xref: /titanic_41/usr/src/uts/i86pc/os/intr.c (revision fc7a376ec44146b66b5d6bc4ed92ba311773e202)
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 /*
23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2012, Joyent, Inc.  All rights reserverd.
25  */
26 
27 /*
28  * To understand the present state of interrupt handling on i86pc, we must
29  * first consider the history of interrupt controllers and our way of handling
30  * interrupts.
31  *
32  * History of Interrupt Controllers on i86pc
33  * -----------------------------------------
34  *
35  *    Intel 8259 and 8259A
36  *
37  * The first interrupt controller that attained widespread use on i86pc was
38  * the Intel 8259(A) Programmable Interrupt Controller that first saw use with
39  * the 8086. It took up to 8 interrupt sources and combined them into one
40  * output wire. Up to 8 8259s could be slaved together providing up to 64 IRQs.
41  * With the switch to the 8259A, level mode interrupts became possible. For a
42  * long time on i86pc the 8259A was the only way to handle interrupts and it
43  * had its own set of quirks. The 8259A and its corresponding interval timer
44  * the 8254 are programmed using outb and inb instructions.
45  *
46  *    Intel Advanced Programmable Interrupt Controller (APIC)
47  *
48  * Starting around the time of the introduction of the P6 family
49  * microarchitecture (i686) Intel introduced a new interrupt controller.
50  * Instead of having the series of slaved 8259A devices, Intel opted to outfit
51  * each processor with a Local APIC (lapic) and to outfit the system with at
52  * least one, but potentially more, I/O APICs (ioapic). The lapics and ioapics
53  * initially communicated over a dedicated bus, but this has since been
54  * replaced. Each physical core and even hyperthread currently contains its
55  * own local apic, which is not shared. There are a few exceptions for
56  * hyperthreads, but that does not usually concern us.
57  *
58  * Instead of talking directly to 8259 for status, sending End Of Interrupt
59  * (EOI), etc. a microprocessor now communicates directly to the lapic. This
60  * also allows for each microprocessor to be able to have independent controls.
61  * The programming method is different from the 8259. Consumers map the lapic
62  * registers into uncacheable memory to read and manipulate the state.
63  *
64  * The number of addressable interrupt vectors was increased to 256. However
65  * vectors 0-31 are reserved for the processor exception handling, leaving the
66  * remaining vectors for general use. In addition to hardware generated
67  * interrupts, the lapic provides a way for generating inter-processor
68  * interrupts (IPI) which are the basis for CPU cross calls and CPU pokes.
69  *
70  * AMD ended up implementing the Intel APIC architecture in lieu of their work
71  * with Cyrix.
72  *
73  *    Intel x2apic
74  *
75  * The x2apic is an extension to the lapic which started showing up around the
76  * same time as the Sandy Bridge chipsets. It provides a new programming mode
77  * as well as new features. The goal of the x2apic is to solve a few problems
78  * with the previous generation of lapic and the x2apic is backwards compatible
79  * with the previous programming and model. The only downsides to using the
80  * backwards compatibility is that you are not able to take advantage of the new
81  * x2apic features.
82  *
83  *    o The APIC ID is increased from an 8-bit value to a 32-bit value. This
84  *    increases the maximum number of addressable physical processors beyond
85  *    256. This new ID is assembled in a similar manner as the information that
86  *    is obtainable by the extended cpuid topology leaves.
87  *
88  *    o A new means of generating IPIs was introduced.
89  *
90  *    o Instead of memory mapping the registers, the x2apic only allows for
91  *    programming it through a series of wrmsrs. This has important semantic
92  *    side effects. Recall that the registers were previously all mapped to
93  *    uncachable memory which meant that all operations to the local apic were
94  *    serializing instructions. With the switch to using wrmsrs this has been
95  *    relaxed and these operations can no longer be assumed to be serializing
96  *    instructions.
97  *
98  * Note for the rest of this we are only going to concern ourselves with the
99  * apic and x2apic which practically all of i86pc has been using now for
100  * quite some time.
101  *
102  * Interrupt Priority Levels
103  * -------------------------
104  *
105  * On i86pc systems there are a total of fifteen interrupt priority levels
106  * (ipls) which range from 1-15. Level 0 is for normal processing and
107  * non-interrupt processing. To manipulate these values the family of spl
108  * functions (which date back to UNIX on the PDP-11) are used. Specifically,
109  * splr() to raise the priority level and splx() to lower it. One should not
110  * generally call setspl() directly.
111  *
112  * Both i86pc and the supported SPARC platforms honor the same conventions for
113  * the meaning behind these IPLs. The most important IPL is the platform's
114  * LOCK_LEVEL (0xa on i86pc). If a thread is above LOCK_LEVEL it _must_ not
115  * sleep on any synchronization object. The only allowed synchronization
116  * primitive is a mutex that has been specifically initialized to be a spin
117  * lock (see mutex_init(9F)). Another important level is DISP_LEVEL (0xb on
118  * i86pc). You must be at DISP_LEVEL if you want to control the dispatcher.
119  * The XC_HI_PIL is the highest level (0xf) and is used during cross-calls.
120  *
121  * Each interrupt that is registered in the system fires at a specific IPL.
122  * Generally most interrupts fire below LOCK_LEVEL.
123  *
124  * PSM Drivers
125  * -----------
126  *
127  * We currently have three sets of PSM (platform specific module) drivers
128  * available. uppc, pcplusmp, and apix. uppc (uni-processor PC) is the original
129  * driver that interacts with the 8259A and 8254. In general, it is not used
130  * anymore given the prevalence of the apic.
131  *
132  * The system prefers to use the apix driver over the pcplusmp driver. The apix
133  * driver requires HW support for an x2apic. If there is no x2apic HW, apix
134  * will not be used. In general we prefer using the apix driver over the
135  * pcplusmp driver because it gives us much more flexibility with respect to
136  * interrupts. In the apix driver each local apic has its own independent set
137  * of  interrupts, whereas the pcplusmp driver only has a single global set of
138  * interrupts. This is why pcplusmp only supports a finite number of interrupts
139  * per IPL -- generally 16, often less. The apix driver supports using either
140  * the x2apic or the local apic programing modes. The programming mode does not
141  * change the number of interrupts available, just the number of processors
142  * that we can address. For the apix driver, the x2apic mode is enabled if the
143  * system supports interrupt re-mapping, otherwise the module manages the
144  * x2apic in local mode.
145  *
146  * When there is no x2apic present, we default back to the pcplusmp PSM driver.
147  * In general, this is not problematic unless you have more than 256
148  * processors in the machine or you do not have enough interrupts available.
149  *
150  * Controlling Interrupt Generation on i86pc
151  * -----------------------------------------
152  *
153  * There are two different ways to manipulate which interrupts will be
154  * generated on i86pc. Each offers different degrees of control.
155  *
156  * The first is through the flags register (eflags and rflags on i386 and amd64
157  * respectively). The IF bit determines whether or not interrupts are enabled
158  * or disabled. This is manipulated in one of several ways. The most common way
159  * is through the cli and sti instructions. These clear the IF flag and set it,
160  * respectively, for the current processor. The other common way is through the
161  * use of the intr_clear and intr_restore functions.
162  *
163  * Assuming interrupts are not blocked by the IF flag, then the second form is
164  * through the Processor-Priority Register (PPR). The PPR is used to determine
165  * whether or not a pending interrupt should be delivered. If the ipl of the
166  * new interrupt is higher than the current value in the PPR, then the lapic
167  * will either deliver it immediately (if interrupts are not in progress) or it
168  * will deliver it once the current interrupt processing has issued an EOI. The
169  * highest unmasked interrupt will be the one delivered.
170  *
171  * The PPR register is based upon the max of the following two registers in the
172  * lapic, the TPR register (also known as CR8 on amd64) that can be used to
173  * mask interrupt levels, and the current vector. Because the pcplusmp module
174  * always sets TPR appropriately early in the do_interrupt path, we can usually
175  * just think that the PPR is the TPR. The pcplusmp module also issues an EOI
176  * once it has set the TPR, so higher priority interrupts can come in while
177  * we're servicing a lower priority interrupt.
178  *
179  * Handling Interrupts
180  * -------------------
181  *
182  * Interrupts can be broken down into three categories based on priority and
183  * source:
184  *
185  *   o High level interrupts
186  *   o Low level hardware interrupts
187  *   o Low level software interrupts
188  *
189  *   High Level Interrupts
190  *
191  * High level interrupts encompasses both hardware-sourced and software-sourced
192  * interrupts. Examples of high level hardware interrupts include the serial
193  * console. High level software-sourced interrupts are still delivered through
194  * the local apic through IPIs. This is primarily cross calls.
195  *
196  * When a high level interrupt comes in, we will raise the SPL and then pin the
197  * current lwp to the processor. We will use its lwp, but our own interrupt
198  * stack and process the high level interrupt in-situ. These handlers are
199  * designed to be very short in nature and cannot go to sleep, only block on a
200  * spin lock. If the interrupt has a lot of work to do, it must generate a
201  * low-priority software interrupt that will be processed later.
202  *
203  *   Low level hardware interrupts
204  *
205  * Low level hardware interrupts start off like their high-level cousins. The
206  * current CPU contains a number of kernel threads (kthread_t) that can be used
207  * to process low level interrupts. These are shared between both low level
208  * hardware and software interrupts. Note that while we run with our
209  * kthread_t, we borrow the pinned threads lwp_t until such a time as we hit a
210  * synchronization object. If we hit one and need to sleep, then the scheduler
211  * will instead create the rest of what we need.
212  *
213  *   Low level software interrupts
214  *
215  * Low level software interrupts are handled in a similar way as hardware
216  * interrupts, but the notification vector is different. Each CPU has a bitmask
217  * of pending software interrupts. We can notify a CPU to process software
218  * interrupts through a specific trap vector as well as through several
219  * checks that are performed throughout the code. These checks will look at
220  * processing software interrupts as we lower our spl.
221  *
222  * We attempt to process the highest pending software interrupt that we can
223  * which is greater than our current IPL. If none currently exist, then we move
224  * on. We process a software interrupt in a similar fashion to a hardware
225  * interrupt.
226  *
227  * Traditional Interrupt Flow
228  * --------------------------
229  *
230  * The following diagram tracks the flow of the traditional uppc and pcplusmp
231  * interrupt handlers. The apix driver has its own version of do_interrupt().
232  * We come into the interrupt handler with all interrupts masked by the IF
233  * flag. This is because we set up the handler using an interrupt-gate, which
234  * is defined architecturally to have cleared the IF flag for us.
235  *
236  * +--------------+    +----------------+    +-----------+
237  * | _interrupt() |--->| do_interrupt() |--->| *setlvl() |
238  * +--------------+    +----------------+    +-----------+
239  *                       |      |     |
240  *                       |      |     |
241  *              low-level|      |     | softint
242  *                HW int |      |     +---------------------------------------+
243  * +--------------+      |      |                                             |
244  * | intr_thread_ |<-----+      | hi-level int                                |
245  * | prolog()     |             |    +----------+                             |
246  * +--------------+             +--->| hilevel_ |      Not on intr stack      |
247  *       |                           | intr_    |-----------------+           |
248  *       |                           | prolog() |                 |           |
249  * +------------+                    +----------+                 |           |
250  * | switch_sp_ |                        | On intr                v           |
251  * | and_call() |                        | Stack          +------------+      |
252  * +------------+                        |                | switch_sp_ |      |
253  *       |                               v                | and_call() |      |
254  *       v                             +-----------+      +------------+      |
255  * +-----------+                       | dispatch_ |             |            |
256  * | dispatch_ |   +-------------------| hilevel() |<------------+            |
257  * | hardint() |   |                   +-----------+                          |
258  * +-----------+   |                                                          |
259  *       |         v                                                          |
260  *       |     +-----+  +----------------------+  +-----+  hi-level           |
261  *       +---->| sti |->| av_dispatch_autovect |->| cli |---------+           |
262  *             +-----+  +----------------------+  +-----+         |           |
263  *                                |                |              |           |
264  *                                v                |              |           |
265  *                         +----------+            |              |           |
266  *                         | for each |            |              |           |
267  *                         | handler  |            |              |           |
268  *                         |  *intr() |            |              v           |
269  * +--------------+        +----------+            |      +----------------+  |
270  * | intr_thread_ |                      low-level |      | hilevel_intr_  |  |
271  * | epilog()     |<-------------------------------+      | epilog()       |  |
272  * +--------------+                                       +----------------+  |
273  *   |       |                                                   |            |
274  *   |       +----------------------v      v---------------------+            |
275  *   |                           +------------+                               |
276  *   |   +---------------------->| *setlvlx() |                               |
277  *   |   |                       +------------+                               |
278  *   |   |                              |                                     |
279  *   |   |                              v                                     |
280  *   |   |      +--------+     +------------------+      +-------------+      |
281  *   |   |      | return |<----| softint pending? |----->| dosoftint() |<-----+
282  *   |   |      +--------+  no +------------------+ yes  +-------------+
283  *   |   |           ^                                      |     |
284  *   |   |           |  softint pil too low                 |     |
285  *   |   |           +--------------------------------------+     |
286  *   |   |                                                        v
287  *   |   |    +-----------+      +------------+          +-----------+
288  *   |   |    | dispatch_ |<-----| switch_sp_ |<---------| *setspl() |
289  *   |   |    | softint() |      | and_call() |          +-----------+
290  *   |   |    +-----------+      +------------+
291  *   |   |        |
292  *   |   |        v
293  *   |   |      +-----+  +----------------------+  +-----+  +------------+
294  *   |   |      | sti |->| av_dispatch_autovect |->| cli |->| dosoftint_ |
295  *   |   |      +-----+  +----------------------+  +-----+  | epilog()   |
296  *   |   |                                                  +------------+
297  *   |   |                                                    |     |
298  *   |   +----------------------------------------------------+     |
299  *   v                                                              |
300  * +-----------+                                                    |
301  * | interrupt |                                                    |
302  * | thread    |<---------------------------------------------------+
303  * | blocked   |
304  * +-----------+
305  *      |
306  *      v
307  *  +----------------+  +------------+  +-----------+  +-------+  +---------+
308  *  | set_base_spl() |->| *setlvlx() |->| splhigh() |->| sti() |->| swtch() |
309  *  +----------------+  +------------+  +-----------+  +-------+  +---------+
310  *
311  *    Calls made on Interrupt Stacks and Epilogue routines
312  *
313  * We use the switch_sp_and_call() assembly routine to switch our sp to the
314  * interrupt stacks and then call the appropriate dispatch function. In the
315  * case of interrupts which may block, softints and hardints, we always ensure
316  * that we are still on the interrupt thread when we call the epilog routine.
317  * This is not just important, it's necessary. If the interrupt thread blocked,
318  * we won't return from our switch_sp_and_call() function and instead we'll go
319  * through and set ourselves up to swtch() directly.
320  *
321  * New Interrupt Flow
322  * ------------------
323  *
324  * The apix module has its own interrupt path. This is done for various
325  * reasons. The first is that rather than having global interrupt vectors, we
326  * now have per-cpu vectors.
327  *
328  * The other substantial change is that the apix design does not use the TPR to
329  * mask interrupts below the current level. In fact, except for one special
330  * case, it does not use the TPR at all. Instead, it only uses the IF flag
331  * (cli/sti) to either block all interrupts or allow any interrupts to come in.
332  * The design is such that when interrupts are allowed to come in, if we are
333  * currently servicing a higher priority interupt, the new interrupt is treated
334  * as pending and serviced later. Specifically, in the pcplusmp module's
335  * apic_intr_enter() the code masks interrupts at or below the current
336  * IPL using the TPR before sending EOI, whereas the apix module's
337  * apix_intr_enter() simply sends EOI.
338  *
339  * The one special case where the apix code uses the TPR is when it calls
340  * through the apic_reg_ops function pointer apic_write_task_reg in
341  * apix_init_intr() to initially mask all levels and then finally to enable all
342  * levels.
343  *
344  * Recall that we come into the interrupt handler with all interrupts masked
345  * by the IF flag. This is because we set up the handler using an
346  * interrupt-gate which is defined architecturally to have cleared the IF flag
347  * for us.
348  *
349  * +--------------+    +---------------------+
350  * | _interrupt() |--->| apix_do_interrupt() |
351  * +--------------+    +---------------------+
352  *                               |
353  *                hard int? +----+--------+ softint?
354  *                          |             | (but no low-level looping)
355  *                   +-----------+        |
356  *                   | *setlvl() |        |
357  * +---------+       +-----------+        +----------------------------------+
358  * |apix_add_|    check IPL |                                                |
359  * |pending_ |<-------------+------+----------------------+                  |
360  * |hardint()|        low-level int|          hi-level int|                  |
361  * +---------+                     v                      v                  |
362  *     | check IPL       +-----------------+     +---------------+           |
363  *  +--+-----+           | apix_intr_      |     | apix_hilevel_ |           |
364  *  |        |           | thread_prolog() |     | intr_prolog() |           |
365  *  |      return        +-----------------+     +---------------+           |
366  *  |                         |                    | On intr                 |
367  *  |                   +------------+             | stack?  +------------+  |
368  *  |                   | switch_sp_ |             +---------| switch_sp_ |  |
369  *  |                   | and_call() |             |         | and_call() |  |
370  *  |                   +------------+             |         +------------+  |
371  *  |                         |                    |          |              |
372  *  |                   +----------------+     +----------------+            |
373  *  |                   | apix_dispatch_ |     | apix_dispatch_ |            |
374  *  |                   | lowlevel()     |     | hilevel()      |            |
375  *  |                   +----------------+     +----------------+            |
376  *  |                                |             |                         |
377  *  |                                v             v                         |
378  *  |                       +-------------------------+                      |
379  *  |                       |apix_dispatch_by_vector()|----+                 |
380  *  |                       +-------------------------+    |                 |
381  *  |               !XC_HI_PIL|         |         |        |                 |
382  *  |                       +---+   +-------+   +---+      |                 |
383  *  |                       |sti|   |*intr()|   |cli|      |                 |
384  *  |                       +---+   +-------+   +---+      |  hi-level?      |
385  *  |                          +---------------------------+----+            |
386  *  |                          v                low-level?      v            |
387  *  |                  +----------------+               +----------------+   |
388  *  |                  | apix_intr_     |               | apix_hilevel_  |   |
389  *  |                  | thread_epilog()|               | intr_epilog()  |   |
390  *  |                  +----------------+               +----------------+   |
391  *  |                          |                                |            |
392  *  |        v-----------------+--------------------------------+            |
393  *  |  +------------+                                                        |
394  *  |  | *setlvlx() |   +----------------------------------------------------+
395  *  |  +------------+   |
396  *  |      |            |            +--------------------------------+ low
397  *  v      v     v------+            v                                | level
398  * +------------------+      +------------------+      +-----------+  | pending?
399  * | apix_do_pending_ |----->| apix_do_pending_ |----->| apix_do_  |--+
400  * | hilevel()        |      | hardint()        |      | softint() |  |
401  * +------------------+      +------------------+      +-----------+    return
402  *     |                       |                         |
403  *     | while pending         | while pending           | while pending
404  *     | hi-level              | low-level               | softint
405  *     |                       |                         |
406  *  +---------------+        +-----------------+       +-----------------+
407  *  | apix_hilevel_ |        | apix_intr_      |       | apix_do_        |
408  *  | intr_prolog() |        | thread_prolog() |       | softint_prolog()|
409  *  +---------------+        +-----------------+       +-----------------+
410  *     | On intr                       |                      |
411  *     | stack? +------------+    +------------+        +------------+
412  *     +--------| switch_sp_ |    | switch_sp_ |        | switch_sp_ |
413  *     |        | and_call() |    | and_call() |        | and_call() |
414  *     |        +------------+    +------------+        +------------+
415  *     |           |                   |                      |
416  *  +------------------+   +------------------+   +------------------------+
417  *  | apix_dispatch_   |   | apix_dispatch_   |   | apix_dispatch_softint()|
418  *  | pending_hilevel()|   | pending_hardint()|   +------------------------+
419  *  +------------------+   +------------------+      |    |      |      |
420  *    |         |           |         |              |    |      |      |
421  *    | +----------------+  | +----------------+     |    |      |      |
422  *    | | apix_hilevel_  |  | | apix_intr_     |     |    |      |      |
423  *    | | intr_epilog()  |  | | thread_epilog()|     |    |      |      |
424  *    | +----------------+  | +----------------+     |    |      |      |
425  *    |         |           |       |                |    |      |      |
426  *    |   +------------+    |  +----------+   +------+    |      |      |
427  *    |   | *setlvlx() |    |  |*setlvlx()|   |           |      |      |
428  *    |   +------------+    |  +----------+   |   +----------+   |   +---------+
429  *    |                     |               +---+ |av_       | +---+ |apix_do_ |
430  * +---------------------------------+      |sti| |dispatch_ | |cli| |softint_ |
431  * | apix_dispatch_pending_autovect()|      +---+ |softvect()| +---+ |epilog() |
432  * +---------------------------------+            +----------+       +---------+
433  *  |!XC_HI_PIL  |       |         |                    |
434  * +---+  +-------+    +---+  +----------+          +-------+
435  * |sti|  |*intr()|    |cli|  |apix_post_|          |*intr()|
436  * +---+  +-------+    +---+  |hardint() |          +-------+
437  *                            +----------+
438  */
439 
440 #include <sys/cpuvar.h>
441 #include <sys/cpu_event.h>
442 #include <sys/regset.h>
443 #include <sys/psw.h>
444 #include <sys/types.h>
445 #include <sys/thread.h>
446 #include <sys/systm.h>
447 #include <sys/segments.h>
448 #include <sys/pcb.h>
449 #include <sys/trap.h>
450 #include <sys/ftrace.h>
451 #include <sys/traptrace.h>
452 #include <sys/clock.h>
453 #include <sys/panic.h>
454 #include <sys/disp.h>
455 #include <vm/seg_kp.h>
456 #include <sys/stack.h>
457 #include <sys/sysmacros.h>
458 #include <sys/cmn_err.h>
459 #include <sys/kstat.h>
460 #include <sys/smp_impldefs.h>
461 #include <sys/pool_pset.h>
462 #include <sys/zone.h>
463 #include <sys/bitmap.h>
464 #include <sys/archsystm.h>
465 #include <sys/machsystm.h>
466 #include <sys/ontrap.h>
467 #include <sys/x86_archext.h>
468 #include <sys/promif.h>
469 #include <vm/hat_i86.h>
470 #if defined(__xpv)
471 #include <sys/hypervisor.h>
472 #endif
473 
474 
475 #if defined(__xpv) && defined(DEBUG)
476 
477 /*
478  * This panic message is intended as an aid to interrupt debugging.
479  *
480  * The associated assertion tests the condition of enabling
481  * events when events are already enabled.  The implication
482  * being that whatever code the programmer thought was
483  * protected by having events disabled until the second
484  * enable happened really wasn't protected at all ..
485  */
486 
487 int stistipanic = 1;	/* controls the debug panic check */
488 const char *stistimsg = "stisti";
489 ulong_t laststi[NCPU];
490 
491 /*
492  * This variable tracks the last place events were disabled on each cpu
493  * it assists in debugging when asserts that interrupts are enabled trip.
494  */
495 ulong_t lastcli[NCPU];
496 
497 #endif
498 
499 void do_interrupt(struct regs *rp, trap_trace_rec_t *ttp);
500 
501 void (*do_interrupt_common)(struct regs *, trap_trace_rec_t *) = do_interrupt;
502 uintptr_t (*get_intr_handler)(int, short) = NULL;
503 
504 /*
505  * Set cpu's base SPL level to the highest active interrupt level
506  */
507 void
508 set_base_spl(void)
509 {
510 	struct cpu *cpu = CPU;
511 	uint16_t active = (uint16_t)cpu->cpu_intr_actv;
512 
513 	cpu->cpu_base_spl = active == 0 ? 0 : bsrw_insn(active);
514 }
515 
516 /*
517  * Do all the work necessary to set up the cpu and thread structures
518  * to dispatch a high-level interrupt.
519  *
520  * Returns 0 if we're -not- already on the high-level interrupt stack,
521  * (and *must* switch to it), non-zero if we are already on that stack.
522  *
523  * Called with interrupts masked.
524  * The 'pil' is already set to the appropriate level for rp->r_trapno.
525  */
526 static int
527 hilevel_intr_prolog(struct cpu *cpu, uint_t pil, uint_t oldpil, struct regs *rp)
528 {
529 	struct machcpu *mcpu = &cpu->cpu_m;
530 	uint_t mask;
531 	hrtime_t intrtime;
532 	hrtime_t now = tsc_read();
533 
534 	ASSERT(pil > LOCK_LEVEL);
535 
536 	if (pil == CBE_HIGH_PIL) {
537 		cpu->cpu_profile_pil = oldpil;
538 		if (USERMODE(rp->r_cs)) {
539 			cpu->cpu_profile_pc = 0;
540 			cpu->cpu_profile_upc = rp->r_pc;
541 			cpu->cpu_cpcprofile_pc = 0;
542 			cpu->cpu_cpcprofile_upc = rp->r_pc;
543 		} else {
544 			cpu->cpu_profile_pc = rp->r_pc;
545 			cpu->cpu_profile_upc = 0;
546 			cpu->cpu_cpcprofile_pc = rp->r_pc;
547 			cpu->cpu_cpcprofile_upc = 0;
548 		}
549 	}
550 
551 	mask = cpu->cpu_intr_actv & CPU_INTR_ACTV_HIGH_LEVEL_MASK;
552 	if (mask != 0) {
553 		int nestpil;
554 
555 		/*
556 		 * We have interrupted another high-level interrupt.
557 		 * Load starting timestamp, compute interval, update
558 		 * cumulative counter.
559 		 */
560 		nestpil = bsrw_insn((uint16_t)mask);
561 		ASSERT(nestpil < pil);
562 		intrtime = now -
563 		    mcpu->pil_high_start[nestpil - (LOCK_LEVEL + 1)];
564 		mcpu->intrstat[nestpil][0] += intrtime;
565 		cpu->cpu_intracct[cpu->cpu_mstate] += intrtime;
566 		/*
567 		 * Another high-level interrupt is active below this one, so
568 		 * there is no need to check for an interrupt thread.  That
569 		 * will be done by the lowest priority high-level interrupt
570 		 * active.
571 		 */
572 	} else {
573 		kthread_t *t = cpu->cpu_thread;
574 
575 		/*
576 		 * See if we are interrupting a low-level interrupt thread.
577 		 * If so, account for its time slice only if its time stamp
578 		 * is non-zero.
579 		 */
580 		if ((t->t_flag & T_INTR_THREAD) != 0 && t->t_intr_start != 0) {
581 			intrtime = now - t->t_intr_start;
582 			mcpu->intrstat[t->t_pil][0] += intrtime;
583 			cpu->cpu_intracct[cpu->cpu_mstate] += intrtime;
584 			t->t_intr_start = 0;
585 		}
586 	}
587 
588 	/*
589 	 * Store starting timestamp in CPU structure for this PIL.
590 	 */
591 	mcpu->pil_high_start[pil - (LOCK_LEVEL + 1)] = now;
592 
593 	ASSERT((cpu->cpu_intr_actv & (1 << pil)) == 0);
594 
595 	if (pil == 15) {
596 		/*
597 		 * To support reentrant level 15 interrupts, we maintain a
598 		 * recursion count in the top half of cpu_intr_actv.  Only
599 		 * when this count hits zero do we clear the PIL 15 bit from
600 		 * the lower half of cpu_intr_actv.
601 		 */
602 		uint16_t *refcntp = (uint16_t *)&cpu->cpu_intr_actv + 1;
603 		(*refcntp)++;
604 	}
605 
606 	mask = cpu->cpu_intr_actv;
607 
608 	cpu->cpu_intr_actv |= (1 << pil);
609 
610 	return (mask & CPU_INTR_ACTV_HIGH_LEVEL_MASK);
611 }
612 
613 /*
614  * Does most of the work of returning from a high level interrupt.
615  *
616  * Returns 0 if there are no more high level interrupts (in which
617  * case we must switch back to the interrupted thread stack) or
618  * non-zero if there are more (in which case we should stay on it).
619  *
620  * Called with interrupts masked
621  */
622 static int
623 hilevel_intr_epilog(struct cpu *cpu, uint_t pil, uint_t oldpil, uint_t vecnum)
624 {
625 	struct machcpu *mcpu = &cpu->cpu_m;
626 	uint_t mask;
627 	hrtime_t intrtime;
628 	hrtime_t now = tsc_read();
629 
630 	ASSERT(mcpu->mcpu_pri == pil);
631 
632 	cpu->cpu_stats.sys.intr[pil - 1]++;
633 
634 	ASSERT(cpu->cpu_intr_actv & (1 << pil));
635 
636 	if (pil == 15) {
637 		/*
638 		 * To support reentrant level 15 interrupts, we maintain a
639 		 * recursion count in the top half of cpu_intr_actv.  Only
640 		 * when this count hits zero do we clear the PIL 15 bit from
641 		 * the lower half of cpu_intr_actv.
642 		 */
643 		uint16_t *refcntp = (uint16_t *)&cpu->cpu_intr_actv + 1;
644 
645 		ASSERT(*refcntp > 0);
646 
647 		if (--(*refcntp) == 0)
648 			cpu->cpu_intr_actv &= ~(1 << pil);
649 	} else {
650 		cpu->cpu_intr_actv &= ~(1 << pil);
651 	}
652 
653 	ASSERT(mcpu->pil_high_start[pil - (LOCK_LEVEL + 1)] != 0);
654 
655 	intrtime = now - mcpu->pil_high_start[pil - (LOCK_LEVEL + 1)];
656 	mcpu->intrstat[pil][0] += intrtime;
657 	cpu->cpu_intracct[cpu->cpu_mstate] += intrtime;
658 
659 	/*
660 	 * Check for lower-pil nested high-level interrupt beneath
661 	 * current one.  If so, place a starting timestamp in its
662 	 * pil_high_start entry.
663 	 */
664 	mask = cpu->cpu_intr_actv & CPU_INTR_ACTV_HIGH_LEVEL_MASK;
665 	if (mask != 0) {
666 		int nestpil;
667 
668 		/*
669 		 * find PIL of nested interrupt
670 		 */
671 		nestpil = bsrw_insn((uint16_t)mask);
672 		ASSERT(nestpil < pil);
673 		mcpu->pil_high_start[nestpil - (LOCK_LEVEL + 1)] = now;
674 		/*
675 		 * (Another high-level interrupt is active below this one,
676 		 * so there is no need to check for an interrupt
677 		 * thread.  That will be done by the lowest priority
678 		 * high-level interrupt active.)
679 		 */
680 	} else {
681 		/*
682 		 * Check to see if there is a low-level interrupt active.
683 		 * If so, place a starting timestamp in the thread
684 		 * structure.
685 		 */
686 		kthread_t *t = cpu->cpu_thread;
687 
688 		if (t->t_flag & T_INTR_THREAD)
689 			t->t_intr_start = now;
690 	}
691 
692 	mcpu->mcpu_pri = oldpil;
693 	(void) (*setlvlx)(oldpil, vecnum);
694 
695 	return (cpu->cpu_intr_actv & CPU_INTR_ACTV_HIGH_LEVEL_MASK);
696 }
697 
698 /*
699  * Set up the cpu, thread and interrupt thread structures for
700  * executing an interrupt thread.  The new stack pointer of the
701  * interrupt thread (which *must* be switched to) is returned.
702  */
703 static caddr_t
704 intr_thread_prolog(struct cpu *cpu, caddr_t stackptr, uint_t pil)
705 {
706 	struct machcpu *mcpu = &cpu->cpu_m;
707 	kthread_t *t, *volatile it;
708 	hrtime_t now = tsc_read();
709 
710 	ASSERT(pil > 0);
711 	ASSERT((cpu->cpu_intr_actv & (1 << pil)) == 0);
712 	cpu->cpu_intr_actv |= (1 << pil);
713 
714 	/*
715 	 * Get set to run an interrupt thread.
716 	 * There should always be an interrupt thread, since we
717 	 * allocate one for each level on each CPU.
718 	 *
719 	 * t_intr_start could be zero due to cpu_intr_swtch_enter.
720 	 */
721 	t = cpu->cpu_thread;
722 	if ((t->t_flag & T_INTR_THREAD) && t->t_intr_start != 0) {
723 		hrtime_t intrtime = now - t->t_intr_start;
724 		mcpu->intrstat[t->t_pil][0] += intrtime;
725 		cpu->cpu_intracct[cpu->cpu_mstate] += intrtime;
726 		t->t_intr_start = 0;
727 	}
728 
729 	ASSERT(SA((uintptr_t)stackptr) == (uintptr_t)stackptr);
730 
731 	t->t_sp = (uintptr_t)stackptr;	/* mark stack in curthread for resume */
732 
733 	/*
734 	 * unlink the interrupt thread off the cpu
735 	 *
736 	 * Note that the code in kcpc_overflow_intr -relies- on the
737 	 * ordering of events here - in particular that t->t_lwp of
738 	 * the interrupt thread is set to the pinned thread *before*
739 	 * curthread is changed.
740 	 */
741 	it = cpu->cpu_intr_thread;
742 	cpu->cpu_intr_thread = it->t_link;
743 	it->t_intr = t;
744 	it->t_lwp = t->t_lwp;
745 
746 	/*
747 	 * (threads on the interrupt thread free list could have state
748 	 * preset to TS_ONPROC, but it helps in debugging if
749 	 * they're TS_FREE.)
750 	 */
751 	it->t_state = TS_ONPROC;
752 
753 	cpu->cpu_thread = it;		/* new curthread on this cpu */
754 	it->t_pil = (uchar_t)pil;
755 	it->t_pri = intr_pri + (pri_t)pil;
756 	it->t_intr_start = now;
757 
758 	return (it->t_stk);
759 }
760 
761 
762 #ifdef DEBUG
763 int intr_thread_cnt;
764 #endif
765 
766 /*
767  * Called with interrupts disabled
768  */
769 static void
770 intr_thread_epilog(struct cpu *cpu, uint_t vec, uint_t oldpil)
771 {
772 	struct machcpu *mcpu = &cpu->cpu_m;
773 	kthread_t *t;
774 	kthread_t *it = cpu->cpu_thread;	/* curthread */
775 	uint_t pil, basespl;
776 	hrtime_t intrtime;
777 	hrtime_t now = tsc_read();
778 
779 	pil = it->t_pil;
780 	cpu->cpu_stats.sys.intr[pil - 1]++;
781 
782 	ASSERT(it->t_intr_start != 0);
783 	intrtime = now - it->t_intr_start;
784 	mcpu->intrstat[pil][0] += intrtime;
785 	cpu->cpu_intracct[cpu->cpu_mstate] += intrtime;
786 
787 	ASSERT(cpu->cpu_intr_actv & (1 << pil));
788 	cpu->cpu_intr_actv &= ~(1 << pil);
789 
790 	/*
791 	 * If there is still an interrupted thread underneath this one
792 	 * then the interrupt was never blocked and the return is
793 	 * fairly simple.  Otherwise it isn't.
794 	 */
795 	if ((t = it->t_intr) == NULL) {
796 		/*
797 		 * The interrupted thread is no longer pinned underneath
798 		 * the interrupt thread.  This means the interrupt must
799 		 * have blocked, and the interrupted thread has been
800 		 * unpinned, and has probably been running around the
801 		 * system for a while.
802 		 *
803 		 * Since there is no longer a thread under this one, put
804 		 * this interrupt thread back on the CPU's free list and
805 		 * resume the idle thread which will dispatch the next
806 		 * thread to run.
807 		 */
808 #ifdef DEBUG
809 		intr_thread_cnt++;
810 #endif
811 		cpu->cpu_stats.sys.intrblk++;
812 		/*
813 		 * Set CPU's base SPL based on active interrupts bitmask
814 		 */
815 		set_base_spl();
816 		basespl = cpu->cpu_base_spl;
817 		mcpu->mcpu_pri = basespl;
818 		(*setlvlx)(basespl, vec);
819 		(void) splhigh();
820 		sti();
821 		it->t_state = TS_FREE;
822 		/*
823 		 * Return interrupt thread to pool
824 		 */
825 		it->t_link = cpu->cpu_intr_thread;
826 		cpu->cpu_intr_thread = it;
827 		swtch();
828 		panic("intr_thread_epilog: swtch returned");
829 		/*NOTREACHED*/
830 	}
831 
832 	/*
833 	 * Return interrupt thread to the pool
834 	 */
835 	it->t_link = cpu->cpu_intr_thread;
836 	cpu->cpu_intr_thread = it;
837 	it->t_state = TS_FREE;
838 
839 	basespl = cpu->cpu_base_spl;
840 	pil = MAX(oldpil, basespl);
841 	mcpu->mcpu_pri = pil;
842 	(*setlvlx)(pil, vec);
843 	t->t_intr_start = now;
844 	cpu->cpu_thread = t;
845 }
846 
847 /*
848  * intr_get_time() is a resource for interrupt handlers to determine how
849  * much time has been spent handling the current interrupt. Such a function
850  * is needed because higher level interrupts can arrive during the
851  * processing of an interrupt.  intr_get_time() only returns time spent in the
852  * current interrupt handler.
853  *
854  * The caller must be calling from an interrupt handler running at a pil
855  * below or at lock level. Timings are not provided for high-level
856  * interrupts.
857  *
858  * The first time intr_get_time() is called while handling an interrupt,
859  * it returns the time since the interrupt handler was invoked. Subsequent
860  * calls will return the time since the prior call to intr_get_time(). Time
861  * is returned as ticks. Use scalehrtimef() to convert ticks to nsec.
862  *
863  * Theory Of Intrstat[][]:
864  *
865  * uint64_t intrstat[pil][0..1] is an array indexed by pil level, with two
866  * uint64_ts per pil.
867  *
868  * intrstat[pil][0] is a cumulative count of the number of ticks spent
869  * handling all interrupts at the specified pil on this CPU. It is
870  * exported via kstats to the user.
871  *
872  * intrstat[pil][1] is always a count of ticks less than or equal to the
873  * value in [0]. The difference between [1] and [0] is the value returned
874  * by a call to intr_get_time(). At the start of interrupt processing,
875  * [0] and [1] will be equal (or nearly so). As the interrupt consumes
876  * time, [0] will increase, but [1] will remain the same. A call to
877  * intr_get_time() will return the difference, then update [1] to be the
878  * same as [0]. Future calls will return the time since the last call.
879  * Finally, when the interrupt completes, [1] is updated to the same as [0].
880  *
881  * Implementation:
882  *
883  * intr_get_time() works much like a higher level interrupt arriving. It
884  * "checkpoints" the timing information by incrementing intrstat[pil][0]
885  * to include elapsed running time, and by setting t_intr_start to rdtsc.
886  * It then sets the return value to intrstat[pil][0] - intrstat[pil][1],
887  * and updates intrstat[pil][1] to be the same as the new value of
888  * intrstat[pil][0].
889  *
890  * In the normal handling of interrupts, after an interrupt handler returns
891  * and the code in intr_thread() updates intrstat[pil][0], it then sets
892  * intrstat[pil][1] to the new value of intrstat[pil][0]. When [0] == [1],
893  * the timings are reset, i.e. intr_get_time() will return [0] - [1] which
894  * is 0.
895  *
896  * Whenever interrupts arrive on a CPU which is handling a lower pil
897  * interrupt, they update the lower pil's [0] to show time spent in the
898  * handler that they've interrupted. This results in a growing discrepancy
899  * between [0] and [1], which is returned the next time intr_get_time() is
900  * called. Time spent in the higher-pil interrupt will not be returned in
901  * the next intr_get_time() call from the original interrupt, because
902  * the higher-pil interrupt's time is accumulated in intrstat[higherpil][].
903  */
904 uint64_t
905 intr_get_time(void)
906 {
907 	struct cpu *cpu;
908 	struct machcpu *mcpu;
909 	kthread_t *t;
910 	uint64_t time, delta, ret;
911 	uint_t pil;
912 
913 	cli();
914 	cpu = CPU;
915 	mcpu = &cpu->cpu_m;
916 	t = cpu->cpu_thread;
917 	pil = t->t_pil;
918 	ASSERT((cpu->cpu_intr_actv & CPU_INTR_ACTV_HIGH_LEVEL_MASK) == 0);
919 	ASSERT(t->t_flag & T_INTR_THREAD);
920 	ASSERT(pil != 0);
921 	ASSERT(t->t_intr_start != 0);
922 
923 	time = tsc_read();
924 	delta = time - t->t_intr_start;
925 	t->t_intr_start = time;
926 
927 	time = mcpu->intrstat[pil][0] + delta;
928 	ret = time - mcpu->intrstat[pil][1];
929 	mcpu->intrstat[pil][0] = time;
930 	mcpu->intrstat[pil][1] = time;
931 	cpu->cpu_intracct[cpu->cpu_mstate] += delta;
932 
933 	sti();
934 	return (ret);
935 }
936 
937 static caddr_t
938 dosoftint_prolog(
939 	struct cpu *cpu,
940 	caddr_t stackptr,
941 	uint32_t st_pending,
942 	uint_t oldpil)
943 {
944 	kthread_t *t, *volatile it;
945 	struct machcpu *mcpu = &cpu->cpu_m;
946 	uint_t pil;
947 	hrtime_t now;
948 
949 top:
950 	ASSERT(st_pending == mcpu->mcpu_softinfo.st_pending);
951 
952 	pil = bsrw_insn((uint16_t)st_pending);
953 	if (pil <= oldpil || pil <= cpu->cpu_base_spl)
954 		return (0);
955 
956 	/*
957 	 * XX64	Sigh.
958 	 *
959 	 * This is a transliteration of the i386 assembler code for
960 	 * soft interrupts.  One question is "why does this need
961 	 * to be atomic?"  One possible race is -other- processors
962 	 * posting soft interrupts to us in set_pending() i.e. the
963 	 * CPU might get preempted just after the address computation,
964 	 * but just before the atomic transaction, so another CPU would
965 	 * actually set the original CPU's st_pending bit.  However,
966 	 * it looks like it would be simpler to disable preemption there.
967 	 * Are there other races for which preemption control doesn't work?
968 	 *
969 	 * The i386 assembler version -also- checks to see if the bit
970 	 * being cleared was actually set; if it wasn't, it rechecks
971 	 * for more.  This seems a bit strange, as the only code that
972 	 * ever clears the bit is -this- code running with interrupts
973 	 * disabled on -this- CPU.  This code would probably be cheaper:
974 	 *
975 	 * atomic_and_32((uint32_t *)&mcpu->mcpu_softinfo.st_pending,
976 	 *   ~(1 << pil));
977 	 *
978 	 * and t->t_preempt--/++ around set_pending() even cheaper,
979 	 * but at this point, correctness is critical, so we slavishly
980 	 * emulate the i386 port.
981 	 */
982 	if (atomic_btr32((uint32_t *)
983 	    &mcpu->mcpu_softinfo.st_pending, pil) == 0) {
984 		st_pending = mcpu->mcpu_softinfo.st_pending;
985 		goto top;
986 	}
987 
988 	mcpu->mcpu_pri = pil;
989 	(*setspl)(pil);
990 
991 	now = tsc_read();
992 
993 	/*
994 	 * Get set to run interrupt thread.
995 	 * There should always be an interrupt thread since we
996 	 * allocate one for each level on the CPU.
997 	 */
998 	it = cpu->cpu_intr_thread;
999 	cpu->cpu_intr_thread = it->t_link;
1000 
1001 	/* t_intr_start could be zero due to cpu_intr_swtch_enter. */
1002 	t = cpu->cpu_thread;
1003 	if ((t->t_flag & T_INTR_THREAD) && t->t_intr_start != 0) {
1004 		hrtime_t intrtime = now - t->t_intr_start;
1005 		mcpu->intrstat[pil][0] += intrtime;
1006 		cpu->cpu_intracct[cpu->cpu_mstate] += intrtime;
1007 		t->t_intr_start = 0;
1008 	}
1009 
1010 	/*
1011 	 * Note that the code in kcpc_overflow_intr -relies- on the
1012 	 * ordering of events here - in particular that t->t_lwp of
1013 	 * the interrupt thread is set to the pinned thread *before*
1014 	 * curthread is changed.
1015 	 */
1016 	it->t_lwp = t->t_lwp;
1017 	it->t_state = TS_ONPROC;
1018 
1019 	/*
1020 	 * Push interrupted thread onto list from new thread.
1021 	 * Set the new thread as the current one.
1022 	 * Set interrupted thread's T_SP because if it is the idle thread,
1023 	 * resume() may use that stack between threads.
1024 	 */
1025 
1026 	ASSERT(SA((uintptr_t)stackptr) == (uintptr_t)stackptr);
1027 	t->t_sp = (uintptr_t)stackptr;
1028 
1029 	it->t_intr = t;
1030 	cpu->cpu_thread = it;
1031 
1032 	/*
1033 	 * Set bit for this pil in CPU's interrupt active bitmask.
1034 	 */
1035 	ASSERT((cpu->cpu_intr_actv & (1 << pil)) == 0);
1036 	cpu->cpu_intr_actv |= (1 << pil);
1037 
1038 	/*
1039 	 * Initialize thread priority level from intr_pri
1040 	 */
1041 	it->t_pil = (uchar_t)pil;
1042 	it->t_pri = (pri_t)pil + intr_pri;
1043 	it->t_intr_start = now;
1044 
1045 	return (it->t_stk);
1046 }
1047 
1048 static void
1049 dosoftint_epilog(struct cpu *cpu, uint_t oldpil)
1050 {
1051 	struct machcpu *mcpu = &cpu->cpu_m;
1052 	kthread_t *t, *it;
1053 	uint_t pil, basespl;
1054 	hrtime_t intrtime;
1055 	hrtime_t now = tsc_read();
1056 
1057 	it = cpu->cpu_thread;
1058 	pil = it->t_pil;
1059 
1060 	cpu->cpu_stats.sys.intr[pil - 1]++;
1061 
1062 	ASSERT(cpu->cpu_intr_actv & (1 << pil));
1063 	cpu->cpu_intr_actv &= ~(1 << pil);
1064 	intrtime = now - it->t_intr_start;
1065 	mcpu->intrstat[pil][0] += intrtime;
1066 	cpu->cpu_intracct[cpu->cpu_mstate] += intrtime;
1067 
1068 	/*
1069 	 * If there is still an interrupted thread underneath this one
1070 	 * then the interrupt was never blocked and the return is
1071 	 * fairly simple.  Otherwise it isn't.
1072 	 */
1073 	if ((t = it->t_intr) == NULL) {
1074 		/*
1075 		 * Put thread back on the interrupt thread list.
1076 		 * This was an interrupt thread, so set CPU's base SPL.
1077 		 */
1078 		set_base_spl();
1079 		it->t_state = TS_FREE;
1080 		it->t_link = cpu->cpu_intr_thread;
1081 		cpu->cpu_intr_thread = it;
1082 		(void) splhigh();
1083 		sti();
1084 		swtch();
1085 		/*NOTREACHED*/
1086 		panic("dosoftint_epilog: swtch returned");
1087 	}
1088 	it->t_link = cpu->cpu_intr_thread;
1089 	cpu->cpu_intr_thread = it;
1090 	it->t_state = TS_FREE;
1091 	cpu->cpu_thread = t;
1092 	if (t->t_flag & T_INTR_THREAD)
1093 		t->t_intr_start = now;
1094 	basespl = cpu->cpu_base_spl;
1095 	pil = MAX(oldpil, basespl);
1096 	mcpu->mcpu_pri = pil;
1097 	(*setspl)(pil);
1098 }
1099 
1100 
1101 /*
1102  * Make the interrupted thread 'to' be runnable.
1103  *
1104  * Since t->t_sp has already been saved, t->t_pc is all
1105  * that needs to be set in this function.
1106  *
1107  * Returns the interrupt level of the interrupt thread.
1108  */
1109 int
1110 intr_passivate(
1111 	kthread_t *it,		/* interrupt thread */
1112 	kthread_t *t)		/* interrupted thread */
1113 {
1114 	extern void _sys_rtt();
1115 
1116 	ASSERT(it->t_flag & T_INTR_THREAD);
1117 	ASSERT(SA(t->t_sp) == t->t_sp);
1118 
1119 	t->t_pc = (uintptr_t)_sys_rtt;
1120 	return (it->t_pil);
1121 }
1122 
1123 /*
1124  * Create interrupt kstats for this CPU.
1125  */
1126 void
1127 cpu_create_intrstat(cpu_t *cp)
1128 {
1129 	int		i;
1130 	kstat_t		*intr_ksp;
1131 	kstat_named_t	*knp;
1132 	char		name[KSTAT_STRLEN];
1133 	zoneid_t	zoneid;
1134 
1135 	ASSERT(MUTEX_HELD(&cpu_lock));
1136 
1137 	if (pool_pset_enabled())
1138 		zoneid = GLOBAL_ZONEID;
1139 	else
1140 		zoneid = ALL_ZONES;
1141 
1142 	intr_ksp = kstat_create_zone("cpu", cp->cpu_id, "intrstat", "misc",
1143 	    KSTAT_TYPE_NAMED, PIL_MAX * 2, NULL, zoneid);
1144 
1145 	/*
1146 	 * Initialize each PIL's named kstat
1147 	 */
1148 	if (intr_ksp != NULL) {
1149 		intr_ksp->ks_update = cpu_kstat_intrstat_update;
1150 		knp = (kstat_named_t *)intr_ksp->ks_data;
1151 		intr_ksp->ks_private = cp;
1152 		for (i = 0; i < PIL_MAX; i++) {
1153 			(void) snprintf(name, KSTAT_STRLEN, "level-%d-time",
1154 			    i + 1);
1155 			kstat_named_init(&knp[i * 2], name, KSTAT_DATA_UINT64);
1156 			(void) snprintf(name, KSTAT_STRLEN, "level-%d-count",
1157 			    i + 1);
1158 			kstat_named_init(&knp[(i * 2) + 1], name,
1159 			    KSTAT_DATA_UINT64);
1160 		}
1161 		kstat_install(intr_ksp);
1162 	}
1163 }
1164 
1165 /*
1166  * Delete interrupt kstats for this CPU.
1167  */
1168 void
1169 cpu_delete_intrstat(cpu_t *cp)
1170 {
1171 	kstat_delete_byname_zone("cpu", cp->cpu_id, "intrstat", ALL_ZONES);
1172 }
1173 
1174 /*
1175  * Convert interrupt statistics from CPU ticks to nanoseconds and
1176  * update kstat.
1177  */
1178 int
1179 cpu_kstat_intrstat_update(kstat_t *ksp, int rw)
1180 {
1181 	kstat_named_t	*knp = ksp->ks_data;
1182 	cpu_t		*cpup = (cpu_t *)ksp->ks_private;
1183 	int		i;
1184 	hrtime_t	hrt;
1185 
1186 	if (rw == KSTAT_WRITE)
1187 		return (EACCES);
1188 
1189 	for (i = 0; i < PIL_MAX; i++) {
1190 		hrt = (hrtime_t)cpup->cpu_m.intrstat[i + 1][0];
1191 		scalehrtimef(&hrt);
1192 		knp[i * 2].value.ui64 = (uint64_t)hrt;
1193 		knp[(i * 2) + 1].value.ui64 = cpup->cpu_stats.sys.intr[i];
1194 	}
1195 
1196 	return (0);
1197 }
1198 
1199 /*
1200  * An interrupt thread is ending a time slice, so compute the interval it
1201  * ran for and update the statistic for its PIL.
1202  */
1203 void
1204 cpu_intr_swtch_enter(kthread_id_t t)
1205 {
1206 	uint64_t	interval;
1207 	uint64_t	start;
1208 	cpu_t		*cpu;
1209 
1210 	ASSERT((t->t_flag & T_INTR_THREAD) != 0);
1211 	ASSERT(t->t_pil > 0 && t->t_pil <= LOCK_LEVEL);
1212 
1213 	/*
1214 	 * We could be here with a zero timestamp. This could happen if:
1215 	 * an interrupt thread which no longer has a pinned thread underneath
1216 	 * it (i.e. it blocked at some point in its past) has finished running
1217 	 * its handler. intr_thread() updated the interrupt statistic for its
1218 	 * PIL and zeroed its timestamp. Since there was no pinned thread to
1219 	 * return to, swtch() gets called and we end up here.
1220 	 *
1221 	 * Note that we use atomic ops below (atomic_cas_64 and
1222 	 * atomic_add_64), which we don't use in the functions above,
1223 	 * because we're not called with interrupts blocked, but the
1224 	 * epilog/prolog functions are.
1225 	 */
1226 	if (t->t_intr_start) {
1227 		do {
1228 			start = t->t_intr_start;
1229 			interval = tsc_read() - start;
1230 		} while (atomic_cas_64(&t->t_intr_start, start, 0) != start);
1231 		cpu = CPU;
1232 		cpu->cpu_m.intrstat[t->t_pil][0] += interval;
1233 
1234 		atomic_add_64((uint64_t *)&cpu->cpu_intracct[cpu->cpu_mstate],
1235 		    interval);
1236 	} else
1237 		ASSERT(t->t_intr == NULL);
1238 }
1239 
1240 /*
1241  * An interrupt thread is returning from swtch(). Place a starting timestamp
1242  * in its thread structure.
1243  */
1244 void
1245 cpu_intr_swtch_exit(kthread_id_t t)
1246 {
1247 	uint64_t ts;
1248 
1249 	ASSERT((t->t_flag & T_INTR_THREAD) != 0);
1250 	ASSERT(t->t_pil > 0 && t->t_pil <= LOCK_LEVEL);
1251 
1252 	do {
1253 		ts = t->t_intr_start;
1254 	} while (atomic_cas_64(&t->t_intr_start, ts, tsc_read()) != ts);
1255 }
1256 
1257 /*
1258  * Dispatch a hilevel interrupt (one above LOCK_LEVEL)
1259  */
1260 /*ARGSUSED*/
1261 static void
1262 dispatch_hilevel(uint_t vector, uint_t arg2)
1263 {
1264 	sti();
1265 	av_dispatch_autovect(vector);
1266 	cli();
1267 }
1268 
1269 /*
1270  * Dispatch a soft interrupt
1271  */
1272 /*ARGSUSED*/
1273 static void
1274 dispatch_softint(uint_t oldpil, uint_t arg2)
1275 {
1276 	struct cpu *cpu = CPU;
1277 
1278 	sti();
1279 	av_dispatch_softvect((int)cpu->cpu_thread->t_pil);
1280 	cli();
1281 
1282 	/*
1283 	 * Must run softint_epilog() on the interrupt thread stack, since
1284 	 * there may not be a return from it if the interrupt thread blocked.
1285 	 */
1286 	dosoftint_epilog(cpu, oldpil);
1287 }
1288 
1289 /*
1290  * Dispatch a normal interrupt
1291  */
1292 static void
1293 dispatch_hardint(uint_t vector, uint_t oldipl)
1294 {
1295 	struct cpu *cpu = CPU;
1296 
1297 	sti();
1298 	av_dispatch_autovect(vector);
1299 	cli();
1300 
1301 	/*
1302 	 * Must run intr_thread_epilog() on the interrupt thread stack, since
1303 	 * there may not be a return from it if the interrupt thread blocked.
1304 	 */
1305 	intr_thread_epilog(cpu, vector, oldipl);
1306 }
1307 
1308 /*
1309  * Deliver any softints the current interrupt priority allows.
1310  * Called with interrupts disabled.
1311  */
1312 void
1313 dosoftint(struct regs *regs)
1314 {
1315 	struct cpu *cpu = CPU;
1316 	int oldipl;
1317 	caddr_t newsp;
1318 
1319 	while (cpu->cpu_softinfo.st_pending) {
1320 		oldipl = cpu->cpu_pri;
1321 		newsp = dosoftint_prolog(cpu, (caddr_t)regs,
1322 		    cpu->cpu_softinfo.st_pending, oldipl);
1323 		/*
1324 		 * If returned stack pointer is NULL, priority is too high
1325 		 * to run any of the pending softints now.
1326 		 * Break out and they will be run later.
1327 		 */
1328 		if (newsp == NULL)
1329 			break;
1330 		switch_sp_and_call(newsp, dispatch_softint, oldipl, 0);
1331 	}
1332 }
1333 
1334 /*
1335  * Interrupt service routine, called with interrupts disabled.
1336  */
1337 /*ARGSUSED*/
1338 void
1339 do_interrupt(struct regs *rp, trap_trace_rec_t *ttp)
1340 {
1341 	struct cpu *cpu = CPU;
1342 	int newipl, oldipl = cpu->cpu_pri;
1343 	uint_t vector;
1344 	caddr_t newsp;
1345 
1346 #ifdef TRAPTRACE
1347 	ttp->ttr_marker = TT_INTERRUPT;
1348 	ttp->ttr_ipl = 0xff;
1349 	ttp->ttr_pri = oldipl;
1350 	ttp->ttr_spl = cpu->cpu_base_spl;
1351 	ttp->ttr_vector = 0xff;
1352 #endif	/* TRAPTRACE */
1353 
1354 	cpu_idle_exit(CPU_IDLE_CB_FLAG_INTR);
1355 
1356 	++*(uint16_t *)&cpu->cpu_m.mcpu_istamp;
1357 
1358 	/*
1359 	 * If it's a softint go do it now.
1360 	 */
1361 	if (rp->r_trapno == T_SOFTINT) {
1362 		dosoftint(rp);
1363 		ASSERT(!interrupts_enabled());
1364 		return;
1365 	}
1366 
1367 	/*
1368 	 * Raise the interrupt priority.
1369 	 */
1370 	newipl = (*setlvl)(oldipl, (int *)&rp->r_trapno);
1371 #ifdef TRAPTRACE
1372 	ttp->ttr_ipl = newipl;
1373 #endif	/* TRAPTRACE */
1374 
1375 	/*
1376 	 * Bail if it is a spurious interrupt
1377 	 */
1378 	if (newipl == -1)
1379 		return;
1380 	cpu->cpu_pri = newipl;
1381 	vector = rp->r_trapno;
1382 #ifdef TRAPTRACE
1383 	ttp->ttr_vector = vector;
1384 #endif	/* TRAPTRACE */
1385 	if (newipl > LOCK_LEVEL) {
1386 		/*
1387 		 * High priority interrupts run on this cpu's interrupt stack.
1388 		 */
1389 		if (hilevel_intr_prolog(cpu, newipl, oldipl, rp) == 0) {
1390 			newsp = cpu->cpu_intr_stack;
1391 			switch_sp_and_call(newsp, dispatch_hilevel, vector, 0);
1392 		} else { /* already on the interrupt stack */
1393 			dispatch_hilevel(vector, 0);
1394 		}
1395 		(void) hilevel_intr_epilog(cpu, newipl, oldipl, vector);
1396 	} else {
1397 		/*
1398 		 * Run this interrupt in a separate thread.
1399 		 */
1400 		newsp = intr_thread_prolog(cpu, (caddr_t)rp, newipl);
1401 		switch_sp_and_call(newsp, dispatch_hardint, vector, oldipl);
1402 	}
1403 
1404 #if !defined(__xpv)
1405 	/*
1406 	 * Deliver any pending soft interrupts.
1407 	 */
1408 	if (cpu->cpu_softinfo.st_pending)
1409 		dosoftint(rp);
1410 #endif	/* !__xpv */
1411 }
1412 
1413 
1414 /*
1415  * Common tasks always done by _sys_rtt, called with interrupts disabled.
1416  * Returns 1 if returning to userland, 0 if returning to system mode.
1417  */
1418 int
1419 sys_rtt_common(struct regs *rp)
1420 {
1421 	kthread_t *tp;
1422 	extern void mutex_exit_critical_start();
1423 	extern long mutex_exit_critical_size;
1424 	extern void mutex_owner_running_critical_start();
1425 	extern long mutex_owner_running_critical_size;
1426 
1427 loop:
1428 
1429 	/*
1430 	 * Check if returning to user
1431 	 */
1432 	tp = CPU->cpu_thread;
1433 	if (USERMODE(rp->r_cs)) {
1434 		/*
1435 		 * Check if AST pending.
1436 		 */
1437 		if (tp->t_astflag) {
1438 			/*
1439 			 * Let trap() handle the AST
1440 			 */
1441 			sti();
1442 			rp->r_trapno = T_AST;
1443 			trap(rp, (caddr_t)0, CPU->cpu_id);
1444 			cli();
1445 			goto loop;
1446 		}
1447 
1448 #if defined(__amd64)
1449 		/*
1450 		 * We are done if segment registers do not need updating.
1451 		 */
1452 		if (tp->t_lwp->lwp_pcb.pcb_rupdate == 0)
1453 			return (1);
1454 
1455 		if (update_sregs(rp, tp->t_lwp)) {
1456 			/*
1457 			 * 1 or more of the selectors is bad.
1458 			 * Deliver a SIGSEGV.
1459 			 */
1460 			proc_t *p = ttoproc(tp);
1461 
1462 			sti();
1463 			mutex_enter(&p->p_lock);
1464 			tp->t_lwp->lwp_cursig = SIGSEGV;
1465 			mutex_exit(&p->p_lock);
1466 			psig();
1467 			tp->t_sig_check = 1;
1468 			cli();
1469 		}
1470 		tp->t_lwp->lwp_pcb.pcb_rupdate = 0;
1471 
1472 #endif	/* __amd64 */
1473 		return (1);
1474 	}
1475 
1476 	/*
1477 	 * Here if we are returning to supervisor mode.
1478 	 * Check for a kernel preemption request.
1479 	 */
1480 	if (CPU->cpu_kprunrun && (rp->r_ps & PS_IE)) {
1481 
1482 		/*
1483 		 * Do nothing if already in kpreempt
1484 		 */
1485 		if (!tp->t_preempt_lk) {
1486 			tp->t_preempt_lk = 1;
1487 			sti();
1488 			kpreempt(1); /* asynchronous kpreempt call */
1489 			cli();
1490 			tp->t_preempt_lk = 0;
1491 		}
1492 	}
1493 
1494 	/*
1495 	 * If we interrupted the mutex_exit() critical region we must
1496 	 * reset the PC back to the beginning to prevent missed wakeups
1497 	 * See the comments in mutex_exit() for details.
1498 	 */
1499 	if ((uintptr_t)rp->r_pc - (uintptr_t)mutex_exit_critical_start <
1500 	    mutex_exit_critical_size) {
1501 		rp->r_pc = (greg_t)mutex_exit_critical_start;
1502 	}
1503 
1504 	/*
1505 	 * If we interrupted the mutex_owner_running() critical region we
1506 	 * must reset the PC back to the beginning to prevent dereferencing
1507 	 * of a freed thread pointer. See the comments in mutex_owner_running
1508 	 * for details.
1509 	 */
1510 	if ((uintptr_t)rp->r_pc -
1511 	    (uintptr_t)mutex_owner_running_critical_start <
1512 	    mutex_owner_running_critical_size) {
1513 		rp->r_pc = (greg_t)mutex_owner_running_critical_start;
1514 	}
1515 
1516 	return (0);
1517 }
1518 
1519 void
1520 send_dirint(int cpuid, int int_level)
1521 {
1522 	(*send_dirintf)(cpuid, int_level);
1523 }
1524 
1525 #define	IS_FAKE_SOFTINT(flag, newpri)		\
1526 	(((flag) & PS_IE) &&				\
1527 	    (((*get_pending_spl)() > (newpri)) ||	\
1528 	    bsrw_insn((uint16_t)cpu->cpu_softinfo.st_pending) > (newpri)))
1529 
1530 /*
1531  * do_splx routine, takes new ipl to set
1532  * returns the old ipl.
1533  * We are careful not to set priority lower than CPU->cpu_base_pri,
1534  * even though it seems we're raising the priority, it could be set
1535  * higher at any time by an interrupt routine, so we must block interrupts
1536  * and look at CPU->cpu_base_pri
1537  */
1538 int
1539 do_splx(int newpri)
1540 {
1541 	ulong_t	flag;
1542 	cpu_t	*cpu;
1543 	int	curpri, basepri;
1544 
1545 	flag = intr_clear();
1546 	cpu = CPU; /* ints are disabled, now safe to cache cpu ptr */
1547 	curpri = cpu->cpu_m.mcpu_pri;
1548 	basepri = cpu->cpu_base_spl;
1549 	if (newpri < basepri)
1550 		newpri = basepri;
1551 	cpu->cpu_m.mcpu_pri = newpri;
1552 	(*setspl)(newpri);
1553 	/*
1554 	 * If we are going to reenable interrupts see if new priority level
1555 	 * allows pending softint delivery.
1556 	 */
1557 	if (IS_FAKE_SOFTINT(flag, newpri))
1558 		fakesoftint();
1559 	ASSERT(!interrupts_enabled());
1560 	intr_restore(flag);
1561 	return (curpri);
1562 }
1563 
1564 /*
1565  * Common spl raise routine, takes new ipl to set
1566  * returns the old ipl, will not lower ipl.
1567  */
1568 int
1569 splr(int newpri)
1570 {
1571 	ulong_t	flag;
1572 	cpu_t	*cpu;
1573 	int	curpri, basepri;
1574 
1575 	flag = intr_clear();
1576 	cpu = CPU; /* ints are disabled, now safe to cache cpu ptr */
1577 	curpri = cpu->cpu_m.mcpu_pri;
1578 	/*
1579 	 * Only do something if new priority is larger
1580 	 */
1581 	if (newpri > curpri) {
1582 		basepri = cpu->cpu_base_spl;
1583 		if (newpri < basepri)
1584 			newpri = basepri;
1585 		cpu->cpu_m.mcpu_pri = newpri;
1586 		(*setspl)(newpri);
1587 		/*
1588 		 * See if new priority level allows pending softint delivery
1589 		 */
1590 		if (IS_FAKE_SOFTINT(flag, newpri))
1591 			fakesoftint();
1592 	}
1593 	intr_restore(flag);
1594 	return (curpri);
1595 }
1596 
1597 int
1598 getpil(void)
1599 {
1600 	return (CPU->cpu_m.mcpu_pri);
1601 }
1602 
1603 int
1604 spl_xcall(void)
1605 {
1606 	return (splr(ipltospl(XCALL_PIL)));
1607 }
1608 
1609 int
1610 interrupts_enabled(void)
1611 {
1612 	ulong_t	flag;
1613 
1614 	flag = getflags();
1615 	return ((flag & PS_IE) == PS_IE);
1616 }
1617 
1618 #ifdef DEBUG
1619 void
1620 assert_ints_enabled(void)
1621 {
1622 	ASSERT(!interrupts_unleashed || interrupts_enabled());
1623 }
1624 #endif	/* DEBUG */
1625