Commit | Line | Data |
---|---|---|
1e23b3ee AG |
1 | /* |
2 | * Copyright (c) 2006 Oracle. 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> | |
5a0e3ad6 | 34 | #include <linux/slab.h> |
1e23b3ee AG |
35 | #include <linux/pci.h> |
36 | #include <linux/dma-mapping.h> | |
37 | #include <rdma/rdma_cm.h> | |
38 | ||
39 | #include "rds.h" | |
40 | #include "ib.h" | |
41 | ||
42 | static struct kmem_cache *rds_ib_incoming_slab; | |
43 | static struct kmem_cache *rds_ib_frag_slab; | |
44 | static atomic_t rds_ib_allocation = ATOMIC_INIT(0); | |
45 | ||
1e23b3ee AG |
46 | void rds_ib_recv_init_ring(struct rds_ib_connection *ic) |
47 | { | |
48 | struct rds_ib_recv_work *recv; | |
49 | u32 i; | |
50 | ||
51 | for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) { | |
52 | struct ib_sge *sge; | |
53 | ||
54 | recv->r_ibinc = NULL; | |
55 | recv->r_frag = NULL; | |
56 | ||
57 | recv->r_wr.next = NULL; | |
58 | recv->r_wr.wr_id = i; | |
59 | recv->r_wr.sg_list = recv->r_sge; | |
60 | recv->r_wr.num_sge = RDS_IB_RECV_SGE; | |
61 | ||
919ced4c | 62 | sge = &recv->r_sge[0]; |
1e23b3ee AG |
63 | sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header)); |
64 | sge->length = sizeof(struct rds_header); | |
65 | sge->lkey = ic->i_mr->lkey; | |
919ced4c AG |
66 | |
67 | sge = &recv->r_sge[1]; | |
68 | sge->addr = 0; | |
69 | sge->length = RDS_FRAG_SIZE; | |
70 | sge->lkey = ic->i_mr->lkey; | |
1e23b3ee AG |
71 | } |
72 | } | |
73 | ||
33244125 CM |
74 | /* |
75 | * The entire 'from' list, including the from element itself, is put on | |
76 | * to the tail of the 'to' list. | |
77 | */ | |
78 | static void list_splice_entire_tail(struct list_head *from, | |
79 | struct list_head *to) | |
80 | { | |
81 | struct list_head *from_last = from->prev; | |
82 | ||
83 | list_splice_tail(from_last, to); | |
84 | list_add_tail(from_last, to); | |
85 | } | |
86 | ||
87 | static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache) | |
88 | { | |
89 | struct list_head *tmp; | |
90 | ||
91 | tmp = xchg(&cache->xfer, NULL); | |
92 | if (tmp) { | |
93 | if (cache->ready) | |
94 | list_splice_entire_tail(tmp, cache->ready); | |
95 | else | |
96 | cache->ready = tmp; | |
97 | } | |
98 | } | |
99 | ||
100 | static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache) | |
101 | { | |
102 | struct rds_ib_cache_head *head; | |
103 | int cpu; | |
104 | ||
105 | cache->percpu = alloc_percpu(struct rds_ib_cache_head); | |
106 | if (!cache->percpu) | |
107 | return -ENOMEM; | |
108 | ||
109 | for_each_possible_cpu(cpu) { | |
110 | head = per_cpu_ptr(cache->percpu, cpu); | |
111 | head->first = NULL; | |
112 | head->count = 0; | |
113 | } | |
114 | cache->xfer = NULL; | |
115 | cache->ready = NULL; | |
116 | ||
117 | return 0; | |
118 | } | |
119 | ||
120 | int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic) | |
121 | { | |
122 | int ret; | |
123 | ||
124 | ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs); | |
125 | if (!ret) { | |
126 | ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags); | |
127 | if (ret) | |
128 | free_percpu(ic->i_cache_incs.percpu); | |
129 | } | |
130 | ||
131 | return ret; | |
132 | } | |
133 | ||
134 | static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache, | |
135 | struct list_head *caller_list) | |
136 | { | |
137 | struct rds_ib_cache_head *head; | |
138 | int cpu; | |
139 | ||
140 | for_each_possible_cpu(cpu) { | |
141 | head = per_cpu_ptr(cache->percpu, cpu); | |
142 | if (head->first) { | |
143 | list_splice_entire_tail(head->first, caller_list); | |
144 | head->first = NULL; | |
145 | } | |
146 | } | |
147 | ||
148 | if (cache->ready) { | |
149 | list_splice_entire_tail(cache->ready, caller_list); | |
150 | cache->ready = NULL; | |
151 | } | |
152 | } | |
153 | ||
154 | void rds_ib_recv_free_caches(struct rds_ib_connection *ic) | |
155 | { | |
156 | struct rds_ib_incoming *inc; | |
157 | struct rds_ib_incoming *inc_tmp; | |
158 | struct rds_page_frag *frag; | |
159 | struct rds_page_frag *frag_tmp; | |
160 | LIST_HEAD(list); | |
161 | ||
162 | rds_ib_cache_xfer_to_ready(&ic->i_cache_incs); | |
163 | rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list); | |
164 | free_percpu(ic->i_cache_incs.percpu); | |
165 | ||
166 | list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) { | |
167 | list_del(&inc->ii_cache_entry); | |
168 | WARN_ON(!list_empty(&inc->ii_frags)); | |
169 | kmem_cache_free(rds_ib_incoming_slab, inc); | |
170 | } | |
171 | ||
172 | rds_ib_cache_xfer_to_ready(&ic->i_cache_frags); | |
173 | rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list); | |
174 | free_percpu(ic->i_cache_frags.percpu); | |
175 | ||
176 | list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) { | |
177 | list_del(&frag->f_cache_entry); | |
178 | WARN_ON(!list_empty(&frag->f_item)); | |
179 | kmem_cache_free(rds_ib_frag_slab, frag); | |
180 | } | |
181 | } | |
182 | ||
183 | /* fwd decl */ | |
184 | static void rds_ib_recv_cache_put(struct list_head *new_item, | |
185 | struct rds_ib_refill_cache *cache); | |
186 | static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache); | |
187 | ||
188 | ||
189 | /* Recycle frag and attached recv buffer f_sg */ | |
190 | static void rds_ib_frag_free(struct rds_ib_connection *ic, | |
191 | struct rds_page_frag *frag) | |
192 | { | |
193 | rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg)); | |
194 | ||
195 | rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags); | |
196 | } | |
197 | ||
198 | /* Recycle inc after freeing attached frags */ | |
199 | void rds_ib_inc_free(struct rds_incoming *inc) | |
200 | { | |
201 | struct rds_ib_incoming *ibinc; | |
202 | struct rds_page_frag *frag; | |
203 | struct rds_page_frag *pos; | |
204 | struct rds_ib_connection *ic = inc->i_conn->c_transport_data; | |
205 | ||
206 | ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); | |
207 | ||
208 | /* Free attached frags */ | |
209 | list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) { | |
210 | list_del_init(&frag->f_item); | |
211 | rds_ib_frag_free(ic, frag); | |
212 | } | |
213 | BUG_ON(!list_empty(&ibinc->ii_frags)); | |
214 | ||
215 | rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc); | |
216 | rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs); | |
217 | } | |
218 | ||
1e23b3ee AG |
219 | static void rds_ib_recv_clear_one(struct rds_ib_connection *ic, |
220 | struct rds_ib_recv_work *recv) | |
221 | { | |
222 | if (recv->r_ibinc) { | |
223 | rds_inc_put(&recv->r_ibinc->ii_inc); | |
224 | recv->r_ibinc = NULL; | |
225 | } | |
226 | if (recv->r_frag) { | |
fc24f780 | 227 | ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE); |
33244125 | 228 | rds_ib_frag_free(ic, recv->r_frag); |
1e23b3ee AG |
229 | recv->r_frag = NULL; |
230 | } | |
231 | } | |
232 | ||
233 | void rds_ib_recv_clear_ring(struct rds_ib_connection *ic) | |
234 | { | |
235 | u32 i; | |
236 | ||
237 | for (i = 0; i < ic->i_recv_ring.w_nr; i++) | |
238 | rds_ib_recv_clear_one(ic, &ic->i_recvs[i]); | |
1e23b3ee AG |
239 | } |
240 | ||
037f18a3 CM |
241 | static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic, |
242 | gfp_t slab_mask) | |
33244125 CM |
243 | { |
244 | struct rds_ib_incoming *ibinc; | |
245 | struct list_head *cache_item; | |
246 | int avail_allocs; | |
247 | ||
248 | cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs); | |
249 | if (cache_item) { | |
250 | ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry); | |
251 | } else { | |
252 | avail_allocs = atomic_add_unless(&rds_ib_allocation, | |
253 | 1, rds_ib_sysctl_max_recv_allocation); | |
254 | if (!avail_allocs) { | |
255 | rds_ib_stats_inc(s_ib_rx_alloc_limit); | |
256 | return NULL; | |
257 | } | |
037f18a3 | 258 | ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask); |
33244125 CM |
259 | if (!ibinc) { |
260 | atomic_dec(&rds_ib_allocation); | |
261 | return NULL; | |
262 | } | |
263 | } | |
264 | INIT_LIST_HEAD(&ibinc->ii_frags); | |
265 | rds_inc_init(&ibinc->ii_inc, ic->conn, ic->conn->c_faddr); | |
266 | ||
267 | return ibinc; | |
268 | } | |
269 | ||
037f18a3 CM |
270 | static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic, |
271 | gfp_t slab_mask, gfp_t page_mask) | |
33244125 CM |
272 | { |
273 | struct rds_page_frag *frag; | |
274 | struct list_head *cache_item; | |
275 | int ret; | |
276 | ||
277 | cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags); | |
278 | if (cache_item) { | |
279 | frag = container_of(cache_item, struct rds_page_frag, f_cache_entry); | |
280 | } else { | |
037f18a3 | 281 | frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask); |
33244125 CM |
282 | if (!frag) |
283 | return NULL; | |
284 | ||
b4e1da3c | 285 | sg_init_table(&frag->f_sg, 1); |
33244125 | 286 | ret = rds_page_remainder_alloc(&frag->f_sg, |
037f18a3 | 287 | RDS_FRAG_SIZE, page_mask); |
33244125 CM |
288 | if (ret) { |
289 | kmem_cache_free(rds_ib_frag_slab, frag); | |
290 | return NULL; | |
291 | } | |
292 | } | |
293 | ||
294 | INIT_LIST_HEAD(&frag->f_item); | |
295 | ||
296 | return frag; | |
297 | } | |
298 | ||
1e23b3ee | 299 | static int rds_ib_recv_refill_one(struct rds_connection *conn, |
73ce4317 | 300 | struct rds_ib_recv_work *recv, gfp_t gfp) |
1e23b3ee AG |
301 | { |
302 | struct rds_ib_connection *ic = conn->c_transport_data; | |
1e23b3ee AG |
303 | struct ib_sge *sge; |
304 | int ret = -ENOMEM; | |
037f18a3 CM |
305 | gfp_t slab_mask = GFP_NOWAIT; |
306 | gfp_t page_mask = GFP_NOWAIT; | |
307 | ||
73ce4317 | 308 | if (gfp & __GFP_WAIT) { |
037f18a3 CM |
309 | slab_mask = GFP_KERNEL; |
310 | page_mask = GFP_HIGHUSER; | |
311 | } | |
1e23b3ee | 312 | |
33244125 CM |
313 | if (!ic->i_cache_incs.ready) |
314 | rds_ib_cache_xfer_to_ready(&ic->i_cache_incs); | |
315 | if (!ic->i_cache_frags.ready) | |
316 | rds_ib_cache_xfer_to_ready(&ic->i_cache_frags); | |
317 | ||
3427e854 AG |
318 | /* |
319 | * ibinc was taken from recv if recv contained the start of a message. | |
320 | * recvs that were continuations will still have this allocated. | |
321 | */ | |
8690bfa1 | 322 | if (!recv->r_ibinc) { |
037f18a3 | 323 | recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask); |
33244125 | 324 | if (!recv->r_ibinc) |
1e23b3ee | 325 | goto out; |
1e23b3ee AG |
326 | } |
327 | ||
3427e854 | 328 | WARN_ON(recv->r_frag); /* leak! */ |
037f18a3 | 329 | recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask); |
3427e854 AG |
330 | if (!recv->r_frag) |
331 | goto out; | |
1e23b3ee | 332 | |
0b088e00 AG |
333 | ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, |
334 | 1, DMA_FROM_DEVICE); | |
335 | WARN_ON(ret != 1); | |
1e23b3ee | 336 | |
919ced4c | 337 | sge = &recv->r_sge[0]; |
1e23b3ee AG |
338 | sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header); |
339 | sge->length = sizeof(struct rds_header); | |
340 | ||
919ced4c | 341 | sge = &recv->r_sge[1]; |
f2e9bd70 MM |
342 | sge->addr = ib_sg_dma_address(ic->i_cm_id->device, &recv->r_frag->f_sg); |
343 | sge->length = ib_sg_dma_len(ic->i_cm_id->device, &recv->r_frag->f_sg); | |
1e23b3ee AG |
344 | |
345 | ret = 0; | |
346 | out: | |
347 | return ret; | |
348 | } | |
349 | ||
73ce4317 | 350 | static int acquire_refill(struct rds_connection *conn) |
351 | { | |
352 | return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0; | |
353 | } | |
354 | ||
355 | static void release_refill(struct rds_connection *conn) | |
356 | { | |
357 | clear_bit(RDS_RECV_REFILL, &conn->c_flags); | |
358 | ||
359 | /* We don't use wait_on_bit()/wake_up_bit() because our waking is in a | |
360 | * hot path and finding waiters is very rare. We don't want to walk | |
361 | * the system-wide hashed waitqueue buckets in the fast path only to | |
362 | * almost never find waiters. | |
363 | */ | |
364 | if (waitqueue_active(&conn->c_waitq)) | |
365 | wake_up_all(&conn->c_waitq); | |
366 | } | |
367 | ||
1e23b3ee AG |
368 | /* |
369 | * This tries to allocate and post unused work requests after making sure that | |
370 | * they have all the allocations they need to queue received fragments into | |
33244125 | 371 | * sockets. |
1e23b3ee AG |
372 | * |
373 | * -1 is returned if posting fails due to temporary resource exhaustion. | |
374 | */ | |
73ce4317 | 375 | void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp) |
1e23b3ee AG |
376 | { |
377 | struct rds_ib_connection *ic = conn->c_transport_data; | |
378 | struct rds_ib_recv_work *recv; | |
379 | struct ib_recv_wr *failed_wr; | |
380 | unsigned int posted = 0; | |
381 | int ret = 0; | |
73ce4317 | 382 | int can_wait = gfp & __GFP_WAIT; |
1e23b3ee AG |
383 | u32 pos; |
384 | ||
73ce4317 | 385 | /* the goal here is to just make sure that someone, somewhere |
386 | * is posting buffers. If we can't get the refill lock, | |
387 | * let them do their thing | |
388 | */ | |
389 | if (!acquire_refill(conn)) | |
390 | return; | |
391 | ||
f64f9e71 JP |
392 | while ((prefill || rds_conn_up(conn)) && |
393 | rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) { | |
1e23b3ee AG |
394 | if (pos >= ic->i_recv_ring.w_nr) { |
395 | printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n", | |
396 | pos); | |
1e23b3ee AG |
397 | break; |
398 | } | |
399 | ||
400 | recv = &ic->i_recvs[pos]; | |
73ce4317 | 401 | ret = rds_ib_recv_refill_one(conn, recv, gfp); |
1e23b3ee | 402 | if (ret) { |
1e23b3ee AG |
403 | break; |
404 | } | |
405 | ||
406 | /* XXX when can this fail? */ | |
407 | ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr); | |
408 | rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv, | |
0b088e00 | 409 | recv->r_ibinc, sg_page(&recv->r_frag->f_sg), |
f2e9bd70 MM |
410 | (long) ib_sg_dma_address( |
411 | ic->i_cm_id->device, | |
412 | &recv->r_frag->f_sg), | |
413 | ret); | |
1e23b3ee AG |
414 | if (ret) { |
415 | rds_ib_conn_error(conn, "recv post on " | |
416 | "%pI4 returned %d, disconnecting and " | |
417 | "reconnecting\n", &conn->c_faddr, | |
418 | ret); | |
1e23b3ee AG |
419 | break; |
420 | } | |
421 | ||
422 | posted++; | |
423 | } | |
424 | ||
425 | /* We're doing flow control - update the window. */ | |
426 | if (ic->i_flowctl && posted) | |
427 | rds_ib_advertise_credits(conn, posted); | |
428 | ||
429 | if (ret) | |
430 | rds_ib_ring_unalloc(&ic->i_recv_ring, 1); | |
73ce4317 | 431 | |
432 | release_refill(conn); | |
433 | ||
434 | /* if we're called from the softirq handler, we'll be GFP_NOWAIT. | |
435 | * in this case the ring being low is going to lead to more interrupts | |
436 | * and we can safely let the softirq code take care of it unless the | |
437 | * ring is completely empty. | |
438 | * | |
439 | * if we're called from krdsd, we'll be GFP_KERNEL. In this case | |
440 | * we might have raced with the softirq code while we had the refill | |
441 | * lock held. Use rds_ib_ring_low() instead of ring_empty to decide | |
442 | * if we should requeue. | |
443 | */ | |
444 | if (rds_conn_up(conn) && | |
445 | ((can_wait && rds_ib_ring_low(&ic->i_recv_ring)) || | |
446 | rds_ib_ring_empty(&ic->i_recv_ring))) { | |
447 | queue_delayed_work(rds_wq, &conn->c_recv_w, 1); | |
448 | } | |
1e23b3ee AG |
449 | } |
450 | ||
33244125 CM |
451 | /* |
452 | * We want to recycle several types of recv allocations, like incs and frags. | |
453 | * To use this, the *_free() function passes in the ptr to a list_head within | |
454 | * the recyclee, as well as the cache to put it on. | |
455 | * | |
456 | * First, we put the memory on a percpu list. When this reaches a certain size, | |
457 | * We move it to an intermediate non-percpu list in a lockless manner, with some | |
458 | * xchg/compxchg wizardry. | |
459 | * | |
460 | * N.B. Instead of a list_head as the anchor, we use a single pointer, which can | |
461 | * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and | |
462 | * list_empty() will return true with one element is actually present. | |
463 | */ | |
464 | static void rds_ib_recv_cache_put(struct list_head *new_item, | |
465 | struct rds_ib_refill_cache *cache) | |
1e23b3ee | 466 | { |
33244125 | 467 | unsigned long flags; |
c196403b | 468 | struct list_head *old, *chpfirst; |
1e23b3ee | 469 | |
33244125 | 470 | local_irq_save(flags); |
1e23b3ee | 471 | |
ae4b46e9 SW |
472 | chpfirst = __this_cpu_read(cache->percpu->first); |
473 | if (!chpfirst) | |
33244125 CM |
474 | INIT_LIST_HEAD(new_item); |
475 | else /* put on front */ | |
ae4b46e9 | 476 | list_add_tail(new_item, chpfirst); |
33244125 | 477 | |
c196403b | 478 | __this_cpu_write(cache->percpu->first, new_item); |
ae4b46e9 SW |
479 | __this_cpu_inc(cache->percpu->count); |
480 | ||
481 | if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT) | |
33244125 CM |
482 | goto end; |
483 | ||
484 | /* | |
485 | * Return our per-cpu first list to the cache's xfer by atomically | |
486 | * grabbing the current xfer list, appending it to our per-cpu list, | |
487 | * and then atomically returning that entire list back to the | |
488 | * cache's xfer list as long as it's still empty. | |
489 | */ | |
490 | do { | |
491 | old = xchg(&cache->xfer, NULL); | |
492 | if (old) | |
ae4b46e9 SW |
493 | list_splice_entire_tail(old, chpfirst); |
494 | old = cmpxchg(&cache->xfer, NULL, chpfirst); | |
33244125 CM |
495 | } while (old); |
496 | ||
ae4b46e9 | 497 | |
c196403b | 498 | __this_cpu_write(cache->percpu->first, NULL); |
ae4b46e9 | 499 | __this_cpu_write(cache->percpu->count, 0); |
33244125 CM |
500 | end: |
501 | local_irq_restore(flags); | |
1e23b3ee AG |
502 | } |
503 | ||
33244125 | 504 | static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache) |
1e23b3ee | 505 | { |
33244125 CM |
506 | struct list_head *head = cache->ready; |
507 | ||
508 | if (head) { | |
509 | if (!list_empty(head)) { | |
510 | cache->ready = head->next; | |
511 | list_del_init(head); | |
512 | } else | |
513 | cache->ready = NULL; | |
514 | } | |
1e23b3ee | 515 | |
33244125 | 516 | return head; |
1e23b3ee AG |
517 | } |
518 | ||
c310e72c | 519 | int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to) |
1e23b3ee AG |
520 | { |
521 | struct rds_ib_incoming *ibinc; | |
522 | struct rds_page_frag *frag; | |
1e23b3ee AG |
523 | unsigned long to_copy; |
524 | unsigned long frag_off = 0; | |
1e23b3ee AG |
525 | int copied = 0; |
526 | int ret; | |
527 | u32 len; | |
528 | ||
529 | ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); | |
530 | frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); | |
531 | len = be32_to_cpu(inc->i_hdr.h_len); | |
532 | ||
c310e72c | 533 | while (iov_iter_count(to) && copied < len) { |
1e23b3ee AG |
534 | if (frag_off == RDS_FRAG_SIZE) { |
535 | frag = list_entry(frag->f_item.next, | |
536 | struct rds_page_frag, f_item); | |
537 | frag_off = 0; | |
538 | } | |
c310e72c AV |
539 | to_copy = min_t(unsigned long, iov_iter_count(to), |
540 | RDS_FRAG_SIZE - frag_off); | |
1e23b3ee AG |
541 | to_copy = min_t(unsigned long, to_copy, len - copied); |
542 | ||
1e23b3ee | 543 | /* XXX needs + offset for multiple recvs per page */ |
c310e72c AV |
544 | rds_stats_add(s_copy_to_user, to_copy); |
545 | ret = copy_page_to_iter(sg_page(&frag->f_sg), | |
546 | frag->f_sg.offset + frag_off, | |
547 | to_copy, | |
548 | to); | |
549 | if (ret != to_copy) | |
550 | return -EFAULT; | |
1e23b3ee | 551 | |
1e23b3ee AG |
552 | frag_off += to_copy; |
553 | copied += to_copy; | |
554 | } | |
555 | ||
556 | return copied; | |
557 | } | |
558 | ||
559 | /* ic starts out kzalloc()ed */ | |
560 | void rds_ib_recv_init_ack(struct rds_ib_connection *ic) | |
561 | { | |
562 | struct ib_send_wr *wr = &ic->i_ack_wr; | |
563 | struct ib_sge *sge = &ic->i_ack_sge; | |
564 | ||
565 | sge->addr = ic->i_ack_dma; | |
566 | sge->length = sizeof(struct rds_header); | |
567 | sge->lkey = ic->i_mr->lkey; | |
568 | ||
569 | wr->sg_list = sge; | |
570 | wr->num_sge = 1; | |
571 | wr->opcode = IB_WR_SEND; | |
572 | wr->wr_id = RDS_IB_ACK_WR_ID; | |
573 | wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED; | |
574 | } | |
575 | ||
576 | /* | |
577 | * You'd think that with reliable IB connections you wouldn't need to ack | |
578 | * messages that have been received. The problem is that IB hardware generates | |
579 | * an ack message before it has DMAed the message into memory. This creates a | |
580 | * potential message loss if the HCA is disabled for any reason between when it | |
581 | * sends the ack and before the message is DMAed and processed. This is only a | |
582 | * potential issue if another HCA is available for fail-over. | |
583 | * | |
584 | * When the remote host receives our ack they'll free the sent message from | |
585 | * their send queue. To decrease the latency of this we always send an ack | |
586 | * immediately after we've received messages. | |
587 | * | |
588 | * For simplicity, we only have one ack in flight at a time. This puts | |
589 | * pressure on senders to have deep enough send queues to absorb the latency of | |
590 | * a single ack frame being in flight. This might not be good enough. | |
591 | * | |
592 | * This is implemented by have a long-lived send_wr and sge which point to a | |
593 | * statically allocated ack frame. This ack wr does not fall under the ring | |
594 | * accounting that the tx and rx wrs do. The QP attribute specifically makes | |
595 | * room for it beyond the ring size. Send completion notices its special | |
596 | * wr_id and avoids working with the ring in that case. | |
597 | */ | |
8cbd9606 | 598 | #ifndef KERNEL_HAS_ATOMIC64 |
1e23b3ee AG |
599 | static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, |
600 | int ack_required) | |
601 | { | |
8cbd9606 AG |
602 | unsigned long flags; |
603 | ||
604 | spin_lock_irqsave(&ic->i_ack_lock, flags); | |
605 | ic->i_ack_next = seq; | |
606 | if (ack_required) | |
607 | set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); | |
608 | spin_unlock_irqrestore(&ic->i_ack_lock, flags); | |
609 | } | |
610 | ||
611 | static u64 rds_ib_get_ack(struct rds_ib_connection *ic) | |
612 | { | |
613 | unsigned long flags; | |
614 | u64 seq; | |
615 | ||
616 | clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); | |
617 | ||
618 | spin_lock_irqsave(&ic->i_ack_lock, flags); | |
619 | seq = ic->i_ack_next; | |
620 | spin_unlock_irqrestore(&ic->i_ack_lock, flags); | |
621 | ||
622 | return seq; | |
623 | } | |
624 | #else | |
625 | static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, | |
626 | int ack_required) | |
627 | { | |
628 | atomic64_set(&ic->i_ack_next, seq); | |
1e23b3ee | 629 | if (ack_required) { |
4e857c58 | 630 | smp_mb__before_atomic(); |
1e23b3ee AG |
631 | set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); |
632 | } | |
633 | } | |
634 | ||
635 | static u64 rds_ib_get_ack(struct rds_ib_connection *ic) | |
636 | { | |
637 | clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); | |
4e857c58 | 638 | smp_mb__after_atomic(); |
1e23b3ee | 639 | |
8cbd9606 | 640 | return atomic64_read(&ic->i_ack_next); |
1e23b3ee | 641 | } |
8cbd9606 AG |
642 | #endif |
643 | ||
1e23b3ee AG |
644 | |
645 | static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits) | |
646 | { | |
647 | struct rds_header *hdr = ic->i_ack; | |
648 | struct ib_send_wr *failed_wr; | |
649 | u64 seq; | |
650 | int ret; | |
651 | ||
652 | seq = rds_ib_get_ack(ic); | |
653 | ||
654 | rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq); | |
655 | rds_message_populate_header(hdr, 0, 0, 0); | |
656 | hdr->h_ack = cpu_to_be64(seq); | |
657 | hdr->h_credit = adv_credits; | |
658 | rds_message_make_checksum(hdr); | |
659 | ic->i_ack_queued = jiffies; | |
660 | ||
661 | ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr); | |
662 | if (unlikely(ret)) { | |
663 | /* Failed to send. Release the WR, and | |
664 | * force another ACK. | |
665 | */ | |
666 | clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); | |
667 | set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); | |
668 | ||
669 | rds_ib_stats_inc(s_ib_ack_send_failure); | |
735f61e6 AG |
670 | |
671 | rds_ib_conn_error(ic->conn, "sending ack failed\n"); | |
1e23b3ee AG |
672 | } else |
673 | rds_ib_stats_inc(s_ib_ack_sent); | |
674 | } | |
675 | ||
676 | /* | |
677 | * There are 3 ways of getting acknowledgements to the peer: | |
678 | * 1. We call rds_ib_attempt_ack from the recv completion handler | |
679 | * to send an ACK-only frame. | |
680 | * However, there can be only one such frame in the send queue | |
681 | * at any time, so we may have to postpone it. | |
682 | * 2. When another (data) packet is transmitted while there's | |
683 | * an ACK in the queue, we piggyback the ACK sequence number | |
684 | * on the data packet. | |
685 | * 3. If the ACK WR is done sending, we get called from the | |
686 | * send queue completion handler, and check whether there's | |
687 | * another ACK pending (postponed because the WR was on the | |
688 | * queue). If so, we transmit it. | |
689 | * | |
690 | * We maintain 2 variables: | |
691 | * - i_ack_flags, which keeps track of whether the ACK WR | |
692 | * is currently in the send queue or not (IB_ACK_IN_FLIGHT) | |
693 | * - i_ack_next, which is the last sequence number we received | |
694 | * | |
695 | * Potentially, send queue and receive queue handlers can run concurrently. | |
8cbd9606 AG |
696 | * It would be nice to not have to use a spinlock to synchronize things, |
697 | * but the one problem that rules this out is that 64bit updates are | |
698 | * not atomic on all platforms. Things would be a lot simpler if | |
699 | * we had atomic64 or maybe cmpxchg64 everywhere. | |
1e23b3ee AG |
700 | * |
701 | * Reconnecting complicates this picture just slightly. When we | |
702 | * reconnect, we may be seeing duplicate packets. The peer | |
703 | * is retransmitting them, because it hasn't seen an ACK for | |
704 | * them. It is important that we ACK these. | |
705 | * | |
706 | * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with | |
707 | * this flag set *MUST* be acknowledged immediately. | |
708 | */ | |
709 | ||
710 | /* | |
711 | * When we get here, we're called from the recv queue handler. | |
712 | * Check whether we ought to transmit an ACK. | |
713 | */ | |
714 | void rds_ib_attempt_ack(struct rds_ib_connection *ic) | |
715 | { | |
716 | unsigned int adv_credits; | |
717 | ||
718 | if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) | |
719 | return; | |
720 | ||
721 | if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) { | |
722 | rds_ib_stats_inc(s_ib_ack_send_delayed); | |
723 | return; | |
724 | } | |
725 | ||
726 | /* Can we get a send credit? */ | |
7b70d033 | 727 | if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) { |
1e23b3ee AG |
728 | rds_ib_stats_inc(s_ib_tx_throttle); |
729 | clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); | |
730 | return; | |
731 | } | |
732 | ||
733 | clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); | |
734 | rds_ib_send_ack(ic, adv_credits); | |
735 | } | |
736 | ||
737 | /* | |
738 | * We get here from the send completion handler, when the | |
739 | * adapter tells us the ACK frame was sent. | |
740 | */ | |
741 | void rds_ib_ack_send_complete(struct rds_ib_connection *ic) | |
742 | { | |
743 | clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); | |
744 | rds_ib_attempt_ack(ic); | |
745 | } | |
746 | ||
747 | /* | |
748 | * This is called by the regular xmit code when it wants to piggyback | |
749 | * an ACK on an outgoing frame. | |
750 | */ | |
751 | u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic) | |
752 | { | |
753 | if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) | |
754 | rds_ib_stats_inc(s_ib_ack_send_piggybacked); | |
755 | return rds_ib_get_ack(ic); | |
756 | } | |
757 | ||
758 | /* | |
759 | * It's kind of lame that we're copying from the posted receive pages into | |
760 | * long-lived bitmaps. We could have posted the bitmaps and rdma written into | |
761 | * them. But receiving new congestion bitmaps should be a *rare* event, so | |
762 | * hopefully we won't need to invest that complexity in making it more | |
763 | * efficient. By copying we can share a simpler core with TCP which has to | |
764 | * copy. | |
765 | */ | |
766 | static void rds_ib_cong_recv(struct rds_connection *conn, | |
767 | struct rds_ib_incoming *ibinc) | |
768 | { | |
769 | struct rds_cong_map *map; | |
770 | unsigned int map_off; | |
771 | unsigned int map_page; | |
772 | struct rds_page_frag *frag; | |
773 | unsigned long frag_off; | |
774 | unsigned long to_copy; | |
775 | unsigned long copied; | |
776 | uint64_t uncongested = 0; | |
777 | void *addr; | |
778 | ||
779 | /* catch completely corrupt packets */ | |
780 | if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES) | |
781 | return; | |
782 | ||
783 | map = conn->c_fcong; | |
784 | map_page = 0; | |
785 | map_off = 0; | |
786 | ||
787 | frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); | |
788 | frag_off = 0; | |
789 | ||
790 | copied = 0; | |
791 | ||
792 | while (copied < RDS_CONG_MAP_BYTES) { | |
793 | uint64_t *src, *dst; | |
794 | unsigned int k; | |
795 | ||
796 | to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off); | |
797 | BUG_ON(to_copy & 7); /* Must be 64bit aligned. */ | |
798 | ||
6114eab5 | 799 | addr = kmap_atomic(sg_page(&frag->f_sg)); |
1e23b3ee AG |
800 | |
801 | src = addr + frag_off; | |
802 | dst = (void *)map->m_page_addrs[map_page] + map_off; | |
803 | for (k = 0; k < to_copy; k += 8) { | |
804 | /* Record ports that became uncongested, ie | |
805 | * bits that changed from 0 to 1. */ | |
806 | uncongested |= ~(*src) & *dst; | |
807 | *dst++ = *src++; | |
808 | } | |
6114eab5 | 809 | kunmap_atomic(addr); |
1e23b3ee AG |
810 | |
811 | copied += to_copy; | |
812 | ||
813 | map_off += to_copy; | |
814 | if (map_off == PAGE_SIZE) { | |
815 | map_off = 0; | |
816 | map_page++; | |
817 | } | |
818 | ||
819 | frag_off += to_copy; | |
820 | if (frag_off == RDS_FRAG_SIZE) { | |
821 | frag = list_entry(frag->f_item.next, | |
822 | struct rds_page_frag, f_item); | |
823 | frag_off = 0; | |
824 | } | |
825 | } | |
826 | ||
827 | /* the congestion map is in little endian order */ | |
828 | uncongested = le64_to_cpu(uncongested); | |
829 | ||
830 | rds_cong_map_updated(map, uncongested); | |
831 | } | |
832 | ||
833 | /* | |
834 | * Rings are posted with all the allocations they'll need to queue the | |
835 | * incoming message to the receiving socket so this can't fail. | |
836 | * All fragments start with a header, so we can make sure we're not receiving | |
837 | * garbage, and we can tell a small 8 byte fragment from an ACK frame. | |
838 | */ | |
839 | struct rds_ib_ack_state { | |
840 | u64 ack_next; | |
841 | u64 ack_recv; | |
842 | unsigned int ack_required:1; | |
843 | unsigned int ack_next_valid:1; | |
844 | unsigned int ack_recv_valid:1; | |
845 | }; | |
846 | ||
847 | static void rds_ib_process_recv(struct rds_connection *conn, | |
597ddd50 | 848 | struct rds_ib_recv_work *recv, u32 data_len, |
1e23b3ee AG |
849 | struct rds_ib_ack_state *state) |
850 | { | |
851 | struct rds_ib_connection *ic = conn->c_transport_data; | |
852 | struct rds_ib_incoming *ibinc = ic->i_ibinc; | |
853 | struct rds_header *ihdr, *hdr; | |
854 | ||
855 | /* XXX shut down the connection if port 0,0 are seen? */ | |
856 | ||
857 | rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv, | |
597ddd50 | 858 | data_len); |
1e23b3ee | 859 | |
597ddd50 | 860 | if (data_len < sizeof(struct rds_header)) { |
1e23b3ee | 861 | rds_ib_conn_error(conn, "incoming message " |
5fd5c44d | 862 | "from %pI4 didn't include a " |
1e23b3ee AG |
863 | "header, disconnecting and " |
864 | "reconnecting\n", | |
865 | &conn->c_faddr); | |
866 | return; | |
867 | } | |
597ddd50 | 868 | data_len -= sizeof(struct rds_header); |
1e23b3ee | 869 | |
f147dd9e | 870 | ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs]; |
1e23b3ee AG |
871 | |
872 | /* Validate the checksum. */ | |
873 | if (!rds_message_verify_checksum(ihdr)) { | |
874 | rds_ib_conn_error(conn, "incoming message " | |
875 | "from %pI4 has corrupted header - " | |
876 | "forcing a reconnect\n", | |
877 | &conn->c_faddr); | |
878 | rds_stats_inc(s_recv_drop_bad_checksum); | |
879 | return; | |
880 | } | |
881 | ||
882 | /* Process the ACK sequence which comes with every packet */ | |
883 | state->ack_recv = be64_to_cpu(ihdr->h_ack); | |
884 | state->ack_recv_valid = 1; | |
885 | ||
886 | /* Process the credits update if there was one */ | |
887 | if (ihdr->h_credit) | |
888 | rds_ib_send_add_credits(conn, ihdr->h_credit); | |
889 | ||
597ddd50 | 890 | if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) { |
1e23b3ee AG |
891 | /* This is an ACK-only packet. The fact that it gets |
892 | * special treatment here is that historically, ACKs | |
893 | * were rather special beasts. | |
894 | */ | |
895 | rds_ib_stats_inc(s_ib_ack_received); | |
896 | ||
897 | /* | |
898 | * Usually the frags make their way on to incs and are then freed as | |
899 | * the inc is freed. We don't go that route, so we have to drop the | |
900 | * page ref ourselves. We can't just leave the page on the recv | |
901 | * because that confuses the dma mapping of pages and each recv's use | |
0b088e00 | 902 | * of a partial page. |
1e23b3ee AG |
903 | * |
904 | * FIXME: Fold this into the code path below. | |
905 | */ | |
33244125 | 906 | rds_ib_frag_free(ic, recv->r_frag); |
0b088e00 | 907 | recv->r_frag = NULL; |
1e23b3ee AG |
908 | return; |
909 | } | |
910 | ||
911 | /* | |
912 | * If we don't already have an inc on the connection then this | |
913 | * fragment has a header and starts a message.. copy its header | |
914 | * into the inc and save the inc so we can hang upcoming fragments | |
915 | * off its list. | |
916 | */ | |
8690bfa1 | 917 | if (!ibinc) { |
1e23b3ee AG |
918 | ibinc = recv->r_ibinc; |
919 | recv->r_ibinc = NULL; | |
920 | ic->i_ibinc = ibinc; | |
921 | ||
922 | hdr = &ibinc->ii_inc.i_hdr; | |
923 | memcpy(hdr, ihdr, sizeof(*hdr)); | |
924 | ic->i_recv_data_rem = be32_to_cpu(hdr->h_len); | |
925 | ||
926 | rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc, | |
927 | ic->i_recv_data_rem, hdr->h_flags); | |
928 | } else { | |
929 | hdr = &ibinc->ii_inc.i_hdr; | |
930 | /* We can't just use memcmp here; fragments of a | |
931 | * single message may carry different ACKs */ | |
f64f9e71 JP |
932 | if (hdr->h_sequence != ihdr->h_sequence || |
933 | hdr->h_len != ihdr->h_len || | |
934 | hdr->h_sport != ihdr->h_sport || | |
935 | hdr->h_dport != ihdr->h_dport) { | |
1e23b3ee AG |
936 | rds_ib_conn_error(conn, |
937 | "fragment header mismatch; forcing reconnect\n"); | |
938 | return; | |
939 | } | |
940 | } | |
941 | ||
942 | list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags); | |
943 | recv->r_frag = NULL; | |
944 | ||
945 | if (ic->i_recv_data_rem > RDS_FRAG_SIZE) | |
946 | ic->i_recv_data_rem -= RDS_FRAG_SIZE; | |
947 | else { | |
948 | ic->i_recv_data_rem = 0; | |
949 | ic->i_ibinc = NULL; | |
950 | ||
951 | if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) | |
952 | rds_ib_cong_recv(conn, ibinc); | |
953 | else { | |
954 | rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr, | |
6114eab5 | 955 | &ibinc->ii_inc, GFP_ATOMIC); |
1e23b3ee AG |
956 | state->ack_next = be64_to_cpu(hdr->h_sequence); |
957 | state->ack_next_valid = 1; | |
958 | } | |
959 | ||
960 | /* Evaluate the ACK_REQUIRED flag *after* we received | |
961 | * the complete frame, and after bumping the next_rx | |
962 | * sequence. */ | |
963 | if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) { | |
964 | rds_stats_inc(s_recv_ack_required); | |
965 | state->ack_required = 1; | |
966 | } | |
967 | ||
968 | rds_inc_put(&ibinc->ii_inc); | |
969 | } | |
970 | } | |
971 | ||
972 | /* | |
973 | * Plucking the oldest entry from the ring can be done concurrently with | |
974 | * the thread refilling the ring. Each ring operation is protected by | |
975 | * spinlocks and the transient state of refilling doesn't change the | |
976 | * recording of which entry is oldest. | |
977 | * | |
978 | * This relies on IB only calling one cq comp_handler for each cq so that | |
979 | * there will only be one caller of rds_recv_incoming() per RDS connection. | |
980 | */ | |
981 | void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context) | |
982 | { | |
983 | struct rds_connection *conn = context; | |
984 | struct rds_ib_connection *ic = conn->c_transport_data; | |
1e23b3ee AG |
985 | |
986 | rdsdebug("conn %p cq %p\n", conn, cq); | |
987 | ||
988 | rds_ib_stats_inc(s_ib_rx_cq_call); | |
989 | ||
d521b63b AG |
990 | tasklet_schedule(&ic->i_recv_tasklet); |
991 | } | |
1e23b3ee | 992 | |
d521b63b AG |
993 | static inline void rds_poll_cq(struct rds_ib_connection *ic, |
994 | struct rds_ib_ack_state *state) | |
995 | { | |
996 | struct rds_connection *conn = ic->conn; | |
997 | struct ib_wc wc; | |
998 | struct rds_ib_recv_work *recv; | |
999 | ||
1000 | while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) { | |
59f740a6 ZB |
1001 | rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n", |
1002 | (unsigned long long)wc.wr_id, wc.status, | |
3c88f3dc | 1003 | ib_wc_status_msg(wc.status), wc.byte_len, |
1e23b3ee AG |
1004 | be32_to_cpu(wc.ex.imm_data)); |
1005 | rds_ib_stats_inc(s_ib_rx_cq_event); | |
1006 | ||
1007 | recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)]; | |
1008 | ||
fc24f780 | 1009 | ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE); |
1e23b3ee AG |
1010 | |
1011 | /* | |
1012 | * Also process recvs in connecting state because it is possible | |
1013 | * to get a recv completion _before_ the rdmacm ESTABLISHED | |
1014 | * event is processed. | |
1015 | */ | |
d455ab64 ZB |
1016 | if (wc.status == IB_WC_SUCCESS) { |
1017 | rds_ib_process_recv(conn, recv, wc.byte_len, state); | |
1018 | } else { | |
1e23b3ee | 1019 | /* We expect errors as the qp is drained during shutdown */ |
d455ab64 | 1020 | if (rds_conn_up(conn) || rds_conn_connecting(conn)) |
59f740a6 ZB |
1021 | rds_ib_conn_error(conn, "recv completion on %pI4 had " |
1022 | "status %u (%s), disconnecting and " | |
d455ab64 | 1023 | "reconnecting\n", &conn->c_faddr, |
59f740a6 | 1024 | wc.status, |
3c88f3dc | 1025 | ib_wc_status_msg(wc.status)); |
1e23b3ee AG |
1026 | } |
1027 | ||
d455ab64 | 1028 | /* |
43962dd7 | 1029 | * rds_ib_process_recv() doesn't always consume the frag, and |
1030 | * we might not have called it at all if the wc didn't indicate | |
1031 | * success. We already unmapped the frag's pages, though, and | |
1032 | * the following rds_ib_ring_free() call tells the refill path | |
1033 | * that it will not find an allocated frag here. Make sure we | |
1034 | * keep that promise by freeing a frag that's still on the ring. | |
d455ab64 | 1035 | */ |
43962dd7 | 1036 | if (recv->r_frag) { |
1037 | rds_ib_frag_free(ic, recv->r_frag); | |
1038 | recv->r_frag = NULL; | |
1039 | } | |
1e23b3ee AG |
1040 | rds_ib_ring_free(&ic->i_recv_ring, 1); |
1041 | } | |
d521b63b AG |
1042 | } |
1043 | ||
1044 | void rds_ib_recv_tasklet_fn(unsigned long data) | |
1045 | { | |
1046 | struct rds_ib_connection *ic = (struct rds_ib_connection *) data; | |
1047 | struct rds_connection *conn = ic->conn; | |
1048 | struct rds_ib_ack_state state = { 0, }; | |
1049 | ||
1050 | rds_poll_cq(ic, &state); | |
1051 | ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); | |
1052 | rds_poll_cq(ic, &state); | |
1e23b3ee AG |
1053 | |
1054 | if (state.ack_next_valid) | |
1055 | rds_ib_set_ack(ic, state.ack_next, state.ack_required); | |
1056 | if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) { | |
1057 | rds_send_drop_acked(conn, state.ack_recv, NULL); | |
1058 | ic->i_ack_recv = state.ack_recv; | |
1059 | } | |
1060 | if (rds_conn_up(conn)) | |
1061 | rds_ib_attempt_ack(ic); | |
1062 | ||
1063 | /* If we ever end up with a really empty receive ring, we're | |
1064 | * in deep trouble, as the sender will definitely see RNR | |
1065 | * timeouts. */ | |
1066 | if (rds_ib_ring_empty(&ic->i_recv_ring)) | |
1067 | rds_ib_stats_inc(s_ib_rx_ring_empty); | |
1068 | ||
1e23b3ee | 1069 | if (rds_ib_ring_low(&ic->i_recv_ring)) |
73ce4317 | 1070 | rds_ib_recv_refill(conn, 0, GFP_NOWAIT); |
1e23b3ee AG |
1071 | } |
1072 | ||
1073 | int rds_ib_recv(struct rds_connection *conn) | |
1074 | { | |
1075 | struct rds_ib_connection *ic = conn->c_transport_data; | |
1076 | int ret = 0; | |
1077 | ||
1078 | rdsdebug("conn %p\n", conn); | |
73ce4317 | 1079 | if (rds_conn_up(conn)) { |
1e23b3ee | 1080 | rds_ib_attempt_ack(ic); |
73ce4317 | 1081 | rds_ib_recv_refill(conn, 0, GFP_KERNEL); |
1082 | } | |
1e23b3ee AG |
1083 | |
1084 | return ret; | |
1085 | } | |
1086 | ||
ef87b7ea | 1087 | int rds_ib_recv_init(void) |
1e23b3ee AG |
1088 | { |
1089 | struct sysinfo si; | |
1090 | int ret = -ENOMEM; | |
1091 | ||
1092 | /* Default to 30% of all available RAM for recv memory */ | |
1093 | si_meminfo(&si); | |
1094 | rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE; | |
1095 | ||
1096 | rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming", | |
1097 | sizeof(struct rds_ib_incoming), | |
c20f5b96 | 1098 | 0, SLAB_HWCACHE_ALIGN, NULL); |
8690bfa1 | 1099 | if (!rds_ib_incoming_slab) |
1e23b3ee AG |
1100 | goto out; |
1101 | ||
1102 | rds_ib_frag_slab = kmem_cache_create("rds_ib_frag", | |
1103 | sizeof(struct rds_page_frag), | |
c20f5b96 | 1104 | 0, SLAB_HWCACHE_ALIGN, NULL); |
8690bfa1 | 1105 | if (!rds_ib_frag_slab) |
1e23b3ee AG |
1106 | kmem_cache_destroy(rds_ib_incoming_slab); |
1107 | else | |
1108 | ret = 0; | |
1109 | out: | |
1110 | return ret; | |
1111 | } | |
1112 | ||
1113 | void rds_ib_recv_exit(void) | |
1114 | { | |
1115 | kmem_cache_destroy(rds_ib_incoming_slab); | |
1116 | kmem_cache_destroy(rds_ib_frag_slab); | |
1117 | } |