ib_recv.c 31 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093
  1. /*
  2. * Copyright (c) 2006, 2019 Oracle and/or its affiliates. All rights reserved.
  3. *
  4. * This software is available to you under a choice of one of two
  5. * licenses. You may choose to be licensed under the terms of the GNU
  6. * General Public License (GPL) Version 2, available from the file
  7. * COPYING in the main directory of this source tree, or the
  8. * OpenIB.org BSD license below:
  9. *
  10. * Redistribution and use in source and binary forms, with or
  11. * without modification, are permitted provided that the following
  12. * conditions are met:
  13. *
  14. * - Redistributions of source code must retain the above
  15. * copyright notice, this list of conditions and the following
  16. * disclaimer.
  17. *
  18. * - Redistributions in binary form must reproduce the above
  19. * copyright notice, this list of conditions and the following
  20. * disclaimer in the documentation and/or other materials
  21. * provided with the distribution.
  22. *
  23. * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24. * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25. * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26. * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27. * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28. * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29. * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30. * SOFTWARE.
  31. *
  32. */
  33. #include <linux/kernel.h>
  34. #include <linux/slab.h>
  35. #include <linux/pci.h>
  36. #include <linux/dma-mapping.h>
  37. #include <rdma/rdma_cm.h>
  38. #include "rds_single_path.h"
  39. #include "rds.h"
  40. #include "ib.h"
  41. static struct kmem_cache *rds_ib_incoming_slab;
  42. static struct kmem_cache *rds_ib_frag_slab;
  43. static atomic_t rds_ib_allocation = ATOMIC_INIT(0);
  44. void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
  45. {
  46. struct rds_ib_recv_work *recv;
  47. u32 i;
  48. for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
  49. struct ib_sge *sge;
  50. recv->r_ibinc = NULL;
  51. recv->r_frag = NULL;
  52. recv->r_wr.next = NULL;
  53. recv->r_wr.wr_id = i;
  54. recv->r_wr.sg_list = recv->r_sge;
  55. recv->r_wr.num_sge = RDS_IB_RECV_SGE;
  56. sge = &recv->r_sge[0];
  57. sge->addr = ic->i_recv_hdrs_dma[i];
  58. sge->length = sizeof(struct rds_header);
  59. sge->lkey = ic->i_pd->local_dma_lkey;
  60. sge = &recv->r_sge[1];
  61. sge->addr = 0;
  62. sge->length = RDS_FRAG_SIZE;
  63. sge->lkey = ic->i_pd->local_dma_lkey;
  64. }
  65. }
  66. /*
  67. * The entire 'from' list, including the from element itself, is put on
  68. * to the tail of the 'to' list.
  69. */
  70. static void list_splice_entire_tail(struct list_head *from,
  71. struct list_head *to)
  72. {
  73. struct list_head *from_last = from->prev;
  74. list_splice_tail(from_last, to);
  75. list_add_tail(from_last, to);
  76. }
  77. static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
  78. {
  79. struct list_head *tmp;
  80. tmp = xchg(&cache->xfer, NULL);
  81. if (tmp) {
  82. if (cache->ready)
  83. list_splice_entire_tail(tmp, cache->ready);
  84. else
  85. cache->ready = tmp;
  86. }
  87. }
  88. static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache, gfp_t gfp)
  89. {
  90. struct rds_ib_cache_head *head;
  91. int cpu;
  92. cache->percpu = alloc_percpu_gfp(struct rds_ib_cache_head, gfp);
  93. if (!cache->percpu)
  94. return -ENOMEM;
  95. for_each_possible_cpu(cpu) {
  96. head = per_cpu_ptr(cache->percpu, cpu);
  97. head->first = NULL;
  98. head->count = 0;
  99. }
  100. cache->xfer = NULL;
  101. cache->ready = NULL;
  102. return 0;
  103. }
  104. int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp)
  105. {
  106. int ret;
  107. ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs, gfp);
  108. if (!ret) {
  109. ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags, gfp);
  110. if (ret)
  111. free_percpu(ic->i_cache_incs.percpu);
  112. }
  113. return ret;
  114. }
  115. static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
  116. struct list_head *caller_list)
  117. {
  118. struct rds_ib_cache_head *head;
  119. int cpu;
  120. for_each_possible_cpu(cpu) {
  121. head = per_cpu_ptr(cache->percpu, cpu);
  122. if (head->first) {
  123. list_splice_entire_tail(head->first, caller_list);
  124. head->first = NULL;
  125. }
  126. }
  127. if (cache->ready) {
  128. list_splice_entire_tail(cache->ready, caller_list);
  129. cache->ready = NULL;
  130. }
  131. }
  132. void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
  133. {
  134. struct rds_ib_incoming *inc;
  135. struct rds_ib_incoming *inc_tmp;
  136. struct rds_page_frag *frag;
  137. struct rds_page_frag *frag_tmp;
  138. LIST_HEAD(list);
  139. rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
  140. rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list);
  141. free_percpu(ic->i_cache_incs.percpu);
  142. list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) {
  143. list_del(&inc->ii_cache_entry);
  144. WARN_ON(!list_empty(&inc->ii_frags));
  145. kmem_cache_free(rds_ib_incoming_slab, inc);
  146. atomic_dec(&rds_ib_allocation);
  147. }
  148. rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
  149. rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
  150. free_percpu(ic->i_cache_frags.percpu);
  151. list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
  152. list_del(&frag->f_cache_entry);
  153. WARN_ON(!list_empty(&frag->f_item));
  154. kmem_cache_free(rds_ib_frag_slab, frag);
  155. }
  156. }
  157. /* fwd decl */
  158. static void rds_ib_recv_cache_put(struct list_head *new_item,
  159. struct rds_ib_refill_cache *cache);
  160. static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
  161. /* Recycle frag and attached recv buffer f_sg */
  162. static void rds_ib_frag_free(struct rds_ib_connection *ic,
  163. struct rds_page_frag *frag)
  164. {
  165. rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
  166. rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
  167. atomic_add(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
  168. rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
  169. }
  170. /* Recycle inc after freeing attached frags */
  171. void rds_ib_inc_free(struct rds_incoming *inc)
  172. {
  173. struct rds_ib_incoming *ibinc;
  174. struct rds_page_frag *frag;
  175. struct rds_page_frag *pos;
  176. struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
  177. ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
  178. /* Free attached frags */
  179. list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
  180. list_del_init(&frag->f_item);
  181. rds_ib_frag_free(ic, frag);
  182. }
  183. BUG_ON(!list_empty(&ibinc->ii_frags));
  184. rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
  185. rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
  186. }
  187. static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
  188. struct rds_ib_recv_work *recv)
  189. {
  190. if (recv->r_ibinc) {
  191. rds_inc_put(&recv->r_ibinc->ii_inc);
  192. recv->r_ibinc = NULL;
  193. }
  194. if (recv->r_frag) {
  195. ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
  196. rds_ib_frag_free(ic, recv->r_frag);
  197. recv->r_frag = NULL;
  198. }
  199. }
  200. void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
  201. {
  202. u32 i;
  203. for (i = 0; i < ic->i_recv_ring.w_nr; i++)
  204. rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
  205. }
  206. static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
  207. gfp_t slab_mask)
  208. {
  209. struct rds_ib_incoming *ibinc;
  210. struct list_head *cache_item;
  211. int avail_allocs;
  212. cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
  213. if (cache_item) {
  214. ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
  215. } else {
  216. avail_allocs = atomic_add_unless(&rds_ib_allocation,
  217. 1, rds_ib_sysctl_max_recv_allocation);
  218. if (!avail_allocs) {
  219. rds_ib_stats_inc(s_ib_rx_alloc_limit);
  220. return NULL;
  221. }
  222. ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
  223. if (!ibinc) {
  224. atomic_dec(&rds_ib_allocation);
  225. return NULL;
  226. }
  227. rds_ib_stats_inc(s_ib_rx_total_incs);
  228. }
  229. INIT_LIST_HEAD(&ibinc->ii_frags);
  230. rds_inc_init(&ibinc->ii_inc, ic->conn, &ic->conn->c_faddr);
  231. return ibinc;
  232. }
  233. static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
  234. gfp_t slab_mask, gfp_t page_mask)
  235. {
  236. struct rds_page_frag *frag;
  237. struct list_head *cache_item;
  238. int ret;
  239. cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
  240. if (cache_item) {
  241. frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
  242. atomic_sub(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
  243. rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
  244. } else {
  245. frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
  246. if (!frag)
  247. return NULL;
  248. sg_init_table(&frag->f_sg, 1);
  249. ret = rds_page_remainder_alloc(&frag->f_sg,
  250. RDS_FRAG_SIZE, page_mask);
  251. if (ret) {
  252. kmem_cache_free(rds_ib_frag_slab, frag);
  253. return NULL;
  254. }
  255. rds_ib_stats_inc(s_ib_rx_total_frags);
  256. }
  257. INIT_LIST_HEAD(&frag->f_item);
  258. return frag;
  259. }
  260. static int rds_ib_recv_refill_one(struct rds_connection *conn,
  261. struct rds_ib_recv_work *recv, gfp_t gfp)
  262. {
  263. struct rds_ib_connection *ic = conn->c_transport_data;
  264. struct ib_sge *sge;
  265. int ret = -ENOMEM;
  266. gfp_t slab_mask = gfp;
  267. gfp_t page_mask = gfp;
  268. if (gfp & __GFP_DIRECT_RECLAIM) {
  269. slab_mask = GFP_KERNEL;
  270. page_mask = GFP_HIGHUSER;
  271. }
  272. if (!ic->i_cache_incs.ready)
  273. rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
  274. if (!ic->i_cache_frags.ready)
  275. rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
  276. /*
  277. * ibinc was taken from recv if recv contained the start of a message.
  278. * recvs that were continuations will still have this allocated.
  279. */
  280. if (!recv->r_ibinc) {
  281. recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
  282. if (!recv->r_ibinc)
  283. goto out;
  284. }
  285. WARN_ON(recv->r_frag); /* leak! */
  286. recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
  287. if (!recv->r_frag)
  288. goto out;
  289. ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
  290. 1, DMA_FROM_DEVICE);
  291. WARN_ON(ret != 1);
  292. sge = &recv->r_sge[0];
  293. sge->addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
  294. sge->length = sizeof(struct rds_header);
  295. sge = &recv->r_sge[1];
  296. sge->addr = sg_dma_address(&recv->r_frag->f_sg);
  297. sge->length = sg_dma_len(&recv->r_frag->f_sg);
  298. ret = 0;
  299. out:
  300. return ret;
  301. }
  302. static int acquire_refill(struct rds_connection *conn)
  303. {
  304. return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0;
  305. }
  306. static void release_refill(struct rds_connection *conn)
  307. {
  308. clear_bit(RDS_RECV_REFILL, &conn->c_flags);
  309. /* We don't use wait_on_bit()/wake_up_bit() because our waking is in a
  310. * hot path and finding waiters is very rare. We don't want to walk
  311. * the system-wide hashed waitqueue buckets in the fast path only to
  312. * almost never find waiters.
  313. */
  314. if (waitqueue_active(&conn->c_waitq))
  315. wake_up_all(&conn->c_waitq);
  316. }
  317. /*
  318. * This tries to allocate and post unused work requests after making sure that
  319. * they have all the allocations they need to queue received fragments into
  320. * sockets.
  321. */
  322. void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
  323. {
  324. struct rds_ib_connection *ic = conn->c_transport_data;
  325. struct rds_ib_recv_work *recv;
  326. unsigned int posted = 0;
  327. int ret = 0;
  328. bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM);
  329. bool must_wake = false;
  330. u32 pos;
  331. /* the goal here is to just make sure that someone, somewhere
  332. * is posting buffers. If we can't get the refill lock,
  333. * let them do their thing
  334. */
  335. if (!acquire_refill(conn))
  336. return;
  337. while ((prefill || rds_conn_up(conn)) &&
  338. rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
  339. if (pos >= ic->i_recv_ring.w_nr) {
  340. printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
  341. pos);
  342. break;
  343. }
  344. recv = &ic->i_recvs[pos];
  345. ret = rds_ib_recv_refill_one(conn, recv, gfp);
  346. if (ret) {
  347. must_wake = true;
  348. break;
  349. }
  350. rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv,
  351. recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
  352. (long)sg_dma_address(&recv->r_frag->f_sg));
  353. /* XXX when can this fail? */
  354. ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, NULL);
  355. if (ret) {
  356. rds_ib_conn_error(conn, "recv post on "
  357. "%pI6c returned %d, disconnecting and "
  358. "reconnecting\n", &conn->c_faddr,
  359. ret);
  360. break;
  361. }
  362. posted++;
  363. if ((posted > 128 && need_resched()) || posted > 8192) {
  364. must_wake = true;
  365. break;
  366. }
  367. }
  368. /* We're doing flow control - update the window. */
  369. if (ic->i_flowctl && posted)
  370. rds_ib_advertise_credits(conn, posted);
  371. if (ret)
  372. rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
  373. release_refill(conn);
  374. /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
  375. * in this case the ring being low is going to lead to more interrupts
  376. * and we can safely let the softirq code take care of it unless the
  377. * ring is completely empty.
  378. *
  379. * if we're called from krdsd, we'll be GFP_KERNEL. In this case
  380. * we might have raced with the softirq code while we had the refill
  381. * lock held. Use rds_ib_ring_low() instead of ring_empty to decide
  382. * if we should requeue.
  383. */
  384. if (rds_conn_up(conn) &&
  385. (must_wake ||
  386. (can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
  387. rds_ib_ring_empty(&ic->i_recv_ring))) {
  388. queue_delayed_work(rds_wq, &conn->c_recv_w, 1);
  389. }
  390. if (can_wait)
  391. cond_resched();
  392. }
  393. /*
  394. * We want to recycle several types of recv allocations, like incs and frags.
  395. * To use this, the *_free() function passes in the ptr to a list_head within
  396. * the recyclee, as well as the cache to put it on.
  397. *
  398. * First, we put the memory on a percpu list. When this reaches a certain size,
  399. * We move it to an intermediate non-percpu list in a lockless manner, with some
  400. * xchg/compxchg wizardry.
  401. *
  402. * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
  403. * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
  404. * list_empty() will return true with one element is actually present.
  405. */
  406. static void rds_ib_recv_cache_put(struct list_head *new_item,
  407. struct rds_ib_refill_cache *cache)
  408. {
  409. unsigned long flags;
  410. struct list_head *old, *chpfirst;
  411. local_irq_save(flags);
  412. chpfirst = __this_cpu_read(cache->percpu->first);
  413. if (!chpfirst)
  414. INIT_LIST_HEAD(new_item);
  415. else /* put on front */
  416. list_add_tail(new_item, chpfirst);
  417. __this_cpu_write(cache->percpu->first, new_item);
  418. __this_cpu_inc(cache->percpu->count);
  419. if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
  420. goto end;
  421. /*
  422. * Return our per-cpu first list to the cache's xfer by atomically
  423. * grabbing the current xfer list, appending it to our per-cpu list,
  424. * and then atomically returning that entire list back to the
  425. * cache's xfer list as long as it's still empty.
  426. */
  427. do {
  428. old = xchg(&cache->xfer, NULL);
  429. if (old)
  430. list_splice_entire_tail(old, chpfirst);
  431. old = cmpxchg(&cache->xfer, NULL, chpfirst);
  432. } while (old);
  433. __this_cpu_write(cache->percpu->first, NULL);
  434. __this_cpu_write(cache->percpu->count, 0);
  435. end:
  436. local_irq_restore(flags);
  437. }
  438. static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
  439. {
  440. struct list_head *head = cache->ready;
  441. if (head) {
  442. if (!list_empty(head)) {
  443. cache->ready = head->next;
  444. list_del_init(head);
  445. } else
  446. cache->ready = NULL;
  447. }
  448. return head;
  449. }
  450. int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
  451. {
  452. struct rds_ib_incoming *ibinc;
  453. struct rds_page_frag *frag;
  454. unsigned long to_copy;
  455. unsigned long frag_off = 0;
  456. int copied = 0;
  457. int ret;
  458. u32 len;
  459. ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
  460. frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
  461. len = be32_to_cpu(inc->i_hdr.h_len);
  462. while (iov_iter_count(to) && copied < len) {
  463. if (frag_off == RDS_FRAG_SIZE) {
  464. frag = list_entry(frag->f_item.next,
  465. struct rds_page_frag, f_item);
  466. frag_off = 0;
  467. }
  468. to_copy = min_t(unsigned long, iov_iter_count(to),
  469. RDS_FRAG_SIZE - frag_off);
  470. to_copy = min_t(unsigned long, to_copy, len - copied);
  471. /* XXX needs + offset for multiple recvs per page */
  472. rds_stats_add(s_copy_to_user, to_copy);
  473. ret = copy_page_to_iter(sg_page(&frag->f_sg),
  474. frag->f_sg.offset + frag_off,
  475. to_copy,
  476. to);
  477. if (ret != to_copy)
  478. return -EFAULT;
  479. frag_off += to_copy;
  480. copied += to_copy;
  481. }
  482. return copied;
  483. }
  484. /* ic starts out kzalloc()ed */
  485. void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
  486. {
  487. struct ib_send_wr *wr = &ic->i_ack_wr;
  488. struct ib_sge *sge = &ic->i_ack_sge;
  489. sge->addr = ic->i_ack_dma;
  490. sge->length = sizeof(struct rds_header);
  491. sge->lkey = ic->i_pd->local_dma_lkey;
  492. wr->sg_list = sge;
  493. wr->num_sge = 1;
  494. wr->opcode = IB_WR_SEND;
  495. wr->wr_id = RDS_IB_ACK_WR_ID;
  496. wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
  497. }
  498. /*
  499. * You'd think that with reliable IB connections you wouldn't need to ack
  500. * messages that have been received. The problem is that IB hardware generates
  501. * an ack message before it has DMAed the message into memory. This creates a
  502. * potential message loss if the HCA is disabled for any reason between when it
  503. * sends the ack and before the message is DMAed and processed. This is only a
  504. * potential issue if another HCA is available for fail-over.
  505. *
  506. * When the remote host receives our ack they'll free the sent message from
  507. * their send queue. To decrease the latency of this we always send an ack
  508. * immediately after we've received messages.
  509. *
  510. * For simplicity, we only have one ack in flight at a time. This puts
  511. * pressure on senders to have deep enough send queues to absorb the latency of
  512. * a single ack frame being in flight. This might not be good enough.
  513. *
  514. * This is implemented by have a long-lived send_wr and sge which point to a
  515. * statically allocated ack frame. This ack wr does not fall under the ring
  516. * accounting that the tx and rx wrs do. The QP attribute specifically makes
  517. * room for it beyond the ring size. Send completion notices its special
  518. * wr_id and avoids working with the ring in that case.
  519. */
  520. #ifndef KERNEL_HAS_ATOMIC64
  521. void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
  522. {
  523. unsigned long flags;
  524. spin_lock_irqsave(&ic->i_ack_lock, flags);
  525. ic->i_ack_next = seq;
  526. if (ack_required)
  527. set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
  528. spin_unlock_irqrestore(&ic->i_ack_lock, flags);
  529. }
  530. static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
  531. {
  532. unsigned long flags;
  533. u64 seq;
  534. clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
  535. spin_lock_irqsave(&ic->i_ack_lock, flags);
  536. seq = ic->i_ack_next;
  537. spin_unlock_irqrestore(&ic->i_ack_lock, flags);
  538. return seq;
  539. }
  540. #else
  541. void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
  542. {
  543. atomic64_set(&ic->i_ack_next, seq);
  544. if (ack_required) {
  545. smp_mb__before_atomic();
  546. set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
  547. }
  548. }
  549. static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
  550. {
  551. clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
  552. smp_mb__after_atomic();
  553. return atomic64_read(&ic->i_ack_next);
  554. }
  555. #endif
  556. static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
  557. {
  558. struct rds_header *hdr = ic->i_ack;
  559. u64 seq;
  560. int ret;
  561. seq = rds_ib_get_ack(ic);
  562. rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
  563. ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, ic->i_ack_dma,
  564. sizeof(*hdr), DMA_TO_DEVICE);
  565. rds_message_populate_header(hdr, 0, 0, 0);
  566. hdr->h_ack = cpu_to_be64(seq);
  567. hdr->h_credit = adv_credits;
  568. rds_message_make_checksum(hdr);
  569. ib_dma_sync_single_for_device(ic->rds_ibdev->dev, ic->i_ack_dma,
  570. sizeof(*hdr), DMA_TO_DEVICE);
  571. ic->i_ack_queued = jiffies;
  572. ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, NULL);
  573. if (unlikely(ret)) {
  574. /* Failed to send. Release the WR, and
  575. * force another ACK.
  576. */
  577. clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
  578. set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
  579. rds_ib_stats_inc(s_ib_ack_send_failure);
  580. rds_ib_conn_error(ic->conn, "sending ack failed\n");
  581. } else
  582. rds_ib_stats_inc(s_ib_ack_sent);
  583. }
  584. /*
  585. * There are 3 ways of getting acknowledgements to the peer:
  586. * 1. We call rds_ib_attempt_ack from the recv completion handler
  587. * to send an ACK-only frame.
  588. * However, there can be only one such frame in the send queue
  589. * at any time, so we may have to postpone it.
  590. * 2. When another (data) packet is transmitted while there's
  591. * an ACK in the queue, we piggyback the ACK sequence number
  592. * on the data packet.
  593. * 3. If the ACK WR is done sending, we get called from the
  594. * send queue completion handler, and check whether there's
  595. * another ACK pending (postponed because the WR was on the
  596. * queue). If so, we transmit it.
  597. *
  598. * We maintain 2 variables:
  599. * - i_ack_flags, which keeps track of whether the ACK WR
  600. * is currently in the send queue or not (IB_ACK_IN_FLIGHT)
  601. * - i_ack_next, which is the last sequence number we received
  602. *
  603. * Potentially, send queue and receive queue handlers can run concurrently.
  604. * It would be nice to not have to use a spinlock to synchronize things,
  605. * but the one problem that rules this out is that 64bit updates are
  606. * not atomic on all platforms. Things would be a lot simpler if
  607. * we had atomic64 or maybe cmpxchg64 everywhere.
  608. *
  609. * Reconnecting complicates this picture just slightly. When we
  610. * reconnect, we may be seeing duplicate packets. The peer
  611. * is retransmitting them, because it hasn't seen an ACK for
  612. * them. It is important that we ACK these.
  613. *
  614. * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
  615. * this flag set *MUST* be acknowledged immediately.
  616. */
  617. /*
  618. * When we get here, we're called from the recv queue handler.
  619. * Check whether we ought to transmit an ACK.
  620. */
  621. void rds_ib_attempt_ack(struct rds_ib_connection *ic)
  622. {
  623. unsigned int adv_credits;
  624. if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
  625. return;
  626. if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
  627. rds_ib_stats_inc(s_ib_ack_send_delayed);
  628. return;
  629. }
  630. /* Can we get a send credit? */
  631. if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
  632. rds_ib_stats_inc(s_ib_tx_throttle);
  633. clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
  634. return;
  635. }
  636. clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
  637. rds_ib_send_ack(ic, adv_credits);
  638. }
  639. /*
  640. * We get here from the send completion handler, when the
  641. * adapter tells us the ACK frame was sent.
  642. */
  643. void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
  644. {
  645. clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
  646. rds_ib_attempt_ack(ic);
  647. }
  648. /*
  649. * This is called by the regular xmit code when it wants to piggyback
  650. * an ACK on an outgoing frame.
  651. */
  652. u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
  653. {
  654. if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
  655. rds_ib_stats_inc(s_ib_ack_send_piggybacked);
  656. return rds_ib_get_ack(ic);
  657. }
  658. /*
  659. * It's kind of lame that we're copying from the posted receive pages into
  660. * long-lived bitmaps. We could have posted the bitmaps and rdma written into
  661. * them. But receiving new congestion bitmaps should be a *rare* event, so
  662. * hopefully we won't need to invest that complexity in making it more
  663. * efficient. By copying we can share a simpler core with TCP which has to
  664. * copy.
  665. */
  666. static void rds_ib_cong_recv(struct rds_connection *conn,
  667. struct rds_ib_incoming *ibinc)
  668. {
  669. struct rds_cong_map *map;
  670. unsigned int map_off;
  671. unsigned int map_page;
  672. struct rds_page_frag *frag;
  673. unsigned long frag_off;
  674. unsigned long to_copy;
  675. unsigned long copied;
  676. __le64 uncongested = 0;
  677. void *addr;
  678. /* catch completely corrupt packets */
  679. if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
  680. return;
  681. map = conn->c_fcong;
  682. map_page = 0;
  683. map_off = 0;
  684. frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
  685. frag_off = 0;
  686. copied = 0;
  687. while (copied < RDS_CONG_MAP_BYTES) {
  688. __le64 *src, *dst;
  689. unsigned int k;
  690. to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
  691. BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
  692. addr = kmap_atomic(sg_page(&frag->f_sg));
  693. src = addr + frag->f_sg.offset + frag_off;
  694. dst = (void *)map->m_page_addrs[map_page] + map_off;
  695. for (k = 0; k < to_copy; k += 8) {
  696. /* Record ports that became uncongested, ie
  697. * bits that changed from 0 to 1. */
  698. uncongested |= ~(*src) & *dst;
  699. *dst++ = *src++;
  700. }
  701. kunmap_atomic(addr);
  702. copied += to_copy;
  703. map_off += to_copy;
  704. if (map_off == PAGE_SIZE) {
  705. map_off = 0;
  706. map_page++;
  707. }
  708. frag_off += to_copy;
  709. if (frag_off == RDS_FRAG_SIZE) {
  710. frag = list_entry(frag->f_item.next,
  711. struct rds_page_frag, f_item);
  712. frag_off = 0;
  713. }
  714. }
  715. /* the congestion map is in little endian order */
  716. rds_cong_map_updated(map, le64_to_cpu(uncongested));
  717. }
  718. static void rds_ib_process_recv(struct rds_connection *conn,
  719. struct rds_ib_recv_work *recv, u32 data_len,
  720. struct rds_ib_ack_state *state)
  721. {
  722. struct rds_ib_connection *ic = conn->c_transport_data;
  723. struct rds_ib_incoming *ibinc = ic->i_ibinc;
  724. struct rds_header *ihdr, *hdr;
  725. dma_addr_t dma_addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
  726. /* XXX shut down the connection if port 0,0 are seen? */
  727. rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
  728. data_len);
  729. if (data_len < sizeof(struct rds_header)) {
  730. rds_ib_conn_error(conn, "incoming message "
  731. "from %pI6c didn't include a "
  732. "header, disconnecting and "
  733. "reconnecting\n",
  734. &conn->c_faddr);
  735. return;
  736. }
  737. data_len -= sizeof(struct rds_header);
  738. ihdr = ic->i_recv_hdrs[recv - ic->i_recvs];
  739. ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, dma_addr,
  740. sizeof(*ihdr), DMA_FROM_DEVICE);
  741. /* Validate the checksum. */
  742. if (!rds_message_verify_checksum(ihdr)) {
  743. rds_ib_conn_error(conn, "incoming message "
  744. "from %pI6c has corrupted header - "
  745. "forcing a reconnect\n",
  746. &conn->c_faddr);
  747. rds_stats_inc(s_recv_drop_bad_checksum);
  748. goto done;
  749. }
  750. /* Process the ACK sequence which comes with every packet */
  751. state->ack_recv = be64_to_cpu(ihdr->h_ack);
  752. state->ack_recv_valid = 1;
  753. /* Process the credits update if there was one */
  754. if (ihdr->h_credit)
  755. rds_ib_send_add_credits(conn, ihdr->h_credit);
  756. if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
  757. /* This is an ACK-only packet. The fact that it gets
  758. * special treatment here is that historically, ACKs
  759. * were rather special beasts.
  760. */
  761. rds_ib_stats_inc(s_ib_ack_received);
  762. /*
  763. * Usually the frags make their way on to incs and are then freed as
  764. * the inc is freed. We don't go that route, so we have to drop the
  765. * page ref ourselves. We can't just leave the page on the recv
  766. * because that confuses the dma mapping of pages and each recv's use
  767. * of a partial page.
  768. *
  769. * FIXME: Fold this into the code path below.
  770. */
  771. rds_ib_frag_free(ic, recv->r_frag);
  772. recv->r_frag = NULL;
  773. goto done;
  774. }
  775. /*
  776. * If we don't already have an inc on the connection then this
  777. * fragment has a header and starts a message.. copy its header
  778. * into the inc and save the inc so we can hang upcoming fragments
  779. * off its list.
  780. */
  781. if (!ibinc) {
  782. ibinc = recv->r_ibinc;
  783. recv->r_ibinc = NULL;
  784. ic->i_ibinc = ibinc;
  785. hdr = &ibinc->ii_inc.i_hdr;
  786. ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] =
  787. local_clock();
  788. memcpy(hdr, ihdr, sizeof(*hdr));
  789. ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
  790. ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] =
  791. local_clock();
  792. rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
  793. ic->i_recv_data_rem, hdr->h_flags);
  794. } else {
  795. hdr = &ibinc->ii_inc.i_hdr;
  796. /* We can't just use memcmp here; fragments of a
  797. * single message may carry different ACKs */
  798. if (hdr->h_sequence != ihdr->h_sequence ||
  799. hdr->h_len != ihdr->h_len ||
  800. hdr->h_sport != ihdr->h_sport ||
  801. hdr->h_dport != ihdr->h_dport) {
  802. rds_ib_conn_error(conn,
  803. "fragment header mismatch; forcing reconnect\n");
  804. goto done;
  805. }
  806. }
  807. list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
  808. recv->r_frag = NULL;
  809. if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
  810. ic->i_recv_data_rem -= RDS_FRAG_SIZE;
  811. else {
  812. ic->i_recv_data_rem = 0;
  813. ic->i_ibinc = NULL;
  814. if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) {
  815. rds_ib_cong_recv(conn, ibinc);
  816. } else {
  817. rds_recv_incoming(conn, &conn->c_faddr, &conn->c_laddr,
  818. &ibinc->ii_inc, GFP_ATOMIC);
  819. state->ack_next = be64_to_cpu(hdr->h_sequence);
  820. state->ack_next_valid = 1;
  821. }
  822. /* Evaluate the ACK_REQUIRED flag *after* we received
  823. * the complete frame, and after bumping the next_rx
  824. * sequence. */
  825. if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
  826. rds_stats_inc(s_recv_ack_required);
  827. state->ack_required = 1;
  828. }
  829. rds_inc_put(&ibinc->ii_inc);
  830. }
  831. done:
  832. ib_dma_sync_single_for_device(ic->rds_ibdev->dev, dma_addr,
  833. sizeof(*ihdr), DMA_FROM_DEVICE);
  834. }
  835. void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
  836. struct ib_wc *wc,
  837. struct rds_ib_ack_state *state)
  838. {
  839. struct rds_connection *conn = ic->conn;
  840. struct rds_ib_recv_work *recv;
  841. rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
  842. (unsigned long long)wc->wr_id, wc->status,
  843. ib_wc_status_msg(wc->status), wc->byte_len,
  844. be32_to_cpu(wc->ex.imm_data));
  845. rds_ib_stats_inc(s_ib_rx_cq_event);
  846. recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
  847. ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1,
  848. DMA_FROM_DEVICE);
  849. /* Also process recvs in connecting state because it is possible
  850. * to get a recv completion _before_ the rdmacm ESTABLISHED
  851. * event is processed.
  852. */
  853. if (wc->status == IB_WC_SUCCESS) {
  854. rds_ib_process_recv(conn, recv, wc->byte_len, state);
  855. } else {
  856. /* We expect errors as the qp is drained during shutdown */
  857. if (rds_conn_up(conn) || rds_conn_connecting(conn))
  858. rds_ib_conn_error(conn, "recv completion on <%pI6c,%pI6c, %d> had status %u (%s), vendor err 0x%x, disconnecting and reconnecting\n",
  859. &conn->c_laddr, &conn->c_faddr,
  860. conn->c_tos, wc->status,
  861. ib_wc_status_msg(wc->status),
  862. wc->vendor_err);
  863. }
  864. /* rds_ib_process_recv() doesn't always consume the frag, and
  865. * we might not have called it at all if the wc didn't indicate
  866. * success. We already unmapped the frag's pages, though, and
  867. * the following rds_ib_ring_free() call tells the refill path
  868. * that it will not find an allocated frag here. Make sure we
  869. * keep that promise by freeing a frag that's still on the ring.
  870. */
  871. if (recv->r_frag) {
  872. rds_ib_frag_free(ic, recv->r_frag);
  873. recv->r_frag = NULL;
  874. }
  875. rds_ib_ring_free(&ic->i_recv_ring, 1);
  876. /* If we ever end up with a really empty receive ring, we're
  877. * in deep trouble, as the sender will definitely see RNR
  878. * timeouts. */
  879. if (rds_ib_ring_empty(&ic->i_recv_ring))
  880. rds_ib_stats_inc(s_ib_rx_ring_empty);
  881. if (rds_ib_ring_low(&ic->i_recv_ring)) {
  882. rds_ib_recv_refill(conn, 0, GFP_NOWAIT | __GFP_NOWARN);
  883. rds_ib_stats_inc(s_ib_rx_refill_from_cq);
  884. }
  885. }
  886. int rds_ib_recv_path(struct rds_conn_path *cp)
  887. {
  888. struct rds_connection *conn = cp->cp_conn;
  889. struct rds_ib_connection *ic = conn->c_transport_data;
  890. rdsdebug("conn %p\n", conn);
  891. if (rds_conn_up(conn)) {
  892. rds_ib_attempt_ack(ic);
  893. rds_ib_recv_refill(conn, 0, GFP_KERNEL);
  894. rds_ib_stats_inc(s_ib_rx_refill_from_thread);
  895. }
  896. return 0;
  897. }
  898. int rds_ib_recv_init(void)
  899. {
  900. struct sysinfo si;
  901. int ret = -ENOMEM;
  902. /* Default to 30% of all available RAM for recv memory */
  903. si_meminfo(&si);
  904. rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
  905. rds_ib_incoming_slab =
  906. kmem_cache_create_usercopy("rds_ib_incoming",
  907. sizeof(struct rds_ib_incoming),
  908. 0, SLAB_HWCACHE_ALIGN,
  909. offsetof(struct rds_ib_incoming,
  910. ii_inc.i_usercopy),
  911. sizeof(struct rds_inc_usercopy),
  912. NULL);
  913. if (!rds_ib_incoming_slab)
  914. goto out;
  915. rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
  916. sizeof(struct rds_page_frag),
  917. 0, SLAB_HWCACHE_ALIGN, NULL);
  918. if (!rds_ib_frag_slab) {
  919. kmem_cache_destroy(rds_ib_incoming_slab);
  920. rds_ib_incoming_slab = NULL;
  921. } else
  922. ret = 0;
  923. out:
  924. return ret;
  925. }
  926. void rds_ib_recv_exit(void)
  927. {
  928. WARN_ON(atomic_read(&rds_ib_allocation));
  929. kmem_cache_destroy(rds_ib_incoming_slab);
  930. kmem_cache_destroy(rds_ib_frag_slab);
  931. }