xref: /linux/include/net/netdev_queues.h (revision 1a9239bb4253f9076b5b4b2a1a4e8d7defd77a95)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _LINUX_NET_QUEUES_H
3 #define _LINUX_NET_QUEUES_H
4 
5 #include <linux/netdevice.h>
6 
7 /**
8  * struct netdev_config - queue-related configuration for a netdev
9  * @hds_thresh:		HDS Threshold value.
10  * @hds_config:		HDS value from userspace.
11  */
12 struct netdev_config {
13 	u32	hds_thresh;
14 	u8	hds_config;
15 };
16 
17 /* See the netdev.yaml spec for definition of each statistic */
18 struct netdev_queue_stats_rx {
19 	u64 bytes;
20 	u64 packets;
21 	u64 alloc_fail;
22 
23 	u64 hw_drops;
24 	u64 hw_drop_overruns;
25 
26 	u64 csum_complete;
27 	u64 csum_unnecessary;
28 	u64 csum_none;
29 	u64 csum_bad;
30 
31 	u64 hw_gro_packets;
32 	u64 hw_gro_bytes;
33 	u64 hw_gro_wire_packets;
34 	u64 hw_gro_wire_bytes;
35 
36 	u64 hw_drop_ratelimits;
37 };
38 
39 struct netdev_queue_stats_tx {
40 	u64 bytes;
41 	u64 packets;
42 
43 	u64 hw_drops;
44 	u64 hw_drop_errors;
45 
46 	u64 csum_none;
47 	u64 needs_csum;
48 
49 	u64 hw_gso_packets;
50 	u64 hw_gso_bytes;
51 	u64 hw_gso_wire_packets;
52 	u64 hw_gso_wire_bytes;
53 
54 	u64 hw_drop_ratelimits;
55 
56 	u64 stop;
57 	u64 wake;
58 };
59 
60 /**
61  * struct netdev_stat_ops - netdev ops for fine grained stats
62  * @get_queue_stats_rx:	get stats for a given Rx queue
63  * @get_queue_stats_tx:	get stats for a given Tx queue
64  * @get_base_stats:	get base stats (not belonging to any live instance)
65  *
66  * Query stats for a given object. The values of the statistics are undefined
67  * on entry (specifically they are *not* zero-initialized). Drivers should
68  * assign values only to the statistics they collect. Statistics which are not
69  * collected must be left undefined.
70  *
71  * Queue objects are not necessarily persistent, and only currently active
72  * queues are queried by the per-queue callbacks. This means that per-queue
73  * statistics will not generally add up to the total number of events for
74  * the device. The @get_base_stats callback allows filling in the delta
75  * between events for currently live queues and overall device history.
76  * @get_base_stats can also be used to report any miscellaneous packets
77  * transferred outside of the main set of queues used by the networking stack.
78  * When the statistics for the entire device are queried, first @get_base_stats
79  * is issued to collect the delta, and then a series of per-queue callbacks.
80  * Only statistics which are set in @get_base_stats will be reported
81  * at the device level, meaning that unlike in queue callbacks, setting
82  * a statistic to zero in @get_base_stats is a legitimate thing to do.
83  * This is because @get_base_stats has a second function of designating which
84  * statistics are in fact correct for the entire device (e.g. when history
85  * for some of the events is not maintained, and reliable "total" cannot
86  * be provided).
87  *
88  * Device drivers can assume that when collecting total device stats,
89  * the @get_base_stats and subsequent per-queue calls are performed
90  * "atomically" (without releasing the rtnl_lock).
91  *
92  * Device drivers are encouraged to reset the per-queue statistics when
93  * number of queues change. This is because the primary use case for
94  * per-queue statistics is currently to detect traffic imbalance.
95  */
96 struct netdev_stat_ops {
97 	void (*get_queue_stats_rx)(struct net_device *dev, int idx,
98 				   struct netdev_queue_stats_rx *stats);
99 	void (*get_queue_stats_tx)(struct net_device *dev, int idx,
100 				   struct netdev_queue_stats_tx *stats);
101 	void (*get_base_stats)(struct net_device *dev,
102 			       struct netdev_queue_stats_rx *rx,
103 			       struct netdev_queue_stats_tx *tx);
104 };
105 
106 /**
107  * struct netdev_queue_mgmt_ops - netdev ops for queue management
108  *
109  * @ndo_queue_mem_size: Size of the struct that describes a queue's memory.
110  *
111  * @ndo_queue_mem_alloc: Allocate memory for an RX queue at the specified index.
112  *			 The new memory is written at the specified address.
113  *
114  * @ndo_queue_mem_free:	Free memory from an RX queue.
115  *
116  * @ndo_queue_start:	Start an RX queue with the specified memory and at the
117  *			specified index.
118  *
119  * @ndo_queue_stop:	Stop the RX queue at the specified index. The stopped
120  *			queue's memory is written at the specified address.
121  *
122  * Note that @ndo_queue_mem_alloc and @ndo_queue_mem_free may be called while
123  * the interface is closed. @ndo_queue_start and @ndo_queue_stop will only
124  * be called for an interface which is open.
125  */
126 struct netdev_queue_mgmt_ops {
127 	size_t			ndo_queue_mem_size;
128 	int			(*ndo_queue_mem_alloc)(struct net_device *dev,
129 						       void *per_queue_mem,
130 						       int idx);
131 	void			(*ndo_queue_mem_free)(struct net_device *dev,
132 						      void *per_queue_mem);
133 	int			(*ndo_queue_start)(struct net_device *dev,
134 						   void *per_queue_mem,
135 						   int idx);
136 	int			(*ndo_queue_stop)(struct net_device *dev,
137 						  void *per_queue_mem,
138 						  int idx);
139 };
140 
141 /**
142  * DOC: Lockless queue stopping / waking helpers.
143  *
144  * The netif_txq_maybe_stop() and __netif_txq_completed_wake()
145  * macros are designed to safely implement stopping
146  * and waking netdev queues without full lock protection.
147  *
148  * We assume that there can be no concurrent stop attempts and no concurrent
149  * wake attempts. The try-stop should happen from the xmit handler,
150  * while wake up should be triggered from NAPI poll context.
151  * The two may run concurrently (single producer, single consumer).
152  *
153  * The try-stop side is expected to run from the xmit handler and therefore
154  * it does not reschedule Tx (netif_tx_start_queue() instead of
155  * netif_tx_wake_queue()). Uses of the ``stop`` macros outside of the xmit
156  * handler may lead to xmit queue being enabled but not run.
157  * The waking side does not have similar context restrictions.
158  *
159  * The macros guarantee that rings will not remain stopped if there's
160  * space available, but they do *not* prevent false wake ups when
161  * the ring is full! Drivers should check for ring full at the start
162  * for the xmit handler.
163  *
164  * All descriptor ring indexes (and other relevant shared state) must
165  * be updated before invoking the macros.
166  */
167 
168 #define netif_txq_try_stop(txq, get_desc, start_thrs)			\
169 	({								\
170 		int _res;						\
171 									\
172 		netif_tx_stop_queue(txq);				\
173 		/* Producer index and stop bit must be visible		\
174 		 * to consumer before we recheck.			\
175 		 * Pairs with a barrier in __netif_txq_completed_wake(). \
176 		 */							\
177 		smp_mb__after_atomic();					\
178 									\
179 		/* We need to check again in a case another		\
180 		 * CPU has just made room available.			\
181 		 */							\
182 		_res = 0;						\
183 		if (unlikely(get_desc >= start_thrs)) {			\
184 			netif_tx_start_queue(txq);			\
185 			_res = -1;					\
186 		}							\
187 		_res;							\
188 	})								\
189 
190 /**
191  * netif_txq_maybe_stop() - locklessly stop a Tx queue, if needed
192  * @txq:	struct netdev_queue to stop/start
193  * @get_desc:	get current number of free descriptors (see requirements below!)
194  * @stop_thrs:	minimal number of available descriptors for queue to be left
195  *		enabled
196  * @start_thrs:	minimal number of descriptors to re-enable the queue, can be
197  *		equal to @stop_thrs or higher to avoid frequent waking
198  *
199  * All arguments may be evaluated multiple times, beware of side effects.
200  * @get_desc must be a formula or a function call, it must always
201  * return up-to-date information when evaluated!
202  * Expected to be used from ndo_start_xmit, see the comment on top of the file.
203  *
204  * Returns:
205  *	 0 if the queue was stopped
206  *	 1 if the queue was left enabled
207  *	-1 if the queue was re-enabled (raced with waking)
208  */
209 #define netif_txq_maybe_stop(txq, get_desc, stop_thrs, start_thrs)	\
210 	({								\
211 		int _res;						\
212 									\
213 		_res = 1;						\
214 		if (unlikely(get_desc < stop_thrs))			\
215 			_res = netif_txq_try_stop(txq, get_desc, start_thrs); \
216 		_res;							\
217 	})								\
218 
219 /* Variant of netdev_tx_completed_queue() which guarantees smp_mb() if
220  * @bytes != 0, regardless of kernel config.
221  */
222 static inline void
netdev_txq_completed_mb(struct netdev_queue * dev_queue,unsigned int pkts,unsigned int bytes)223 netdev_txq_completed_mb(struct netdev_queue *dev_queue,
224 			unsigned int pkts, unsigned int bytes)
225 {
226 	if (IS_ENABLED(CONFIG_BQL))
227 		netdev_tx_completed_queue(dev_queue, pkts, bytes);
228 	else if (bytes)
229 		smp_mb();
230 }
231 
232 /**
233  * __netif_txq_completed_wake() - locklessly wake a Tx queue, if needed
234  * @txq:	struct netdev_queue to stop/start
235  * @pkts:	number of packets completed
236  * @bytes:	number of bytes completed
237  * @get_desc:	get current number of free descriptors (see requirements below!)
238  * @start_thrs:	minimal number of descriptors to re-enable the queue
239  * @down_cond:	down condition, predicate indicating that the queue should
240  *		not be woken up even if descriptors are available
241  *
242  * All arguments may be evaluated multiple times.
243  * @get_desc must be a formula or a function call, it must always
244  * return up-to-date information when evaluated!
245  * Reports completed pkts/bytes to BQL.
246  *
247  * Returns:
248  *	 0 if the queue was woken up
249  *	 1 if the queue was already enabled (or disabled but @down_cond is true)
250  *	-1 if the queue was left unchanged (@start_thrs not reached)
251  */
252 #define __netif_txq_completed_wake(txq, pkts, bytes,			\
253 				   get_desc, start_thrs, down_cond)	\
254 	({								\
255 		int _res;						\
256 									\
257 		/* Report to BQL and piggy back on its barrier.		\
258 		 * Barrier makes sure that anybody stopping the queue	\
259 		 * after this point sees the new consumer index.	\
260 		 * Pairs with barrier in netif_txq_try_stop().		\
261 		 */							\
262 		netdev_txq_completed_mb(txq, pkts, bytes);		\
263 									\
264 		_res = -1;						\
265 		if (pkts && likely(get_desc >= start_thrs)) {		\
266 			_res = 1;					\
267 			if (unlikely(netif_tx_queue_stopped(txq)) &&	\
268 			    !(down_cond)) {				\
269 				netif_tx_wake_queue(txq);		\
270 				_res = 0;				\
271 			}						\
272 		}							\
273 		_res;							\
274 	})
275 
276 #define netif_txq_completed_wake(txq, pkts, bytes, get_desc, start_thrs) \
277 	__netif_txq_completed_wake(txq, pkts, bytes, get_desc, start_thrs, false)
278 
279 /* subqueue variants follow */
280 
281 #define netif_subqueue_try_stop(dev, idx, get_desc, start_thrs)		\
282 	({								\
283 		struct netdev_queue *txq;				\
284 									\
285 		txq = netdev_get_tx_queue(dev, idx);			\
286 		netif_txq_try_stop(txq, get_desc, start_thrs);		\
287 	})
288 
289 #define netif_subqueue_maybe_stop(dev, idx, get_desc, stop_thrs, start_thrs) \
290 	({								\
291 		struct netdev_queue *txq;				\
292 									\
293 		txq = netdev_get_tx_queue(dev, idx);			\
294 		netif_txq_maybe_stop(txq, get_desc, stop_thrs, start_thrs); \
295 	})
296 
297 #define netif_subqueue_completed_wake(dev, idx, pkts, bytes,		\
298 				      get_desc, start_thrs)		\
299 	({								\
300 		struct netdev_queue *txq;				\
301 									\
302 		txq = netdev_get_tx_queue(dev, idx);			\
303 		netif_txq_completed_wake(txq, pkts, bytes,		\
304 					 get_desc, start_thrs);		\
305 	})
306 
307 #endif
308