1 // SPDX-License-Identifier: GPL-2.0
2
3 #include "io_uring.h"
4 #include "napi.h"
5
6 #ifdef CONFIG_NET_RX_BUSY_POLL
7
8 /* Timeout for cleanout of stale entries. */
9 #define NAPI_TIMEOUT (60 * SEC_CONVERSION)
10
11 struct io_napi_entry {
12 unsigned int napi_id;
13 struct list_head list;
14
15 unsigned long timeout;
16 struct hlist_node node;
17
18 struct rcu_head rcu;
19 };
20
io_napi_hash_find(struct hlist_head * hash_list,unsigned int napi_id)21 static struct io_napi_entry *io_napi_hash_find(struct hlist_head *hash_list,
22 unsigned int napi_id)
23 {
24 struct io_napi_entry *e;
25
26 hlist_for_each_entry_rcu(e, hash_list, node) {
27 if (e->napi_id != napi_id)
28 continue;
29 return e;
30 }
31
32 return NULL;
33 }
34
net_to_ktime(unsigned long t)35 static inline ktime_t net_to_ktime(unsigned long t)
36 {
37 /* napi approximating usecs, reverse busy_loop_current_time */
38 return ns_to_ktime(t << 10);
39 }
40
__io_napi_add(struct io_ring_ctx * ctx,struct socket * sock)41 void __io_napi_add(struct io_ring_ctx *ctx, struct socket *sock)
42 {
43 struct hlist_head *hash_list;
44 unsigned int napi_id;
45 struct sock *sk;
46 struct io_napi_entry *e;
47
48 sk = sock->sk;
49 if (!sk)
50 return;
51
52 napi_id = READ_ONCE(sk->sk_napi_id);
53
54 /* Non-NAPI IDs can be rejected. */
55 if (napi_id < MIN_NAPI_ID)
56 return;
57
58 hash_list = &ctx->napi_ht[hash_min(napi_id, HASH_BITS(ctx->napi_ht))];
59
60 rcu_read_lock();
61 e = io_napi_hash_find(hash_list, napi_id);
62 if (e) {
63 e->timeout = jiffies + NAPI_TIMEOUT;
64 rcu_read_unlock();
65 return;
66 }
67 rcu_read_unlock();
68
69 e = kmalloc(sizeof(*e), GFP_NOWAIT);
70 if (!e)
71 return;
72
73 e->napi_id = napi_id;
74 e->timeout = jiffies + NAPI_TIMEOUT;
75
76 spin_lock(&ctx->napi_lock);
77 if (unlikely(io_napi_hash_find(hash_list, napi_id))) {
78 spin_unlock(&ctx->napi_lock);
79 kfree(e);
80 return;
81 }
82
83 hlist_add_tail_rcu(&e->node, hash_list);
84 list_add_tail(&e->list, &ctx->napi_list);
85 spin_unlock(&ctx->napi_lock);
86 }
87
__io_napi_remove_stale(struct io_ring_ctx * ctx)88 static void __io_napi_remove_stale(struct io_ring_ctx *ctx)
89 {
90 struct io_napi_entry *e;
91 unsigned int i;
92
93 spin_lock(&ctx->napi_lock);
94 hash_for_each(ctx->napi_ht, i, e, node) {
95 if (time_after(jiffies, e->timeout)) {
96 list_del(&e->list);
97 hash_del_rcu(&e->node);
98 kfree_rcu(e, rcu);
99 }
100 }
101 spin_unlock(&ctx->napi_lock);
102 }
103
io_napi_remove_stale(struct io_ring_ctx * ctx,bool is_stale)104 static inline void io_napi_remove_stale(struct io_ring_ctx *ctx, bool is_stale)
105 {
106 if (is_stale)
107 __io_napi_remove_stale(ctx);
108 }
109
io_napi_busy_loop_timeout(ktime_t start_time,ktime_t bp)110 static inline bool io_napi_busy_loop_timeout(ktime_t start_time,
111 ktime_t bp)
112 {
113 if (bp) {
114 ktime_t end_time = ktime_add(start_time, bp);
115 ktime_t now = net_to_ktime(busy_loop_current_time());
116
117 return ktime_after(now, end_time);
118 }
119
120 return true;
121 }
122
io_napi_busy_loop_should_end(void * data,unsigned long start_time)123 static bool io_napi_busy_loop_should_end(void *data,
124 unsigned long start_time)
125 {
126 struct io_wait_queue *iowq = data;
127
128 if (signal_pending(current))
129 return true;
130 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
131 return true;
132 if (io_napi_busy_loop_timeout(net_to_ktime(start_time),
133 iowq->napi_busy_poll_dt))
134 return true;
135
136 return false;
137 }
138
__io_napi_do_busy_loop(struct io_ring_ctx * ctx,void * loop_end_arg)139 static bool __io_napi_do_busy_loop(struct io_ring_ctx *ctx,
140 void *loop_end_arg)
141 {
142 struct io_napi_entry *e;
143 bool (*loop_end)(void *, unsigned long) = NULL;
144 bool is_stale = false;
145
146 if (loop_end_arg)
147 loop_end = io_napi_busy_loop_should_end;
148
149 list_for_each_entry_rcu(e, &ctx->napi_list, list) {
150 napi_busy_loop_rcu(e->napi_id, loop_end, loop_end_arg,
151 ctx->napi_prefer_busy_poll, BUSY_POLL_BUDGET);
152
153 if (time_after(jiffies, e->timeout))
154 is_stale = true;
155 }
156
157 return is_stale;
158 }
159
io_napi_blocking_busy_loop(struct io_ring_ctx * ctx,struct io_wait_queue * iowq)160 static void io_napi_blocking_busy_loop(struct io_ring_ctx *ctx,
161 struct io_wait_queue *iowq)
162 {
163 unsigned long start_time = busy_loop_current_time();
164 void *loop_end_arg = NULL;
165 bool is_stale = false;
166
167 /* Singular lists use a different napi loop end check function and are
168 * only executed once.
169 */
170 if (list_is_singular(&ctx->napi_list))
171 loop_end_arg = iowq;
172
173 rcu_read_lock();
174 do {
175 is_stale = __io_napi_do_busy_loop(ctx, loop_end_arg);
176 } while (!io_napi_busy_loop_should_end(iowq, start_time) && !loop_end_arg);
177 rcu_read_unlock();
178
179 io_napi_remove_stale(ctx, is_stale);
180 }
181
182 /*
183 * io_napi_init() - Init napi settings
184 * @ctx: pointer to io-uring context structure
185 *
186 * Init napi settings in the io-uring context.
187 */
io_napi_init(struct io_ring_ctx * ctx)188 void io_napi_init(struct io_ring_ctx *ctx)
189 {
190 u64 sys_dt = READ_ONCE(sysctl_net_busy_poll) * NSEC_PER_USEC;
191
192 INIT_LIST_HEAD(&ctx->napi_list);
193 spin_lock_init(&ctx->napi_lock);
194 ctx->napi_prefer_busy_poll = false;
195 ctx->napi_busy_poll_dt = ns_to_ktime(sys_dt);
196 }
197
198 /*
199 * io_napi_free() - Deallocate napi
200 * @ctx: pointer to io-uring context structure
201 *
202 * Free the napi list and the hash table in the io-uring context.
203 */
io_napi_free(struct io_ring_ctx * ctx)204 void io_napi_free(struct io_ring_ctx *ctx)
205 {
206 struct io_napi_entry *e;
207 unsigned int i;
208
209 spin_lock(&ctx->napi_lock);
210 hash_for_each(ctx->napi_ht, i, e, node) {
211 hash_del_rcu(&e->node);
212 kfree_rcu(e, rcu);
213 }
214 spin_unlock(&ctx->napi_lock);
215 }
216
217 /*
218 * io_napi_register() - Register napi with io-uring
219 * @ctx: pointer to io-uring context structure
220 * @arg: pointer to io_uring_napi structure
221 *
222 * Register napi in the io-uring context.
223 */
io_register_napi(struct io_ring_ctx * ctx,void __user * arg)224 int io_register_napi(struct io_ring_ctx *ctx, void __user *arg)
225 {
226 const struct io_uring_napi curr = {
227 .busy_poll_to = ktime_to_us(ctx->napi_busy_poll_dt),
228 .prefer_busy_poll = ctx->napi_prefer_busy_poll
229 };
230 struct io_uring_napi napi;
231
232 if (ctx->flags & IORING_SETUP_IOPOLL)
233 return -EINVAL;
234 if (copy_from_user(&napi, arg, sizeof(napi)))
235 return -EFAULT;
236 if (napi.pad[0] || napi.pad[1] || napi.pad[2] || napi.resv)
237 return -EINVAL;
238
239 if (copy_to_user(arg, &curr, sizeof(curr)))
240 return -EFAULT;
241
242 WRITE_ONCE(ctx->napi_busy_poll_dt, napi.busy_poll_to * NSEC_PER_USEC);
243 WRITE_ONCE(ctx->napi_prefer_busy_poll, !!napi.prefer_busy_poll);
244 WRITE_ONCE(ctx->napi_enabled, true);
245 return 0;
246 }
247
248 /*
249 * io_napi_unregister() - Unregister napi with io-uring
250 * @ctx: pointer to io-uring context structure
251 * @arg: pointer to io_uring_napi structure
252 *
253 * Unregister napi. If arg has been specified copy the busy poll timeout and
254 * prefer busy poll setting to the passed in structure.
255 */
io_unregister_napi(struct io_ring_ctx * ctx,void __user * arg)256 int io_unregister_napi(struct io_ring_ctx *ctx, void __user *arg)
257 {
258 const struct io_uring_napi curr = {
259 .busy_poll_to = ktime_to_us(ctx->napi_busy_poll_dt),
260 .prefer_busy_poll = ctx->napi_prefer_busy_poll
261 };
262
263 if (arg && copy_to_user(arg, &curr, sizeof(curr)))
264 return -EFAULT;
265
266 WRITE_ONCE(ctx->napi_busy_poll_dt, 0);
267 WRITE_ONCE(ctx->napi_prefer_busy_poll, false);
268 WRITE_ONCE(ctx->napi_enabled, false);
269 return 0;
270 }
271
272 /*
273 * __io_napi_busy_loop() - execute busy poll loop
274 * @ctx: pointer to io-uring context structure
275 * @iowq: pointer to io wait queue
276 *
277 * Execute the busy poll loop and merge the spliced off list.
278 */
__io_napi_busy_loop(struct io_ring_ctx * ctx,struct io_wait_queue * iowq)279 void __io_napi_busy_loop(struct io_ring_ctx *ctx, struct io_wait_queue *iowq)
280 {
281 if (ctx->flags & IORING_SETUP_SQPOLL)
282 return;
283
284 iowq->napi_busy_poll_dt = READ_ONCE(ctx->napi_busy_poll_dt);
285 if (iowq->timeout != KTIME_MAX) {
286 ktime_t dt = ktime_sub(iowq->timeout, io_get_time(ctx));
287
288 iowq->napi_busy_poll_dt = min_t(u64, iowq->napi_busy_poll_dt, dt);
289 }
290
291 iowq->napi_prefer_busy_poll = READ_ONCE(ctx->napi_prefer_busy_poll);
292 io_napi_blocking_busy_loop(ctx, iowq);
293 }
294
295 /*
296 * io_napi_sqpoll_busy_poll() - busy poll loop for sqpoll
297 * @ctx: pointer to io-uring context structure
298 *
299 * Splice of the napi list and execute the napi busy poll loop.
300 */
io_napi_sqpoll_busy_poll(struct io_ring_ctx * ctx)301 int io_napi_sqpoll_busy_poll(struct io_ring_ctx *ctx)
302 {
303 bool is_stale = false;
304
305 if (!READ_ONCE(ctx->napi_busy_poll_dt))
306 return 0;
307 if (list_empty_careful(&ctx->napi_list))
308 return 0;
309
310 rcu_read_lock();
311 is_stale = __io_napi_do_busy_loop(ctx, NULL);
312 rcu_read_unlock();
313
314 io_napi_remove_stale(ctx, is_stale);
315 return 1;
316 }
317
318 #endif
319