs, pollcnt, count);
2306 
2307 			/*
2308 			 * If there are any promiscuous mode callbacks
2309 			 * defined for this MAC client, pass them a copy
2310 			 * if appropriate and also update the counters.
2311 			 */
2312 			if (smcip != NULL) {
2313 				if (smcip->mci_mip->mi_promisc_list != NULL) {
2314 					mutex_exit(lock);
2315 					mac_promisc_dispatch(smcip->mci_mip,
2316 					    head, NULL);
2317 					mutex_enter(lock);
2318 				}
2319 			}
2320 			if (mac_srs->srs_type & SRST_BW_CONTROL) {
2321 				mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2322 				mac_srs->srs_bw->mac_bw_polled += sz;
2323 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2324 			}
2325 			MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail,
2326 			    count, sz);
2327 			if (count <= 10)
2328 				srs_rx->sr_stat.mrs_chaincntundr10++;
2329 			else if (count > 10 && count <= 50)
2330 				srs_rx->sr_stat.mrs_chaincnt10to50++;
2331 			else
2332 				srs_rx->sr_stat.mrs_chaincntover50++;
2333 		}
2334 
2335 		/*
2336 		 * We are guaranteed that SRS_PROC will be set if we
2337 		 * are here. Also, poll thread gets to run only if
2338 		 * the drain was being done by a worker thread although
2339 		 * its possible that worker thread is still running
2340 		 * and poll thread was sent down to keep the pipeline
2341 		 * going instead of doing a complete drain and then
2342 		 * trying to poll the NIC.
2343 		 *
2344 		 * So we need to check SRS_WORKER flag to make sure
2345 		 * that the worker thread is not processing the queue
2346 		 * in parallel to us. The flags and conditions are
2347 		 * protected by the srs_lock to prevent any race. We
2348 		 * ensure that we don't drop the srs_lock from now
2349 		 * till the end and similarly we don't drop the srs_lock
2350 		 * in mac_rx_srs_drain() till similar condition check
2351 		 * are complete. The mac_rx_srs_drain() needs to ensure
2352 		 * that SRS_WORKER flag remains set as long as its
2353 		 * processing the queue.
2354 		 */
2355 		if (!(mac_srs->srs_state & SRS_WORKER) &&
2356 		    (mac_srs->srs_first != NULL)) {
2357 			/*
2358 			 * We have packets to process and worker thread
2359 			 * is not running. Check to see if poll thread is
2360 			 * allowed to process.
2361 			 */
2362 			if (mac_srs->srs_state & SRS_LATENCY_OPT) {
2363 				mac_srs->srs_drain_func(mac_srs, SRS_POLL_PROC);
2364 				if (!(mac_srs->srs_state & SRS_PAUSE) &&
2365 				    srs_rx->sr_poll_pkt_cnt <=
2366 				    srs_rx->sr_lowat) {
2367 					srs_rx->sr_poll_again++;
2368 					goto check_again;
2369 				}
2370 				/*
2371 				 * We are already above low water mark
2372 				 * so stay in the polling mode but no
2373 				 * need to poll. Once we dip below
2374 				 * the polling threshold, the processing
2375 				 * thread (soft ring) will signal us
2376 				 * to poll again (MAC_UPDATE_SRS_COUNT)
2377 				 */
2378 				srs_rx->sr_poll_drain_no_poll++;
2379 				mac_srs->srs_state &= ~(SRS_PROC|SRS_GET_PKTS);
2380 				/*
2381 				 * In B/W control case, its possible
2382 				 * that the backlog built up due to
2383 				 * B/W limit being reached and packets
2384 				 * are queued only in SRS. In this case,
2385 				 * we should schedule worker thread
2386 				 * since no one else will wake us up.
2387 				 */
2388 				if ((mac_srs->srs_type & SRST_BW_CONTROL) &&
2389 				    (mac_srs->srs_tid == NULL)) {
2390 					mac_srs->srs_tid =
2391 					    timeout(mac_srs_fire, mac_srs, 1);
2392 					srs_rx->sr_poll_worker_wakeup++;
2393 				}
2394 			} else {
2395 				/*
2396 				 * Wakeup the worker thread for more processing.
2397 				 * We optimize for throughput in this case.
2398 				 */
2399 				mac_srs->srs_state &= ~(SRS_PROC|SRS_GET_PKTS);
2400 				MAC_SRS_WORKER_WAKEUP(mac_srs);
2401 				srs_rx->sr_poll_sig_worker++;
2402 			}
2403 		} else if ((mac_srs->srs_first == NULL) &&
2404 		    !(mac_srs->srs_state & SRS_WORKER)) {
2405 			/*
2406 			 * There is nothing queued in SRS and
2407 			 * no worker thread running. Plus we
2408 			 * didn't get anything from the H/W
2409 			 * as well (head == NULL);
2410 			 */
2411 			ASSERT(head == NULL);
2412 			mac_srs->srs_state &=
2413 			    ~(SRS_PROC|SRS_GET_PKTS);
2414 
2415 			/*
2416 			 * If we have a packets in soft ring, don't allow
2417 			 * more packets to come into this SRS by keeping the
2418 			 * interrupts off but not polling the H/W. The
2419 			 * poll thread will get signaled as soon as
2420 			 * srs_poll_pkt_cnt dips below poll threshold.
2421 			 */
2422 			if (srs_rx->sr_poll_pkt_cnt == 0) {
2423 				srs_rx->sr_poll_intr_enable++;
2424 				MAC_SRS_POLLING_OFF(mac_srs);
2425 			} else {
2426 				/*
2427 				 * We know nothing is queued in SRS
2428 				 * since we are here after checking
2429 				 * srs_first is NULL. The backlog
2430 				 * is entirely due to packets queued
2431 				 * in Soft ring which will wake us up
2432 				 * and get the interface out of polling
2433 				 * mode once the backlog dips below
2434 				 * sr_poll_thres.
2435 				 */
2436 				srs_rx->sr_poll_no_poll++;
2437 			}
2438 		} else {
2439 			/*
2440 			 * Worker thread is already running.
2441 			 * Nothing much to do. If the polling
2442 			 * was enabled, worker thread will deal
2443 			 * with that.
2444 			 */
2445 			mac_srs->srs_state &= ~SRS_GET_PKTS;
2446 			srs_rx->sr_poll_goto_sleep++;
2447 		}
2448 	}
2449 done:
2450 	mac_srs->srs_state |= SRS_POLL_THR_QUIESCED;
2451 	cv_signal(&mac_srs->srs_async);
2452 	/*
2453 	 * If this is a temporary quiesce then wait for the restart signal
2454 	 * from the srs worker. Then clear the flags and signal the srs worker
2455 	 * to ensure a positive handshake and go back to start.
2456 	 */
2457 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_POLL_THR_RESTART)))
2458 		cv_wait(async, lock);
2459 	if (mac_srs->srs_state & SRS_POLL_THR_RESTART) {
2460 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
2461 		mac_srs->srs_state &=
2462 		    ~(SRS_POLL_THR_QUIESCED | SRS_POLL_THR_RESTART);
2463 		cv_signal(&mac_srs->srs_async);
2464 		goto start;
2465 	} else {
2466 		mac_srs->srs_state |= SRS_POLL_THR_EXITED;
2467 		cv_signal(&mac_srs->srs_async);
2468 		CALLB_CPR_EXIT(&cprinfo);
2469 		thread_exit();
2470 	}
2471 }
2472 
2473 /*
2474  * mac_srs_pick_chain
2475  *
2476  * In Bandwidth control case, checks how many packets can be processed
2477  * and return them in a sub chain.
2478  */
2479 static mblk_t *
2480 mac_srs_pick_chain(mac_soft_ring_set_t *mac_srs, mblk_t **chain_tail,
2481     size_t *chain_sz, int *chain_cnt)
2482 {
2483 	mblk_t			*head = NULL;
2484 	mblk_t			*tail = NULL;
2485 	size_t			sz;
2486 	size_t			tsz = 0;
2487 	int			cnt = 0;
2488 	mblk_t			*mp;
2489 
2490 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2491 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2492 	if (((mac_srs->srs_bw->mac_bw_used + mac_srs->srs_size) <=
2493 	    mac_srs->srs_bw->mac_bw_limit) ||
2494 	    (mac_srs->srs_bw->mac_bw_limit == 0)) {
2495 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2496 		head = mac_srs->srs_first;
2497 		mac_srs->srs_first = NULL;
2498 		*chain_tail = mac_srs->srs_last;
2499 		mac_srs->srs_last = NULL;
2500 		*chain_sz = mac_srs->srs_size;
2501 		*chain_cnt = mac_srs->srs_count;
2502 		mac_srs->srs_count = 0;
2503 		mac_srs->srs_size = 0;
2504 		return (head);
2505 	}
2506 
2507 	/*
2508 	 * Can't clear the entire backlog.
2509 	 * Need to find how many packets to pick
2510 	 */
2511 	ASSERT(MUTEX_HELD(&mac_srs->srs_bw->mac_bw_lock));
2512 	while ((mp = mac_srs->srs_first) != NULL) {
2513 		sz = msgdsize(mp);
2514 		if ((tsz + sz + mac_srs->srs_bw->mac_bw_used) >
2515 		    mac_srs->srs_bw->mac_bw_limit) {
2516 			if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED))
2517 				mac_srs->srs_bw->mac_bw_state |=
2518 				    SRS_BW_ENFORCED;
2519 			break;
2520 		}
2521 
2522 		/*
2523 		 * The _size & cnt is  decremented from the softrings
2524 		 * when they send up the packet for polling to work
2525 		 * properly.
2526 		 */
2527 		tsz += sz;
2528 		cnt++;
2529 		mac_srs->srs_count--;
2530 		mac_srs->srs_size -= sz;
2531 		if (tail != NULL)
2532 			tail->b_next = mp;
2533 		else
2534 			head = mp;
2535 		tail = mp;
2536 		mac_srs->srs_first = mac_srs->srs_first->b_next;
2537 	}
2538 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2539 	if (mac_srs->srs_first == NULL)
2540 		mac_srs->srs_last = NULL;
2541 
2542 	if (tail != NULL)
2543 		tail->b_next = NULL;
2544 	*chain_tail = tail;
2545 	*chain_cnt = cnt;
2546 	*chain_sz = tsz;
2547 
2548 	return (head);
2549 }
2550 
2551 /*
2552  * mac_rx_srs_drain
2553  *
2554  * The SRS drain routine. Gets to run to clear the queue. Any thread
2555  * (worker, interrupt, poll) can call this based on processing model.
2556  * The first thing we do is disable interrupts if possible and then
2557  * drain the queue. we also try to poll the underlying hardware if
2558  * there is a dedicated hardware Rx ring assigned to this SRS.
2559  *
2560  * There is a equivalent drain routine in bandwidth control mode
2561  * mac_rx_srs_drain_bw. There is some code duplication between the two
2562  * routines but they are highly performance sensitive and are easier
2563  * to read/debug if they stay separate. Any code changes here might
2564  * also apply to mac_rx_srs_drain_bw as well.
2565  */
2566 void
2567 mac_rx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
2568 {
2569 	mblk_t			*head;
2570 	mblk_t			*tail;
2571 	timeout_id_t		tid;
2572 	int			cnt = 0;
2573 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
2574 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
2575 
2576 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2577 	ASSERT(!(mac_srs->srs_type & SRST_BW_CONTROL));
2578 
2579 	/* If we are blanked i.e. can't do upcalls, then we are done */
2580 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
2581 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
2582 		    (mac_srs->srs_state & SRS_PAUSE));
2583 		goto out;
2584 	}
2585 
2586 	if (mac_srs->srs_first == NULL)
2587 		goto out;
2588 
2589 	if (!(mac_srs->srs_state & SRS_LATENCY_OPT) &&
2590 	    (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)) {
2591 		/*
2592 		 * In the normal case, the SRS worker thread does no
2593 		 * work and we wait for a backlog to build up before
2594 		 * we switch into polling mode. In case we are
2595 		 * optimizing for throughput, we use the worker thread
2596 		 * as well. The goal is to let worker thread process
2597 		 * the queue and poll thread to feed packets into
2598 		 * the queue. As such, we should signal the poll
2599 		 * thread to try and get more packets.
2600 		 *
2601 		 * We could have pulled this check in the POLL_RING
2602 		 * macro itself but keeping it explicit here makes
2603 		 * the architecture more human understandable.
2604 		 */
2605 		MAC_SRS_POLL_RING(mac_srs);
2606 	}
2607 
2608 again:
2609 	head = mac_srs->srs_first;
2610 	mac_srs->srs_first = NULL;
2611 	tail = mac_srs->srs_last;
2612 	mac_srs->srs_last = NULL;
2613 	cnt = mac_srs->srs_count;
2614 	mac_srs->srs_count = 0;
2615 
2616 	ASSERT(head != NULL);
2617 	ASSERT(tail != NULL);
2618 
2619 	if ((tid = mac_srs->srs_tid) != NULL)
2620 		mac_srs->srs_tid = NULL;
2621 
2622 	mac_srs->srs_state |= (SRS_PROC|proc_type);
2623 
2624 
2625 	/*
2626 	 * mcip is NULL for broadcast and multicast flows. The promisc
2627 	 * callbacks for broadcast and multicast packets are delivered from
2628 	 * mac_rx() and we don't need to worry about that case in this path
2629 	 */
2630 	if (mcip != NULL) {
2631 		if (mcip->mci_promisc_list != NULL) {
2632 			mutex_exit(&mac_srs->srs_lock);
2633 			mac_promisc_client_dispatch(mcip, head);
2634 			mutex_enter(&mac_srs->srs_lock);
2635 		}
2636 		if (MAC_PROTECT_ENABLED(mcip, MPT_IPNOSPOOF)) {
2637 			mutex_exit(&mac_srs->srs_lock);
2638 			mac_protect_intercept_dynamic(mcip, head);
2639 			mutex_enter(&mac_srs->srs_lock);
2640 		}
2641 	}
2642 
2643 	/*
2644 	 * Check if SRS itself is doing the processing
2645 	 * This direct path does not apply when subflows are present. In this
2646 	 * case, packets need to be dispatched to a soft ring according to the
2647 	 * flow's bandwidth and other resources contraints.
2648 	 */
2649 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
2650 		mac_direct_rx_t		proc;
2651 		void			*arg1;
2652 		mac_resource_handle_t	arg2;
2653 
2654 		/*
2655 		 * This is the case when a Rx is directly
2656 		 * assigned and we have a fully classified
2657 		 * protocol chain. We can deal with it in
2658 		 * one shot.
2659 		 */
2660 		proc = srs_rx->sr_func;
2661 		arg1 = srs_rx->sr_arg1;
2662 		arg2 = srs_rx->sr_arg2;
2663 
2664 		mac_srs->srs_state |= SRS_CLIENT_PROC;
2665 		mutex_exit(&mac_srs->srs_lock);
2666 		if (tid != NULL) {
2667 			(void) untimeout(tid);
2668 			tid = NULL;
2669 		}
2670 
2671 		proc(arg1, arg2, head, NULL);
2672 		/*
2673 		 * Decrement the size and count here itelf
2674 		 * since the packet has been processed.
2675 		 */
2676 		mutex_enter(&mac_srs->srs_lock);
2677 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
2678 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
2679 			cv_signal(&mac_srs->srs_client_cv);
2680 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
2681 	} else {
2682 		/* Some kind of softrings based fanout is required */
2683 		mutex_exit(&mac_srs->srs_lock);
2684 		if (tid != NULL) {
2685 			(void) untimeout(tid);
2686 			tid = NULL;
2687 		}
2688 
2689 		/*
2690 		 * Since the fanout routines can deal with chains,
2691 		 * shoot the entire chain up.
2692 		 */
2693 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
2694 			mac_rx_srs_fanout(mac_srs, head);
2695 		else
2696 			mac_rx_srs_proto_fanout(mac_srs, head);
2697 		mutex_enter(&mac_srs->srs_lock);
2698 	}
2699 
2700 	if (!(mac_srs->srs_state & (SRS_BLANK|SRS_PAUSE)) &&
2701 	    (mac_srs->srs_first != NULL)) {
2702 		/*
2703 		 * More packets arrived while we were clearing the
2704 		 * SRS. This can be possible because of one of
2705 		 * three conditions below:
2706 		 * 1) The driver is using multiple worker threads
2707 		 *    to send the packets to us.
2708 		 * 2) The driver has a race in switching
2709 		 *    between interrupt and polling mode or
2710 		 * 3) Packets are arriving in this SRS via the
2711 		 *    S/W classification as well.
2712 		 *
2713 		 * We should switch to polling mode and see if we
2714 		 * need to send the poll thread down. Also, signal
2715 		 * the worker thread to process whats just arrived.
2716 		 */
2717 		MAC_SRS_POLLING_ON(mac_srs);
2718 		if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat) {
2719 			srs_rx->sr_drain_poll_sig++;
2720 			MAC_SRS_POLL_RING(mac_srs);
2721 		}
2722 
2723 		/*
2724 		 * If we didn't signal the poll thread, we need
2725 		 * to deal with the pending packets ourselves.
2726 		 */
2727 		if (proc_type == SRS_WORKER) {
2728 			srs_rx->sr_drain_again++;
2729 			goto again;
2730 		} else {
2731 			srs_rx->sr_drain_worker_sig++;
2732 			cv_signal(&mac_srs->srs_async);
2733 		}
2734 	}
2735 
2736 out:
2737 	if (mac_srs->srs_state & SRS_GET_PKTS) {
2738 		/*
2739 		 * Poll thread is already running. Leave the
2740 		 * SRS_RPOC set and hand over the control to
2741 		 * poll thread.
2742 		 */
2743 		mac_srs->srs_state &= ~proc_type;
2744 		srs_rx->sr_drain_poll_running++;
2745 		return;
2746 	}
2747 
2748 	/*
2749 	 * Even if there are no packets queued in SRS, we
2750 	 * need to make sure that the shared counter is
2751 	 * clear and any associated softrings have cleared
2752 	 * all the backlog. Otherwise, leave the interface
2753 	 * in polling mode and the poll thread will get
2754 	 * signalled once the count goes down to zero.
2755 	 *
2756 	 * If someone is already draining the queue (SRS_PROC is
2757 	 * set) when the srs_poll_pkt_cnt goes down to zero,
2758 	 * then it means that drain is already running and we
2759 	 * will turn off polling at that time if there is
2760 	 * no backlog.
2761 	 *
2762 	 * As long as there are packets queued either
2763 	 * in soft ring set or its soft rings, we will leave
2764 	 * the interface in polling mode (even if the drain
2765 	 * was done being the interrupt thread). We signal
2766 	 * the poll thread as well if we have dipped below
2767 	 * low water mark.
2768 	 *
2769 	 * NOTE: We can't use the MAC_SRS_POLLING_ON macro
2770 	 * since that turn polling on only for worker thread.
2771 	 * Its not worth turning polling on for interrupt
2772 	 * thread (since NIC will not issue another interrupt)
2773 	 * unless a backlog builds up.
2774 	 */
2775 	if ((srs_rx->sr_poll_pkt_cnt > 0) &&
2776 	    (mac_srs->srs_state & SRS_POLLING_CAPAB)) {
2777 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2778 		srs_rx->sr_drain_keep_polling++;
2779 		MAC_SRS_POLLING_ON(mac_srs);
2780 		if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)
2781 			MAC_SRS_POLL_RING(mac_srs);
2782 		return;
2783 	}
2784 
2785 	/* Nothing else to do. Get out of poll mode */
2786 	MAC_SRS_POLLING_OFF(mac_srs);
2787 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2788 	srs_rx->sr_drain_finish_intr++;
2789 }
2790 
2791 /*
2792  * mac_rx_srs_drain_bw
2793  *
2794  * The SRS BW drain routine. Gets to run to clear the queue. Any thread
2795  * (worker, interrupt, poll) can call this based on processing model.
2796  * The first thing we do is disable interrupts if possible and then
2797  * drain the queue. we also try to poll the underlying hardware if
2798  * there is a dedicated hardware Rx ring assigned to this SRS.
2799  *
2800  * There is a equivalent drain routine in non bandwidth control mode
2801  * mac_rx_srs_drain. There is some code duplication between the two
2802  * routines but they are highly performance sensitive and are easier
2803  * to read/debug if they stay separate. Any code changes here might
2804  * also apply to mac_rx_srs_drain as well.
2805  */
2806 void
2807 mac_rx_srs_drain_bw(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
2808 {
2809 	mblk_t			*head;
2810 	mblk_t			*tail;
2811 	timeout_id_t		tid;
2812 	size_t			sz = 0;
2813 	int			cnt = 0;
2814 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
2815 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
2816 	clock_t			now;
2817 
2818 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2819 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
2820 again:
2821 	/* Check if we are doing B/W control */
2822 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2823 	now = ddi_get_lbolt();
2824 	if (mac_srs->srs_bw->mac_bw_curr_time != now) {
2825 		mac_srs->srs_bw->mac_bw_curr_time = now;
2826 		mac_srs->srs_bw->mac_bw_used = 0;
2827 		if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
2828 			mac_srs->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED;
2829 	} else if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) {
2830 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2831 		goto done;
2832 	} else if (mac_srs->srs_bw->mac_bw_used >
2833 	    mac_srs->srs_bw->mac_bw_limit) {
2834 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
2835 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2836 		goto done;
2837 	}
2838 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2839 
2840 	/* If we are blanked i.e. can't do upcalls, then we are done */
2841 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
2842 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
2843 		    (mac_srs->srs_state & SRS_PAUSE));
2844 		goto done;
2845 	}
2846 
2847 	sz = 0;
2848 	cnt = 0;
2849 	if ((head = mac_srs_pick_chain(mac_srs, &tail, &sz, &cnt)) == NULL) {
2850 		/*
2851 		 * We couldn't pick up a single packet.
2852 		 */
2853 		mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2854 		if ((mac_srs->srs_bw->mac_bw_used == 0) &&
2855 		    (mac_srs->srs_size != 0) &&
2856 		    !(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
2857 			/*
2858 			 * Seems like configured B/W doesn't
2859 			 * even allow processing of 1 packet
2860 			 * per tick.
2861 			 *
2862 			 * XXX: raise the limit to processing
2863 			 * at least 1 packet per tick.
2864 			 */
2865 			mac_srs->srs_bw->mac_bw_limit +=
2866 			    mac_srs->srs_bw->mac_bw_limit;
2867 			mac_srs->srs_bw->mac_bw_drop_threshold +=
2868 			    mac_srs->srs_bw->mac_bw_drop_threshold;
2869 			cmn_err(CE_NOTE, "mac_rx_srs_drain: srs(%p) "
2870 			    "raised B/W limit to %d since not even a "
2871 			    "single packet can be processed per "
2872 			    "tick %d\n", (void *)mac_srs,
2873 			    (int)mac_srs->srs_bw->mac_bw_limit,
2874 			    (int)msgdsize(mac_srs->srs_first));
2875 		}
2876 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2877 		goto done;
2878 	}
2879 
2880 	ASSERT(head != NULL);
2881 	ASSERT(tail != NULL);
2882 
2883 	/* zero bandwidth: drop all and return to interrupt mode */
2884 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2885 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
2886 		srs_rx->sr_stat.mrs_sdrops += cnt;
2887 		ASSERT(mac_srs->srs_bw->mac_bw_sz >= sz);
2888 		mac_srs->srs_bw->mac_bw_sz -= sz;
2889 		mac_srs->srs_bw->mac_bw_drop_bytes += sz;
2890 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2891 		mac_pkt_drop(NULL, NULL, head, B_FALSE);
2892 		goto leave_poll;
2893 	} else {
2894 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2895 	}
2896 
2897 	if ((tid = mac_srs->srs_tid) != NULL)
2898 		mac_srs->srs_tid = NULL;
2899 
2900 	mac_srs->srs_state |= (SRS_PROC|proc_type);
2901 	MAC_SRS_WORKER_POLLING_ON(mac_srs);
2902 
2903 	/*
2904 	 * mcip is NULL for broadcast and multicast flows. The promisc
2905 	 * callbacks for broadcast and multicast packets are delivered from
2906 	 * mac_rx() and we don't need to worry about that case in this path
2907 	 */
2908 	if (mcip != NULL) {
2909 		if (mcip->mci_promisc_list != NULL) {
2910 			mutex_exit(&mac_srs->srs_lock);
2911 			mac_promisc_client_dispatch(mcip, head);
2912 			mutex_enter(&mac_srs->srs_lock);
2913 		}
2914 		if (MAC_PROTECT_ENABLED(mcip, MPT_IPNOSPOOF)) {
2915 			mutex_exit(&mac_srs->srs_lock);
2916 			mac_protect_intercept_dynamic(mcip, head);
2917 			mutex_enter(&mac_srs->srs_lock);
2918 		}
2919 	}
2920 
2921 	/*
2922 	 * Check if SRS itself is doing the processing
2923 	 * This direct path does not apply when subflows are present. In this
2924 	 * case, packets need to be dispatched to a soft ring according to the
2925 	 * flow's bandwidth and other resources contraints.
2926 	 */
2927 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
2928 		mac_direct_rx_t		proc;
2929 		void			*arg1;
2930 		mac_resource_handle_t	arg2;
2931 
2932 		/*
2933 		 * This is the case when a Rx is directly
2934 		 * assigned and we have a fully classified
2935 		 * protocol chain. We can deal with it in
2936 		 * one shot.
2937 		 */
2938 		proc = srs_rx->sr_func;
2939 		arg1 = srs_rx->sr_arg1;
2940 		arg2 = srs_rx->sr_arg2;
2941 
2942 		mac_srs->srs_state |= SRS_CLIENT_PROC;
2943 		mutex_exit(&mac_srs->srs_lock);
2944 		if (tid != NULL) {
2945 			(void) untimeout(tid);
2946 			tid = NULL;
2947 		}
2948 
2949 		proc(arg1, arg2, head, NULL);
2950 		/*
2951 		 * Decrement the size and count here itelf
2952 		 * since the packet has been processed.
2953 		 */
2954 		mutex_enter(&mac_srs->srs_lock);
2955 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
2956 		MAC_UPDATE_SRS_SIZE_LOCKED(mac_srs, sz);
2957 
2958 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
2959 			cv_signal(&mac_srs->srs_client_cv);
2960 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
2961 	} else {
2962 		/* Some kind of softrings based fanout is required */
2963 		mutex_exit(&mac_srs->srs_lock);
2964 		if (tid != NULL) {
2965 			(void) untimeout(tid);
2966 			tid = NULL;
2967 		}
2968 
2969 		/*
2970 		 * Since the fanout routines can deal with chains,
2971 		 * shoot the entire chain up.
2972 		 */
2973 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
2974 			mac_rx_srs_fanout(mac_srs, head);
2975 		else
2976 			mac_rx_srs_proto_fanout(mac_srs, head);
2977 		mutex_enter(&mac_srs->srs_lock);
2978 	}
2979 
2980 	/*
2981 	 * Send the poll thread to pick up any packets arrived
2982 	 * so far. This also serves as the last check in case
2983 	 * nothing else is queued in the SRS. The poll thread
2984 	 * is signalled only in the case the drain was done
2985 	 * by the worker thread and SRS_WORKER is set. The
2986 	 * worker thread can run in parallel as long as the
2987 	 * SRS_WORKER flag is set. We we have nothing else to
2988 	 * process, we can exit while leaving SRS_PROC set
2989 	 * which gives the poll thread control to process and
2990 	 * cleanup once it returns from the NIC.
2991 	 *
2992 	 * If we have nothing else to process, we need to
2993 	 * ensure that we keep holding the srs_lock till
2994 	 * all the checks below are done and control is
2995 	 * handed to the poll thread if it was running.
2996 	 */
2997 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2998 	if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
2999 		if (mac_srs->srs_first != NULL) {
3000 			if (proc_type == SRS_WORKER) {
3001 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
3002 				if (srs_rx->sr_poll_pkt_cnt <=
3003 				    srs_rx->sr_lowat)
3004 					MAC_SRS_POLL_RING(mac_srs);
3005 				goto again;
3006 			} else {
3007 				cv_signal(&mac_srs->srs_async);
3008 			}
3009 		}
3010 	}
3011 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
3012 
3013 done:
3014 
3015 	if (mac_srs->srs_state & SRS_GET_PKTS) {
3016 		/*
3017 		 * Poll thread is already running. Leave the
3018 		 * SRS_RPOC set and hand over the control to
3019 		 * poll thread.
3020 		 */
3021 		mac_srs->srs_state &= ~proc_type;
3022 		return;
3023 	}
3024 
3025 	/*
3026 	 * If we can't process packets because we have exceeded
3027 	 * B/W limit for this tick, just set the timeout
3028 	 * and leave.
3029 	 *
3030 	 * Even if there are no packets queued in SRS, we
3031 	 * need to make sure that the shared counter is
3032 	 * clear and any associated softrings have cleared
3033 	 * all the backlog. Otherwise, leave the interface
3034 	 * in polling mode and the poll thread will get
3035 	 * signalled once the count goes down to zero.
3036 	 *
3037 	 * If someone is already draining the queue (SRS_PROC is
3038 	 * set) when the srs_poll_pkt_cnt goes down to zero,
3039 	 * then it means that drain is already running and we
3040 	 * will turn off polling at that time if there is
3041 	 * no backlog. As long as there are packets queued either
3042 	 * is soft ring set or its soft rings, we will leave
3043 	 * the interface in polling mode.
3044 	 */
3045 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
3046 	if ((mac_srs->srs_state & SRS_POLLING_CAPAB) &&
3047 	    ((mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) ||
3048 	    (srs_rx->sr_poll_pkt_cnt > 0))) {
3049 		MAC_SRS_POLLING_ON(mac_srs);
3050 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
3051 		if ((mac_srs->srs_first != NULL) &&
3052 		    (mac_srs->srs_tid == NULL))
3053 			mac_srs->srs_tid = timeout(mac_srs_fire,
3054 			    mac_srs, 1);
3055 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
3056 		return;
3057 	}
3058 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
3059 
3060 leave_poll:
3061 
3062 	/* Nothing else to do. Get out of poll mode */
3063 	MAC_SRS_POLLING_OFF(mac_srs);
3064 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
3065 }
3066 
3067 /*
3068  * mac_srs_worker
3069  *
3070  * The SRS worker routine. Drains the queue when no one else is
3071  * processing it.
3072  */
3073 void
3074 mac_srs_worker(mac_soft_ring_set_t *mac_srs)
3075 {
3076 	kmutex_t		*lock = &mac_srs->srs_lock;
3077 	kcondvar_t		*async = &mac_srs->srs_async;
3078 	callb_cpr_t		cprinfo;
3079 	boolean_t		bw_ctl_flag;
3080 
3081 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "srs_worker");
3082 	mutex_enter(lock);
3083 
3084 start:
3085 	for (;;) {
3086 		bw_ctl_flag = B_FALSE;
3087 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
3088 			MAC_SRS_BW_LOCK(mac_srs);
3089 			MAC_SRS_CHECK_BW_CONTROL(mac_srs);
3090 			if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
3091 				bw_ctl_flag = B_TRUE;
3092 			MAC_SRS_BW_UNLOCK(mac_srs);
3093 		}
3094 		/*
3095 		 * The SRS_BW_ENFORCED flag may change since we have dropped
3096 		 * the mac_bw_lock. However the drain function can handle both
3097 		 * a drainable SRS or a bandwidth controlled SRS, and the
3098 		 * effect of scheduling a timeout is to wakeup the worker
3099 		 * thread which in turn will call the drain function. Since
3100 		 * we release the srs_lock atomically only in the cv_wait there
3101 		 * isn't a fear of waiting for ever.
3102 		 */
3103 		while (((mac_srs->srs_state & SRS_PROC) ||
3104 		    (mac_srs->srs_first == NULL) || bw_ctl_flag ||
3105 		    (mac_srs->srs_state & SRS_TX_BLOCKED)) &&
3106 		    !(mac_srs->srs_state & SRS_PAUSE)) {
3107 			/*
3108 			 * If we have packets queued and we are here
3109 			 * because B/W control is in place, we better
3110 			 * schedule the worker wakeup after 1 tick
3111 			 * to see if bandwidth control can be relaxed.
3112 			 */
3113 			if (bw_ctl_flag && mac_srs->srs_tid == NULL) {
3114 				/*
3115 				 * We need to ensure that a timer  is already
3116 				 * scheduled or we force  schedule one for
3117 				 * later so that we can continue processing
3118 				 * after this  quanta is over.
3119 				 */
3120 				mac_srs->srs_tid = timeout(mac_srs_fire,
3121 				    mac_srs, 1);
3122 			}
3123 wait:
3124 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3125 			cv_wait(async, lock);
3126 			CALLB_CPR_SAFE_END(&cprinfo, lock);
3127 
3128 			if (mac_srs->srs_state & SRS_PAUSE)
3129 				goto done;
3130 			if (mac_srs->srs_state & SRS_PROC)
3131 				goto wait;
3132 
3133 			if (mac_srs->srs_first != NULL &&
3134 			    mac_srs->srs_type & SRST_BW_CONTROL) {
3135 				MAC_SRS_BW_LOCK(mac_srs);
3136 				if (mac_srs->srs_bw->mac_bw_state &
3137 				    SRS_BW_ENFORCED) {
3138 					MAC_SRS_CHECK_BW_CONTROL(mac_srs);
3139 				}
3140 				bw_ctl_flag = mac_srs->srs_bw->mac_bw_state &
3141 				    SRS_BW_ENFORCED;
3142 				MAC_SRS_BW_UNLOCK(mac_srs);
3143 			}
3144 		}
3145 
3146 		if (mac_srs->srs_state & SRS_PAUSE)
3147 			goto done;
3148 		mac_srs->srs_drain_func(mac_srs, SRS_WORKER);
3149 	}
3150 done:
3151 	/*
3152 	 * The Rx SRS quiesce logic first cuts off packet supply to the SRS
3153 	 * from both hard and soft classifications and waits for such threads
3154 	 * to finish before signaling the worker. So at this point the only
3155 	 * thread left that could be competing with the worker is the poll
3156 	 * thread. In the case of Tx, there shouldn't be any thread holding
3157 	 * SRS_PROC at this point.
3158 	 */
3159 	if (!(mac_srs->srs_state & SRS_PROC)) {
3160 		mac_srs->srs_state |= SRS_PROC;
3161 	} else {
3162 		ASSERT((mac_srs->srs_type & SRST_TX) == 0);
3163 		/*
3164 		 * Poll thread still owns the SRS and is still running
3165 		 */
3166 		ASSERT((mac_srs->srs_poll_thr == NULL) ||
3167 		    ((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
3168 		    SRS_POLL_THR_OWNER));
3169 	}
3170 	mac_srs_worker_quiesce(mac_srs);
3171 	/*
3172 	 * Wait for the SRS_RESTART or SRS_CONDEMNED signal from the initiator
3173 	 * of the quiesce operation
3174 	 */
3175 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_RESTART)))
3176 		cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3177 
3178 	if (mac_srs->srs_state & SRS_RESTART) {
3179 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
3180 		mac_srs_worker_restart(mac_srs);
3181 		mac_srs->srs_state &= ~SRS_PROC;
3182 		goto start;
3183 	}
3184 
3185 	if (!(mac_srs->srs_state & SRS_CONDEMNED_DONE))
3186 		mac_srs_worker_quiesce(mac_srs);
3187 
3188 	mac_srs->srs_state &= ~SRS_PROC;
3189 	/* The macro drops the srs_lock */
3190 	CALLB_CPR_EXIT(&cprinfo);
3191 	thread_exit();
3192 }
3193 
3194 /*
3195  * mac_rx_srs_subflow_process
3196  *
3197  * Receive side routine called from interrupt path when there are
3198  * sub flows present on this SRS.
3199  */
3200 /* ARGSUSED */
3201 void
3202 mac_rx_srs_subflow_process(void *arg, mac_resource_handle_t srs,
3203     mblk_t *mp_chain, boolean_t loopback)
3204 {
3205 	flow_entry_t		*flent = NULL;
3206 	flow_entry_t		*prev_flent = NULL;
3207 	mblk_t			*mp = NULL;
3208 	mblk_t			*tail = NULL;
3209 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
3210 	mac_client_impl_t	*mcip;
3211 
3212 	mcip = mac_srs->srs_mcip;
3213 	ASSERT(mcip != NULL);
3214 
3215 	/*
3216 	 * We need to determine the SRS for every packet
3217 	 * by walking the flow table, if we don't get any,
3218 	 * then we proceed using the SRS we came with.
3219 	 */
3220 	mp = tail = mp_chain;
3221 	while (mp != NULL) {
3222 
3223 		/*
3224 		 * We will increment the stats for the mactching subflow.
3225 		 * when we get the bytes/pkt count for the classified packets
3226 		 * later in mac_rx_srs_process.
3227 		 */
3228 		(void) mac_flow_lookup(mcip->mci_subflow_tab, mp,
3229 		    FLOW_INBOUND, &flent);
3230 
3231 		if (mp == mp_chain || flent == prev_flent) {
3232 			if (prev_flent != NULL)
3233 				FLOW_REFRELE(prev_flent);
3234 			prev_flent = flent;
3235 			flent = NULL;
3236 			tail = mp;
3237 			mp = mp->b_next;
3238 			continue;
3239 		}
3240 		tail->b_next = NULL;
3241 		/*
3242 		 * A null indicates, this is for the mac_srs itself.
3243 		 * XXX-venu : probably assert for fe_rx_srs_cnt == 0.
3244 		 */
3245 		if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
3246 			mac_rx_srs_process(arg,
3247 			    (mac_resource_handle_t)mac_srs, mp_chain,
3248 			    loopback);
3249 		} else {
3250 			(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
3251 			    prev_flent->fe_cb_arg2, mp_chain, loopback);
3252 			FLOW_REFRELE(prev_flent);
3253 		}
3254 		prev_flent = flent;
3255 		flent = NULL;
3256 		mp_chain = mp;
3257 		tail = mp;
3258 		mp = mp->b_next;
3259 	}
3260 	/* Last chain */
3261 	ASSERT(mp_chain != NULL);
3262 	if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
3263 		mac_rx_srs_process(arg,
3264 		    (mac_resource_handle_t)mac_srs, mp_chain, loopback);
3265 	} else {
3266 		(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
3267 		    prev_flent->fe_cb_arg2, mp_chain, loopback);
3268 		FLOW_REFRELE(prev_flent);
3269 	}
3270 }
3271 
3272 /*
3273  * mac_rx_srs_process
3274  *
3275  * Receive side routine called from the interrupt path.
3276  *
3277  * loopback is set to force a context switch on the loopback
3278  * path between MAC clients.
3279  */
3280 /* ARGSUSED */
3281 void
3282 mac_rx_srs_process(void *arg, mac_resource_handle_t srs, mblk_t *mp_chain,
3283     boolean_t loopback)
3284 {
3285 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
3286 	mblk_t			*mp, *tail, *head;
3287 	int			count = 0;
3288 	int			count1;
3289 	size_t			sz = 0;
3290 	size_t			chain_sz, sz1;
3291 	mac_bw_ctl_t		*mac_bw;
3292 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
3293 
3294 	/*
3295 	 * Set the tail, count and sz. We set the sz irrespective
3296 	 * of whether we are doing B/W control or not for the
3297 	 * purpose of updating the stats.
3298 	 */
3299 	mp = tail = mp_chain;
3300 	while (mp != NULL) {
3301 		tail = mp;
3302 		count++;
3303 		sz += msgdsize(mp);
3304 		mp = mp->b_next;
3305 	}
3306 
3307 	mutex_enter(&mac_srs->srs_lock);
3308 
3309 	if (loopback) {
3310 		SRS_RX_STAT_UPDATE(mac_srs, lclbytes, sz);
3311 		SRS_RX_STAT_UPDATE(mac_srs, lclcnt, count);
3312 
3313 	} else {
3314 		SRS_RX_STAT_UPDATE(mac_srs, intrbytes, sz);
3315 		SRS_RX_STAT_UPDATE(mac_srs, intrcnt, count);
3316 	}
3317 
3318 	/*
3319 	 * If the SRS in already being processed; has been blanked;
3320 	 * can be processed by worker thread only; or the B/W limit
3321 	 * has been reached, then queue the chain and check if
3322 	 * worker thread needs to be awakend.
3323 	 */
3324 	if (mac_srs->srs_type & SRST_BW_CONTROL) {
3325 		mac_bw = mac_srs->srs_bw;
3326 		ASSERT(mac_bw != NULL);
3327 		mutex_enter(&mac_bw->mac_bw_lock);
3328 		mac_bw->mac_bw_intr += sz;
3329 		if (mac_bw->mac_bw_limit == 0) {
3330 			/* zero bandwidth: drop all */
3331 			srs_rx->sr_stat.mrs_sdrops += count;
3332 			mac_bw->mac_bw_drop_bytes += sz;
3333 			mutex_exit(&mac_bw->mac_bw_lock);
3334 			mutex_exit(&mac_srs->srs_lock);
3335 			mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);
3336 			return;
3337 		} else {
3338 			if ((mac_bw->mac_bw_sz + sz) <=
3339 			    mac_bw->mac_bw_drop_threshold) {
3340 				mutex_exit(&mac_bw->mac_bw_lock);
3341 				MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain,
3342 				    tail, count, sz);
3343 			} else {
3344 				mp = mp_chain;
3345 				chain_sz = 0;
3346 				count1 = 0;
3347 				tail = NULL;
3348 				head = NULL;
3349 				while (mp != NULL) {
3350 					sz1 = msgdsize(mp);
3351 					if (mac_bw->mac_bw_sz + chain_sz + sz1 >
3352 					    mac_bw->mac_bw_drop_threshold)
3353 						break;
3354 					chain_sz += sz1;
3355 					count1++;
3356 					tail = mp;
3357 					mp = mp->b_next;
3358 				}
3359 				mutex_exit(&mac_bw->mac_bw_lock);
3360 				if (tail != NULL) {
3361 					head = tail->b_next;
3362 					tail->b_next = NULL;
3363 					MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs,
3364 					    mp_chain, tail, count1, chain_sz);
3365 					sz -= chain_sz;
3366 					count -= count1;
3367 				} else {
3368 					/* Can't pick up any */
3369 					head = mp_chain;
3370 				}
3371 				if (head != NULL) {
3372 					/* Drop any packet over the threshold */
3373 					srs_rx->sr_stat.mrs_sdrops += count;
3374 					mutex_enter(&mac_bw->mac_bw_lock);
3375 					mac_bw->mac_bw_drop_bytes += sz;
3376 					mutex_exit(&mac_bw->mac_bw_lock);
3377 					freemsgchain(head);
3378 				}
3379 			}
3380 			MAC_SRS_WORKER_WAKEUP(mac_srs);
3381 			mutex_exit(&mac_srs->srs_lock);
3382 			return;
3383 		}
3384 	}
3385 
3386 	/*
3387 	 * If the total number of packets queued in the SRS and
3388 	 * its associated soft rings exceeds the max allowed,
3389 	 * then drop the chain. If we are polling capable, this
3390 	 * shouldn't be happening.
3391 	 */
3392 	if (!(mac_srs->srs_type & SRST_BW_CONTROL) &&
3393 	    (srs_rx->sr_poll_pkt_cnt > srs_rx->sr_hiwat)) {
3394 		mac_bw = mac_srs->srs_bw;
3395 		srs_rx->sr_stat.mrs_sdrops += count;
3396 		mutex_enter(&mac_bw->mac_bw_lock);
3397 		mac_bw->mac_bw_drop_bytes += sz;
3398 		mutex_exit(&mac_bw->mac_bw_lock);
3399 		freemsgchain(mp_chain);
3400 		mutex_exit(&mac_srs->srs_lock);
3401 		return;
3402 	}
3403 
3404 	MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain, tail, count, sz);
3405 
3406 	if (!(mac_srs->srs_state & SRS_PROC)) {
3407 		/*
3408 		 * If we are coming via loopback, if we are not optimizing for
3409 		 * latency, or if our stack is running deep, we should signal
3410 		 * the worker thread.
3411 		 */
3412 		if (loopback || !(mac_srs->srs_state & SRS_LATENCY_OPT) ||
3413 		    MAC_RX_SRS_TOODEEP()) {
3414 			/*
3415 			 * For loopback, We need to let the worker take
3416 			 * over as we don't want to continue in the same
3417 			 * thread even if we can. This could lead to stack
3418 			 * overflows and may also end up using
3419 			 * resources (cpu) incorrectly.
3420 			 */
3421 			cv_signal(&mac_srs->srs_async);
3422 		} else {
3423 			/*
3424 			 * Seems like no one is processing the SRS and
3425 			 * there is no backlog. We also inline process
3426 			 * our packet if its a single packet in non
3427 			 * latency optimized case (in latency optimized
3428 			 * case, we inline process chains of any size).
3429 			 */
3430 			mac_srs->srs_drain_func(mac_srs, SRS_PROC_FAST);
3431 		}
3432 	}
3433 	mutex_exit(&mac_srs->srs_lock);
3434 }
3435 
3436 /* TX SIDE ROUTINES (RUNTIME) */
3437 
3438 /*
3439  * mac_tx_srs_no_desc
3440  *
3441  * This routine is called by Tx single ring default mode
3442  * when Tx ring runs out of descs.
3443  */
3444 mac_tx_cookie_t
3445 mac_tx_srs_no_desc(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3446     uint16_t flag, mblk_t **ret_mp)
3447 {
3448 	mac_tx_cookie_t cookie = 0;
3449 	mac_srs_tx_t *srs_tx = &mac_srs->srs_tx;
3450 	boolean_t wakeup_worker = B_TRUE;
3451 	uint32_t tx_mode = srs_tx->st_mode;
3452 	int cnt, sz;
3453 	mblk_t *tail;
3454 
3455 	ASSERT(tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_BW);
3456 	if (flag & MAC_DROP_ON_NO_DESC) {
3457 		MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
3458 	} else {
3459 		if (mac_srs->srs_first != NULL)
3460 			wakeup_worker = B_FALSE;
3461 		MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3462 		if (flag & MAC_TX_NO_ENQUEUE) {
3463 			/*
3464 			 * If TX_QUEUED is not set, queue the
3465 			 * packet and let mac_tx_srs_drain()
3466 			 * set the TX_BLOCKED bit for the
3467 			 * reasons explained above. Otherwise,
3468 			 * return the mblks.
3469 			 */
3470 			if (wakeup_worker) {
3471 				MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
3472 				    mp_chain, tail, cnt, sz);
3473 			} else {
3474 				MAC_TX_SET_NO_ENQUEUE(mac_srs,
3475 				    mp_chain, ret_mp, cookie);
3476 			}
3477 		} else {
3478 			MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
3479 			    tail, cnt, sz, cookie);
3480 		}
3481 		if (wakeup_worker)
3482 			cv_signal(&mac_srs->srs_async);
3483 	}
3484 	return (cookie);
3485 }
3486 
3487 /*
3488  * mac_tx_srs_enqueue
3489  *
3490  * This routine is called when Tx SRS is operating in either serializer
3491  * or bandwidth mode. In serializer mode, a packet will get enqueued
3492  * when a thread cannot enter SRS exclusively. In bandwidth mode,
3493  * packets gets queued if allowed byte-count limit for a tick is
3494  * exceeded. The action that gets taken when MAC_DROP_ON_NO_DESC and
3495  * MAC_TX_NO_ENQUEUE is set is different than when operaing in either
3496  * the default mode or fanout mode. Here packets get dropped or
3497  * returned back to the caller only after hi-watermark worth of data
3498  * is queued.
3499  */
3500 static mac_tx_cookie_t
3501 mac_tx_srs_enqueue(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3502     uint16_t flag, uintptr_t fanout_hint, mblk_t **ret_mp)
3503 {
3504 	mac_tx_cookie_t cookie = 0;
3505 	int cnt, sz;
3506 	mblk_t *tail;
3507 	boolean_t wakeup_worker = B_TRUE;
3508 
3509 	/*
3510 	 * Ignore fanout hint if we don't have multiple tx rings.
3511 	 */
3512 	if (!MAC_TX_SOFT_RINGS(mac_srs))
3513 		fanout_hint = 0;
3514 
3515 	if (mac_srs->srs_first != NULL)
3516 		wakeup_worker = B_FALSE;
3517 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3518 	if (flag & MAC_DROP_ON_NO_DESC) {
3519 		if (mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) {
3520 			MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
3521 		} else {
3522 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
3523 			    mp_chain, tail, cnt, sz);
3524 		}
3525 	} else if (flag & MAC_TX_NO_ENQUEUE) {
3526 		if ((mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) ||
3527 		    (mac_srs->srs_state & SRS_TX_WAKEUP_CLIENT)) {
3528 			MAC_TX_SET_NO_ENQUEUE(mac_srs, mp_chain,
3529 			    ret_mp, cookie);
3530 		} else {
3531 			mp_chain->b_prev = (mblk_t *)fanout_hint;
3532 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
3533 			    mp_chain, tail, cnt, sz);
3534 		}
3535 	} else {
3536 		/*
3537 		 * If you are BW_ENFORCED, just enqueue the
3538 		 * packet. srs_worker will drain it at the
3539 		 * prescribed rate. Before enqueueing, save
3540 		 * the fanout hint.
3541 		 */
3542 		mp_chain->b_prev = (mblk_t *)fanout_hint;
3543 		MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
3544 		    tail, cnt, sz, cookie);
3545 	}
3546 	if (wakeup_worker)
3547 		cv_signal(&mac_srs->srs_async);
3548 	return (cookie);
3549 }
3550 
3551 /*
3552  * There are seven tx modes:
3553  *
3554  * 1) Default mode (SRS_TX_DEFAULT)
3555  * 2) Serialization mode (SRS_TX_SERIALIZE)
3556  * 3) Fanout mode (SRS_TX_FANOUT)
3557  * 4) Bandwdith mode (SRS_TX_BW)
3558  * 5) Fanout and Bandwidth mode (SRS_TX_BW_FANOUT)
3559  * 6) aggr Tx mode (SRS_TX_AGGR)
3560  * 7) aggr Tx bw mode (SRS_TX_BW_AGGR)
3561  *
3562  * The tx mode in which an SRS operates is decided in mac_tx_srs_setup()
3563  * based on the number of Tx rings requested for an SRS and whether
3564  * bandwidth control is requested or not.
3565  *
3566  * The default mode (i.e., no fanout/no bandwidth) is used when the
3567  * underlying NIC does not have Tx rings or just one Tx ring. In this mode,
3568  * the SRS acts as a pass-thru. Packets will go directly to mac_tx_send().
3569  * When the underlying Tx ring runs out of Tx descs, it starts queueing up
3570  * packets in SRS. When flow-control is relieved, the srs_worker drains
3571  * the queued packets and informs blocked clients to restart sending
3572  * packets.
3573  *
3574  * In the SRS_TX_SERIALIZE mode, all calls to mac_tx() are serialized. This
3575  * mode is used when the link has no Tx rings or only one Tx ring.
3576  *
3577  * In the SRS_TX_FANOUT mode, packets will be fanned out to multiple
3578  * Tx rings. Each Tx ring will have a soft ring associated with it.
3579  * These soft rings will be hung off the Tx SRS. Queueing if it happens
3580  * due to lack of Tx desc will be in individual soft ring (and not srs)
3581  * associated with Tx ring.
3582  *
3583  * In the TX_BW mode, tx srs will allow packets to go down to Tx ring
3584  * only if bw is available. Otherwise the packets will be queued in
3585  * SRS. If fanout to multiple Tx rings is configured, the packets will
3586  * be fanned out among the soft rings associated with the Tx rings.
3587  *
3588  * In SRS_TX_AGGR mode, mac_tx_aggr_mode() routine is called. This routine
3589  * invokes an aggr function, aggr_find_tx_ring(), to find a pseudo Tx ring
3590  * belonging to a port on which the packet has to be sent. Aggr will
3591  * always have a pseudo Tx ring associated with it even when it is an
3592  * aggregation over a single NIC that has no Tx rings. Even in such a
3593  * case, the single pseudo Tx ring will have a soft ring associated with
3594  * it and the soft ring will hang off the SRS.
3595  *
3596  * If a bandwidth is specified for an aggr, SRS_TX_BW_AGGR mode is used.
3597  * In this mode, the bandwidth is first applied on the outgoing packets
3598  * and later mac_tx_addr_mode() function is called to send the packet out
3599  * of one of the pseudo Tx rings.
3600  *
3601  * Four flags are used in srs_state for indicating flow control
3602  * conditions : SRS_TX_BLOCKED, SRS_TX_HIWAT, SRS_TX_WAKEUP_CLIENT.
3603  * SRS_TX_BLOCKED indicates out of Tx descs. SRS expects a wakeup from the
3604  * driver below.
3605  * SRS_TX_HIWAT indicates packet count enqueued in Tx SRS exceeded Tx hiwat
3606  * and flow-control pressure is applied back to clients. The clients expect
3607  * wakeup when flow-control is relieved.
3608  * SRS_TX_WAKEUP_CLIENT get set when (flag == MAC_TX_NO_ENQUEUE) and mblk
3609  * got returned back to client either due to lack of Tx descs or due to bw
3610  * control reasons. The clients expect a wakeup when condition is relieved.
3611  *
3612  * The fourth argument to mac_tx() is the flag. Normally it will be 0 but
3613  * some clients set the following values too: MAC_DROP_ON_NO_DESC,
3614  * MAC_TX_NO_ENQUEUE
3615  * Mac clients that do not want packets to be enqueued in the mac layer set
3616  * MAC_DROP_ON_NO_DESC value. The packets won't be queued in the Tx SRS or
3617  * Tx soft rings but instead get dropped when the NIC runs out of desc. The
3618  * behaviour of this flag is different when the Tx is running in serializer
3619  * or bandwidth mode. Under these (Serializer, bandwidth) modes, the packet
3620  * get dropped when Tx high watermark is reached.
3621  * There are some mac clients like vsw, aggr that want the mblks to be
3622  * returned back to clients instead of being queued in Tx SRS (or Tx soft
3623  * rings) under flow-control (i.e., out of desc or exceeding bw limits)
3624  * conditions. These clients call mac_tx() with MAC_TX_NO_ENQUEUE flag set.
3625  * In the default and Tx fanout mode, the un-transmitted mblks will be
3626  * returned back to the clients when the driver runs out of Tx descs.
3627  * SRS_TX_WAKEUP_CLIENT (or S_RING_WAKEUP_CLIENT) will be set in SRS (or
3628  * soft ring) so that the clients can be woken up when Tx desc become
3629  * available. When running in serializer or bandwidth mode mode,
3630  * SRS_TX_WAKEUP_CLIENT will be set when tx hi-watermark is reached.
3631  */
3632 
3633 mac_tx_func_t
3634 mac_tx_get_func(uint32_t mode)
3635 {
3636 	return (mac_tx_mode_list[mode].mac_tx_func);
3637 }
3638 
3639 /* ARGSUSED */
3640 static mac_tx_cookie_t
3641 mac_tx_single_ring_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3642     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
3643 {
3644 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
3645 	mac_tx_stats_t		stats;
3646 	mac_tx_cookie_t		cookie = 0;
3647 
3648 	ASSERT(srs_tx->st_mode == SRS_TX_DEFAULT);
3649 
3650 	/* Regular case with a single Tx ring */
3651 	/*
3652 	 * SRS_TX_BLOCKED is set when underlying NIC runs
3653 	 * out of Tx descs and messages start getting
3654 	 * queued. It won't get reset until
3655 	 * tx_srs_drain() completely drains out the
3656 	 * messages.
3657 	 */
3658 	if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
3659 		/* Tx descs/resources not available */
3660 		mutex_enter(&mac_srs->srs_lock);
3661 		if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
3662 			cookie = mac_tx_srs_no_desc(mac_srs, mp_chain,
3663 			    flag, ret_mp);
3664 			mutex_exit(&mac_srs->srs_lock);
3665 			return (cookie);
3666 		}
3667 		/*
3668 		 * While we were computing mblk count, the
3669 		 * flow control condition got relieved.
3670 		 * Continue with the transmission.
3671 		 */
3672 		mutex_exit(&mac_srs->srs_lock);
3673 	}
3674 
3675 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
3676 	    mp_chain, &stats);
3677 
3678 	/*
3679 	 * Multiple threads could be here sending packets.
3680 	 * Under such conditions, it is not possible to
3681 	 * automically set SRS_TX_BLOCKED bit to indicate
3682 	 * out of tx desc condition. To atomically set
3683 	 * this, we queue the returned packet and do
3684 	 * the setting of SRS_TX_BLOCKED in
3685 	 * mac_tx_srs_drain().
3686 	 */
3687 	if (mp_chain != NULL) {
3688 		mutex_enter(&mac_srs->srs_lock);
3689 		cookie = mac_tx_srs_no_desc(mac_srs, mp_chain, flag, ret_mp);
3690 		mutex_exit(&mac_srs->srs_lock);
3691 		return (cookie);
3692 	}
3693 	SRS_TX_STATS_UPDATE(mac_srs, &stats);
3694 
3695 	return (0);
3696 }
3697 
3698 /*
3699  * mac_tx_serialize_mode
3700  *
3701  * This is an experimental mode implemented as per the request of PAE.
3702  * In this mode, all callers attempting to send a packet to the NIC
3703  * will get serialized. Only one thread at any time will access the
3704  * NIC to send the packet out.
3705  */
3706 /* ARGSUSED */
3707 static mac_tx_cookie_t
3708 mac_tx_serializer_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3709     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
3710 {
3711 	mac_tx_stats_t		stats;
3712 	mac_tx_cookie_t		cookie = 0;
3713 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
3714 
3715 	/* Single ring, serialize below */
3716 	ASSERT(srs_tx->st_mode == SRS_TX_SERIALIZE);
3717 	mutex_enter(&mac_srs->srs_lock);
3718 	if ((mac_srs->srs_first != NULL) ||
3719 	    (mac_srs->srs_state & SRS_PROC)) {
3720 		/*
3721 		 * In serialization mode, queue all packets until
3722 		 * TX_HIWAT is set.
3723 		 * If drop bit is set, drop if TX_HIWAT is set.
3724 		 * If no_enqueue is set, still enqueue until hiwat
3725 		 * is set and return mblks after TX_HIWAT is set.
3726 		 */
3727 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain,
3728 		    flag, 0, ret_mp);
3729 		mutex_exit(&mac_srs->srs_lock);
3730 		return (cookie);
3731 	}
3732 	/*
3733 	 * No packets queued, nothing on proc and no flow
3734 	 * control condition. Fast-path, ok. Do inline
3735 	 * processing.
3736 	 */
3737 	mac_srs->srs_state |= SRS_PROC;
3738 	mutex_exit(&mac_srs->srs_lock);
3739 
3740 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
3741 	    mp_chain, &stats);
3742 
3743 	mutex_enter(&mac_srs->srs_lock);
3744 	mac_srs->srs_state &= ~SRS_PROC;
3745 	if (mp_chain != NULL) {
3746 		cookie = mac_tx_srs_enqueue(mac_srs,
3747 		    mp_chain, flag, 0, ret_mp);
3748 	}
3749 	if (mac_srs->srs_first != NULL) {
3750 		/*
3751 		 * We processed inline our packet and a new
3752 		 * packet/s got queued while we were
3753 		 * processing. Wakeup srs worker
3754 		 */
3755 		cv_signal(&mac_srs->srs_async);
3756 	}
3757 	mutex_exit(&mac_srs->srs_lock);
3758 
3759 	if (cookie == 0)
3760 		SRS_TX_STATS_UPDATE(mac_srs, &stats);
3761 
3762 	return (cookie);
3763 }
3764 
3765 /*
3766  * mac_tx_fanout_mode
3767  *
3768  * In this mode, the SRS will have access to multiple Tx rings to send
3769  * the packet out. The fanout hint that is passed as an argument is
3770  * used to find an appropriate ring to fanout the traffic. Each Tx
3771  * ring, in turn,  will have a soft ring associated with it. If a Tx
3772  * ring runs out of Tx desc's the returned packet will be queued in
3773  * the soft ring associated with that Tx ring. The srs itself will not
3774  * queue any packets.
3775  */
3776 
3777 #define	MAC_TX_SOFT_RING_PROCESS(chain) {				\
3778 	index = COMPUTE_INDEX(hash, mac_srs->srs_tx_ring_count),	\
3779 	softring = mac_srs->srs_tx_soft_rings[index];			\
3780 	cookie = mac_tx_soft_ring_process(softring, chain, flag, ret_mp); \
3781 	DTRACE_PROBE2(tx__fanout, uint64_t, hash, uint_t, index);	\
3782 }
3783 
3784 static mac_tx_cookie_t
3785 mac_tx_fanout_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3786     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
3787 {
3788 	mac_soft_ring_t		*softring;
3789 	uint64_t		hash;
3790 	uint_t			index;
3791 	mac_tx_cookie_t		cookie = 0;
3792 
3793 	ASSERT(mac_srs->srs_tx.st_mode == SRS_TX_FANOUT ||
3794 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_FANOUT);
3795 	if (fanout_hint != 0) {
3796 		/*
3797 		 * The hint is specified by the caller, simply pass the
3798 		 * whole chain to the soft ring.
3799 		 */
3800 		hash = HASH_HINT(fanout_hint);
3801 		MAC_TX_SOFT_RING_PROCESS(mp_chain);
3802 	} else {
3803 		mblk_t *last_mp, *cur_mp, *sub_chain;
3804 		uint64_t last_hash = 0;
3805 		uint_t media = mac_srs->srs_mcip->mci_mip->mi_info.mi_media;
3806 
3807 		/*
3808 		 * Compute the hash from the contents (headers) of the
3809 		 * packets of the mblk chain. Split the chains into
3810 		 * subchains of the same conversation.
3811 		 *
3812 		 * Since there may be more than one ring used for
3813 		 * sub-chains of the same call, and since the caller
3814 		 * does not maintain per conversation state since it
3815 		 * passed a zero hint, unsent subchains will be
3816 		 * dropped.
3817 		 */
3818 
3819 		flag |= MAC_DROP_ON_NO_DESC;
3820 		ret_mp = NULL;
3821 
3822 		ASSERT(ret_mp == NULL);
3823 
3824 		sub_chain = NULL;
3825 		last_mp = NULL;
3826 
3827 		for (cur_mp = mp_chain; cur_mp != NULL;
3828 		    cur_mp = cur_mp->b_next) {
3829 			hash = mac_pkt_hash(media, cur_mp, MAC_PKT_HASH_L4,
3830 			    B_TRUE);
3831 			if (last_hash != 0 && hash != last_hash) {
3832 				/*
3833 				 * Starting a different subchain, send current
3834 				 * chain out.
3835 				 */
3836 				ASSERT(last_mp != NULL);
3837 				last_mp->b_next = NULL;
3838 				MAC_TX_SOFT_RING_PROCESS(sub_chain);
3839 				sub_chain = NULL;
3840 			}
3841 
3842 			/* add packet to subchain */
3843 			if (sub_chain == NULL)
3844 				sub_chain = cur_mp;
3845 			last_mp = cur_mp;
3846 			last_hash = hash;
3847 		}
3848 
3849 		if (sub_chain != NULL) {
3850 			/* send last subchain */
3851 			ASSERT(last_mp != NULL);
3852 			last_mp->b_next = NULL;
3853 			MAC_TX_SOFT_RING_PROCESS(sub_chain);
3854 		}
3855 
3856 		cookie = 0;
3857 	}
3858 
3859 	return (cookie);
3860 }
3861 
3862 /*
3863  * mac_tx_bw_mode
3864  *
3865  * In the bandwidth mode, Tx srs will allow packets to go down to Tx ring
3866  * only if bw is available. Otherwise the packets will be queued in
3867  * SRS. If the SRS has multiple Tx rings, then packets will get fanned
3868  * out to a Tx rings.
3869  */
3870 static mac_tx_cookie_t
3871 mac_tx_bw_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3872     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
3873 {
3874 	int			cnt, sz;
3875 	mblk_t			*tail;
3876 	mac_tx_cookie_t		cookie = 0;
3877 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
3878 	clock_t			now;
3879 
3880 	ASSERT(TX_BANDWIDTH_MODE(mac_srs));
3881 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
3882 	mutex_enter(&mac_srs->srs_lock);
3883 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
3884 		/*
3885 		 * zero bandwidth, no traffic is sent: drop the packets,
3886 		 * or return the whole chain if the caller requests all
3887 		 * unsent packets back.
3888 		 */
3889 		if (flag & MAC_TX_NO_ENQUEUE) {
3890 			cookie = (mac_tx_cookie_t)mac_srs;
3891 			*ret_mp = mp_chain;
3892 		} else {
3893 			MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
3894 		}
3895 		mutex_exit(&mac_srs->srs_lock);
3896 		return (cookie);
3897 	} else if ((mac_srs->srs_first != NULL) ||
3898 	    (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
3899 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
3900 		    fanout_hint, ret_mp);
3901 		mutex_exit(&mac_srs->srs_lock);
3902 		return (cookie);
3903 	}
3904 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3905 	now = ddi_get_lbolt();
3906 	if (mac_srs->srs_bw->mac_bw_curr_time != now) {
3907 		mac_srs->srs_bw->mac_bw_curr_time = now;
3908 		mac_srs->srs_bw->mac_bw_used = 0;
3909 	} else if (mac_srs->srs_bw->mac_bw_used >
3910 	    mac_srs->srs_bw->mac_bw_limit) {
3911 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
3912 		MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
3913 		    mp_chain, tail, cnt, sz);
3914 		/*
3915 		 * Wakeup worker thread. Note that worker
3916 		 * thread has to be woken up so that it
3917 		 * can fire up the timer to be woken up
3918 		 * on the next tick. Also once
3919 		 * BW_ENFORCED is set, it can only be
3920 		 * reset by srs_worker thread. Until then
3921 		 * all packets will get queued up in SRS
3922 		 * and hence this this code path won't be
3923 		 * entered until BW_ENFORCED is reset.
3924 		 */
3925 		cv_signal(&mac_srs->srs_async);
3926 		mutex_exit(&mac_srs->srs_lock);
3927 		return (cookie);
3928 	}
3929 
3930 	mac_srs->srs_bw->mac_bw_used += sz;
3931 	mutex_exit(&mac_srs->srs_lock);
3932 
3933 	if (srs_tx->st_mode == SRS_TX_BW_FANOUT) {
3934 		mac_soft_ring_t *softring;
3935 		uint_t indx, hash;
3936 
3937 		hash = HASH_HINT(fanout_hint);
3938 		indx = COMPUTE_INDEX(hash,
3939 		    mac_srs->srs_tx_ring_count);
3940 		softring = mac_srs->srs_tx_soft_rings[indx];
3941 		return (mac_tx_soft_ring_process(softring, mp_chain, flag,
3942 		    ret_mp));
3943 	} else if (srs_tx->st_mode == SRS_TX_BW_AGGR) {
3944 		return (mac_tx_aggr_mode(mac_srs, mp_chain,
3945 		    fanout_hint, flag, ret_mp));
3946 	} else {
3947 		mac_tx_stats_t		stats;
3948 
3949 		mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
3950 		    mp_chain, &stats);
3951 
3952 		if (mp_chain != NULL) {
3953 			mutex_enter(&mac_srs->srs_lock);
3954 			MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3955 			if (mac_srs->srs_bw->mac_bw_used > sz)
3956 				mac_srs->srs_bw->mac_bw_used -= sz;
3957 			else
3958 				mac_srs->srs_bw->mac_bw_used = 0;
3959 			cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
3960 			    fanout_hint, ret_mp);
3961 			mutex_exit(&mac_srs->srs_lock);
3962 			return (cookie);
3963 		}
3964 		SRS_TX_STATS_UPDATE(mac_srs, &stats);
3965 
3966 		return (0);
3967 	}
3968 }
3969 
3970 /*
3971  * mac_tx_aggr_mode
3972  *
3973  * This routine invokes an aggr function, aggr_find_tx_ring(), to find
3974  * a (pseudo) Tx ring belonging to a port on which the packet has to
3975  * be sent. aggr_find_tx_ring() first finds the outgoing port based on
3976  * L2/L3/L4 policy and then uses the fanout_hint passed to it to pick
3977  * a Tx ring from the selected port.
3978  *
3979  * Note that a port can be deleted from the aggregation. In such a case,
3980  * the aggregation layer first separates the port from the rest of the
3981  * ports making sure that port (and thus any Tx rings associated with
3982  * it) won't get selected in the call to aggr_find_tx_ring() function.
3983  * Later calls are made to mac_group_rem_ring() passing pseudo Tx ring
3984  * handles one by one which in turn will quiesce the Tx SRS and remove
3985  * the soft ring associated with the pseudo Tx ring. Unlike Rx side
3986  * where a cookie is used to protect against mac_rx_ring() calls on
3987  * rings that have been removed, no such cookie is needed on the Tx
3988  * side as the pseudo Tx ring won't be available anymore to
3989  * aggr_find_tx_ring() once the port has been removed.
3990  */
3991 static mac_tx_cookie_t
3992 mac_tx_aggr_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3993     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
3994 {
3995 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
3996 	mac_tx_ring_fn_t	find_tx_ring_fn;
3997 	mac_ring_handle_t	ring = NULL;
3998 	void			*arg;
3999 	mac_soft_ring_t		*sringp;
4000 
4001 	find_tx_ring_fn = srs_tx->st_capab_aggr.mca_find_tx_ring_fn;
4002 	arg = srs_tx->st_capab_aggr.mca_arg;
4003 	if (find_tx_ring_fn(arg, mp_chain, fanout_hint, &ring) == NULL)
4004 		return (0);
4005 	sringp = srs_tx->st_soft_rings[((mac_ring_t *)ring)->mr_index];
4006 	return (mac_tx_soft_ring_process(sringp, mp_chain, flag, ret_mp));
4007 }
4008 
4009 void
4010 mac_tx_invoke_callbacks(mac_client_impl_t *mcip, mac_tx_cookie_t cookie)
4011 {
4012 	mac_cb_t *mcb;
4013 	mac_tx_notify_cb_t *mtnfp;
4014 
4015 	/* Wakeup callback registered clients */
4016 	MAC_CALLBACK_WALKER_INC(&mcip->mci_tx_notify_cb_info);
4017 	for (mcb = mcip->mci_tx_notify_cb_list; mcb != NULL;
4018 	    mcb = mcb->mcb_nextp) {
4019 		mtnfp = (mac_tx_notify_cb_t *)mcb->mcb_objp;
4020 		mtnfp->mtnf_fn(mtnfp->mtnf_arg, cookie);
4021 	}
4022 	MAC_CALLBACK_WALKER_DCR(&mcip->mci_tx_notify_cb_info,
4023 	    &mcip->mci_tx_notify_cb_list);
4024 }
4025 
4026 /* ARGSUSED */
4027 void
4028 mac_tx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
4029 {
4030 	mblk_t			*head, *tail;
4031 	size_t			sz;
4032 	uint32_t		tx_mode;
4033 	uint_t			saved_pkt_count;
4034 	mac_tx_stats_t		stats;
4035 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
4036 	clock_t			now;
4037 
4038 	saved_pkt_count = 0;
4039 	ASSERT(mutex_owned(&mac_srs->srs_lock));
4040 	ASSERT(!(mac_srs->srs_state & SRS_PROC));
4041 
4042 	mac_srs->srs_state |= SRS_PROC;
4043 
4044 	tx_mode = srs_tx->st_mode;
4045 	if (tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_SERIALIZE) {
4046 		if (mac_srs->srs_first != NULL) {
4047 			head = mac_srs->srs_first;
4048 			tail = mac_srs->srs_last;
4049 			saved_pkt_count = mac_srs->srs_count;
4050 			mac_srs->srs_first = NULL;
4051 			mac_srs->srs_last = NULL;
4052 			mac_srs->srs_count = 0;
4053 			mutex_exit(&mac_srs->srs_lock);
4054 
4055 			head = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
4056 			    head, &stats);
4057 
4058 			mutex_enter(&mac_srs->srs_lock);
4059 			if (head != NULL) {
4060 				/* Device out of tx desc, set block */
4061 				if (head->b_next == NULL)
4062 					VERIFY(head == tail);
4063 				tail->b_next = mac_srs->srs_first;
4064 				mac_srs->srs_first = head;
4065 				mac_srs->srs_count +=
4066 				    (saved_pkt_count - stats.mts_opackets);
4067 				if (mac_srs->srs_last == NULL)
4068 					mac_srs->srs_last = tail;
4069 				MAC_TX_SRS_BLOCK(mac_srs, head);
4070 			} else {
4071 				srs_tx->st_woken_up = B_FALSE;
4072 				SRS_TX_STATS_UPDATE(mac_srs, &stats);
4073 			}
4074 		}
4075 	} else if (tx_mode == SRS_TX_BW) {
4076 		/*
4077 		 * We are here because the timer fired and we have some data
4078 		 * to tranmit. Also mac_tx_srs_worker should have reset
4079 		 * SRS_BW_ENFORCED flag
4080 		 */
4081 		ASSERT(!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED));
4082 		head = tail = mac_srs->srs_first;
4083 		while (mac_srs->srs_first != NULL) {
4084 			tail = mac_srs->srs_first;
4085 			tail->b_prev = NULL;
4086 			mac_srs->srs_first = tail->b_next;
4087 			if (mac_srs->srs_first == NULL)
4088 				mac_srs->srs_last = NULL;
4089 			mac_srs->srs_count--;
4090 			sz = msgdsize(tail);
4091 			mac_srs->srs_size -= sz;
4092 			saved_pkt_count++;
4093 			MAC_TX_UPDATE_BW_INFO(mac_srs, sz);
4094 
4095 			if (mac_srs->srs_bw->mac_bw_used <
4096 			    mac_srs->srs_bw->mac_bw_limit)
4097 				continue;
4098 
4099 			now = ddi_get_lbolt();
4100 			if (mac_srs->srs_bw->mac_bw_curr_time != now) {
4101 				mac_srs->srs_bw->mac_bw_curr_time = now;
4102 				mac_srs->srs_bw->mac_bw_used = sz;
4103 				continue;
4104 			}
4105 			mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
4106 			break;
4107 		}
4108 
4109 		ASSERT((head == NULL && tail == NULL) ||
4110 		    (head != NULL && tail != NULL));
4111 		if (tail != NULL) {
4112 			tail->b_next = NULL;
4113 			mutex_exit(&mac_srs->srs_lock);
4114 
4115 			head = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
4116 			    head, &stats);
4117 
4118 			mutex_enter(&mac_srs->srs_lock);
4119 			if (head != NULL) {
4120 				uint_t size_sent;
4121 
4122 				/* Device out of tx desc, set block */
4123 				if (head->b_next == NULL)
4124 					VERIFY(head == tail);
4125 				tail->b_next = mac_srs->srs_first;
4126 				mac_srs->srs_first = head;
4127 				mac_srs->srs_count +=
4128 				    (saved_pkt_count - stats.mts_opackets);
4129 				if (mac_srs->srs_last == NULL)
4130 					mac_srs->srs_last = tail;
4131 				size_sent = sz - stats.mts_obytes;
4132 				mac_srs->srs_size += size_sent;
4133 				mac_srs->srs_bw->mac_bw_sz += size_sent;
4134 				if (mac_srs->srs_bw->mac_bw_used > size_sent) {
4135 					mac_srs->srs_bw->mac_bw_used -=
4136 					    size_sent;
4137 				} else {
4138 					mac_srs->srs_bw->mac_bw_used = 0;
4139 				}
4140 				MAC_TX_SRS_BLOCK(mac_srs, head);
4141 			} else {
4142 				srs_tx->st_woken_up = B_FALSE;
4143 				SRS_TX_STATS_UPDATE(mac_srs, &stats);
4144 			}
4145 		}
4146 	} else if (tx_mode == SRS_TX_BW_FANOUT || tx_mode == SRS_TX_BW_AGGR) {
4147 		mblk_t *prev;
4148 		uint64_t hint;
4149 
4150 		/*
4151 		 * We are here because the timer fired and we
4152 		 * have some quota to tranmit.
4153 		 */
4154 		prev = NULL;
4155 		head = tail = mac_srs->srs_first;
4156 		while (mac_srs->srs_first != NULL) {
4157 			tail = mac_srs->srs_first;
4158 			mac_srs->srs_first = tail->b_next;
4159 			if (mac_srs->srs_first == NULL)
4160 				mac_srs->srs_last = NULL;
4161 			mac_srs->srs_count--;
4162 			sz = msgdsize(tail);
4163 			mac_srs->srs_size -= sz;
4164 			mac_srs->srs_bw->mac_bw_used += sz;
4165 			if (prev == NULL)
4166 				hint = (ulong_t)tail->b_prev;
4167 			if (hint != (ulong_t)tail->b_prev) {
4168 				prev->b_next = NULL;
4169 				mutex_exit(&mac_srs->srs_lock);
4170 				TX_SRS_TO_SOFT_RING(mac_srs, head, hint);
4171 				head = tail;
4172 				hint = (ulong_t)tail->b_prev;
4173 				mutex_enter(&mac_srs->srs_lock);
4174 			}
4175 
4176 			prev = tail;
4177 			tail->b_prev = NULL;
4178 			if (mac_srs->srs_bw->mac_bw_used <
4179 			    mac_srs->srs_bw->mac_bw_limit)
4180 				continue;
4181 
4182 			now = ddi_get_lbolt();
4183 			if (mac_srs->srs_bw->mac_bw_curr_time != now) {
4184 				mac_srs->srs_bw->mac_bw_curr_time = now;
4185 				mac_srs->srs_bw->mac_bw_used = 0;
4186 				continue;
4187 			}
4188 			mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
4189 			break;
4190 		}
4191 		ASSERT((head == NULL && tail == NULL) ||
4192 		    (head != NULL && tail != NULL));
4193 		if (tail != NULL) {
4194 			tail->b_next = NULL;
4195 			mutex_exit(&mac_srs->srs_lock);
4196 			TX_SRS_TO_SOFT_RING(mac_srs, head, hint);
4197 			mutex_enter(&mac_srs->srs_lock);
4198 		}
4199 	}
4200 	/*
4201 	 * SRS_TX_FANOUT case not considered here because packets
4202 	 * won't be queued in the SRS for this case. Packets will
4203 	 * be sent directly to soft rings underneath and if there
4204 	 * is any queueing at all, it would be in Tx side soft
4205 	 * rings.
4206 	 */
4207 
4208 	/*
4209 	 * When srs_count becomes 0, reset SRS_TX_HIWAT and
4210 	 * SRS_TX_WAKEUP_CLIENT and wakeup registered clients.
4211 	 */
4212 	if (mac_srs->srs_count == 0 && (mac_srs->srs_state &
4213 	    (SRS_TX_HIWAT | SRS_TX_WAKEUP_CLIENT | SRS_ENQUEUED))) {
4214 		mac_client_impl_t *mcip = mac_srs->srs_mcip;
4215 		boolean_t wakeup_required = B_FALSE;
4216 
4217 		if (mac_srs->srs_state &
4218 		    (SRS_TX_HIWAT|SRS_TX_WAKEUP_CLIENT)) {
4219 			wakeup_required = B_TRUE;
4220 		}
4221 		mac_srs->srs_state &= ~(SRS_TX_HIWAT |
4222 		    SRS_TX_WAKEUP_CLIENT | SRS_ENQUEUED);
4223 		mutex_exit(&mac_srs->srs_lock);
4224 		if (wakeup_required) {
4225 			mac_tx_invoke_callbacks(mcip, (mac_tx_cookie_t)mac_srs);
4226 			/*
4227 			 * If the client is not the primary MAC client, then we
4228 			 * need to send the notification to the clients upper
4229 			 * MAC, i.e. mci_upper_mip.
4230 			 */
4231 			mac_tx_notify(mcip->mci_upper_mip != NULL ?
4232 			    mcip->mci_upper_mip : mcip->mci_mip);
4233 		}
4234 		mutex_enter(&mac_srs->srs_lock);
4235 	}
4236 	mac_srs->srs_state &= ~SRS_PROC;
4237 }
4238 
4239 /*
4240  * Given a packet, get the flow_entry that identifies the flow
4241  * to which that packet belongs. The flow_entry will contain
4242  * the transmit function to be used to send the packet. If the
4243  * function returns NULL, the packet should be sent using the
4244  * underlying NIC.
4245  */
4246 static flow_entry_t *
4247 mac_tx_classify(mac_impl_t *mip, mblk_t *mp)
4248 {
4249 	flow_entry_t		*flent = NULL;
4250 	mac_client_impl_t	*mcip;
4251 	int	err;
4252 
4253 	/*
4254 	 * Do classification on the packet.
4255 	 */
4256 	err = mac_flow_lookup(mip->mi_flow_tab, mp, FLOW_OUTBOUND, &flent);
4257 	if (err != 0)
4258 		return (NULL);
4259 
4260 	/*
4261 	 * This flent might just be an additional one on the MAC client,
4262 	 * i.e. for classification purposes (different fdesc), however
4263 	 * the resources, SRS et. al., are in the mci_flent, so if
4264 	 * this isn't the mci_flent, we need to get it.
4265 	 */
4266 	if ((mcip = flent->fe_mcip) != NULL && mcip->mci_flent != flent) {
4267 		FLOW_REFRELE(flent);
4268 		flent = mcip->mci_flent;
4269 		FLOW_TRY_REFHOLD(flent, err);
4270 		if (err != 0)
4271 			return (NULL);
4272 	}
4273 
4274 	return (flent);
4275 }
4276 
4277 /*
4278  * This macro is only meant to be used by mac_tx_send().
4279  */
4280 #define	CHECK_VID_AND_ADD_TAG(mp) {			\
4281 	if (vid_check) {				\
4282 		int err = 0;				\
4283 							\
4284 		MAC_VID_CHECK(src_mcip, (mp), err);	\
4285 		if (err != 0) {				\
4286 			freemsg((mp));			\
4287 			(mp) = next;			\
4288 			oerrors++;			\
4289 			continue;			\
4290 		}					\
4291 	}						\
4292 	if (add_tag) {					\
4293 		(mp) = mac_add_vlan_tag((mp), 0, vid);	\
4294 		if ((mp) == NULL) {			\
4295 			(mp) = next;			\
4296 			oerrors++;			\
4297 			continue;			\
4298 		}					\
4299 	}						\
4300 }
4301 
4302 mblk_t *
4303 mac_tx_send(mac_client_handle_t mch, mac_ring_handle_t ring, mblk_t *mp_chain,
4304     mac_tx_stats_t *stats)
4305 {
4306 	mac_client_impl_t *src_mcip = (mac_client_impl_t *)mch;
4307 	mac_impl_t *mip = src_mcip->mci_mip;
4308 	uint_t obytes = 0, opackets = 0, oerrors = 0;
4309 	mblk_t *mp = NULL, *next;
4310 	boolean_t vid_check, add_tag;
4311 	uint16_t vid = 0;
4312 
4313 	if (mip->mi_nclients > 1) {
4314 		vid_check = MAC_VID_CHECK_NEEDED(src_mcip);
4315 		add_tag = MAC_TAG_NEEDED(src_mcip);
4316 		if (add_tag)
4317 			vid = mac_client_vid(mch);
4318 	} else {
4319 		ASSERT(mip->mi_nclients == 1);
4320 		vid_check = add_tag = B_FALSE;
4321 	}
4322 
4323 	/*
4324 	 * Fastpath: if there's only one client, we simply send
4325 	 * the packet down to the underlying NIC.
4326 	 */
4327 	if (mip->mi_nactiveclients == 1) {
4328 		DTRACE_PROBE2(fastpath,
4329 		    mac_client_impl_t *, src_mcip, mblk_t *, mp_chain);
4330 
4331 		mp = mp_chain;
4332 		while (mp != NULL) {
4333 			next = mp->b_next;
4334 			mp->b_next = NULL;
4335 			opackets++;
4336 			obytes += (mp->b_cont == NULL ? MBLKL(mp) :
4337 			    msgdsize(mp));
4338 
4339 			CHECK_VID_AND_ADD_TAG(mp);
4340 			MAC_TX(mip, ring, mp, src_mcip);
4341 
4342 			/*
4343 			 * If the driver is out of descriptors and does a
4344 			 * partial send it will return a chain of unsent
4345 			 * mblks. Adjust the accounting stats.
4346 			 */
4347 			if (mp != NULL) {
4348 				opackets--;
4349 				obytes -= msgdsize(mp);
4350 				mp->b_next = next;
4351 				break;
4352 			}
4353 			mp = next;
4354 		}
4355 		goto done;
4356 	}
4357 
4358 	/*
4359 	 * No fastpath, we either have more than one MAC client
4360 	 * defined on top of the same MAC, or one or more MAC
4361 	 * client promiscuous callbacks.
4362 	 */
4363 	DTRACE_PROBE3(slowpath, mac_client_impl_t *,
4364 	    src_mcip, int, mip->mi_nclients, mblk_t *, mp_chain);
4365 
4366 	mp = mp_chain;
4367 	while (mp != NULL) {
4368 		flow_entry_t *dst_flow_ent;
4369 		void *flow_cookie;
4370 		size_t	pkt_size;
4371 		mblk_t *mp1;
4372 
4373 		next = mp->b_next;
4374 		mp->b_next = NULL;
4375 		opackets++;
4376 		pkt_size = (mp->b_cont == NULL ? MBLKL(mp) : msgdsize(mp));
4377 		obytes += pkt_size;
4378 		CHECK_VID_AND_ADD_TAG(mp);
4379 
4380 		/*
4381 		 * Find the destination.
4382 		 */
4383 		dst_flow_ent = mac_tx_classify(mip, mp);
4384 
4385 		if (dst_flow_ent != NULL) {
4386 			size_t	hdrsize;
4387 			int	err = 0;
4388 
4389 			if (mip->mi_info.mi_nativemedia == DL_ETHER) {
4390 				struct ether_vlan_header *evhp =
4391 				    (struct ether_vlan_header *)mp->b_rptr;
4392 
4393 				if (ntohs(evhp->ether_tpid) == ETHERTYPE_VLAN)
4394 					hdrsize = sizeof (*evhp);
4395 				else
4396 					hdrsize = sizeof (struct ether_header);
4397 			} else {
4398 				mac_header_info_t	mhi;
4399 
4400 				err = mac_header_info((mac_handle_t)mip,
4401 				    mp, &mhi);
4402 				if (err == 0)
4403 					hdrsize = mhi.mhi_hdrsize;
4404 			}
4405 
4406 			/*
4407 			 * Got a matching flow. It's either another
4408 			 * MAC client, or a broadcast/multicast flow.
4409 			 * Make sure the packet size is within the
4410 			 * allowed size. If not drop the packet and
4411 			 * move to next packet.
4412 			 */
4413 			if (err != 0 ||
4414 			    (pkt_size - hdrsize) > mip->mi_sdu_max) {
4415 				oerrors++;
4416 				DTRACE_PROBE2(loopback__drop, size_t, pkt_size,
4417 				    mblk_t *, mp);
4418 				freemsg(mp);
4419 				mp = next;
4420 				FLOW_REFRELE(dst_flow_ent);
4421 				continue;
4422 			}
4423 			flow_cookie = mac_flow_get_client_cookie(dst_flow_ent);
4424 			if (flow_cookie != NULL) {
4425 				/*
4426 				 * The vnic_bcast_send function expects
4427 				 * to receive the sender MAC client
4428 				 * as value for arg2.
4429 				 */
4430 				mac_bcast_send(flow_cookie, src_mcip, mp,
4431 				    B_TRUE);
4432 			} else {
4433 				/*
4434 				 * loopback the packet to a local MAC
4435 				 * client. We force a context switch
4436 				 * if both source and destination MAC
4437 				 * clients are used by IP, i.e.
4438 				 * bypass is set.
4439 				 */
4440 				boolean_t do_switch;
4441 				mac_client_impl_t *dst_mcip =
4442 				    dst_flow_ent->fe_mcip;
4443 
4444 				/*
4445 				 * Check if there are promiscuous mode
4446 				 * callbacks defined. This check is
4447 				 * done here in the 'else' case and
4448 				 * not in other cases because this
4449 				 * path is for local loopback
4450 				 * communication which does not go
4451 				 * through MAC_TX(). For paths that go
4452 				 * through MAC_TX(), the promisc_list
4453 				 * check is done inside the MAC_TX()
4454 				 * macro.
4455 				 */
4456 				if (mip->mi_promisc_list != NULL)
4457 					mac_promisc_dispatch(mip, mp, src_mcip);
4458 
4459 				do_switch = ((src_mcip->mci_state_flags &
4460 				    dst_mcip->mci_state_flags &
4461 				    MCIS_CLIENT_POLL_CAPABLE) != 0);
4462 
4463 				if ((mp1 = mac_fix_cksum(mp)) != NULL) {
4464 					(dst_flow_ent->fe_cb_fn)(
4465 					    dst_flow_ent->fe_cb_arg1,
4466 					    dst_flow_ent->fe_cb_arg2,
4467 					    mp1, do_switch);
4468 				}
4469 			}
4470 			FLOW_REFRELE(dst_flow_ent);
4471 		} else {
4472 			/*
4473 			 * Unknown destination, send via the underlying
4474 			 * NIC.
4475 			 */
4476 			MAC_TX(mip, ring, mp, src_mcip);
4477 			if (mp != NULL) {
4478 				/*
4479 				 * Adjust for the last packet that
4480 				 * could not be transmitted
4481 				 */
4482 				opackets--;
4483 				obytes -= pkt_size;
4484 				mp->b_next = next;
4485 				break;
4486 			}
4487 		}
4488 		mp = next;
4489 	}
4490 
4491 done:
4492 	stats->mts_obytes = obytes;
4493 	stats->mts_opackets = opackets;
4494 	stats->mts_oerrors = oerrors;
4495 	return (mp);
4496 }
4497 
4498 /*
4499  * mac_tx_srs_ring_present
4500  *
4501  * Returns whether the specified ring is part of the specified SRS.
4502  */
4503 boolean_t
4504 mac_tx_srs_ring_present(mac_soft_ring_set_t *srs, mac_ring_t *tx_ring)
4505 {
4506 	int i;
4507 	mac_soft_ring_t *soft_ring;
4508 
4509 	if (srs->srs_tx.st_arg2 == tx_ring)
4510 		return (B_TRUE);
4511 
4512 	for (i = 0; i < srs->srs_tx_ring_count; i++) {
4513 		soft_ring =  srs->srs_tx_soft_rings[i];
4514 		if (soft_ring->s_ring_tx_arg2 == tx_ring)
4515 			return (B_TRUE);
4516 	}
4517 
4518 	return (B_FALSE);
4519 }
4520 
4521 /*
4522  * mac_tx_srs_get_soft_ring
4523  *
4524  * Returns the TX soft ring associated with the given ring, if present.
4525  */
4526 mac_soft_ring_t *
4527 mac_tx_srs_get_soft_ring(mac_soft_ring_set_t *srs, mac_ring_t *tx_ring)
4528 {
4529 	int		i;
4530 	mac_soft_ring_t	*soft_ring;
4531 
4532 	if (srs->srs_tx.st_arg2 == tx_ring)
4533 		return (NULL);
4534 
4535 	for (i = 0; i < srs->srs_tx_ring_count; i++) {
4536 		soft_ring =  srs->srs_tx_soft_rings[i];
4537 		if (soft_ring->s_ring_tx_arg2 == tx_ring)
4538 			return (soft_ring);
4539 	}
4540 
4541 	return (NULL);
4542 }
4543 
4544 /*
4545  * mac_tx_srs_wakeup
4546  *
4547  * Called when Tx desc become available. Wakeup the appropriate worker
4548  * thread after resetting the SRS_TX_BLOCKED/S_RING_BLOCK bit in the
4549  * state field.
4550  */
4551 void
4552 mac_tx_srs_wakeup(mac_soft_ring_set_t *mac_srs, mac_ring_handle_t ring)
4553 {
4554 	int i;
4555 	mac_soft_ring_t *sringp;
4556 	mac_srs_tx_t *srs_tx = &mac_srs->srs_tx;
4557 
4558 	mutex_enter(&mac_srs->srs_lock);
4559 	/*
4560 	 * srs_tx_ring_count == 0 is the single ring mode case. In
4561 	 * this mode, there will not be Tx soft rings associated
4562 	 * with the SRS.
4563 	 */
4564 	if (!MAC_TX_SOFT_RINGS(mac_srs)) {
4565 		if (srs_tx->st_arg2 == ring &&
4566 		    mac_srs->srs_state & SRS_TX_BLOCKED) {
4567 			mac_srs->srs_state &= ~SRS_TX_BLOCKED;
4568 			srs_tx->st_stat.mts_unblockcnt++;
4569 			cv_signal(&mac_srs->srs_async);
4570 		}
4571 		/*
4572 		 * A wakeup can come before tx_srs_drain() could
4573 		 * grab srs lock and set SRS_TX_BLOCKED. So
4574 		 * always set woken_up flag when we come here.
4575 		 */
4576 		srs_tx->st_woken_up = B_TRUE;
4577 		mutex_exit(&mac_srs->srs_lock);
4578 		return;
4579 	}
4580 
4581 	/*
4582 	 * If you are here, it is for FANOUT, BW_FANOUT,
4583 	 * AGGR_MODE or AGGR_BW_MODE case
4584 	 */
4585 	for (i = 0; i < mac_srs->srs_tx_ring_count; i++) {
4586 		sringp = mac_srs->srs_tx_soft_rings[i];
4587 		mutex_enter(&sringp->s_ring_lock);
4588 		if (sringp->s_ring_tx_arg2 == ring) {
4589 			if (sringp->s_ring_state & S_RING_BLOCK) {
4590 				sringp->s_ring_state &= ~S_RING_BLOCK;
4591 				sringp->s_st_stat.mts_unblockcnt++;
4592 				cv_signal(&sringp->s_ring_async);
4593 			}
4594 			sringp->s_ring_tx_woken_up = B_TRUE;
4595 		}
4596 		mutex_exit(&sringp->s_ring_lock);
4597 	}
4598 	mutex_exit(&mac_srs->srs_lock);
4599 }
4600 
4601 /*
4602  * Once the driver is done draining, send a MAC_NOTE_TX notification to unleash
4603  * the blocked clients again.
4604  */
4605 void
4606 mac_tx_notify(mac_impl_t *mip)
4607 {
4608 	i_mac_notify(mip, MAC_NOTE_TX);
4609 }
4610 
4611 /*
4612  * RX SOFTRING RELATED FUNCTIONS
4613  *
4614  * These functions really belong in mac_soft_ring.c and here for
4615  * a short period.
4616  */
4617 
4618 #define	SOFT_RING_ENQUEUE_CHAIN(ringp, mp, tail, cnt, sz) {		\
4619 	/*								\
4620 	 * Enqueue our mblk chain.					\
4621 	 */								\
4622 	ASSERT(MUTEX_HELD(&(ringp)->s_ring_lock));			\
4623 									\
4624 	if ((ringp)->s_ring_last != NULL)				\
4625 		(ringp)->s_ring_last->b_next = (mp);			\
4626 	else								\
4627 		(ringp)->s_ring_first = (mp);				\
4628 	(ringp)->s_ring_last = (tail);					\
4629 	(ringp)->s_ring_count += (cnt);					\
4630 	ASSERT((ringp)->s_ring_count > 0);				\
4631 	if ((ringp)->s_ring_type & ST_RING_BW_CTL) {			\
4632 		(ringp)->s_ring_size += sz;				\
4633 	}								\
4634 }
4635 
4636 /*
4637  * Default entry point to deliver a packet chain to a MAC client.
4638  * If the MAC client has flows, do the classification with these
4639  * flows as well.
4640  */
4641 /* ARGSUSED */
4642 void
4643 mac_rx_deliver(void *arg1, mac_resource_handle_t mrh, mblk_t *mp_chain,
4644     mac_header_info_t *arg3)
4645 {
4646 	mac_client_impl_t *mcip = arg1;
4647 
4648 	if (mcip->mci_nvids == 1 &&
4649 	    !(mcip->mci_state_flags & MCIS_STRIP_DISABLE)) {
4650 		/*
4651 		 * If the client has exactly one VID associated with it
4652 		 * and striping of VLAN header is not disabled,
4653 		 * remove the VLAN tag from the packet before
4654 		 * passing it on to the client's receive callback.
4655 		 * Note that this needs to be done after we dispatch
4656 		 * the packet to the promiscuous listeners of the
4657 		 * client, since they expect to see the whole
4658 		 * frame including the VLAN headers.
4659 		 */
4660 		mp_chain = mac_strip_vlan_tag_chain(mp_chain);
4661 	}
4662 
4663 	mcip->mci_rx_fn(mcip->mci_rx_arg, mrh, mp_chain, B_FALSE);
4664 }
4665 
4666 /*
4667  * mac_rx_soft_ring_process
4668  *
4669  * process a chain for a given soft ring. The number of packets queued
4670  * in the SRS and its associated soft rings (including this one) is
4671  * very small (tracked by srs_poll_pkt_cnt), then allow the entering
4672  * thread (interrupt or poll thread) to do inline processing. This
4673  * helps keep the latency down under low load.
4674  *
4675  * The proc and arg for each mblk is already stored in the mblk in
4676  * appropriate places.
4677  */
4678 /* ARGSUSED */
4679 void
4680 mac_rx_soft_ring_process(mac_client_impl_t *mcip, mac_soft_ring_t *ringp,
4681     mblk_t *mp_chain, mblk_t *tail, int cnt, size_t sz)
4682 {
4683 	mac_direct_rx_t		proc;
4684 	void			*arg1;
4685 	mac_resource_handle_t	arg2;
4686 	mac_soft_ring_set_t	*mac_srs = ringp->s_ring_set;
4687 
4688 	ASSERT(ringp != NULL);
4689 	ASSERT(mp_chain != NULL);
4690 	ASSERT(tail != NULL);
4691 	ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
4692 
4693 	mutex_enter(&ringp->s_ring_lock);
4694 	ringp->s_ring_total_inpkt += cnt;
4695 	ringp->s_ring_total_rbytes += sz;
4696 	if ((mac_srs->srs_rx.sr_poll_pkt_cnt <= 1) &&
4697 	    !(ringp->s_ring_type & ST_RING_WORKER_ONLY)) {
4698 		/* If on processor or blanking on, then enqueue and return */
4699 		if (ringp->s_ring_state & S_RING_BLANK ||
4700 		    ringp->s_ring_state & S_RING_PROC) {
4701 			SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
4702 			mutex_exit(&ringp->s_ring_lock);
4703 			return;
4704 		}
4705 		proc = ringp->s_ring_rx_func;
4706 		arg1 = ringp->s_ring_rx_arg1;
4707 		arg2 = ringp->s_ring_rx_arg2;
4708 		/*
4709 		 * See if anything is already queued. If we are the
4710 		 * first packet, do inline processing else queue the
4711 		 * packet and do the drain.
4712 		 */
4713 		if (ringp->s_ring_first == NULL) {
4714 			/*
4715 			 * Fast-path, ok to process and nothing queued.
4716 			 */
4717 			ringp->s_ring_run = curthread;
4718 			ringp->s_ring_state |= (S_RING_PROC);
4719 
4720 			mutex_exit(&ringp->s_ring_lock);
4721 
4722 			/*
4723 			 * We are the chain of 1 packet so
4724 			 * go through this fast path.
4725 			 */
4726 			ASSERT(mp_chain->b_next == NULL);
4727 
4728 			(*proc)(arg1, arg2, mp_chain, NULL);
4729 
4730 			ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
4731 			/*
4732 			 * If we have a soft ring set which is doing
4733 			 * bandwidth control, we need to decrement
4734 			 * srs_size and count so it the SRS can have a
4735 			 * accurate idea of what is the real data
4736 			 * queued between SRS and its soft rings. We
4737 			 * decrement the counters only when the packet
4738 			 * gets processed by both SRS and the soft ring.
4739 			 */
4740 			mutex_enter(&mac_srs->srs_lock);
4741 			MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
4742 			MAC_UPDATE_SRS_SIZE_LOCKED(mac_srs, sz);
4743 			mutex_exit(&mac_srs->srs_lock);
4744 
4745 			mutex_enter(&ringp->s_ring_lock);
4746 			ringp->s_ring_run = NULL;
4747 			ringp->s_ring_state &= ~S_RING_PROC;
4748 			if (ringp->s_ring_state & S_RING_CLIENT_WAIT)
4749 				cv_signal(&ringp->s_ring_client_cv);
4750 
4751 			if ((ringp->s_ring_first == NULL) ||
4752 			    (ringp->s_ring_state & S_RING_BLANK)) {
4753 				/*
4754 				 * We processed inline our packet and
4755 				 * nothing new has arrived or our
4756 				 * receiver doesn't want to receive
4757 				 * any packets. We are done.
4758 				 */
4759 				mutex_exit(&ringp->s_ring_lock);
4760 				return;
4761 			}
4762 		} else {
4763 			SOFT_RING_ENQUEUE_CHAIN(ringp,
4764 			    mp_chain, tail, cnt, sz);
4765 		}
4766 
4767 		/*
4768 		 * We are here because either we couldn't do inline
4769 		 * processing (because something was already
4770 		 * queued), or we had a chain of more than one
4771 		 * packet, or something else arrived after we were
4772 		 * done with inline processing.
4773 		 */
4774 		ASSERT(MUTEX_HELD(&ringp->s_ring_lock));
4775 		ASSERT(ringp->s_ring_first != NULL);
4776 
4777 		ringp->s_ring_drain_func(ringp);
4778 		mutex_exit(&ringp->s_ring_lock);
4779 		return;
4780 	} else {
4781 		/* ST_RING_WORKER_ONLY case */
4782 		SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
4783 		mac_soft_ring_worker_wakeup(ringp);
4784 		mutex_exit(&ringp->s_ring_lock);
4785 	}
4786 }
4787 
4788 /*
4789  * TX SOFTRING RELATED FUNCTIONS
4790  *
4791  * These functions really belong in mac_soft_ring.c and here for
4792  * a short period.
4793  */
4794 
4795 #define	TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp, tail, cnt, sz) {		\
4796 	ASSERT(MUTEX_HELD(&ringp->s_ring_lock));			\
4797 	ringp->s_ring_state |= S_RING_ENQUEUED;				\
4798 	SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);	\
4799 }
4800 
4801 /*
4802  * mac_tx_sring_queued
4803  *
4804  * When we are out of transmit descriptors and we already have a
4805  * queue that exceeds hiwat (or the client called us with
4806  * MAC_TX_NO_ENQUEUE or MAC_DROP_ON_NO_DESC flag), return the
4807  * soft ring pointer as the opaque cookie for the client enable
4808  * flow control.
4809  */
4810 static mac_tx_cookie_t
4811 mac_tx_sring_enqueue(mac_soft_ring_t *ringp, mblk_t *mp_chain, uint16_t flag,
4812     mblk_t **ret_mp)
4813 {
4814 	int cnt;
4815 	size_t sz;
4816 	mblk_t *tail;
4817 	mac_soft_ring_set_t *mac_srs = ringp->s_ring_set;
4818 	mac_tx_cookie_t cookie = 0;
4819 	boolean_t wakeup_worker = B_TRUE;
4820 
4821 	ASSERT(MUTEX_HELD(&ringp->s_ring_lock));
4822 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
4823 	if (flag & MAC_DROP_ON_NO_DESC) {
4824 		mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);
4825 		/* increment freed stats */
4826 		ringp->s_ring_drops += cnt;
4827 		cookie = (mac_tx_cookie_t)ringp;
4828 	} else {
4829 		if (ringp->s_ring_first != NULL)
4830 			wakeup_worker = B_FALSE;
4831 
4832 		if (flag & MAC_TX_NO_ENQUEUE) {
4833 			/*
4834 			 * If QUEUED is not set, queue the packet
4835 			 * and let mac_tx_soft_ring_drain() set
4836 			 * the TX_BLOCKED bit for the reasons
4837 			 * explained above. Otherwise, return the
4838 			 * mblks.
4839 			 */
4840 			if (wakeup_worker) {
4841 				TX_SOFT_RING_ENQUEUE_CHAIN(ringp,
4842 				    mp_chain, tail, cnt, sz);
4843 			} else {
4844 				ringp->s_ring_state |= S_RING_WAKEUP_CLIENT;
4845 				cookie = (mac_tx_cookie_t)ringp;
4846 				*ret_mp = mp_chain;
4847 			}
4848 		} else {
4849 			boolean_t enqueue = B_TRUE;
4850 
4851 			if (ringp->s_ring_count > ringp->s_ring_tx_hiwat) {
4852 				/*
4853 				 * flow-controlled. Store ringp in cookie
4854 				 * so that it can be returned as
4855 				 * mac_tx_cookie_t to client
4856 				 */
4857 				ringp->s_ring_state |= S_RING_TX_HIWAT;
4858 				cookie = (mac_tx_cookie_t)ringp;
4859 				ringp->s_ring_hiwat_cnt++;
4860 				if (ringp->s_ring_count >
4861 				    ringp->s_ring_tx_max_q_cnt) {
4862 					/* increment freed stats */
4863 					ringp->s_ring_drops += cnt;
4864 					/*
4865 					 * b_prev may be set to the fanout hint
4866 					 * hence can't use freemsg directly
4867 					 */
4868 					mac_pkt_drop(NULL, NULL,
4869 					    mp_chain, B_FALSE);
4870 					DTRACE_PROBE1(tx_queued_hiwat,
4871 					    mac_soft_ring_t *, ringp);
4872 					enqueue = B_FALSE;
4873 				}
4874 			}
4875 			if (enqueue) {
4876 				TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain,
4877 				    tail, cnt, sz);
4878 			}
4879 		}
4880 		if (wakeup_worker)
4881 			cv_signal(&ringp->s_ring_async);
4882 	}
4883 	return (cookie);
4884 }
4885 
4886 
4887 /*
4888  * mac_tx_soft_ring_process
4889  *
4890  * This routine is called when fanning out outgoing traffic among
4891  * multipe Tx rings.
4892  * Note that a soft ring is associated with a h/w Tx ring.
4893  */
4894 mac_tx_cookie_t
4895 mac_tx_soft_ring_process(mac_soft_ring_t *ringp, mblk_t *mp_chain,
4896     uint16_t flag, mblk_t **ret_mp)
4897 {
4898 	mac_soft_ring_set_t *mac_srs = ringp->s_ring_set;
4899 	int	cnt;
4900 	size_t	sz;
4901 	mblk_t	*tail;
4902 	mac_tx_cookie_t cookie = 0;
4903 
4904 	ASSERT(ringp != NULL);
4905 	ASSERT(mp_chain != NULL);
4906 	ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
4907 	/*
4908 	 * The following modes can come here: SRS_TX_BW_FANOUT,
4909 	 * SRS_TX_FANOUT, SRS_TX_AGGR, SRS_TX_BW_AGGR.
4910 	 */
4911 	ASSERT(MAC_TX_SOFT_RINGS(mac_srs));
4912 	ASSERT(mac_srs->srs_tx.st_mode == SRS_TX_FANOUT ||
4913 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_FANOUT ||
4914 	    mac_srs->srs_tx.st_mode == SRS_TX_AGGR ||
4915 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_AGGR);
4916 
4917 	if (ringp->s_ring_type & ST_RING_WORKER_ONLY) {
4918 		/* Serialization mode */
4919 
4920 		mutex_enter(&ringp->s_ring_lock);
4921 		if (ringp->s_ring_count > ringp->s_ring_tx_hiwat) {
4922 			cookie = mac_tx_sring_enqueue(ringp, mp_chain,
4923 			    flag, ret_mp);
4924 			mutex_exit(&ringp->s_ring_lock);
4925 			return (cookie);
4926 		}
4927 		MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
4928 		TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
4929 		if (ringp->s_ring_state & (S_RING_BLOCK | S_RING_PROC)) {
4930 			/*
4931 			 * If ring is blocked due to lack of Tx
4932 			 * descs, just return. Worker thread
4933 			 * will get scheduled when Tx desc's
4934 			 * become available.
4935 			 */
4936 			mutex_exit(&ringp->s_ring_lock);
4937 			return (cookie);
4938 		}
4939 		mac_soft_ring_worker_wakeup(ringp);
4940 		mutex_exit(&ringp->s_ring_lock);
4941 		return (cookie);
4942 	} else {
4943 		/* Default fanout mode */
4944 		/*
4945 		 * S_RING_BLOCKED is set when underlying NIC runs
4946 		 * out of Tx descs and messages start getting
4947 		 * queued. It won't get reset until
4948 		 * tx_srs_drain() completely drains out the
4949 		 * messages.
4950 		 */
4951 		mac_tx_stats_t		stats;
4952 
4953 		if (ringp->s_ring_state & S_RING_ENQUEUED) {
4954 			/* Tx descs/resources not available */
4955 			mutex_enter(&ringp->s_ring_lock);
4956 			if (ringp->s_ring_state & S_RING_ENQUEUED) {
4957 				cookie = mac_tx_sring_enqueue(ringp, mp_chain,
4958 				    flag, ret_mp);
4959 				mutex_exit(&ringp->s_ring_lock);
4960 				return (cookie);
4961 			}
4962 			/*
4963 			 * While we were computing mblk count, the
4964 			 * flow control condition got relieved.
4965 			 * Continue with the transmission.
4966 			 */
4967 			mutex_exit(&ringp->s_ring_lock);
4968 		}
4969 
4970 		mp_chain = mac_tx_send(ringp->s_ring_tx_arg1,
4971 		    ringp->s_ring_tx_arg2, mp_chain, &stats);
4972 
4973 		/*
4974 		 * Multiple threads could be here sending packets.
4975 		 * Under such conditions, it is not possible to
4976 		 * automically set S_RING_BLOCKED bit to indicate
4977 		 * out of tx desc condition. To atomically set
4978 		 * this, we queue the returned packet and do
4979 		 * the setting of S_RING_BLOCKED in
4980 		 * mac_tx_soft_ring_drain().
4981 		 */
4982 		if (mp_chain != NULL) {
4983 			mutex_enter(&ringp->s_ring_lock);
4984 			cookie =
4985 			    mac_tx_sring_enqueue(ringp, mp_chain, flag, ret_mp);
4986 			mutex_exit(&ringp->s_ring_lock);
4987 			return (cookie);
4988 		}
4989 		SRS_TX_STATS_UPDATE(mac_srs, &stats);
4990 		SOFTRING_TX_STATS_UPDATE(ringp, &stats);
4991 
4992 		return (0);
4993 	}
4994 }
4995