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