Line data Source code
1 : /* SPDX-License-Identifier: BSD-3-Clause
2 : * Copyright (C) 2019 Intel Corporation.
3 : * All rights reserved.
4 : */
5 :
6 : #include "spdk/stdinc.h"
7 : #include "spdk/config.h"
8 :
9 : #include <linux/errqueue.h>
10 : #include <liburing.h>
11 :
12 : #include "spdk/barrier.h"
13 : #include "spdk/env.h"
14 : #include "spdk/log.h"
15 : #include "spdk/pipe.h"
16 : #include "spdk/sock.h"
17 : #include "spdk/string.h"
18 : #include "spdk/util.h"
19 : #include "spdk/net.h"
20 : #include "spdk/file.h"
21 :
22 : #include "spdk_internal/sock.h"
23 : #include "spdk_internal/assert.h"
24 : #include "spdk/net.h"
25 :
26 : #define MAX_TMPBUF 1024
27 : #define PORTNUMLEN 32
28 : #define SPDK_SOCK_GROUP_QUEUE_DEPTH 4096
29 : #define SPDK_SOCK_CMG_INFO_SIZE (sizeof(struct cmsghdr) + sizeof(struct sock_extended_err))
30 :
31 : enum uring_task_type {
32 : URING_TASK_READ = 0,
33 : URING_TASK_ERRQUEUE,
34 : URING_TASK_WRITE,
35 : URING_TASK_CANCEL,
36 : };
37 :
38 : #if defined(SO_ZEROCOPY) && defined(MSG_ZEROCOPY)
39 : #define SPDK_ZEROCOPY
40 : #endif
41 :
42 : /* We don't know how big the buffers that the user posts will be, but this
43 : * is the maximum we'll ever allow it to receive in a single command.
44 : * If the user buffers are smaller, it will just receive less. */
45 : #define URING_MAX_RECV_SIZE (128 * 1024)
46 :
47 : /* We don't know how many buffers the user will post, but this is the
48 : * maximum number we'll take from the pool to post per group. */
49 : #define URING_BUF_POOL_SIZE 128
50 :
51 : /* We use 1 just so it's not zero and we can validate it's right. */
52 : #define URING_BUF_GROUP_ID 1
53 :
54 : enum spdk_uring_sock_task_status {
55 : SPDK_URING_SOCK_TASK_NOT_IN_USE = 0,
56 : SPDK_URING_SOCK_TASK_IN_PROCESS,
57 : };
58 :
59 : struct spdk_uring_task {
60 : enum spdk_uring_sock_task_status status;
61 : enum uring_task_type type;
62 : struct spdk_uring_sock *sock;
63 : struct msghdr msg;
64 : struct iovec iovs[IOV_BATCH_SIZE];
65 : int iov_cnt;
66 : struct spdk_sock_request *last_req;
67 : bool is_zcopy;
68 : STAILQ_ENTRY(spdk_uring_task) link;
69 : };
70 :
71 : struct spdk_uring_sock {
72 : struct spdk_sock base;
73 : int fd;
74 : uint32_t sendmsg_idx;
75 : struct spdk_uring_sock_group_impl *group;
76 : STAILQ_HEAD(, spdk_uring_buf_tracker) recv_stream;
77 : size_t recv_offset;
78 : struct spdk_uring_task write_task;
79 : struct spdk_uring_task errqueue_task;
80 : struct spdk_uring_task read_task;
81 : struct spdk_uring_task cancel_task;
82 : struct spdk_pipe *recv_pipe;
83 : void *recv_buf;
84 : int recv_buf_sz;
85 : bool zcopy;
86 : bool pending_recv;
87 : bool pending_group_remove;
88 : int zcopy_send_flags;
89 : int connection_status;
90 : int placement_id;
91 : uint8_t reserved[4];
92 : uint8_t buf[SPDK_SOCK_CMG_INFO_SIZE];
93 : TAILQ_ENTRY(spdk_uring_sock) link;
94 : char interface_name[IFNAMSIZ];
95 : };
96 : /* 'struct cmsghdr' is mapped to the buffer 'buf', and while first element
97 : * of this control message header has a size of 8 bytes, 'buf'
98 : * must be 8-byte aligned.
99 : */
100 : SPDK_STATIC_ASSERT(offsetof(struct spdk_uring_sock, buf) % 8 == 0,
101 : "Incorrect alignment: `buf` must be aligned to 8 bytes");
102 :
103 : TAILQ_HEAD(pending_recv_list, spdk_uring_sock);
104 :
105 : struct spdk_uring_buf_tracker {
106 : void *buf;
107 : size_t buflen;
108 : size_t len;
109 : void *ctx;
110 : int id;
111 : STAILQ_ENTRY(spdk_uring_buf_tracker) link;
112 : };
113 :
114 : struct spdk_uring_sock_group_impl {
115 : struct spdk_sock_group_impl base;
116 : struct io_uring uring;
117 : uint32_t io_inflight;
118 : uint32_t io_queued;
119 : uint32_t io_avail;
120 : struct pending_recv_list pending_recv;
121 :
122 : struct io_uring_buf_ring *buf_ring;
123 : uint32_t buf_ring_count;
124 : struct spdk_uring_buf_tracker *trackers;
125 : STAILQ_HEAD(, spdk_uring_buf_tracker) free_trackers;
126 : };
127 :
128 : static struct spdk_sock_impl_opts g_spdk_uring_sock_impl_opts = {
129 : .recv_buf_size = DEFAULT_SO_RCVBUF_SIZE,
130 : .send_buf_size = DEFAULT_SO_SNDBUF_SIZE,
131 : .enable_recv_pipe = true,
132 : .enable_quickack = false,
133 : .enable_placement_id = PLACEMENT_NONE,
134 : .enable_zerocopy_send_server = false,
135 : .enable_zerocopy_send_client = false,
136 : .zerocopy_threshold = 0,
137 : .tls_version = 0,
138 : .enable_ktls = false,
139 : .psk_key = NULL,
140 : .psk_identity = NULL
141 : };
142 :
143 : static struct spdk_sock_map g_map = {
144 : .entries = STAILQ_HEAD_INITIALIZER(g_map.entries),
145 : .mtx = PTHREAD_MUTEX_INITIALIZER
146 : };
147 :
148 : __attribute((destructor)) static void
149 1 : uring_sock_map_cleanup(void)
150 : {
151 1 : spdk_sock_map_cleanup(&g_map);
152 1 : }
153 :
154 : #define SPDK_URING_SOCK_REQUEST_IOV(req) ((struct iovec *)((uint8_t *)req + sizeof(struct spdk_sock_request)))
155 :
156 : #define __uring_sock(sock) (struct spdk_uring_sock *)sock
157 : #define __uring_group_impl(group) (struct spdk_uring_sock_group_impl *)group
158 :
159 : static void
160 0 : uring_sock_copy_impl_opts(struct spdk_sock_impl_opts *dest, const struct spdk_sock_impl_opts *src,
161 : size_t len)
162 : {
163 : #define FIELD_OK(field) \
164 : offsetof(struct spdk_sock_impl_opts, field) + sizeof(src->field) <= len
165 :
166 : #define SET_FIELD(field) \
167 : if (FIELD_OK(field)) { \
168 : dest->field = src->field; \
169 : }
170 :
171 0 : SET_FIELD(recv_buf_size);
172 0 : SET_FIELD(send_buf_size);
173 0 : SET_FIELD(enable_recv_pipe);
174 0 : SET_FIELD(enable_quickack);
175 0 : SET_FIELD(enable_placement_id);
176 0 : SET_FIELD(enable_zerocopy_send_server);
177 0 : SET_FIELD(enable_zerocopy_send_client);
178 0 : SET_FIELD(zerocopy_threshold);
179 0 : SET_FIELD(tls_version);
180 0 : SET_FIELD(enable_ktls);
181 0 : SET_FIELD(psk_key);
182 0 : SET_FIELD(psk_identity);
183 :
184 : #undef SET_FIELD
185 : #undef FIELD_OK
186 0 : }
187 :
188 : static int
189 0 : uring_sock_impl_get_opts(struct spdk_sock_impl_opts *opts, size_t *len)
190 : {
191 0 : if (!opts || !len) {
192 0 : errno = EINVAL;
193 0 : return -1;
194 : }
195 :
196 0 : assert(sizeof(*opts) >= *len);
197 0 : memset(opts, 0, *len);
198 :
199 0 : uring_sock_copy_impl_opts(opts, &g_spdk_uring_sock_impl_opts, *len);
200 0 : *len = spdk_min(*len, sizeof(g_spdk_uring_sock_impl_opts));
201 :
202 0 : return 0;
203 : }
204 :
205 : static int
206 0 : uring_sock_impl_set_opts(const struct spdk_sock_impl_opts *opts, size_t len)
207 : {
208 0 : if (!opts) {
209 0 : errno = EINVAL;
210 0 : return -1;
211 : }
212 :
213 0 : assert(sizeof(*opts) >= len);
214 0 : uring_sock_copy_impl_opts(&g_spdk_uring_sock_impl_opts, opts, len);
215 :
216 0 : return 0;
217 : }
218 :
219 : static void
220 0 : uring_opts_get_impl_opts(const struct spdk_sock_opts *opts, struct spdk_sock_impl_opts *dest)
221 : {
222 : /* Copy the default impl_opts first to cover cases when user's impl_opts is smaller */
223 0 : memcpy(dest, &g_spdk_uring_sock_impl_opts, sizeof(*dest));
224 :
225 0 : if (opts->impl_opts != NULL) {
226 0 : assert(sizeof(*dest) >= opts->impl_opts_size);
227 0 : uring_sock_copy_impl_opts(dest, opts->impl_opts, opts->impl_opts_size);
228 : }
229 0 : }
230 :
231 : static int
232 0 : uring_sock_getaddr(struct spdk_sock *_sock, char *saddr, int slen, uint16_t *sport,
233 : char *caddr, int clen, uint16_t *cport)
234 : {
235 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
236 :
237 0 : assert(sock != NULL);
238 0 : return spdk_net_getaddr(sock->fd, saddr, slen, sport, caddr, clen, cport);
239 : }
240 :
241 : static const char *
242 0 : uring_sock_get_interface_name(struct spdk_sock *_sock)
243 : {
244 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
245 : char saddr[64];
246 : int rc;
247 :
248 0 : rc = spdk_net_getaddr(sock->fd, saddr, sizeof(saddr), NULL, NULL, 0, NULL);
249 0 : if (rc != 0) {
250 0 : return NULL;
251 : }
252 :
253 0 : rc = spdk_net_get_interface_name(saddr, sock->interface_name,
254 : sizeof(sock->interface_name));
255 0 : if (rc != 0) {
256 0 : return NULL;
257 : }
258 :
259 0 : return sock->interface_name;
260 : }
261 :
262 : static int32_t
263 0 : uring_sock_get_numa_id(struct spdk_sock *sock)
264 : {
265 : const char *interface_name;
266 : uint32_t numa_id;
267 : int rc;
268 :
269 0 : interface_name = uring_sock_get_interface_name(sock);
270 0 : if (interface_name == NULL) {
271 0 : return SPDK_ENV_NUMA_ID_ANY;
272 : }
273 :
274 0 : rc = spdk_read_sysfs_attribute_uint32(&numa_id,
275 : "/sys/class/net/%s/device/numa_node", interface_name);
276 0 : if (rc == 0 && numa_id <= INT32_MAX) {
277 0 : return (int32_t)numa_id;
278 : } else {
279 0 : return SPDK_ENV_NUMA_ID_ANY;
280 : }
281 : }
282 :
283 : enum uring_sock_create_type {
284 : SPDK_SOCK_CREATE_LISTEN,
285 : SPDK_SOCK_CREATE_CONNECT,
286 : };
287 :
288 : static int
289 0 : uring_sock_alloc_pipe(struct spdk_uring_sock *sock, int sz)
290 : {
291 : uint8_t *new_buf;
292 : struct spdk_pipe *new_pipe;
293 : struct iovec siov[2];
294 : struct iovec diov[2];
295 : int sbytes;
296 : ssize_t bytes;
297 : int rc;
298 :
299 0 : if (sock->recv_buf_sz == sz) {
300 0 : return 0;
301 : }
302 :
303 : /* If the new size is 0, just free the pipe */
304 0 : if (sz == 0) {
305 0 : spdk_pipe_destroy(sock->recv_pipe);
306 0 : free(sock->recv_buf);
307 0 : sock->recv_pipe = NULL;
308 0 : sock->recv_buf = NULL;
309 0 : return 0;
310 0 : } else if (sz < MIN_SOCK_PIPE_SIZE) {
311 0 : SPDK_ERRLOG("The size of the pipe must be larger than %d\n", MIN_SOCK_PIPE_SIZE);
312 0 : return -1;
313 : }
314 :
315 : /* Round up to next 64 byte multiple */
316 0 : rc = posix_memalign((void **)&new_buf, 64, sz);
317 0 : if (rc != 0) {
318 0 : SPDK_ERRLOG("socket recv buf allocation failed\n");
319 0 : return -ENOMEM;
320 : }
321 0 : memset(new_buf, 0, sz);
322 :
323 0 : new_pipe = spdk_pipe_create(new_buf, sz);
324 0 : if (new_pipe == NULL) {
325 0 : SPDK_ERRLOG("socket pipe allocation failed\n");
326 0 : free(new_buf);
327 0 : return -ENOMEM;
328 : }
329 :
330 0 : if (sock->recv_pipe != NULL) {
331 : /* Pull all of the data out of the old pipe */
332 0 : sbytes = spdk_pipe_reader_get_buffer(sock->recv_pipe, sock->recv_buf_sz, siov);
333 0 : if (sbytes > sz) {
334 : /* Too much data to fit into the new pipe size */
335 0 : spdk_pipe_destroy(new_pipe);
336 0 : free(new_buf);
337 0 : return -EINVAL;
338 : }
339 :
340 0 : sbytes = spdk_pipe_writer_get_buffer(new_pipe, sz, diov);
341 0 : assert(sbytes == sz);
342 :
343 0 : bytes = spdk_iovcpy(siov, 2, diov, 2);
344 0 : spdk_pipe_writer_advance(new_pipe, bytes);
345 :
346 0 : spdk_pipe_destroy(sock->recv_pipe);
347 0 : free(sock->recv_buf);
348 : }
349 :
350 0 : sock->recv_buf_sz = sz;
351 0 : sock->recv_buf = new_buf;
352 0 : sock->recv_pipe = new_pipe;
353 :
354 0 : return 0;
355 : }
356 :
357 : static int
358 0 : uring_sock_set_recvbuf(struct spdk_sock *_sock, int sz)
359 : {
360 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
361 : int min_size;
362 : int rc;
363 :
364 0 : assert(sock != NULL);
365 :
366 0 : if (_sock->impl_opts.enable_recv_pipe) {
367 0 : rc = uring_sock_alloc_pipe(sock, sz);
368 0 : if (rc) {
369 0 : SPDK_ERRLOG("unable to allocate sufficient recvbuf with sz=%d on sock=%p\n", sz, _sock);
370 0 : return rc;
371 : }
372 : }
373 :
374 : /* Set kernel buffer size to be at least MIN_SO_RCVBUF_SIZE and
375 : * g_spdk_uring_sock_impl_opts.recv_buf_size. */
376 0 : min_size = spdk_max(MIN_SO_RCVBUF_SIZE, g_spdk_uring_sock_impl_opts.recv_buf_size);
377 :
378 0 : if (sz < min_size) {
379 0 : sz = min_size;
380 : }
381 :
382 0 : rc = setsockopt(sock->fd, SOL_SOCKET, SO_RCVBUF, &sz, sizeof(sz));
383 0 : if (rc < 0) {
384 0 : return rc;
385 : }
386 :
387 0 : _sock->impl_opts.recv_buf_size = sz;
388 :
389 0 : return 0;
390 : }
391 :
392 : static int
393 0 : uring_sock_set_sendbuf(struct spdk_sock *_sock, int sz)
394 : {
395 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
396 : int min_size;
397 : int rc;
398 :
399 0 : assert(sock != NULL);
400 :
401 : /* Set kernel buffer size to be at least MIN_SO_SNDBUF_SIZE and
402 : * g_spdk_uring_sock_impl_opts.seend_buf_size. */
403 0 : min_size = spdk_max(MIN_SO_SNDBUF_SIZE, g_spdk_uring_sock_impl_opts.send_buf_size);
404 :
405 0 : if (sz < min_size) {
406 0 : sz = min_size;
407 : }
408 :
409 0 : rc = setsockopt(sock->fd, SOL_SOCKET, SO_SNDBUF, &sz, sizeof(sz));
410 0 : if (rc < 0) {
411 0 : return rc;
412 : }
413 :
414 0 : _sock->impl_opts.send_buf_size = sz;
415 :
416 0 : return 0;
417 : }
418 :
419 : static struct spdk_uring_sock *
420 0 : uring_sock_alloc(int fd, struct spdk_sock_impl_opts *impl_opts, bool enable_zero_copy)
421 : {
422 : struct spdk_uring_sock *sock;
423 : #if defined(__linux__)
424 : int flag;
425 : int rc;
426 : #endif
427 :
428 0 : sock = calloc(1, sizeof(*sock));
429 0 : if (sock == NULL) {
430 0 : SPDK_ERRLOG("sock allocation failed\n");
431 0 : return NULL;
432 : }
433 :
434 0 : sock->fd = fd;
435 0 : memcpy(&sock->base.impl_opts, impl_opts, sizeof(*impl_opts));
436 :
437 0 : STAILQ_INIT(&sock->recv_stream);
438 :
439 : #if defined(__linux__)
440 0 : flag = 1;
441 :
442 0 : if (sock->base.impl_opts.enable_quickack) {
443 0 : rc = setsockopt(sock->fd, IPPROTO_TCP, TCP_QUICKACK, &flag, sizeof(flag));
444 0 : if (rc != 0) {
445 0 : SPDK_ERRLOG("quickack was failed to set\n");
446 : }
447 : }
448 :
449 0 : spdk_sock_get_placement_id(sock->fd, sock->base.impl_opts.enable_placement_id,
450 : &sock->placement_id);
451 : #ifdef SPDK_ZEROCOPY
452 : /* Try to turn on zero copy sends */
453 0 : flag = 1;
454 :
455 0 : if (enable_zero_copy) {
456 0 : rc = setsockopt(sock->fd, SOL_SOCKET, SO_ZEROCOPY, &flag, sizeof(flag));
457 0 : if (rc == 0) {
458 0 : sock->zcopy = true;
459 0 : sock->zcopy_send_flags = MSG_ZEROCOPY;
460 : }
461 : }
462 : #endif
463 : #endif
464 :
465 0 : return sock;
466 : }
467 :
468 : static struct spdk_sock *
469 0 : uring_sock_create(const char *ip, int port,
470 : enum uring_sock_create_type type,
471 : struct spdk_sock_opts *opts)
472 : {
473 : struct spdk_uring_sock *sock;
474 : struct spdk_sock_impl_opts impl_opts;
475 : char buf[MAX_TMPBUF];
476 : char portnum[PORTNUMLEN];
477 : char *p;
478 : const char *src_addr;
479 : uint16_t src_port;
480 : struct addrinfo hints, *res, *res0, *src_ai;
481 : int fd, flag;
482 0 : int val = 1;
483 : int rc;
484 0 : bool enable_zcopy_impl_opts = false;
485 0 : bool enable_zcopy_user_opts = true;
486 :
487 0 : assert(opts != NULL);
488 0 : uring_opts_get_impl_opts(opts, &impl_opts);
489 :
490 0 : if (ip == NULL) {
491 0 : return NULL;
492 : }
493 0 : if (ip[0] == '[') {
494 0 : snprintf(buf, sizeof(buf), "%s", ip + 1);
495 0 : p = strchr(buf, ']');
496 0 : if (p != NULL) {
497 0 : *p = '\0';
498 : }
499 0 : ip = (const char *) &buf[0];
500 : }
501 :
502 0 : snprintf(portnum, sizeof portnum, "%d", port);
503 0 : memset(&hints, 0, sizeof hints);
504 0 : hints.ai_family = PF_UNSPEC;
505 0 : hints.ai_socktype = SOCK_STREAM;
506 0 : hints.ai_flags = AI_NUMERICSERV;
507 0 : hints.ai_flags |= AI_PASSIVE;
508 0 : hints.ai_flags |= AI_NUMERICHOST;
509 0 : rc = getaddrinfo(ip, portnum, &hints, &res0);
510 0 : if (rc != 0) {
511 0 : SPDK_ERRLOG("getaddrinfo() failed %s (%d)\n", gai_strerror(rc), rc);
512 0 : return NULL;
513 : }
514 :
515 : /* try listen */
516 0 : fd = -1;
517 0 : for (res = res0; res != NULL; res = res->ai_next) {
518 0 : retry:
519 0 : fd = socket(res->ai_family, res->ai_socktype, res->ai_protocol);
520 0 : if (fd < 0) {
521 : /* error */
522 0 : continue;
523 : }
524 :
525 0 : val = impl_opts.recv_buf_size;
526 0 : rc = setsockopt(fd, SOL_SOCKET, SO_RCVBUF, &val, sizeof val);
527 : if (rc) {
528 : /* Not fatal */
529 : }
530 :
531 0 : val = impl_opts.send_buf_size;
532 0 : rc = setsockopt(fd, SOL_SOCKET, SO_SNDBUF, &val, sizeof val);
533 : if (rc) {
534 : /* Not fatal */
535 : }
536 :
537 0 : rc = setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &val, sizeof val);
538 0 : if (rc != 0) {
539 0 : close(fd);
540 0 : fd = -1;
541 : /* error */
542 0 : continue;
543 : }
544 0 : rc = setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &val, sizeof val);
545 0 : if (rc != 0) {
546 0 : close(fd);
547 0 : fd = -1;
548 : /* error */
549 0 : continue;
550 : }
551 :
552 0 : if (opts->ack_timeout) {
553 : #if defined(__linux__)
554 0 : val = opts->ack_timeout;
555 0 : rc = setsockopt(fd, IPPROTO_TCP, TCP_USER_TIMEOUT, &val, sizeof val);
556 0 : if (rc != 0) {
557 0 : close(fd);
558 0 : fd = -1;
559 : /* error */
560 0 : continue;
561 : }
562 : #else
563 : SPDK_WARNLOG("TCP_USER_TIMEOUT is not supported.\n");
564 : #endif
565 : }
566 :
567 :
568 :
569 : #if defined(SO_PRIORITY)
570 0 : if (opts != NULL && opts->priority) {
571 0 : rc = setsockopt(fd, SOL_SOCKET, SO_PRIORITY, &opts->priority, sizeof val);
572 0 : if (rc != 0) {
573 0 : close(fd);
574 0 : fd = -1;
575 : /* error */
576 0 : continue;
577 : }
578 : }
579 : #endif
580 0 : if (res->ai_family == AF_INET6) {
581 0 : rc = setsockopt(fd, IPPROTO_IPV6, IPV6_V6ONLY, &val, sizeof val);
582 0 : if (rc != 0) {
583 0 : close(fd);
584 0 : fd = -1;
585 : /* error */
586 0 : continue;
587 : }
588 : }
589 :
590 0 : if (type == SPDK_SOCK_CREATE_LISTEN) {
591 0 : rc = bind(fd, res->ai_addr, res->ai_addrlen);
592 0 : if (rc != 0) {
593 0 : SPDK_ERRLOG("bind() failed at port %d, errno = %d\n", port, errno);
594 0 : switch (errno) {
595 0 : case EINTR:
596 : /* interrupted? */
597 0 : close(fd);
598 0 : goto retry;
599 0 : case EADDRNOTAVAIL:
600 0 : SPDK_ERRLOG("IP address %s not available. "
601 : "Verify IP address in config file "
602 : "and make sure setup script is "
603 : "run before starting spdk app.\n", ip);
604 : /* FALLTHROUGH */
605 0 : default:
606 : /* try next family */
607 0 : close(fd);
608 0 : fd = -1;
609 0 : continue;
610 : }
611 : }
612 : /* bind OK */
613 0 : rc = listen(fd, 512);
614 0 : if (rc != 0) {
615 0 : SPDK_ERRLOG("listen() failed, errno = %d\n", errno);
616 0 : close(fd);
617 0 : fd = -1;
618 0 : break;
619 : }
620 :
621 0 : flag = fcntl(fd, F_GETFL);
622 0 : if (fcntl(fd, F_SETFL, flag | O_NONBLOCK) < 0) {
623 0 : SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%d)\n", fd, errno);
624 0 : close(fd);
625 0 : fd = -1;
626 0 : break;
627 : }
628 :
629 0 : enable_zcopy_impl_opts = impl_opts.enable_zerocopy_send_server;
630 0 : } else if (type == SPDK_SOCK_CREATE_CONNECT) {
631 0 : src_addr = SPDK_GET_FIELD(opts, src_addr, NULL, opts->opts_size);
632 0 : src_port = SPDK_GET_FIELD(opts, src_port, 0, opts->opts_size);
633 0 : if (src_addr != NULL || src_port != 0) {
634 0 : snprintf(portnum, sizeof(portnum), "%"PRIu16, src_port);
635 0 : memset(&hints, 0, sizeof hints);
636 0 : hints.ai_family = AF_UNSPEC;
637 0 : hints.ai_socktype = SOCK_STREAM;
638 0 : hints.ai_flags = AI_NUMERICSERV | AI_NUMERICHOST | AI_PASSIVE;
639 0 : rc = getaddrinfo(src_addr, src_port > 0 ? portnum : NULL,
640 : &hints, &src_ai);
641 0 : if (rc != 0 || src_ai == NULL) {
642 0 : SPDK_ERRLOG("getaddrinfo() failed %s (%d)\n",
643 : rc != 0 ? gai_strerror(rc) : "", rc);
644 0 : close(fd);
645 0 : fd = -1;
646 0 : break;
647 : }
648 0 : rc = bind(fd, src_ai->ai_addr, src_ai->ai_addrlen);
649 0 : if (rc != 0) {
650 0 : SPDK_ERRLOG("bind() failed errno %d (%s:%s)\n", errno,
651 : src_addr ? src_addr : "", portnum);
652 0 : close(fd);
653 0 : fd = -1;
654 0 : freeaddrinfo(src_ai);
655 0 : src_ai = NULL;
656 0 : break;
657 : }
658 0 : freeaddrinfo(src_ai);
659 0 : src_ai = NULL;
660 : }
661 :
662 0 : rc = connect(fd, res->ai_addr, res->ai_addrlen);
663 0 : if (rc != 0) {
664 0 : SPDK_ERRLOG("connect() failed, errno = %d\n", errno);
665 : /* try next family */
666 0 : close(fd);
667 0 : fd = -1;
668 0 : continue;
669 : }
670 :
671 0 : flag = fcntl(fd, F_GETFL);
672 0 : if (fcntl(fd, F_SETFL, flag & ~O_NONBLOCK) < 0) {
673 0 : SPDK_ERRLOG("fcntl can't set blocking mode for socket, fd: %d (%d)\n", fd, errno);
674 0 : close(fd);
675 0 : fd = -1;
676 0 : break;
677 : }
678 :
679 0 : enable_zcopy_impl_opts = impl_opts.enable_zerocopy_send_client;
680 : }
681 0 : break;
682 : }
683 0 : freeaddrinfo(res0);
684 :
685 0 : if (fd < 0) {
686 0 : return NULL;
687 : }
688 :
689 0 : enable_zcopy_user_opts = opts->zcopy && !spdk_net_is_loopback(fd);
690 0 : sock = uring_sock_alloc(fd, &impl_opts, enable_zcopy_user_opts && enable_zcopy_impl_opts);
691 0 : if (sock == NULL) {
692 0 : SPDK_ERRLOG("sock allocation failed\n");
693 0 : close(fd);
694 0 : return NULL;
695 : }
696 :
697 0 : return &sock->base;
698 : }
699 :
700 : static struct spdk_sock *
701 0 : uring_sock_listen(const char *ip, int port, struct spdk_sock_opts *opts)
702 : {
703 0 : if (spdk_interrupt_mode_is_enabled()) {
704 0 : SPDK_ERRLOG("Interrupt mode is not supported in the uring sock implementation.");
705 0 : return NULL;
706 : }
707 :
708 0 : return uring_sock_create(ip, port, SPDK_SOCK_CREATE_LISTEN, opts);
709 : }
710 :
711 : static struct spdk_sock *
712 0 : uring_sock_connect(const char *ip, int port, struct spdk_sock_opts *opts)
713 : {
714 0 : if (spdk_interrupt_mode_is_enabled()) {
715 0 : SPDK_ERRLOG("Interrupt mode is not supported in the uring sock implementation.");
716 0 : return NULL;
717 : }
718 :
719 0 : return uring_sock_create(ip, port, SPDK_SOCK_CREATE_CONNECT, opts);
720 : }
721 :
722 : static struct spdk_sock *
723 0 : uring_sock_accept(struct spdk_sock *_sock)
724 : {
725 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
726 : struct sockaddr_storage sa;
727 : socklen_t salen;
728 : int rc, fd;
729 : struct spdk_uring_sock *new_sock;
730 : int flag;
731 :
732 0 : memset(&sa, 0, sizeof(sa));
733 0 : salen = sizeof(sa);
734 :
735 0 : assert(sock != NULL);
736 :
737 0 : rc = accept(sock->fd, (struct sockaddr *)&sa, &salen);
738 :
739 0 : if (rc == -1) {
740 0 : return NULL;
741 : }
742 :
743 0 : fd = rc;
744 :
745 0 : flag = fcntl(fd, F_GETFL);
746 0 : if ((flag & O_NONBLOCK) && (fcntl(fd, F_SETFL, flag & ~O_NONBLOCK) < 0)) {
747 0 : SPDK_ERRLOG("fcntl can't set blocking mode for socket, fd: %d (%d)\n", fd, errno);
748 0 : close(fd);
749 0 : return NULL;
750 : }
751 :
752 : #if defined(SO_PRIORITY)
753 : /* The priority is not inherited, so call this function again */
754 0 : if (sock->base.opts.priority) {
755 0 : rc = setsockopt(fd, SOL_SOCKET, SO_PRIORITY, &sock->base.opts.priority, sizeof(int));
756 0 : if (rc != 0) {
757 0 : close(fd);
758 0 : return NULL;
759 : }
760 : }
761 : #endif
762 :
763 0 : new_sock = uring_sock_alloc(fd, &sock->base.impl_opts, sock->zcopy);
764 0 : if (new_sock == NULL) {
765 0 : close(fd);
766 0 : return NULL;
767 : }
768 :
769 0 : return &new_sock->base;
770 : }
771 :
772 : static int
773 0 : uring_sock_close(struct spdk_sock *_sock)
774 : {
775 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
776 :
777 0 : assert(TAILQ_EMPTY(&_sock->pending_reqs));
778 0 : assert(sock->group == NULL);
779 :
780 : /* If the socket fails to close, the best choice is to
781 : * leak the fd but continue to free the rest of the sock
782 : * memory. */
783 0 : close(sock->fd);
784 :
785 0 : spdk_pipe_destroy(sock->recv_pipe);
786 0 : free(sock->recv_buf);
787 0 : free(sock);
788 :
789 0 : return 0;
790 : }
791 :
792 : static ssize_t
793 0 : uring_sock_recv_from_pipe(struct spdk_uring_sock *sock, struct iovec *diov, int diovcnt)
794 : {
795 : struct iovec siov[2];
796 : int sbytes;
797 : ssize_t bytes;
798 : struct spdk_uring_sock_group_impl *group;
799 :
800 0 : sbytes = spdk_pipe_reader_get_buffer(sock->recv_pipe, sock->recv_buf_sz, siov);
801 0 : if (sbytes < 0) {
802 0 : errno = EINVAL;
803 0 : return -1;
804 0 : } else if (sbytes == 0) {
805 0 : errno = EAGAIN;
806 0 : return -1;
807 : }
808 :
809 0 : bytes = spdk_iovcpy(siov, 2, diov, diovcnt);
810 :
811 0 : if (bytes == 0) {
812 : /* The only way this happens is if diov is 0 length */
813 0 : errno = EINVAL;
814 0 : return -1;
815 : }
816 :
817 0 : spdk_pipe_reader_advance(sock->recv_pipe, bytes);
818 :
819 : /* If we drained the pipe, take it off the level-triggered list */
820 0 : if (sock->base.group_impl && spdk_pipe_reader_bytes_available(sock->recv_pipe) == 0) {
821 0 : group = __uring_group_impl(sock->base.group_impl);
822 0 : TAILQ_REMOVE(&group->pending_recv, sock, link);
823 0 : sock->pending_recv = false;
824 : }
825 :
826 0 : return bytes;
827 : }
828 :
829 : static inline ssize_t
830 0 : sock_readv(int fd, struct iovec *iov, int iovcnt)
831 : {
832 0 : struct msghdr msg = {
833 : .msg_iov = iov,
834 : .msg_iovlen = iovcnt,
835 : };
836 :
837 0 : return recvmsg(fd, &msg, MSG_DONTWAIT);
838 : }
839 :
840 : static inline ssize_t
841 0 : uring_sock_read(struct spdk_uring_sock *sock)
842 : {
843 : struct iovec iov[2];
844 : int bytes;
845 : struct spdk_uring_sock_group_impl *group;
846 :
847 0 : bytes = spdk_pipe_writer_get_buffer(sock->recv_pipe, sock->recv_buf_sz, iov);
848 :
849 0 : if (bytes > 0) {
850 0 : bytes = sock_readv(sock->fd, iov, 2);
851 0 : if (bytes > 0) {
852 0 : spdk_pipe_writer_advance(sock->recv_pipe, bytes);
853 0 : if (sock->base.group_impl && !sock->pending_recv) {
854 0 : group = __uring_group_impl(sock->base.group_impl);
855 0 : TAILQ_INSERT_TAIL(&group->pending_recv, sock, link);
856 0 : sock->pending_recv = true;
857 : }
858 : }
859 : }
860 :
861 0 : return bytes;
862 : }
863 :
864 : static int
865 0 : uring_sock_recv_next(struct spdk_sock *_sock, void **_buf, void **ctx)
866 : {
867 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
868 : struct spdk_uring_sock_group_impl *group;
869 : struct spdk_uring_buf_tracker *tr;
870 :
871 0 : if (sock->connection_status < 0) {
872 0 : errno = -sock->connection_status;
873 0 : return -1;
874 : }
875 :
876 0 : if (sock->recv_pipe != NULL) {
877 0 : errno = ENOTSUP;
878 0 : return -1;
879 : }
880 :
881 0 : group = __uring_group_impl(_sock->group_impl);
882 :
883 0 : tr = STAILQ_FIRST(&sock->recv_stream);
884 0 : if (tr == NULL) {
885 0 : if (sock->group->buf_ring_count > 0) {
886 : /* There are buffers posted, but data hasn't arrived. */
887 0 : errno = EAGAIN;
888 : } else {
889 : /* There are no buffers posted, so this won't ever
890 : * make forward progress. */
891 0 : errno = ENOBUFS;
892 : }
893 0 : return -1;
894 : }
895 0 : assert(sock->pending_recv == true);
896 0 : assert(tr->buf != NULL);
897 :
898 0 : *_buf = tr->buf + sock->recv_offset;
899 0 : *ctx = tr->ctx;
900 :
901 0 : STAILQ_REMOVE_HEAD(&sock->recv_stream, link);
902 0 : STAILQ_INSERT_HEAD(&group->free_trackers, tr, link);
903 :
904 0 : if (STAILQ_EMPTY(&sock->recv_stream)) {
905 0 : sock->pending_recv = false;
906 0 : TAILQ_REMOVE(&group->pending_recv, sock, link);
907 : }
908 :
909 0 : return tr->len - sock->recv_offset;
910 : }
911 :
912 : static ssize_t
913 0 : uring_sock_readv_no_pipe(struct spdk_sock *_sock, struct iovec *iovs, int iovcnt)
914 : {
915 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
916 : struct spdk_uring_buf_tracker *tr;
917 : struct iovec iov;
918 : ssize_t total, len;
919 : int i;
920 :
921 0 : if (sock->connection_status < 0) {
922 0 : errno = -sock->connection_status;
923 0 : return -1;
924 : }
925 :
926 0 : if (_sock->group_impl == NULL) {
927 : /* If not in a group just read from the socket the regular way. */
928 0 : return sock_readv(sock->fd, iovs, iovcnt);
929 : }
930 :
931 0 : if (STAILQ_EMPTY(&sock->recv_stream)) {
932 0 : if (sock->group->buf_ring_count == 0) {
933 : /* If the user hasn't posted any buffers, read from the socket
934 : * directly. */
935 :
936 0 : if (sock->pending_recv) {
937 0 : sock->pending_recv = false;
938 0 : TAILQ_REMOVE(&(__uring_group_impl(_sock->group_impl))->pending_recv, sock, link);
939 : }
940 :
941 0 : return sock_readv(sock->fd, iovs, iovcnt);
942 : }
943 :
944 0 : errno = EAGAIN;
945 0 : return -1;
946 : }
947 :
948 0 : total = 0;
949 0 : for (i = 0; i < iovcnt; i++) {
950 : /* Copy to stack so we can change it */
951 0 : iov = iovs[i];
952 :
953 0 : tr = STAILQ_FIRST(&sock->recv_stream);
954 0 : while (tr != NULL) {
955 0 : len = spdk_min(iov.iov_len, tr->len - sock->recv_offset);
956 0 : memcpy(iov.iov_base, tr->buf + sock->recv_offset, len);
957 :
958 0 : total += len;
959 0 : sock->recv_offset += len;
960 0 : iov.iov_base += len;
961 0 : iov.iov_len -= len;
962 :
963 0 : if (sock->recv_offset == tr->len) {
964 0 : sock->recv_offset = 0;
965 0 : STAILQ_REMOVE_HEAD(&sock->recv_stream, link);
966 0 : STAILQ_INSERT_HEAD(&sock->group->free_trackers, tr, link);
967 0 : spdk_sock_group_provide_buf(sock->group->base.group, tr->buf, tr->buflen, tr->ctx);
968 0 : tr = STAILQ_FIRST(&sock->recv_stream);
969 : }
970 :
971 0 : if (iov.iov_len == 0) {
972 0 : break;
973 : }
974 : }
975 : }
976 :
977 0 : if (STAILQ_EMPTY(&sock->recv_stream)) {
978 : struct spdk_uring_sock_group_impl *group;
979 :
980 0 : group = __uring_group_impl(_sock->group_impl);
981 0 : sock->pending_recv = false;
982 0 : TAILQ_REMOVE(&group->pending_recv, sock, link);
983 : }
984 :
985 0 : assert(total > 0);
986 0 : return total;
987 : }
988 :
989 : static ssize_t
990 0 : uring_sock_readv(struct spdk_sock *_sock, struct iovec *iov, int iovcnt)
991 : {
992 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
993 : int rc, i;
994 : size_t len;
995 :
996 0 : if (sock->connection_status < 0) {
997 0 : errno = -sock->connection_status;
998 0 : return -1;
999 : }
1000 :
1001 0 : if (sock->recv_pipe == NULL) {
1002 0 : return uring_sock_readv_no_pipe(_sock, iov, iovcnt);
1003 : }
1004 :
1005 0 : len = 0;
1006 0 : for (i = 0; i < iovcnt; i++) {
1007 0 : len += iov[i].iov_len;
1008 : }
1009 :
1010 0 : if (spdk_pipe_reader_bytes_available(sock->recv_pipe) == 0) {
1011 : /* If the user is receiving a sufficiently large amount of data,
1012 : * receive directly to their buffers. */
1013 0 : if (len >= MIN_SOCK_PIPE_SIZE) {
1014 0 : return sock_readv(sock->fd, iov, iovcnt);
1015 : }
1016 :
1017 : /* Otherwise, do a big read into our pipe */
1018 0 : rc = uring_sock_read(sock);
1019 0 : if (rc <= 0) {
1020 0 : return rc;
1021 : }
1022 : }
1023 :
1024 0 : return uring_sock_recv_from_pipe(sock, iov, iovcnt);
1025 : }
1026 :
1027 : static ssize_t
1028 0 : uring_sock_recv(struct spdk_sock *sock, void *buf, size_t len)
1029 : {
1030 : struct iovec iov[1];
1031 :
1032 0 : iov[0].iov_base = buf;
1033 0 : iov[0].iov_len = len;
1034 :
1035 0 : return uring_sock_readv(sock, iov, 1);
1036 : }
1037 :
1038 : static ssize_t
1039 0 : uring_sock_writev(struct spdk_sock *_sock, struct iovec *iov, int iovcnt)
1040 : {
1041 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1042 0 : struct msghdr msg = {
1043 : .msg_iov = iov,
1044 : .msg_iovlen = iovcnt,
1045 : };
1046 :
1047 0 : if (sock->write_task.status != SPDK_URING_SOCK_TASK_NOT_IN_USE) {
1048 0 : errno = EAGAIN;
1049 0 : return -1;
1050 : }
1051 :
1052 0 : return sendmsg(sock->fd, &msg, MSG_DONTWAIT);
1053 : }
1054 :
1055 : static ssize_t
1056 13 : sock_request_advance_offset(struct spdk_sock_request *req, ssize_t rc)
1057 : {
1058 : unsigned int offset;
1059 : size_t len;
1060 : int i;
1061 :
1062 13 : offset = req->internal.offset;
1063 37 : for (i = 0; i < req->iovcnt; i++) {
1064 : /* Advance by the offset first */
1065 28 : if (offset >= SPDK_SOCK_REQUEST_IOV(req, i)->iov_len) {
1066 2 : offset -= SPDK_SOCK_REQUEST_IOV(req, i)->iov_len;
1067 2 : continue;
1068 : }
1069 :
1070 : /* Calculate the remaining length of this element */
1071 26 : len = SPDK_SOCK_REQUEST_IOV(req, i)->iov_len - offset;
1072 :
1073 26 : if (len > (size_t)rc) {
1074 4 : req->internal.offset += rc;
1075 4 : return -1;
1076 : }
1077 :
1078 22 : offset = 0;
1079 22 : req->internal.offset += len;
1080 22 : rc -= len;
1081 : }
1082 :
1083 9 : return rc;
1084 : }
1085 :
1086 : static int
1087 11 : sock_complete_write_reqs(struct spdk_sock *_sock, ssize_t rc, bool is_zcopy)
1088 : {
1089 11 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1090 : struct spdk_sock_request *req;
1091 : int retval;
1092 :
1093 11 : if (is_zcopy) {
1094 : /* Handling overflow case, because we use psock->sendmsg_idx - 1 for the
1095 : * req->internal.offset, so sendmsg_idx should not be zero */
1096 0 : if (spdk_unlikely(sock->sendmsg_idx == UINT32_MAX)) {
1097 0 : sock->sendmsg_idx = 1;
1098 : } else {
1099 0 : sock->sendmsg_idx++;
1100 : }
1101 : }
1102 :
1103 : /* Consume the requests that were actually written */
1104 11 : req = TAILQ_FIRST(&_sock->queued_reqs);
1105 13 : while (req) {
1106 : /* req->internal.is_zcopy is true when the whole req or part of it is sent with zerocopy */
1107 13 : req->internal.is_zcopy = is_zcopy;
1108 :
1109 13 : rc = sock_request_advance_offset(req, rc);
1110 13 : if (rc < 0) {
1111 : /* This element was partially sent. */
1112 4 : return 0;
1113 : }
1114 :
1115 : /* Handled a full request. */
1116 9 : spdk_sock_request_pend(_sock, req);
1117 :
1118 9 : if (!req->internal.is_zcopy && req == TAILQ_FIRST(&_sock->pending_reqs)) {
1119 9 : retval = spdk_sock_request_put(_sock, req, 0);
1120 9 : if (retval) {
1121 0 : return retval;
1122 : }
1123 : } else {
1124 : /* Re-use the offset field to hold the sendmsg call index. The
1125 : * index is 0 based, so subtract one here because we've already
1126 : * incremented above. */
1127 0 : req->internal.offset = sock->sendmsg_idx - 1;
1128 : }
1129 :
1130 9 : if (rc == 0) {
1131 7 : break;
1132 : }
1133 :
1134 2 : req = TAILQ_FIRST(&_sock->queued_reqs);
1135 : }
1136 :
1137 7 : return 0;
1138 : }
1139 :
1140 : #ifdef SPDK_ZEROCOPY
1141 : static int
1142 0 : _sock_check_zcopy(struct spdk_sock *_sock, int status)
1143 : {
1144 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1145 : ssize_t rc;
1146 : struct sock_extended_err *serr;
1147 : struct cmsghdr *cm;
1148 : uint32_t idx;
1149 : struct spdk_sock_request *req, *treq;
1150 : bool found;
1151 :
1152 0 : assert(sock->zcopy == true);
1153 0 : if (spdk_unlikely(status) < 0) {
1154 0 : if (!TAILQ_EMPTY(&_sock->pending_reqs)) {
1155 0 : SPDK_ERRLOG("Attempting to receive from ERRQUEUE yielded error, but pending list still has orphaned entries, status =%d\n",
1156 : status);
1157 : } else {
1158 0 : SPDK_WARNLOG("Recvmsg yielded an error!\n");
1159 : }
1160 0 : return 0;
1161 : }
1162 :
1163 0 : cm = CMSG_FIRSTHDR(&sock->errqueue_task.msg);
1164 0 : if (!((cm->cmsg_level == SOL_IP && cm->cmsg_type == IP_RECVERR) ||
1165 0 : (cm->cmsg_level == SOL_IPV6 && cm->cmsg_type == IPV6_RECVERR))) {
1166 0 : SPDK_WARNLOG("Unexpected cmsg level or type!\n");
1167 0 : return 0;
1168 : }
1169 :
1170 0 : serr = (struct sock_extended_err *)CMSG_DATA(cm);
1171 0 : if (serr->ee_errno != 0 || serr->ee_origin != SO_EE_ORIGIN_ZEROCOPY) {
1172 0 : SPDK_WARNLOG("Unexpected extended error origin\n");
1173 0 : return 0;
1174 : }
1175 :
1176 : /* Most of the time, the pending_reqs array is in the exact
1177 : * order we need such that all of the requests to complete are
1178 : * in order, in the front. It is guaranteed that all requests
1179 : * belonging to the same sendmsg call are sequential, so once
1180 : * we encounter one match we can stop looping as soon as a
1181 : * non-match is found.
1182 : */
1183 0 : for (idx = serr->ee_info; idx <= serr->ee_data; idx++) {
1184 0 : found = false;
1185 0 : TAILQ_FOREACH_SAFE(req, &_sock->pending_reqs, internal.link, treq) {
1186 0 : if (!req->internal.is_zcopy) {
1187 : /* This wasn't a zcopy request. It was just waiting in line to complete */
1188 0 : rc = spdk_sock_request_put(_sock, req, 0);
1189 0 : if (rc < 0) {
1190 0 : return rc;
1191 : }
1192 0 : } else if (req->internal.offset == idx) {
1193 0 : found = true;
1194 0 : rc = spdk_sock_request_put(_sock, req, 0);
1195 0 : if (rc < 0) {
1196 0 : return rc;
1197 : }
1198 0 : } else if (found) {
1199 0 : break;
1200 : }
1201 : }
1202 : }
1203 :
1204 0 : return 0;
1205 : }
1206 :
1207 : static void
1208 0 : _sock_prep_errqueue(struct spdk_sock *_sock)
1209 : {
1210 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1211 0 : struct spdk_uring_task *task = &sock->errqueue_task;
1212 : struct io_uring_sqe *sqe;
1213 :
1214 0 : if (task->status == SPDK_URING_SOCK_TASK_IN_PROCESS) {
1215 0 : return;
1216 : }
1217 :
1218 0 : if (sock->pending_group_remove) {
1219 0 : return;
1220 : }
1221 :
1222 0 : assert(sock->group != NULL);
1223 0 : sock->group->io_queued++;
1224 :
1225 0 : sqe = io_uring_get_sqe(&sock->group->uring);
1226 0 : io_uring_prep_recvmsg(sqe, sock->fd, &task->msg, MSG_ERRQUEUE);
1227 0 : io_uring_sqe_set_data(sqe, task);
1228 0 : task->status = SPDK_URING_SOCK_TASK_IN_PROCESS;
1229 : }
1230 :
1231 : #endif
1232 :
1233 : static void
1234 0 : _sock_flush(struct spdk_sock *_sock)
1235 : {
1236 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1237 0 : struct spdk_uring_task *task = &sock->write_task;
1238 : uint32_t iovcnt;
1239 : struct io_uring_sqe *sqe;
1240 : int flags;
1241 :
1242 0 : if (task->status == SPDK_URING_SOCK_TASK_IN_PROCESS) {
1243 0 : return;
1244 : }
1245 :
1246 : #ifdef SPDK_ZEROCOPY
1247 0 : if (sock->zcopy) {
1248 0 : flags = MSG_DONTWAIT | sock->zcopy_send_flags;
1249 : } else
1250 : #endif
1251 : {
1252 0 : flags = MSG_DONTWAIT;
1253 : }
1254 :
1255 0 : iovcnt = spdk_sock_prep_reqs(&sock->base, task->iovs, task->iov_cnt, &task->last_req, &flags);
1256 0 : if (!iovcnt) {
1257 0 : return;
1258 : }
1259 :
1260 0 : task->iov_cnt = iovcnt;
1261 0 : assert(sock->group != NULL);
1262 0 : task->msg.msg_iov = task->iovs;
1263 0 : task->msg.msg_iovlen = task->iov_cnt;
1264 : #ifdef SPDK_ZEROCOPY
1265 0 : task->is_zcopy = (flags & MSG_ZEROCOPY) ? true : false;
1266 : #endif
1267 0 : sock->group->io_queued++;
1268 :
1269 0 : sqe = io_uring_get_sqe(&sock->group->uring);
1270 0 : io_uring_prep_sendmsg(sqe, sock->fd, &sock->write_task.msg, flags);
1271 0 : io_uring_sqe_set_data(sqe, task);
1272 0 : task->status = SPDK_URING_SOCK_TASK_IN_PROCESS;
1273 : }
1274 :
1275 : static void
1276 0 : _sock_prep_read(struct spdk_sock *_sock)
1277 : {
1278 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1279 0 : struct spdk_uring_task *task = &sock->read_task;
1280 : struct io_uring_sqe *sqe;
1281 :
1282 : /* Do not prepare read event */
1283 0 : if (task->status == SPDK_URING_SOCK_TASK_IN_PROCESS) {
1284 0 : return;
1285 : }
1286 :
1287 0 : if (sock->pending_group_remove) {
1288 0 : return;
1289 : }
1290 :
1291 0 : assert(sock->group != NULL);
1292 0 : sock->group->io_queued++;
1293 :
1294 0 : sqe = io_uring_get_sqe(&sock->group->uring);
1295 0 : io_uring_prep_recv(sqe, sock->fd, NULL, URING_MAX_RECV_SIZE, 0);
1296 0 : sqe->buf_group = URING_BUF_GROUP_ID;
1297 0 : sqe->flags |= IOSQE_BUFFER_SELECT;
1298 0 : io_uring_sqe_set_data(sqe, task);
1299 0 : task->status = SPDK_URING_SOCK_TASK_IN_PROCESS;
1300 : }
1301 :
1302 : static void
1303 0 : _sock_prep_cancel_task(struct spdk_sock *_sock, void *user_data)
1304 : {
1305 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1306 0 : struct spdk_uring_task *task = &sock->cancel_task;
1307 : struct io_uring_sqe *sqe;
1308 :
1309 0 : if (task->status == SPDK_URING_SOCK_TASK_IN_PROCESS) {
1310 0 : return;
1311 : }
1312 :
1313 0 : assert(sock->group != NULL);
1314 0 : sock->group->io_queued++;
1315 :
1316 0 : sqe = io_uring_get_sqe(&sock->group->uring);
1317 0 : io_uring_prep_cancel(sqe, user_data, 0);
1318 0 : io_uring_sqe_set_data(sqe, task);
1319 0 : task->status = SPDK_URING_SOCK_TASK_IN_PROCESS;
1320 : }
1321 :
1322 : static void
1323 0 : uring_sock_fail(struct spdk_uring_sock *sock, int status)
1324 : {
1325 0 : struct spdk_uring_sock_group_impl *group = sock->group;
1326 : int rc;
1327 :
1328 0 : sock->connection_status = status;
1329 0 : rc = spdk_sock_abort_requests(&sock->base);
1330 :
1331 : /* The user needs to be notified that this socket is dead. */
1332 0 : if (rc == 0 && sock->base.cb_fn != NULL &&
1333 0 : sock->pending_recv == false) {
1334 0 : sock->pending_recv = true;
1335 0 : TAILQ_INSERT_TAIL(&group->pending_recv, sock, link);
1336 : }
1337 0 : }
1338 :
1339 : static int
1340 0 : sock_uring_group_reap(struct spdk_uring_sock_group_impl *group, int max, int max_read_events,
1341 : struct spdk_sock **socks)
1342 : {
1343 : int i, count, ret;
1344 : struct io_uring_cqe *cqe;
1345 : struct spdk_uring_sock *sock, *tmp;
1346 : struct spdk_uring_task *task;
1347 : int status, bid, flags;
1348 : bool is_zcopy;
1349 :
1350 0 : for (i = 0; i < max; i++) {
1351 0 : ret = io_uring_peek_cqe(&group->uring, &cqe);
1352 0 : if (ret != 0) {
1353 0 : break;
1354 : }
1355 :
1356 0 : if (cqe == NULL) {
1357 0 : break;
1358 : }
1359 :
1360 0 : task = (struct spdk_uring_task *)cqe->user_data;
1361 0 : assert(task != NULL);
1362 0 : sock = task->sock;
1363 0 : assert(sock != NULL);
1364 0 : assert(sock->group != NULL);
1365 0 : assert(sock->group == group);
1366 0 : sock->group->io_inflight--;
1367 0 : sock->group->io_avail++;
1368 0 : status = cqe->res;
1369 0 : flags = cqe->flags;
1370 0 : io_uring_cqe_seen(&group->uring, cqe);
1371 :
1372 0 : task->status = SPDK_URING_SOCK_TASK_NOT_IN_USE;
1373 :
1374 0 : switch (task->type) {
1375 0 : case URING_TASK_READ:
1376 0 : if (status == -EAGAIN || status == -EWOULDBLOCK) {
1377 : /* This likely shouldn't happen, but would indicate that the
1378 : * kernel didn't have enough resources to queue a task internally. */
1379 0 : _sock_prep_read(&sock->base);
1380 0 : } else if (status == -ECANCELED) {
1381 0 : continue;
1382 0 : } else if (status == -ENOBUFS) {
1383 : /* There's data in the socket but the user hasn't provided any buffers.
1384 : * We need to notify the user that the socket has data pending. */
1385 0 : if (sock->base.cb_fn != NULL &&
1386 0 : sock->pending_recv == false) {
1387 0 : sock->pending_recv = true;
1388 0 : TAILQ_INSERT_TAIL(&group->pending_recv, sock, link);
1389 : }
1390 :
1391 0 : _sock_prep_read(&sock->base);
1392 0 : } else if (spdk_unlikely(status <= 0)) {
1393 0 : uring_sock_fail(sock, status < 0 ? status : -ECONNRESET);
1394 : } else {
1395 : struct spdk_uring_buf_tracker *tracker;
1396 :
1397 0 : assert((flags & IORING_CQE_F_BUFFER) != 0);
1398 :
1399 0 : bid = flags >> IORING_CQE_BUFFER_SHIFT;
1400 0 : tracker = &group->trackers[bid];
1401 :
1402 0 : assert(tracker->buf != NULL);
1403 0 : assert(tracker->len != 0);
1404 :
1405 : /* Append this data to the stream */
1406 0 : tracker->len = status;
1407 0 : STAILQ_INSERT_TAIL(&sock->recv_stream, tracker, link);
1408 0 : assert(group->buf_ring_count > 0);
1409 0 : group->buf_ring_count--;
1410 :
1411 0 : if (sock->base.cb_fn != NULL &&
1412 0 : sock->pending_recv == false) {
1413 0 : sock->pending_recv = true;
1414 0 : TAILQ_INSERT_TAIL(&group->pending_recv, sock, link);
1415 : }
1416 :
1417 0 : _sock_prep_read(&sock->base);
1418 : }
1419 0 : break;
1420 0 : case URING_TASK_WRITE:
1421 0 : if (status == -EAGAIN || status == -EWOULDBLOCK ||
1422 0 : (status == -ENOBUFS && sock->zcopy) ||
1423 : status == -ECANCELED) {
1424 0 : continue;
1425 0 : } else if (spdk_unlikely(status) < 0) {
1426 0 : uring_sock_fail(sock, status);
1427 : } else {
1428 0 : task->last_req = NULL;
1429 0 : task->iov_cnt = 0;
1430 0 : is_zcopy = task->is_zcopy;
1431 0 : task->is_zcopy = false;
1432 0 : sock_complete_write_reqs(&sock->base, status, is_zcopy);
1433 : }
1434 :
1435 0 : break;
1436 : #ifdef SPDK_ZEROCOPY
1437 0 : case URING_TASK_ERRQUEUE:
1438 0 : if (status == -EAGAIN || status == -EWOULDBLOCK) {
1439 0 : _sock_prep_errqueue(&sock->base);
1440 0 : } else if (status == -ECANCELED) {
1441 0 : continue;
1442 0 : } else if (spdk_unlikely(status < 0)) {
1443 0 : uring_sock_fail(sock, status);
1444 : } else {
1445 0 : _sock_check_zcopy(&sock->base, status);
1446 0 : _sock_prep_errqueue(&sock->base);
1447 : }
1448 0 : break;
1449 : #endif
1450 0 : case URING_TASK_CANCEL:
1451 : /* Do nothing */
1452 0 : break;
1453 0 : default:
1454 0 : SPDK_UNREACHABLE();
1455 : }
1456 : }
1457 :
1458 0 : if (!socks) {
1459 0 : return 0;
1460 : }
1461 0 : count = 0;
1462 0 : TAILQ_FOREACH_SAFE(sock, &group->pending_recv, link, tmp) {
1463 0 : if (count == max_read_events) {
1464 0 : break;
1465 : }
1466 :
1467 : /* If the socket's cb_fn is NULL, do not add it to socks array */
1468 0 : if (spdk_unlikely(sock->base.cb_fn == NULL)) {
1469 0 : assert(sock->pending_recv == true);
1470 0 : sock->pending_recv = false;
1471 0 : TAILQ_REMOVE(&group->pending_recv, sock, link);
1472 0 : continue;
1473 : }
1474 :
1475 0 : socks[count++] = &sock->base;
1476 : }
1477 :
1478 :
1479 : /* Cycle the pending_recv list so that each time we poll things aren't
1480 : * in the same order. Say we have 6 sockets in the list, named as follows:
1481 : * A B C D E F
1482 : * And all 6 sockets had the poll events, but max_events is only 3. That means
1483 : * psock currently points at D. We want to rearrange the list to the following:
1484 : * D E F A B C
1485 : *
1486 : * The variables below are named according to this example to make it easier to
1487 : * follow the swaps.
1488 : */
1489 0 : if (sock != NULL) {
1490 : struct spdk_uring_sock *ua, *uc, *ud, *uf;
1491 :
1492 : /* Capture pointers to the elements we need */
1493 0 : ud = sock;
1494 :
1495 0 : ua = TAILQ_FIRST(&group->pending_recv);
1496 0 : if (ua == ud) {
1497 0 : goto end;
1498 : }
1499 :
1500 0 : uf = TAILQ_LAST(&group->pending_recv, pending_recv_list);
1501 0 : if (uf == ud) {
1502 0 : TAILQ_REMOVE(&group->pending_recv, ud, link);
1503 0 : TAILQ_INSERT_HEAD(&group->pending_recv, ud, link);
1504 0 : goto end;
1505 : }
1506 :
1507 0 : uc = TAILQ_PREV(ud, pending_recv_list, link);
1508 0 : assert(uc != NULL);
1509 :
1510 : /* Break the link between C and D */
1511 0 : uc->link.tqe_next = NULL;
1512 :
1513 : /* Connect F to A */
1514 0 : uf->link.tqe_next = ua;
1515 0 : ua->link.tqe_prev = &uf->link.tqe_next;
1516 :
1517 : /* Fix up the list first/last pointers */
1518 0 : group->pending_recv.tqh_first = ud;
1519 0 : group->pending_recv.tqh_last = &uc->link.tqe_next;
1520 :
1521 : /* D is in front of the list, make tqe prev pointer point to the head of list */
1522 0 : ud->link.tqe_prev = &group->pending_recv.tqh_first;
1523 : }
1524 :
1525 0 : end:
1526 0 : return count;
1527 : }
1528 :
1529 : static int uring_sock_flush(struct spdk_sock *_sock);
1530 :
1531 : static void
1532 0 : uring_sock_writev_async(struct spdk_sock *_sock, struct spdk_sock_request *req)
1533 : {
1534 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1535 : int rc;
1536 :
1537 0 : if (spdk_unlikely(sock->connection_status)) {
1538 0 : req->cb_fn(req->cb_arg, sock->connection_status);
1539 0 : return;
1540 : }
1541 :
1542 0 : spdk_sock_request_queue(_sock, req);
1543 :
1544 0 : if (!sock->group) {
1545 0 : if (_sock->queued_iovcnt >= IOV_BATCH_SIZE) {
1546 0 : rc = uring_sock_flush(_sock);
1547 0 : if (rc < 0 && errno != EAGAIN) {
1548 0 : spdk_sock_abort_requests(_sock);
1549 : }
1550 : }
1551 : }
1552 : }
1553 :
1554 : static int
1555 0 : uring_sock_set_recvlowat(struct spdk_sock *_sock, int nbytes)
1556 : {
1557 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1558 : int val;
1559 : int rc;
1560 :
1561 0 : assert(sock != NULL);
1562 :
1563 0 : val = nbytes;
1564 0 : rc = setsockopt(sock->fd, SOL_SOCKET, SO_RCVLOWAT, &val, sizeof val);
1565 0 : if (rc != 0) {
1566 0 : return -1;
1567 : }
1568 0 : return 0;
1569 : }
1570 :
1571 : static bool
1572 0 : uring_sock_is_ipv6(struct spdk_sock *_sock)
1573 : {
1574 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1575 : struct sockaddr_storage sa;
1576 : socklen_t salen;
1577 : int rc;
1578 :
1579 0 : assert(sock != NULL);
1580 :
1581 0 : memset(&sa, 0, sizeof sa);
1582 0 : salen = sizeof sa;
1583 0 : rc = getsockname(sock->fd, (struct sockaddr *) &sa, &salen);
1584 0 : if (rc != 0) {
1585 0 : SPDK_ERRLOG("getsockname() failed (errno=%d)\n", errno);
1586 0 : return false;
1587 : }
1588 :
1589 0 : return (sa.ss_family == AF_INET6);
1590 : }
1591 :
1592 : static bool
1593 0 : uring_sock_is_ipv4(struct spdk_sock *_sock)
1594 : {
1595 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1596 : struct sockaddr_storage sa;
1597 : socklen_t salen;
1598 : int rc;
1599 :
1600 0 : assert(sock != NULL);
1601 :
1602 0 : memset(&sa, 0, sizeof sa);
1603 0 : salen = sizeof sa;
1604 0 : rc = getsockname(sock->fd, (struct sockaddr *) &sa, &salen);
1605 0 : if (rc != 0) {
1606 0 : SPDK_ERRLOG("getsockname() failed (errno=%d)\n", errno);
1607 0 : return false;
1608 : }
1609 :
1610 0 : return (sa.ss_family == AF_INET);
1611 : }
1612 :
1613 : static bool
1614 0 : uring_sock_is_connected(struct spdk_sock *_sock)
1615 : {
1616 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1617 : uint8_t byte;
1618 : int rc;
1619 :
1620 0 : rc = recv(sock->fd, &byte, 1, MSG_PEEK | MSG_DONTWAIT);
1621 0 : if (rc == 0) {
1622 0 : return false;
1623 : }
1624 :
1625 0 : if (rc < 0) {
1626 0 : if (errno == EAGAIN || errno == EWOULDBLOCK) {
1627 0 : return true;
1628 : }
1629 :
1630 0 : return false;
1631 : }
1632 :
1633 0 : return true;
1634 : }
1635 :
1636 : static struct spdk_sock_group_impl *
1637 0 : uring_sock_group_impl_get_optimal(struct spdk_sock *_sock, struct spdk_sock_group_impl *hint)
1638 : {
1639 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1640 : struct spdk_sock_group_impl *group;
1641 :
1642 0 : if (sock->placement_id != -1) {
1643 0 : spdk_sock_map_lookup(&g_map, sock->placement_id, &group, hint);
1644 0 : return group;
1645 : }
1646 :
1647 0 : return NULL;
1648 : }
1649 :
1650 : static int
1651 0 : uring_sock_group_impl_buf_pool_free(struct spdk_uring_sock_group_impl *group_impl)
1652 : {
1653 0 : if (group_impl->buf_ring) {
1654 0 : io_uring_unregister_buf_ring(&group_impl->uring, URING_BUF_GROUP_ID);
1655 0 : free(group_impl->buf_ring);
1656 : }
1657 :
1658 0 : free(group_impl->trackers);
1659 :
1660 0 : return 0;
1661 : }
1662 :
1663 : static int
1664 1 : uring_sock_group_impl_buf_pool_alloc(struct spdk_uring_sock_group_impl *group_impl)
1665 : {
1666 1 : struct io_uring_buf_reg buf_reg = {};
1667 : struct io_uring_buf_ring *buf_ring;
1668 : int i, rc;
1669 :
1670 1 : rc = posix_memalign((void **)&buf_ring, 0x1000, URING_BUF_POOL_SIZE * sizeof(struct io_uring_buf));
1671 1 : if (rc != 0) {
1672 : /* posix_memalign returns positive errno values */
1673 0 : return -rc;
1674 : }
1675 :
1676 1 : buf_reg.ring_addr = (unsigned long long)buf_ring;
1677 1 : buf_reg.ring_entries = URING_BUF_POOL_SIZE;
1678 1 : buf_reg.bgid = URING_BUF_GROUP_ID;
1679 :
1680 1 : rc = io_uring_register_buf_ring(&group_impl->uring, &buf_reg, 0);
1681 1 : if (rc != 0) {
1682 1 : free(buf_ring);
1683 1 : return rc;
1684 : }
1685 :
1686 0 : group_impl->buf_ring = buf_ring;
1687 0 : io_uring_buf_ring_init(group_impl->buf_ring);
1688 0 : group_impl->buf_ring_count = 0;
1689 :
1690 0 : group_impl->trackers = calloc(URING_BUF_POOL_SIZE, sizeof(struct spdk_uring_buf_tracker));
1691 0 : if (group_impl->trackers == NULL) {
1692 0 : uring_sock_group_impl_buf_pool_free(group_impl);
1693 0 : return -ENOMEM;
1694 : }
1695 :
1696 0 : STAILQ_INIT(&group_impl->free_trackers);
1697 :
1698 0 : for (i = 0; i < URING_BUF_POOL_SIZE; i++) {
1699 0 : struct spdk_uring_buf_tracker *tracker = &group_impl->trackers[i];
1700 :
1701 0 : tracker->buf = NULL;
1702 0 : tracker->len = 0;
1703 0 : tracker->ctx = NULL;
1704 0 : tracker->id = i;
1705 :
1706 0 : STAILQ_INSERT_TAIL(&group_impl->free_trackers, tracker, link);
1707 : }
1708 :
1709 0 : return 0;
1710 : }
1711 :
1712 : static struct spdk_sock_group_impl *
1713 1 : uring_sock_group_impl_create(void)
1714 : {
1715 : struct spdk_uring_sock_group_impl *group_impl;
1716 :
1717 1 : group_impl = calloc(1, sizeof(*group_impl));
1718 1 : if (group_impl == NULL) {
1719 0 : SPDK_ERRLOG("group_impl allocation failed\n");
1720 0 : return NULL;
1721 : }
1722 :
1723 1 : group_impl->io_avail = SPDK_SOCK_GROUP_QUEUE_DEPTH;
1724 :
1725 1 : if (io_uring_queue_init(SPDK_SOCK_GROUP_QUEUE_DEPTH, &group_impl->uring, 0) < 0) {
1726 0 : SPDK_ERRLOG("uring I/O context setup failure\n");
1727 0 : free(group_impl);
1728 0 : return NULL;
1729 : }
1730 :
1731 1 : TAILQ_INIT(&group_impl->pending_recv);
1732 :
1733 1 : if (uring_sock_group_impl_buf_pool_alloc(group_impl) < 0) {
1734 1 : SPDK_ERRLOG("Failed to create buffer ring."
1735 : "uring sock implementation is likely not supported on this kernel.\n");
1736 1 : io_uring_queue_exit(&group_impl->uring);
1737 1 : free(group_impl);
1738 1 : return NULL;
1739 : }
1740 :
1741 0 : if (g_spdk_uring_sock_impl_opts.enable_placement_id == PLACEMENT_CPU) {
1742 0 : spdk_sock_map_insert(&g_map, spdk_env_get_current_core(), &group_impl->base);
1743 : }
1744 :
1745 0 : return &group_impl->base;
1746 : }
1747 :
1748 : static int
1749 0 : uring_sock_group_impl_add_sock(struct spdk_sock_group_impl *_group,
1750 : struct spdk_sock *_sock)
1751 : {
1752 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1753 0 : struct spdk_uring_sock_group_impl *group = __uring_group_impl(_group);
1754 : int rc;
1755 :
1756 0 : sock->group = group;
1757 0 : sock->write_task.sock = sock;
1758 0 : sock->write_task.type = URING_TASK_WRITE;
1759 :
1760 0 : sock->read_task.sock = sock;
1761 0 : sock->read_task.type = URING_TASK_READ;
1762 :
1763 0 : sock->errqueue_task.sock = sock;
1764 0 : sock->errqueue_task.type = URING_TASK_ERRQUEUE;
1765 0 : sock->errqueue_task.msg.msg_control = sock->buf;
1766 0 : sock->errqueue_task.msg.msg_controllen = sizeof(sock->buf);
1767 :
1768 0 : sock->cancel_task.sock = sock;
1769 0 : sock->cancel_task.type = URING_TASK_CANCEL;
1770 :
1771 : /* switched from another polling group due to scheduling */
1772 0 : if (spdk_unlikely(sock->recv_pipe != NULL &&
1773 : (spdk_pipe_reader_bytes_available(sock->recv_pipe) > 0))) {
1774 0 : assert(sock->pending_recv == false);
1775 0 : sock->pending_recv = true;
1776 0 : TAILQ_INSERT_TAIL(&group->pending_recv, sock, link);
1777 : }
1778 :
1779 0 : if (sock->placement_id != -1) {
1780 0 : rc = spdk_sock_map_insert(&g_map, sock->placement_id, &group->base);
1781 0 : if (rc != 0) {
1782 0 : SPDK_ERRLOG("Failed to insert sock group into map: %d", rc);
1783 : /* Do not treat this as an error. The system will continue running. */
1784 : }
1785 : }
1786 :
1787 : /* We get an async read going immediately */
1788 0 : _sock_prep_read(&sock->base);
1789 : #ifdef SPDK_ZEROCOPY
1790 0 : if (sock->zcopy) {
1791 0 : _sock_prep_errqueue(_sock);
1792 : }
1793 : #endif
1794 :
1795 0 : return 0;
1796 : }
1797 :
1798 : static void
1799 0 : uring_sock_group_populate_buf_ring(struct spdk_uring_sock_group_impl *group)
1800 : {
1801 : struct spdk_uring_buf_tracker *tracker;
1802 : int count, mask;
1803 :
1804 0 : if (g_spdk_uring_sock_impl_opts.enable_recv_pipe) {
1805 : /* If recv_pipe is enabled, we do not post buffers. */
1806 0 : return;
1807 : }
1808 :
1809 : /* Try to re-populate the io_uring's buffer pool using user-provided buffers */
1810 0 : tracker = STAILQ_FIRST(&group->free_trackers);
1811 0 : count = 0;
1812 0 : mask = io_uring_buf_ring_mask(URING_BUF_POOL_SIZE);
1813 0 : while (tracker != NULL) {
1814 0 : tracker->buflen = spdk_sock_group_get_buf(group->base.group, &tracker->buf, &tracker->ctx);
1815 0 : if (tracker->buflen == 0) {
1816 0 : break;
1817 : }
1818 :
1819 0 : assert(tracker->buf != NULL);
1820 0 : STAILQ_REMOVE_HEAD(&group->free_trackers, link);
1821 0 : assert(STAILQ_FIRST(&group->free_trackers) != tracker);
1822 :
1823 0 : io_uring_buf_ring_add(group->buf_ring, tracker->buf, tracker->buflen, tracker->id, mask, count);
1824 0 : count++;
1825 0 : tracker = STAILQ_FIRST(&group->free_trackers);
1826 : }
1827 :
1828 0 : if (count > 0) {
1829 0 : group->buf_ring_count += count;
1830 0 : io_uring_buf_ring_advance(group->buf_ring, count);
1831 : }
1832 : }
1833 :
1834 : static int
1835 0 : uring_sock_group_impl_poll(struct spdk_sock_group_impl *_group, int max_events,
1836 : struct spdk_sock **socks)
1837 : {
1838 0 : struct spdk_uring_sock_group_impl *group = __uring_group_impl(_group);
1839 : int count, ret;
1840 : int to_complete, to_submit;
1841 : struct spdk_sock *_sock, *tmp;
1842 : struct spdk_uring_sock *sock;
1843 :
1844 0 : if (spdk_likely(socks)) {
1845 0 : TAILQ_FOREACH_SAFE(_sock, &group->base.socks, link, tmp) {
1846 0 : sock = __uring_sock(_sock);
1847 0 : if (spdk_unlikely(sock->connection_status)) {
1848 0 : continue;
1849 : }
1850 0 : _sock_flush(_sock);
1851 : }
1852 : }
1853 :
1854 : /* Try to re-populate the io_uring's buffer pool using user-provided buffers */
1855 0 : uring_sock_group_populate_buf_ring(group);
1856 :
1857 0 : to_submit = group->io_queued;
1858 :
1859 : /* For network I/O, it cannot be set with O_DIRECT, so we do not need to call spdk_io_uring_enter */
1860 0 : if (to_submit > 0) {
1861 : /* If there are I/O to submit, use io_uring_submit here.
1862 : * It will automatically call io_uring_enter appropriately. */
1863 0 : ret = io_uring_submit(&group->uring);
1864 0 : if (ret < 0) {
1865 0 : return 1;
1866 : }
1867 0 : group->io_queued = 0;
1868 0 : group->io_inflight += to_submit;
1869 0 : group->io_avail -= to_submit;
1870 : }
1871 :
1872 0 : count = 0;
1873 0 : to_complete = group->io_inflight;
1874 0 : if (to_complete > 0 || !TAILQ_EMPTY(&group->pending_recv)) {
1875 0 : count = sock_uring_group_reap(group, to_complete, max_events, socks);
1876 : }
1877 :
1878 0 : return count;
1879 : }
1880 :
1881 : static int
1882 0 : uring_sock_group_impl_remove_sock(struct spdk_sock_group_impl *_group,
1883 : struct spdk_sock *_sock)
1884 : {
1885 0 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1886 0 : struct spdk_uring_sock_group_impl *group = __uring_group_impl(_group);
1887 :
1888 0 : sock->pending_group_remove = true;
1889 :
1890 0 : if (sock->write_task.status != SPDK_URING_SOCK_TASK_NOT_IN_USE) {
1891 0 : _sock_prep_cancel_task(_sock, &sock->write_task);
1892 : /* Since spdk_sock_group_remove_sock is not asynchronous interface, so
1893 : * currently can use a while loop here. */
1894 0 : while ((sock->write_task.status != SPDK_URING_SOCK_TASK_NOT_IN_USE) ||
1895 0 : (sock->cancel_task.status != SPDK_URING_SOCK_TASK_NOT_IN_USE)) {
1896 0 : uring_sock_group_impl_poll(_group, 32, NULL);
1897 : }
1898 : }
1899 :
1900 0 : if (sock->read_task.status != SPDK_URING_SOCK_TASK_NOT_IN_USE) {
1901 0 : _sock_prep_cancel_task(_sock, &sock->read_task);
1902 : /* Since spdk_sock_group_remove_sock is not asynchronous interface, so
1903 : * currently can use a while loop here. */
1904 0 : while ((sock->read_task.status != SPDK_URING_SOCK_TASK_NOT_IN_USE) ||
1905 0 : (sock->cancel_task.status != SPDK_URING_SOCK_TASK_NOT_IN_USE)) {
1906 0 : uring_sock_group_impl_poll(_group, 32, NULL);
1907 : }
1908 : }
1909 :
1910 0 : if (sock->errqueue_task.status != SPDK_URING_SOCK_TASK_NOT_IN_USE) {
1911 0 : _sock_prep_cancel_task(_sock, &sock->errqueue_task);
1912 : /* Since spdk_sock_group_remove_sock is not asynchronous interface, so
1913 : * currently can use a while loop here. */
1914 0 : while ((sock->errqueue_task.status != SPDK_URING_SOCK_TASK_NOT_IN_USE) ||
1915 0 : (sock->cancel_task.status != SPDK_URING_SOCK_TASK_NOT_IN_USE)) {
1916 0 : uring_sock_group_impl_poll(_group, 32, NULL);
1917 : }
1918 : }
1919 :
1920 : /* Make sure the cancelling the tasks above didn't cause sending new requests */
1921 0 : assert(sock->write_task.status == SPDK_URING_SOCK_TASK_NOT_IN_USE);
1922 0 : assert(sock->read_task.status == SPDK_URING_SOCK_TASK_NOT_IN_USE);
1923 0 : assert(sock->errqueue_task.status == SPDK_URING_SOCK_TASK_NOT_IN_USE);
1924 :
1925 0 : if (sock->pending_recv) {
1926 0 : TAILQ_REMOVE(&group->pending_recv, sock, link);
1927 0 : sock->pending_recv = false;
1928 : }
1929 0 : assert(sock->pending_recv == false);
1930 :
1931 : /* We have no way to handle this case. We could let the user read this
1932 : * buffer, but the buffer came from a group and we have lost the association
1933 : * to that so we couldn't release it. */
1934 0 : assert(STAILQ_EMPTY(&sock->recv_stream));
1935 :
1936 0 : if (sock->placement_id != -1) {
1937 0 : spdk_sock_map_release(&g_map, sock->placement_id);
1938 : }
1939 :
1940 0 : sock->pending_group_remove = false;
1941 0 : sock->group = NULL;
1942 0 : return 0;
1943 : }
1944 :
1945 : static int
1946 0 : uring_sock_group_impl_close(struct spdk_sock_group_impl *_group)
1947 : {
1948 0 : struct spdk_uring_sock_group_impl *group = __uring_group_impl(_group);
1949 :
1950 : /* try to reap all the active I/O */
1951 0 : while (group->io_inflight) {
1952 0 : uring_sock_group_impl_poll(_group, 32, NULL);
1953 : }
1954 0 : assert(group->io_inflight == 0);
1955 0 : assert(group->io_avail == SPDK_SOCK_GROUP_QUEUE_DEPTH);
1956 :
1957 0 : uring_sock_group_impl_buf_pool_free(group);
1958 :
1959 0 : io_uring_queue_exit(&group->uring);
1960 :
1961 0 : if (g_spdk_uring_sock_impl_opts.enable_placement_id == PLACEMENT_CPU) {
1962 0 : spdk_sock_map_release(&g_map, spdk_env_get_current_core());
1963 : }
1964 :
1965 0 : free(group);
1966 0 : return 0;
1967 : }
1968 :
1969 : static int
1970 6 : uring_sock_flush(struct spdk_sock *_sock)
1971 : {
1972 6 : struct spdk_uring_sock *sock = __uring_sock(_sock);
1973 6 : struct msghdr msg = {};
1974 : struct iovec iovs[IOV_BATCH_SIZE];
1975 : int iovcnt;
1976 : ssize_t rc;
1977 6 : int flags = sock->zcopy_send_flags;
1978 : int retval;
1979 6 : bool is_zcopy = false;
1980 6 : struct spdk_uring_task *task = &sock->errqueue_task;
1981 :
1982 : /* Can't flush from within a callback or we end up with recursive calls */
1983 6 : if (_sock->cb_cnt > 0) {
1984 0 : errno = EAGAIN;
1985 0 : return -1;
1986 : }
1987 :
1988 : /* Can't flush while a write is already outstanding */
1989 6 : if (sock->write_task.status != SPDK_URING_SOCK_TASK_NOT_IN_USE) {
1990 0 : errno = EAGAIN;
1991 0 : return -1;
1992 : }
1993 :
1994 : /* Gather an iov */
1995 6 : iovcnt = spdk_sock_prep_reqs(_sock, iovs, 0, NULL, &flags);
1996 6 : if (iovcnt == 0) {
1997 : /* Nothing to send */
1998 0 : return 0;
1999 : }
2000 :
2001 : /* Perform the vectored write */
2002 6 : msg.msg_iov = iovs;
2003 6 : msg.msg_iovlen = iovcnt;
2004 6 : rc = sendmsg(sock->fd, &msg, flags | MSG_DONTWAIT);
2005 6 : if (rc <= 0) {
2006 0 : if (rc == 0 || errno == EAGAIN || errno == EWOULDBLOCK || (errno == ENOBUFS && sock->zcopy)) {
2007 0 : errno = EAGAIN;
2008 : }
2009 0 : return -1;
2010 : }
2011 :
2012 : #ifdef SPDK_ZEROCOPY
2013 6 : is_zcopy = flags & MSG_ZEROCOPY;
2014 : #endif
2015 6 : retval = sock_complete_write_reqs(_sock, rc, is_zcopy);
2016 6 : if (retval < 0) {
2017 : /* if the socket is closed, return to avoid heap-use-after-free error */
2018 0 : errno = ENOTCONN;
2019 0 : return -1;
2020 : }
2021 :
2022 : #ifdef SPDK_ZEROCOPY
2023 : /* At least do once to check zero copy case */
2024 6 : if (sock->zcopy && !TAILQ_EMPTY(&_sock->pending_reqs)) {
2025 0 : retval = recvmsg(sock->fd, &task->msg, MSG_ERRQUEUE);
2026 0 : if (retval < 0) {
2027 0 : if (errno == EWOULDBLOCK || errno == EAGAIN) {
2028 0 : return rc;
2029 : }
2030 : }
2031 0 : _sock_check_zcopy(_sock, retval);;
2032 : }
2033 : #endif
2034 :
2035 6 : return rc;
2036 : }
2037 :
2038 : static int
2039 0 : uring_sock_group_impl_register_interrupt(struct spdk_sock_group_impl *_group, uint32_t events,
2040 : spdk_interrupt_fn fn, void *arg, const char *name)
2041 : {
2042 0 : SPDK_ERRLOG("Interrupt mode is not supported in the uring sock implementation.");
2043 :
2044 0 : return -ENOTSUP;
2045 : }
2046 :
2047 : static void
2048 0 : uring_sock_group_impl_unregister_interrupt(struct spdk_sock_group_impl *_group)
2049 : {
2050 0 : }
2051 :
2052 : static struct spdk_net_impl g_uring_net_impl = {
2053 : .name = "uring",
2054 : .getaddr = uring_sock_getaddr,
2055 : .get_interface_name = uring_sock_get_interface_name,
2056 : .get_numa_id = uring_sock_get_numa_id,
2057 : .connect = uring_sock_connect,
2058 : .listen = uring_sock_listen,
2059 : .accept = uring_sock_accept,
2060 : .close = uring_sock_close,
2061 : .recv = uring_sock_recv,
2062 : .readv = uring_sock_readv,
2063 : .writev = uring_sock_writev,
2064 : .recv_next = uring_sock_recv_next,
2065 : .writev_async = uring_sock_writev_async,
2066 : .flush = uring_sock_flush,
2067 : .set_recvlowat = uring_sock_set_recvlowat,
2068 : .set_recvbuf = uring_sock_set_recvbuf,
2069 : .set_sendbuf = uring_sock_set_sendbuf,
2070 : .is_ipv6 = uring_sock_is_ipv6,
2071 : .is_ipv4 = uring_sock_is_ipv4,
2072 : .is_connected = uring_sock_is_connected,
2073 : .group_impl_get_optimal = uring_sock_group_impl_get_optimal,
2074 : .group_impl_create = uring_sock_group_impl_create,
2075 : .group_impl_add_sock = uring_sock_group_impl_add_sock,
2076 : .group_impl_remove_sock = uring_sock_group_impl_remove_sock,
2077 : .group_impl_poll = uring_sock_group_impl_poll,
2078 : .group_impl_register_interrupt = uring_sock_group_impl_register_interrupt,
2079 : .group_impl_unregister_interrupt = uring_sock_group_impl_unregister_interrupt,
2080 : .group_impl_close = uring_sock_group_impl_close,
2081 : .get_opts = uring_sock_impl_get_opts,
2082 : .set_opts = uring_sock_impl_set_opts,
2083 : };
2084 :
2085 : __attribute__((constructor)) static void
2086 1 : net_impl_register_uring(void)
2087 : {
2088 : struct spdk_sock_group_impl *impl;
2089 :
2090 : /* Check if we can create a uring sock group before we register
2091 : * it as a valid impl. */
2092 1 : impl = uring_sock_group_impl_create();
2093 1 : if (impl) {
2094 0 : uring_sock_group_impl_close(impl);
2095 0 : spdk_net_impl_register(&g_uring_net_impl);
2096 : }
2097 1 : }
|
