xref: /openbmc/linux/net/rds/ib_recv.c (revision 2870c4d6)
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/sched/clock.h>
35 #include <linux/slab.h>
36 #include <linux/pci.h>
37 #include <linux/dma-mapping.h>
38 #include <rdma/rdma_cm.h>
39 
40 #include "rds_single_path.h"
41 #include "rds.h"
42 #include "ib.h"
43 
44 static struct kmem_cache *rds_ib_incoming_slab;
45 static struct kmem_cache *rds_ib_frag_slab;
46 static atomic_t	rds_ib_allocation = ATOMIC_INIT(0);
47 
rds_ib_recv_init_ring(struct rds_ib_connection * ic)48 void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
49 {
50 	struct rds_ib_recv_work *recv;
51 	u32 i;
52 
53 	for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
54 		struct ib_sge *sge;
55 
56 		recv->r_ibinc = NULL;
57 		recv->r_frag = NULL;
58 
59 		recv->r_wr.next = NULL;
60 		recv->r_wr.wr_id = i;
61 		recv->r_wr.sg_list = recv->r_sge;
62 		recv->r_wr.num_sge = RDS_IB_RECV_SGE;
63 
64 		sge = &recv->r_sge[0];
65 		sge->addr = ic->i_recv_hdrs_dma[i];
66 		sge->length = sizeof(struct rds_header);
67 		sge->lkey = ic->i_pd->local_dma_lkey;
68 
69 		sge = &recv->r_sge[1];
70 		sge->addr = 0;
71 		sge->length = RDS_FRAG_SIZE;
72 		sge->lkey = ic->i_pd->local_dma_lkey;
73 	}
74 }
75 
76 /*
77  * The entire 'from' list, including the from element itself, is put on
78  * to the tail of the 'to' list.
79  */
list_splice_entire_tail(struct list_head * from,struct list_head * to)80 static void list_splice_entire_tail(struct list_head *from,
81 				    struct list_head *to)
82 {
83 	struct list_head *from_last = from->prev;
84 
85 	list_splice_tail(from_last, to);
86 	list_add_tail(from_last, to);
87 }
88 
rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache * cache)89 static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
90 {
91 	struct list_head *tmp;
92 
93 	tmp = xchg(&cache->xfer, NULL);
94 	if (tmp) {
95 		if (cache->ready)
96 			list_splice_entire_tail(tmp, cache->ready);
97 		else
98 			cache->ready = tmp;
99 	}
100 }
101 
rds_ib_recv_alloc_cache(struct rds_ib_refill_cache * cache,gfp_t gfp)102 static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache, gfp_t gfp)
103 {
104 	struct rds_ib_cache_head *head;
105 	int cpu;
106 
107 	cache->percpu = alloc_percpu_gfp(struct rds_ib_cache_head, gfp);
108 	if (!cache->percpu)
109 	       return -ENOMEM;
110 
111 	for_each_possible_cpu(cpu) {
112 		head = per_cpu_ptr(cache->percpu, cpu);
113 		head->first = NULL;
114 		head->count = 0;
115 	}
116 	cache->xfer = NULL;
117 	cache->ready = NULL;
118 
119 	return 0;
120 }
121 
rds_ib_recv_alloc_caches(struct rds_ib_connection * ic,gfp_t gfp)122 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp)
123 {
124 	int ret;
125 
126 	ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs, gfp);
127 	if (!ret) {
128 		ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags, gfp);
129 		if (ret)
130 			free_percpu(ic->i_cache_incs.percpu);
131 	}
132 
133 	return ret;
134 }
135 
rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache * cache,struct list_head * caller_list)136 static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
137 					  struct list_head *caller_list)
138 {
139 	struct rds_ib_cache_head *head;
140 	int cpu;
141 
142 	for_each_possible_cpu(cpu) {
143 		head = per_cpu_ptr(cache->percpu, cpu);
144 		if (head->first) {
145 			list_splice_entire_tail(head->first, caller_list);
146 			head->first = NULL;
147 		}
148 	}
149 
150 	if (cache->ready) {
151 		list_splice_entire_tail(cache->ready, caller_list);
152 		cache->ready = NULL;
153 	}
154 }
155 
rds_ib_recv_free_caches(struct rds_ib_connection * ic)156 void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
157 {
158 	struct rds_ib_incoming *inc;
159 	struct rds_ib_incoming *inc_tmp;
160 	struct rds_page_frag *frag;
161 	struct rds_page_frag *frag_tmp;
162 	LIST_HEAD(list);
163 
164 	rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
165 	rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list);
166 	free_percpu(ic->i_cache_incs.percpu);
167 
168 	list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) {
169 		list_del(&inc->ii_cache_entry);
170 		WARN_ON(!list_empty(&inc->ii_frags));
171 		kmem_cache_free(rds_ib_incoming_slab, inc);
172 		atomic_dec(&rds_ib_allocation);
173 	}
174 
175 	rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
176 	rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
177 	free_percpu(ic->i_cache_frags.percpu);
178 
179 	list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
180 		list_del(&frag->f_cache_entry);
181 		WARN_ON(!list_empty(&frag->f_item));
182 		kmem_cache_free(rds_ib_frag_slab, frag);
183 	}
184 }
185 
186 /* fwd decl */
187 static void rds_ib_recv_cache_put(struct list_head *new_item,
188 				  struct rds_ib_refill_cache *cache);
189 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
190 
191 
192 /* Recycle frag and attached recv buffer f_sg */
rds_ib_frag_free(struct rds_ib_connection * ic,struct rds_page_frag * frag)193 static void rds_ib_frag_free(struct rds_ib_connection *ic,
194 			     struct rds_page_frag *frag)
195 {
196 	rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
197 
198 	rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
199 	atomic_add(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
200 	rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
201 }
202 
203 /* Recycle inc after freeing attached frags */
rds_ib_inc_free(struct rds_incoming * inc)204 void rds_ib_inc_free(struct rds_incoming *inc)
205 {
206 	struct rds_ib_incoming *ibinc;
207 	struct rds_page_frag *frag;
208 	struct rds_page_frag *pos;
209 	struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
210 
211 	ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
212 
213 	/* Free attached frags */
214 	list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
215 		list_del_init(&frag->f_item);
216 		rds_ib_frag_free(ic, frag);
217 	}
218 	BUG_ON(!list_empty(&ibinc->ii_frags));
219 
220 	rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
221 	rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
222 }
223 
rds_ib_recv_clear_one(struct rds_ib_connection * ic,struct rds_ib_recv_work * recv)224 static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
225 				  struct rds_ib_recv_work *recv)
226 {
227 	if (recv->r_ibinc) {
228 		rds_inc_put(&recv->r_ibinc->ii_inc);
229 		recv->r_ibinc = NULL;
230 	}
231 	if (recv->r_frag) {
232 		ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
233 		rds_ib_frag_free(ic, recv->r_frag);
234 		recv->r_frag = NULL;
235 	}
236 }
237 
rds_ib_recv_clear_ring(struct rds_ib_connection * ic)238 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
239 {
240 	u32 i;
241 
242 	for (i = 0; i < ic->i_recv_ring.w_nr; i++)
243 		rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
244 }
245 
rds_ib_refill_one_inc(struct rds_ib_connection * ic,gfp_t slab_mask)246 static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
247 						     gfp_t slab_mask)
248 {
249 	struct rds_ib_incoming *ibinc;
250 	struct list_head *cache_item;
251 	int avail_allocs;
252 
253 	cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
254 	if (cache_item) {
255 		ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
256 	} else {
257 		avail_allocs = atomic_add_unless(&rds_ib_allocation,
258 						 1, rds_ib_sysctl_max_recv_allocation);
259 		if (!avail_allocs) {
260 			rds_ib_stats_inc(s_ib_rx_alloc_limit);
261 			return NULL;
262 		}
263 		ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
264 		if (!ibinc) {
265 			atomic_dec(&rds_ib_allocation);
266 			return NULL;
267 		}
268 		rds_ib_stats_inc(s_ib_rx_total_incs);
269 	}
270 	INIT_LIST_HEAD(&ibinc->ii_frags);
271 	rds_inc_init(&ibinc->ii_inc, ic->conn, &ic->conn->c_faddr);
272 
273 	return ibinc;
274 }
275 
rds_ib_refill_one_frag(struct rds_ib_connection * ic,gfp_t slab_mask,gfp_t page_mask)276 static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
277 						    gfp_t slab_mask, gfp_t page_mask)
278 {
279 	struct rds_page_frag *frag;
280 	struct list_head *cache_item;
281 	int ret;
282 
283 	cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
284 	if (cache_item) {
285 		frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
286 		atomic_sub(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
287 		rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
288 	} else {
289 		frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
290 		if (!frag)
291 			return NULL;
292 
293 		sg_init_table(&frag->f_sg, 1);
294 		ret = rds_page_remainder_alloc(&frag->f_sg,
295 					       RDS_FRAG_SIZE, page_mask);
296 		if (ret) {
297 			kmem_cache_free(rds_ib_frag_slab, frag);
298 			return NULL;
299 		}
300 		rds_ib_stats_inc(s_ib_rx_total_frags);
301 	}
302 
303 	INIT_LIST_HEAD(&frag->f_item);
304 
305 	return frag;
306 }
307 
rds_ib_recv_refill_one(struct rds_connection * conn,struct rds_ib_recv_work * recv,gfp_t gfp)308 static int rds_ib_recv_refill_one(struct rds_connection *conn,
309 				  struct rds_ib_recv_work *recv, gfp_t gfp)
310 {
311 	struct rds_ib_connection *ic = conn->c_transport_data;
312 	struct ib_sge *sge;
313 	int ret = -ENOMEM;
314 	gfp_t slab_mask = gfp;
315 	gfp_t page_mask = gfp;
316 
317 	if (gfp & __GFP_DIRECT_RECLAIM) {
318 		slab_mask = GFP_KERNEL;
319 		page_mask = GFP_HIGHUSER;
320 	}
321 
322 	if (!ic->i_cache_incs.ready)
323 		rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
324 	if (!ic->i_cache_frags.ready)
325 		rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
326 
327 	/*
328 	 * ibinc was taken from recv if recv contained the start of a message.
329 	 * recvs that were continuations will still have this allocated.
330 	 */
331 	if (!recv->r_ibinc) {
332 		recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
333 		if (!recv->r_ibinc)
334 			goto out;
335 	}
336 
337 	WARN_ON(recv->r_frag); /* leak! */
338 	recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
339 	if (!recv->r_frag)
340 		goto out;
341 
342 	ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
343 			    1, DMA_FROM_DEVICE);
344 	WARN_ON(ret != 1);
345 
346 	sge = &recv->r_sge[0];
347 	sge->addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
348 	sge->length = sizeof(struct rds_header);
349 
350 	sge = &recv->r_sge[1];
351 	sge->addr = sg_dma_address(&recv->r_frag->f_sg);
352 	sge->length = sg_dma_len(&recv->r_frag->f_sg);
353 
354 	ret = 0;
355 out:
356 	return ret;
357 }
358 
acquire_refill(struct rds_connection * conn)359 static int acquire_refill(struct rds_connection *conn)
360 {
361 	return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0;
362 }
363 
release_refill(struct rds_connection * conn)364 static void release_refill(struct rds_connection *conn)
365 {
366 	clear_bit(RDS_RECV_REFILL, &conn->c_flags);
367 	smp_mb__after_atomic();
368 
369 	/* We don't use wait_on_bit()/wake_up_bit() because our waking is in a
370 	 * hot path and finding waiters is very rare.  We don't want to walk
371 	 * the system-wide hashed waitqueue buckets in the fast path only to
372 	 * almost never find waiters.
373 	 */
374 	if (waitqueue_active(&conn->c_waitq))
375 		wake_up_all(&conn->c_waitq);
376 }
377 
378 /*
379  * This tries to allocate and post unused work requests after making sure that
380  * they have all the allocations they need to queue received fragments into
381  * sockets.
382  */
rds_ib_recv_refill(struct rds_connection * conn,int prefill,gfp_t gfp)383 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
384 {
385 	struct rds_ib_connection *ic = conn->c_transport_data;
386 	struct rds_ib_recv_work *recv;
387 	unsigned int posted = 0;
388 	int ret = 0;
389 	bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM);
390 	bool must_wake = false;
391 	u32 pos;
392 
393 	/* the goal here is to just make sure that someone, somewhere
394 	 * is posting buffers.  If we can't get the refill lock,
395 	 * let them do their thing
396 	 */
397 	if (!acquire_refill(conn))
398 		return;
399 
400 	while ((prefill || rds_conn_up(conn)) &&
401 	       rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
402 		if (pos >= ic->i_recv_ring.w_nr) {
403 			printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
404 					pos);
405 			break;
406 		}
407 
408 		recv = &ic->i_recvs[pos];
409 		ret = rds_ib_recv_refill_one(conn, recv, gfp);
410 		if (ret) {
411 			must_wake = true;
412 			break;
413 		}
414 
415 		rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv,
416 			 recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
417 			 (long)sg_dma_address(&recv->r_frag->f_sg));
418 
419 		/* XXX when can this fail? */
420 		ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, NULL);
421 		if (ret) {
422 			rds_ib_conn_error(conn, "recv post on "
423 			       "%pI6c returned %d, disconnecting and "
424 			       "reconnecting\n", &conn->c_faddr,
425 			       ret);
426 			break;
427 		}
428 
429 		posted++;
430 
431 		if ((posted > 128 && need_resched()) || posted > 8192) {
432 			must_wake = true;
433 			break;
434 		}
435 	}
436 
437 	/* We're doing flow control - update the window. */
438 	if (ic->i_flowctl && posted)
439 		rds_ib_advertise_credits(conn, posted);
440 
441 	if (ret)
442 		rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
443 
444 	release_refill(conn);
445 
446 	/* if we're called from the softirq handler, we'll be GFP_NOWAIT.
447 	 * in this case the ring being low is going to lead to more interrupts
448 	 * and we can safely let the softirq code take care of it unless the
449 	 * ring is completely empty.
450 	 *
451 	 * if we're called from krdsd, we'll be GFP_KERNEL.  In this case
452 	 * we might have raced with the softirq code while we had the refill
453 	 * lock held.  Use rds_ib_ring_low() instead of ring_empty to decide
454 	 * if we should requeue.
455 	 */
456 	if (rds_conn_up(conn) &&
457 	    (must_wake ||
458 	    (can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
459 	    rds_ib_ring_empty(&ic->i_recv_ring))) {
460 		queue_delayed_work(rds_wq, &conn->c_recv_w, 1);
461 	}
462 	if (can_wait)
463 		cond_resched();
464 }
465 
466 /*
467  * We want to recycle several types of recv allocations, like incs and frags.
468  * To use this, the *_free() function passes in the ptr to a list_head within
469  * the recyclee, as well as the cache to put it on.
470  *
471  * First, we put the memory on a percpu list. When this reaches a certain size,
472  * We move it to an intermediate non-percpu list in a lockless manner, with some
473  * xchg/compxchg wizardry.
474  *
475  * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
476  * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
477  * list_empty() will return true with one element is actually present.
478  */
rds_ib_recv_cache_put(struct list_head * new_item,struct rds_ib_refill_cache * cache)479 static void rds_ib_recv_cache_put(struct list_head *new_item,
480 				 struct rds_ib_refill_cache *cache)
481 {
482 	unsigned long flags;
483 	struct list_head *old, *chpfirst;
484 
485 	local_irq_save(flags);
486 
487 	chpfirst = __this_cpu_read(cache->percpu->first);
488 	if (!chpfirst)
489 		INIT_LIST_HEAD(new_item);
490 	else /* put on front */
491 		list_add_tail(new_item, chpfirst);
492 
493 	__this_cpu_write(cache->percpu->first, new_item);
494 	__this_cpu_inc(cache->percpu->count);
495 
496 	if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
497 		goto end;
498 
499 	/*
500 	 * Return our per-cpu first list to the cache's xfer by atomically
501 	 * grabbing the current xfer list, appending it to our per-cpu list,
502 	 * and then atomically returning that entire list back to the
503 	 * cache's xfer list as long as it's still empty.
504 	 */
505 	do {
506 		old = xchg(&cache->xfer, NULL);
507 		if (old)
508 			list_splice_entire_tail(old, chpfirst);
509 		old = cmpxchg(&cache->xfer, NULL, chpfirst);
510 	} while (old);
511 
512 
513 	__this_cpu_write(cache->percpu->first, NULL);
514 	__this_cpu_write(cache->percpu->count, 0);
515 end:
516 	local_irq_restore(flags);
517 }
518 
rds_ib_recv_cache_get(struct rds_ib_refill_cache * cache)519 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
520 {
521 	struct list_head *head = cache->ready;
522 
523 	if (head) {
524 		if (!list_empty(head)) {
525 			cache->ready = head->next;
526 			list_del_init(head);
527 		} else
528 			cache->ready = NULL;
529 	}
530 
531 	return head;
532 }
533 
rds_ib_inc_copy_to_user(struct rds_incoming * inc,struct iov_iter * to)534 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
535 {
536 	struct rds_ib_incoming *ibinc;
537 	struct rds_page_frag *frag;
538 	unsigned long to_copy;
539 	unsigned long frag_off = 0;
540 	int copied = 0;
541 	int ret;
542 	u32 len;
543 
544 	ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
545 	frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
546 	len = be32_to_cpu(inc->i_hdr.h_len);
547 
548 	while (iov_iter_count(to) && copied < len) {
549 		if (frag_off == RDS_FRAG_SIZE) {
550 			frag = list_entry(frag->f_item.next,
551 					  struct rds_page_frag, f_item);
552 			frag_off = 0;
553 		}
554 		to_copy = min_t(unsigned long, iov_iter_count(to),
555 				RDS_FRAG_SIZE - frag_off);
556 		to_copy = min_t(unsigned long, to_copy, len - copied);
557 
558 		/* XXX needs + offset for multiple recvs per page */
559 		rds_stats_add(s_copy_to_user, to_copy);
560 		ret = copy_page_to_iter(sg_page(&frag->f_sg),
561 					frag->f_sg.offset + frag_off,
562 					to_copy,
563 					to);
564 		if (ret != to_copy)
565 			return -EFAULT;
566 
567 		frag_off += to_copy;
568 		copied += to_copy;
569 	}
570 
571 	return copied;
572 }
573 
574 /* ic starts out kzalloc()ed */
rds_ib_recv_init_ack(struct rds_ib_connection * ic)575 void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
576 {
577 	struct ib_send_wr *wr = &ic->i_ack_wr;
578 	struct ib_sge *sge = &ic->i_ack_sge;
579 
580 	sge->addr = ic->i_ack_dma;
581 	sge->length = sizeof(struct rds_header);
582 	sge->lkey = ic->i_pd->local_dma_lkey;
583 
584 	wr->sg_list = sge;
585 	wr->num_sge = 1;
586 	wr->opcode = IB_WR_SEND;
587 	wr->wr_id = RDS_IB_ACK_WR_ID;
588 	wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
589 }
590 
591 /*
592  * You'd think that with reliable IB connections you wouldn't need to ack
593  * messages that have been received.  The problem is that IB hardware generates
594  * an ack message before it has DMAed the message into memory.  This creates a
595  * potential message loss if the HCA is disabled for any reason between when it
596  * sends the ack and before the message is DMAed and processed.  This is only a
597  * potential issue if another HCA is available for fail-over.
598  *
599  * When the remote host receives our ack they'll free the sent message from
600  * their send queue.  To decrease the latency of this we always send an ack
601  * immediately after we've received messages.
602  *
603  * For simplicity, we only have one ack in flight at a time.  This puts
604  * pressure on senders to have deep enough send queues to absorb the latency of
605  * a single ack frame being in flight.  This might not be good enough.
606  *
607  * This is implemented by have a long-lived send_wr and sge which point to a
608  * statically allocated ack frame.  This ack wr does not fall under the ring
609  * accounting that the tx and rx wrs do.  The QP attribute specifically makes
610  * room for it beyond the ring size.  Send completion notices its special
611  * wr_id and avoids working with the ring in that case.
612  */
613 #ifndef KERNEL_HAS_ATOMIC64
rds_ib_set_ack(struct rds_ib_connection * ic,u64 seq,int ack_required)614 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
615 {
616 	unsigned long flags;
617 
618 	spin_lock_irqsave(&ic->i_ack_lock, flags);
619 	ic->i_ack_next = seq;
620 	if (ack_required)
621 		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
622 	spin_unlock_irqrestore(&ic->i_ack_lock, flags);
623 }
624 
rds_ib_get_ack(struct rds_ib_connection * ic)625 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
626 {
627 	unsigned long flags;
628 	u64 seq;
629 
630 	clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
631 
632 	spin_lock_irqsave(&ic->i_ack_lock, flags);
633 	seq = ic->i_ack_next;
634 	spin_unlock_irqrestore(&ic->i_ack_lock, flags);
635 
636 	return seq;
637 }
638 #else
rds_ib_set_ack(struct rds_ib_connection * ic,u64 seq,int ack_required)639 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
640 {
641 	atomic64_set(&ic->i_ack_next, seq);
642 	if (ack_required) {
643 		smp_mb__before_atomic();
644 		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
645 	}
646 }
647 
rds_ib_get_ack(struct rds_ib_connection * ic)648 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
649 {
650 	clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
651 	smp_mb__after_atomic();
652 
653 	return atomic64_read(&ic->i_ack_next);
654 }
655 #endif
656 
657 
rds_ib_send_ack(struct rds_ib_connection * ic,unsigned int adv_credits)658 static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
659 {
660 	struct rds_header *hdr = ic->i_ack;
661 	u64 seq;
662 	int ret;
663 
664 	seq = rds_ib_get_ack(ic);
665 
666 	rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
667 
668 	ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, ic->i_ack_dma,
669 				   sizeof(*hdr), DMA_TO_DEVICE);
670 	rds_message_populate_header(hdr, 0, 0, 0);
671 	hdr->h_ack = cpu_to_be64(seq);
672 	hdr->h_credit = adv_credits;
673 	rds_message_make_checksum(hdr);
674 	ib_dma_sync_single_for_device(ic->rds_ibdev->dev, ic->i_ack_dma,
675 				      sizeof(*hdr), DMA_TO_DEVICE);
676 
677 	ic->i_ack_queued = jiffies;
678 
679 	ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, NULL);
680 	if (unlikely(ret)) {
681 		/* Failed to send. Release the WR, and
682 		 * force another ACK.
683 		 */
684 		clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
685 		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
686 
687 		rds_ib_stats_inc(s_ib_ack_send_failure);
688 
689 		rds_ib_conn_error(ic->conn, "sending ack failed\n");
690 	} else
691 		rds_ib_stats_inc(s_ib_ack_sent);
692 }
693 
694 /*
695  * There are 3 ways of getting acknowledgements to the peer:
696  *  1.	We call rds_ib_attempt_ack from the recv completion handler
697  *	to send an ACK-only frame.
698  *	However, there can be only one such frame in the send queue
699  *	at any time, so we may have to postpone it.
700  *  2.	When another (data) packet is transmitted while there's
701  *	an ACK in the queue, we piggyback the ACK sequence number
702  *	on the data packet.
703  *  3.	If the ACK WR is done sending, we get called from the
704  *	send queue completion handler, and check whether there's
705  *	another ACK pending (postponed because the WR was on the
706  *	queue). If so, we transmit it.
707  *
708  * We maintain 2 variables:
709  *  -	i_ack_flags, which keeps track of whether the ACK WR
710  *	is currently in the send queue or not (IB_ACK_IN_FLIGHT)
711  *  -	i_ack_next, which is the last sequence number we received
712  *
713  * Potentially, send queue and receive queue handlers can run concurrently.
714  * It would be nice to not have to use a spinlock to synchronize things,
715  * but the one problem that rules this out is that 64bit updates are
716  * not atomic on all platforms. Things would be a lot simpler if
717  * we had atomic64 or maybe cmpxchg64 everywhere.
718  *
719  * Reconnecting complicates this picture just slightly. When we
720  * reconnect, we may be seeing duplicate packets. The peer
721  * is retransmitting them, because it hasn't seen an ACK for
722  * them. It is important that we ACK these.
723  *
724  * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
725  * this flag set *MUST* be acknowledged immediately.
726  */
727 
728 /*
729  * When we get here, we're called from the recv queue handler.
730  * Check whether we ought to transmit an ACK.
731  */
rds_ib_attempt_ack(struct rds_ib_connection * ic)732 void rds_ib_attempt_ack(struct rds_ib_connection *ic)
733 {
734 	unsigned int adv_credits;
735 
736 	if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
737 		return;
738 
739 	if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
740 		rds_ib_stats_inc(s_ib_ack_send_delayed);
741 		return;
742 	}
743 
744 	/* Can we get a send credit? */
745 	if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
746 		rds_ib_stats_inc(s_ib_tx_throttle);
747 		clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
748 		return;
749 	}
750 
751 	clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
752 	rds_ib_send_ack(ic, adv_credits);
753 }
754 
755 /*
756  * We get here from the send completion handler, when the
757  * adapter tells us the ACK frame was sent.
758  */
rds_ib_ack_send_complete(struct rds_ib_connection * ic)759 void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
760 {
761 	clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
762 	rds_ib_attempt_ack(ic);
763 }
764 
765 /*
766  * This is called by the regular xmit code when it wants to piggyback
767  * an ACK on an outgoing frame.
768  */
rds_ib_piggyb_ack(struct rds_ib_connection * ic)769 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
770 {
771 	if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
772 		rds_ib_stats_inc(s_ib_ack_send_piggybacked);
773 	return rds_ib_get_ack(ic);
774 }
775 
776 /*
777  * It's kind of lame that we're copying from the posted receive pages into
778  * long-lived bitmaps.  We could have posted the bitmaps and rdma written into
779  * them.  But receiving new congestion bitmaps should be a *rare* event, so
780  * hopefully we won't need to invest that complexity in making it more
781  * efficient.  By copying we can share a simpler core with TCP which has to
782  * copy.
783  */
rds_ib_cong_recv(struct rds_connection * conn,struct rds_ib_incoming * ibinc)784 static void rds_ib_cong_recv(struct rds_connection *conn,
785 			      struct rds_ib_incoming *ibinc)
786 {
787 	struct rds_cong_map *map;
788 	unsigned int map_off;
789 	unsigned int map_page;
790 	struct rds_page_frag *frag;
791 	unsigned long frag_off;
792 	unsigned long to_copy;
793 	unsigned long copied;
794 	__le64 uncongested = 0;
795 	void *addr;
796 
797 	/* catch completely corrupt packets */
798 	if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
799 		return;
800 
801 	map = conn->c_fcong;
802 	map_page = 0;
803 	map_off = 0;
804 
805 	frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
806 	frag_off = 0;
807 
808 	copied = 0;
809 
810 	while (copied < RDS_CONG_MAP_BYTES) {
811 		__le64 *src, *dst;
812 		unsigned int k;
813 
814 		to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
815 		BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
816 
817 		addr = kmap_atomic(sg_page(&frag->f_sg));
818 
819 		src = addr + frag->f_sg.offset + frag_off;
820 		dst = (void *)map->m_page_addrs[map_page] + map_off;
821 		for (k = 0; k < to_copy; k += 8) {
822 			/* Record ports that became uncongested, ie
823 			 * bits that changed from 0 to 1. */
824 			uncongested |= ~(*src) & *dst;
825 			*dst++ = *src++;
826 		}
827 		kunmap_atomic(addr);
828 
829 		copied += to_copy;
830 
831 		map_off += to_copy;
832 		if (map_off == PAGE_SIZE) {
833 			map_off = 0;
834 			map_page++;
835 		}
836 
837 		frag_off += to_copy;
838 		if (frag_off == RDS_FRAG_SIZE) {
839 			frag = list_entry(frag->f_item.next,
840 					  struct rds_page_frag, f_item);
841 			frag_off = 0;
842 		}
843 	}
844 
845 	/* the congestion map is in little endian order */
846 	rds_cong_map_updated(map, le64_to_cpu(uncongested));
847 }
848 
rds_ib_process_recv(struct rds_connection * conn,struct rds_ib_recv_work * recv,u32 data_len,struct rds_ib_ack_state * state)849 static void rds_ib_process_recv(struct rds_connection *conn,
850 				struct rds_ib_recv_work *recv, u32 data_len,
851 				struct rds_ib_ack_state *state)
852 {
853 	struct rds_ib_connection *ic = conn->c_transport_data;
854 	struct rds_ib_incoming *ibinc = ic->i_ibinc;
855 	struct rds_header *ihdr, *hdr;
856 	dma_addr_t dma_addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
857 
858 	/* XXX shut down the connection if port 0,0 are seen? */
859 
860 	rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
861 		 data_len);
862 
863 	if (data_len < sizeof(struct rds_header)) {
864 		rds_ib_conn_error(conn, "incoming message "
865 		       "from %pI6c didn't include a "
866 		       "header, disconnecting and "
867 		       "reconnecting\n",
868 		       &conn->c_faddr);
869 		return;
870 	}
871 	data_len -= sizeof(struct rds_header);
872 
873 	ihdr = ic->i_recv_hdrs[recv - ic->i_recvs];
874 
875 	ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, dma_addr,
876 				   sizeof(*ihdr), DMA_FROM_DEVICE);
877 	/* Validate the checksum. */
878 	if (!rds_message_verify_checksum(ihdr)) {
879 		rds_ib_conn_error(conn, "incoming message "
880 		       "from %pI6c has corrupted header - "
881 		       "forcing a reconnect\n",
882 		       &conn->c_faddr);
883 		rds_stats_inc(s_recv_drop_bad_checksum);
884 		goto done;
885 	}
886 
887 	/* Process the ACK sequence which comes with every packet */
888 	state->ack_recv = be64_to_cpu(ihdr->h_ack);
889 	state->ack_recv_valid = 1;
890 
891 	/* Process the credits update if there was one */
892 	if (ihdr->h_credit)
893 		rds_ib_send_add_credits(conn, ihdr->h_credit);
894 
895 	if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
896 		/* This is an ACK-only packet. The fact that it gets
897 		 * special treatment here is that historically, ACKs
898 		 * were rather special beasts.
899 		 */
900 		rds_ib_stats_inc(s_ib_ack_received);
901 
902 		/*
903 		 * Usually the frags make their way on to incs and are then freed as
904 		 * the inc is freed.  We don't go that route, so we have to drop the
905 		 * page ref ourselves.  We can't just leave the page on the recv
906 		 * because that confuses the dma mapping of pages and each recv's use
907 		 * of a partial page.
908 		 *
909 		 * FIXME: Fold this into the code path below.
910 		 */
911 		rds_ib_frag_free(ic, recv->r_frag);
912 		recv->r_frag = NULL;
913 		goto done;
914 	}
915 
916 	/*
917 	 * If we don't already have an inc on the connection then this
918 	 * fragment has a header and starts a message.. copy its header
919 	 * into the inc and save the inc so we can hang upcoming fragments
920 	 * off its list.
921 	 */
922 	if (!ibinc) {
923 		ibinc = recv->r_ibinc;
924 		recv->r_ibinc = NULL;
925 		ic->i_ibinc = ibinc;
926 
927 		hdr = &ibinc->ii_inc.i_hdr;
928 		ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] =
929 				local_clock();
930 		memcpy(hdr, ihdr, sizeof(*hdr));
931 		ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
932 		ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] =
933 				local_clock();
934 
935 		rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
936 			 ic->i_recv_data_rem, hdr->h_flags);
937 	} else {
938 		hdr = &ibinc->ii_inc.i_hdr;
939 		/* We can't just use memcmp here; fragments of a
940 		 * single message may carry different ACKs */
941 		if (hdr->h_sequence != ihdr->h_sequence ||
942 		    hdr->h_len != ihdr->h_len ||
943 		    hdr->h_sport != ihdr->h_sport ||
944 		    hdr->h_dport != ihdr->h_dport) {
945 			rds_ib_conn_error(conn,
946 				"fragment header mismatch; forcing reconnect\n");
947 			goto done;
948 		}
949 	}
950 
951 	list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
952 	recv->r_frag = NULL;
953 
954 	if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
955 		ic->i_recv_data_rem -= RDS_FRAG_SIZE;
956 	else {
957 		ic->i_recv_data_rem = 0;
958 		ic->i_ibinc = NULL;
959 
960 		if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) {
961 			rds_ib_cong_recv(conn, ibinc);
962 		} else {
963 			rds_recv_incoming(conn, &conn->c_faddr, &conn->c_laddr,
964 					  &ibinc->ii_inc, GFP_ATOMIC);
965 			state->ack_next = be64_to_cpu(hdr->h_sequence);
966 			state->ack_next_valid = 1;
967 		}
968 
969 		/* Evaluate the ACK_REQUIRED flag *after* we received
970 		 * the complete frame, and after bumping the next_rx
971 		 * sequence. */
972 		if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
973 			rds_stats_inc(s_recv_ack_required);
974 			state->ack_required = 1;
975 		}
976 
977 		rds_inc_put(&ibinc->ii_inc);
978 	}
979 done:
980 	ib_dma_sync_single_for_device(ic->rds_ibdev->dev, dma_addr,
981 				      sizeof(*ihdr), DMA_FROM_DEVICE);
982 }
983 
rds_ib_recv_cqe_handler(struct rds_ib_connection * ic,struct ib_wc * wc,struct rds_ib_ack_state * state)984 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
985 			     struct ib_wc *wc,
986 			     struct rds_ib_ack_state *state)
987 {
988 	struct rds_connection *conn = ic->conn;
989 	struct rds_ib_recv_work *recv;
990 
991 	rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
992 		 (unsigned long long)wc->wr_id, wc->status,
993 		 ib_wc_status_msg(wc->status), wc->byte_len,
994 		 be32_to_cpu(wc->ex.imm_data));
995 
996 	rds_ib_stats_inc(s_ib_rx_cq_event);
997 	recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
998 	ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1,
999 			DMA_FROM_DEVICE);
1000 
1001 	/* Also process recvs in connecting state because it is possible
1002 	 * to get a recv completion _before_ the rdmacm ESTABLISHED
1003 	 * event is processed.
1004 	 */
1005 	if (wc->status == IB_WC_SUCCESS) {
1006 		rds_ib_process_recv(conn, recv, wc->byte_len, state);
1007 	} else {
1008 		/* We expect errors as the qp is drained during shutdown */
1009 		if (rds_conn_up(conn) || rds_conn_connecting(conn))
1010 			rds_ib_conn_error(conn, "recv completion on <%pI6c,%pI6c, %d> had status %u (%s), vendor err 0x%x, disconnecting and reconnecting\n",
1011 					  &conn->c_laddr, &conn->c_faddr,
1012 					  conn->c_tos, wc->status,
1013 					  ib_wc_status_msg(wc->status),
1014 					  wc->vendor_err);
1015 	}
1016 
1017 	/* rds_ib_process_recv() doesn't always consume the frag, and
1018 	 * we might not have called it at all if the wc didn't indicate
1019 	 * success. We already unmapped the frag's pages, though, and
1020 	 * the following rds_ib_ring_free() call tells the refill path
1021 	 * that it will not find an allocated frag here. Make sure we
1022 	 * keep that promise by freeing a frag that's still on the ring.
1023 	 */
1024 	if (recv->r_frag) {
1025 		rds_ib_frag_free(ic, recv->r_frag);
1026 		recv->r_frag = NULL;
1027 	}
1028 	rds_ib_ring_free(&ic->i_recv_ring, 1);
1029 
1030 	/* If we ever end up with a really empty receive ring, we're
1031 	 * in deep trouble, as the sender will definitely see RNR
1032 	 * timeouts. */
1033 	if (rds_ib_ring_empty(&ic->i_recv_ring))
1034 		rds_ib_stats_inc(s_ib_rx_ring_empty);
1035 
1036 	if (rds_ib_ring_low(&ic->i_recv_ring)) {
1037 		rds_ib_recv_refill(conn, 0, GFP_NOWAIT | __GFP_NOWARN);
1038 		rds_ib_stats_inc(s_ib_rx_refill_from_cq);
1039 	}
1040 }
1041 
rds_ib_recv_path(struct rds_conn_path * cp)1042 int rds_ib_recv_path(struct rds_conn_path *cp)
1043 {
1044 	struct rds_connection *conn = cp->cp_conn;
1045 	struct rds_ib_connection *ic = conn->c_transport_data;
1046 
1047 	rdsdebug("conn %p\n", conn);
1048 	if (rds_conn_up(conn)) {
1049 		rds_ib_attempt_ack(ic);
1050 		rds_ib_recv_refill(conn, 0, GFP_KERNEL);
1051 		rds_ib_stats_inc(s_ib_rx_refill_from_thread);
1052 	}
1053 
1054 	return 0;
1055 }
1056 
rds_ib_recv_init(void)1057 int rds_ib_recv_init(void)
1058 {
1059 	struct sysinfo si;
1060 	int ret = -ENOMEM;
1061 
1062 	/* Default to 30% of all available RAM for recv memory */
1063 	si_meminfo(&si);
1064 	rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1065 
1066 	rds_ib_incoming_slab =
1067 		kmem_cache_create_usercopy("rds_ib_incoming",
1068 					   sizeof(struct rds_ib_incoming),
1069 					   0, SLAB_HWCACHE_ALIGN,
1070 					   offsetof(struct rds_ib_incoming,
1071 						    ii_inc.i_usercopy),
1072 					   sizeof(struct rds_inc_usercopy),
1073 					   NULL);
1074 	if (!rds_ib_incoming_slab)
1075 		goto out;
1076 
1077 	rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1078 					sizeof(struct rds_page_frag),
1079 					0, SLAB_HWCACHE_ALIGN, NULL);
1080 	if (!rds_ib_frag_slab) {
1081 		kmem_cache_destroy(rds_ib_incoming_slab);
1082 		rds_ib_incoming_slab = NULL;
1083 	} else
1084 		ret = 0;
1085 out:
1086 	return ret;
1087 }
1088 
rds_ib_recv_exit(void)1089 void rds_ib_recv_exit(void)
1090 {
1091 	WARN_ON(atomic_read(&rds_ib_allocation));
1092 
1093 	kmem_cache_destroy(rds_ib_incoming_slab);
1094 	kmem_cache_destroy(rds_ib_frag_slab);
1095 }
1096