projects/linux·blob master · refs

598 lines · 15.4 KiB

1/*
2 * Copyright (c) 2006, 2020 Oracle and/or its affiliates.
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/export.h>
36#include <linux/skbuff.h>
37#include <linux/list.h>
38#include <linux/errqueue.h>
39
40#include "rds.h"
41
42static unsigned int rds_exthdr_size[__RDS_EXTHDR_MAX] = {
43[RDS_EXTHDR_NONE] = 0,
44[RDS_EXTHDR_VERSION] = sizeof(struct rds_ext_header_version),
45[RDS_EXTHDR_RDMA] = sizeof(struct rds_ext_header_rdma),
46[RDS_EXTHDR_RDMA_DEST] = sizeof(struct rds_ext_header_rdma_dest),
47[RDS_EXTHDR_RDMA_BYTES] = sizeof(struct rds_ext_header_rdma_bytes),
48[RDS_EXTHDR_NPATHS] = sizeof(__be16),
49[RDS_EXTHDR_GEN_NUM] = sizeof(__be32),
50[RDS_EXTHDR_SPORT_IDX] = 1,
51};
52
53void rds_message_addref(struct rds_message *rm)
54{
55 rdsdebug("addref rm %p ref %d\n", rm, refcount_read(&rm->m_refcount));
56 refcount_inc(&rm->m_refcount);
57}
58EXPORT_SYMBOL_GPL(rds_message_addref);
59
60static inline bool rds_zcookie_add(struct rds_msg_zcopy_info *info, u32 cookie)
61{
62 struct rds_zcopy_cookies *ck = &info->zcookies;
63 int ncookies = ck->num;
64
65 if (ncookies == RDS_MAX_ZCOOKIES)
66 return false;
67 ck->cookies[ncookies] = cookie;
68 ck->num = ++ncookies;
69 return true;
70}
71
72static struct rds_msg_zcopy_info *rds_info_from_znotifier(struct rds_znotifier *znotif)
73{
74 return container_of(znotif, struct rds_msg_zcopy_info, znotif);
75}
76
77void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *q)
78{
79 unsigned long flags;
80 LIST_HEAD(copy);
81 struct rds_msg_zcopy_info *info, *tmp;
82
83 spin_lock_irqsave(&q->lock, flags);
84 list_splice(&q->zcookie_head, &copy);
85 INIT_LIST_HEAD(&q->zcookie_head);
86 spin_unlock_irqrestore(&q->lock, flags);
87
88 list_for_each_entry_safe(info, tmp, &copy, rs_zcookie_next) {
89 list_del(&info->rs_zcookie_next);
90 kfree(info);
91 }
92}
93
94static void rds_rm_zerocopy_callback(struct rds_sock *rs,
95 struct rds_znotifier *znotif)
96{
97 struct rds_msg_zcopy_info *info;
98 struct rds_msg_zcopy_queue *q;
99 u32 cookie = znotif->z_cookie;
100 struct rds_zcopy_cookies *ck;
101 struct list_head *head;
102 unsigned long flags;
103
104 mm_unaccount_pinned_pages(&znotif->z_mmp);
105 q = &rs->rs_zcookie_queue;
106 spin_lock_irqsave(&q->lock, flags);
107 head = &q->zcookie_head;
108 if (!list_empty(head)) {
109 info = list_first_entry(head, struct rds_msg_zcopy_info,
110 rs_zcookie_next);
111 if (rds_zcookie_add(info, cookie)) {
112 spin_unlock_irqrestore(&q->lock, flags);
113 kfree(rds_info_from_znotifier(znotif));
114 /* caller invokes rds_wake_sk_sleep() */
115 return;
116 }
117 }
118
119 info = rds_info_from_znotifier(znotif);
120 ck = &info->zcookies;
121 memset(ck, 0, sizeof(*ck));
122 WARN_ON(!rds_zcookie_add(info, cookie));
123 list_add_tail(&info->rs_zcookie_next, &q->zcookie_head);
124
125 spin_unlock_irqrestore(&q->lock, flags);
126 /* caller invokes rds_wake_sk_sleep() */
127}
128
129/*
130 * This relies on dma_map_sg() not touching sg[].page during merging.
131 */
132static void rds_message_purge(struct rds_message *rm)
133{
134 struct rds_znotifier *znotifier;
135 unsigned long i, flags;
136 bool zcopy;
137
138 if (unlikely(test_bit(RDS_MSG_PAGEVEC, &rm->m_flags)))
139 return;
140
141 spin_lock_irqsave(&rm->m_rs_lock, flags);
142 znotifier = rm->data.op_mmp_znotifier;
143 rm->data.op_mmp_znotifier = NULL;
144 zcopy = !!znotifier;
145
146 if (rm->m_rs) {
147 struct rds_sock *rs = rm->m_rs;
148
149 if (znotifier) {
150 rds_rm_zerocopy_callback(rs, znotifier);
151 rds_wake_sk_sleep(rs);
152 }
153 sock_put(rds_rs_to_sk(rs));
154 rm->m_rs = NULL;
155 } else if (znotifier) {
156 /*
157 * Zerocopy can fail before the message is queued on the
158 * socket, so there is no rs to carry the notification.
159 */
160 mm_unaccount_pinned_pages(&znotifier->z_mmp);
161 kfree(rds_info_from_znotifier(znotifier));
162 }
163 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
164
165 for (i = 0; i < rm->data.op_nents; i++) {
166 /* XXX will have to put_page for page refs */
167 if (!zcopy)
168 __free_page(sg_page(&rm->data.op_sg[i]));
169 else
170 put_page(sg_page(&rm->data.op_sg[i]));
171 }
172 rm->data.op_nents = 0;
173
174 if (rm->rdma.op_active)
175 rds_rdma_free_op(&rm->rdma);
176 if (rm->rdma.op_rdma_mr)
177 kref_put(&rm->rdma.op_rdma_mr->r_kref, __rds_put_mr_final);
178
179 if (rm->atomic.op_active)
180 rds_atomic_free_op(&rm->atomic);
181 if (rm->atomic.op_rdma_mr)
182 kref_put(&rm->atomic.op_rdma_mr->r_kref, __rds_put_mr_final);
183}
184
185static void rds_message_unpin_worker(struct work_struct *work)
186{
187 struct rds_message *rm = container_of(work, struct rds_message,
188 m_unpin_work);
189
190 if (rm->rdma.op_unpin_deferred)
191 rds_rdma_op_unpin_pages(&rm->rdma);
192 if (rm->atomic.op_unpin_deferred)
193 rds_atomic_op_unpin_page(&rm->atomic);
194
195 kfree(rm);
196}
197
198void rds_message_put(struct rds_message *rm)
199{
200 rdsdebug("put rm %p ref %d\n", rm, refcount_read(&rm->m_refcount));
201 WARN(!refcount_read(&rm->m_refcount), "danger refcount zero on %p\n", rm);
202 if (refcount_dec_and_test(&rm->m_refcount)) {
203 BUG_ON(!list_empty(&rm->m_sock_item));
204 BUG_ON(!list_empty(&rm->m_conn_item));
205
206 rds_message_purge(rm);
207
208 /* A final put in atomic context cannot dirty the ops'
209 * user pages on unpin, so rds_rdma_free_op() and
210 * rds_atomic_free_op() deferred it. Finish the unpin,
211 * and the free, from process context.
212 */
213 if (rm->rdma.op_unpin_deferred ||
214 rm->atomic.op_unpin_deferred) {
215 INIT_WORK(&rm->m_unpin_work, rds_message_unpin_worker);
216 queue_work(rds_wq, &rm->m_unpin_work);
217 return;
218 }
219
220 kfree(rm);
221 }
222}
223EXPORT_SYMBOL_GPL(rds_message_put);
224
225void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
226 __be16 dport, u64 seq)
227{
228 hdr->h_flags = 0;
229 hdr->h_sport = sport;
230 hdr->h_dport = dport;
231 hdr->h_sequence = cpu_to_be64(seq);
232 /* see rds_find_next_ext_space for reason why we memset the
233 * ext header
234 */
235 memset(hdr->h_exthdr, RDS_EXTHDR_NONE, RDS_HEADER_EXT_SPACE);
236}
237EXPORT_SYMBOL_GPL(rds_message_populate_header);
238
239/*
240 * Find the next place we can add an RDS header extension with
241 * specific length. Extension headers are pushed one after the
242 * other. In the following, the number after the colon is the number
243 * of bytes:
244 *
245 * [ type1:1 dta1:len1 [ type2:1 dta2:len2 ] ... ] RDS_EXTHDR_NONE
246 *
247 * If the extension headers fill the complete extension header space
248 * (16 bytes), the trailing RDS_EXTHDR_NONE is omitted.
249 */
250static int rds_find_next_ext_space(struct rds_header *hdr, unsigned int len,
251 u8 **ext_start)
252{
253 unsigned int ext_len;
254 unsigned int type;
255 int ind = 0;
256
257 while ((ind + 1 + len) <= RDS_HEADER_EXT_SPACE) {
258 if (hdr->h_exthdr[ind] == RDS_EXTHDR_NONE) {
259 *ext_start = hdr->h_exthdr + ind;
260 return 0;
261 }
262
263 type = hdr->h_exthdr[ind];
264
265 ext_len = (type < __RDS_EXTHDR_MAX) ? rds_exthdr_size[type] : 0;
266 WARN_ONCE(!ext_len, "Unknown ext hdr type %d\n", type);
267 if (!ext_len)
268 return -EINVAL;
269
270 /* ind points to a valid ext hdr with known length */
271 ind += 1 + ext_len;
272 }
273
274 /* no room for extension */
275 return -ENOSPC;
276}
277
278/* The ext hdr space is prefilled with zero from the kzalloc() */
279int rds_message_add_extension(struct rds_header *hdr,
280 unsigned int type, const void *data)
281{
282 unsigned char *dst;
283 unsigned int len;
284
285 len = (type < __RDS_EXTHDR_MAX) ? rds_exthdr_size[type] : 0;
286 if (!len)
287 return 0;
288
289 if (rds_find_next_ext_space(hdr, len, &dst))
290 return 0;
291
292 *dst++ = type;
293 memcpy(dst, data, len);
294
295 return 1;
296}
297EXPORT_SYMBOL_GPL(rds_message_add_extension);
298
299/*
300 * If a message has extension headers, retrieve them here.
301 * Call like this:
302 *
303 * unsigned int pos = 0;
304 *
305 * while (1) {
306 * buflen = sizeof(buffer);
307 * type = rds_message_next_extension(hdr, &pos, buffer, &buflen);
308 * if (type == RDS_EXTHDR_NONE)
309 * break;
310 * ...
311 * }
312 */
313int rds_message_next_extension(struct rds_header *hdr,
314 unsigned int *pos, void *buf, unsigned int *buflen)
315{
316 unsigned int offset, ext_type, ext_len;
317 u8 *src = hdr->h_exthdr;
318
319 offset = *pos;
320 if (offset >= RDS_HEADER_EXT_SPACE)
321 goto none;
322
323 /* Get the extension type and length. For now, the
324 * length is implied by the extension type. */
325 ext_type = src[offset++];
326
327 if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX)
328 goto none;
329 ext_len = rds_exthdr_size[ext_type];
330 if (offset + ext_len > RDS_HEADER_EXT_SPACE)
331 goto none;
332
333 *pos = offset + ext_len;
334 if (ext_len < *buflen)
335 *buflen = ext_len;
336 memcpy(buf, src + offset, *buflen);
337 return ext_type;
338
339none:
340 *pos = RDS_HEADER_EXT_SPACE;
341 *buflen = 0;
342 return RDS_EXTHDR_NONE;
343}
344
345int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset)
346{
347 struct rds_ext_header_rdma_dest ext_hdr;
348
349 ext_hdr.h_rdma_rkey = cpu_to_be32(r_key);
350 ext_hdr.h_rdma_offset = cpu_to_be32(offset);
351 return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr);
352}
353EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension);
354
355/*
356 * Each rds_message is allocated with extra space for the scatterlist entries
357 * rds ops will need. This is to minimize memory allocation count. Then, each rds op
358 * can grab SGs when initializing its part of the rds_message.
359 */
360struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
361{
362 struct rds_message *rm;
363
364 if (extra_len > KMALLOC_MAX_SIZE - sizeof(struct rds_message))
365 return NULL;
366
367 rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp);
368 if (!rm)
369 goto out;
370
371 rm->m_used_sgs = 0;
372 rm->m_total_sgs = extra_len / sizeof(struct scatterlist);
373
374 refcount_set(&rm->m_refcount, 1);
375 INIT_LIST_HEAD(&rm->m_sock_item);
376 INIT_LIST_HEAD(&rm->m_conn_item);
377 spin_lock_init(&rm->m_rs_lock);
378 init_waitqueue_head(&rm->m_flush_wait);
379
380out:
381 return rm;
382}
383
384/*
385 * RDS ops use this to grab SG entries from the rm's sg pool.
386 */
387struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents)
388{
389 struct scatterlist *sg_first = (struct scatterlist *) &rm[1];
390 struct scatterlist *sg_ret;
391
392 if (nents <= 0) {
393 pr_warn("rds: alloc sgs failed! nents <= 0\n");
394 return ERR_PTR(-EINVAL);
395 }
396
397 if (rm->m_used_sgs + nents > rm->m_total_sgs) {
398 pr_warn("rds: alloc sgs failed! total %d used %d nents %d\n",
399 rm->m_total_sgs, rm->m_used_sgs, nents);
400 return ERR_PTR(-ENOMEM);
401 }
402
403 sg_ret = &sg_first[rm->m_used_sgs];
404 sg_init_table(sg_ret, nents);
405 rm->m_used_sgs += nents;
406
407 return sg_ret;
408}
409
410struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
411{
412 struct rds_message *rm;
413 unsigned int i;
414 int num_sgs = DIV_ROUND_UP(total_len, PAGE_SIZE);
415 int extra_bytes = num_sgs * sizeof(struct scatterlist);
416
417 rm = rds_message_alloc(extra_bytes, GFP_NOWAIT);
418 if (!rm)
419 return ERR_PTR(-ENOMEM);
420
421 set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
422 rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
423 rm->data.op_nents = DIV_ROUND_UP(total_len, PAGE_SIZE);
424 rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
425 if (IS_ERR(rm->data.op_sg)) {
426 void *err = ERR_CAST(rm->data.op_sg);
427 rds_message_put(rm);
428 return err;
429 }
430
431 for (i = 0; i < rm->data.op_nents; ++i) {
432 sg_set_page(&rm->data.op_sg[i],
433 virt_to_page((void *)page_addrs[i]),
434 i == rm->data.op_nents - 1
435 ? total_len - (i * PAGE_SIZE)
436 : PAGE_SIZE, 0);
437 }
438
439 return rm;
440}
441
442static int rds_message_zcopy_from_user(struct rds_message *rm, struct iov_iter *from)
443{
444 struct scatterlist *sg;
445 int ret = 0;
446 int length = iov_iter_count(from);
447 struct rds_msg_zcopy_info *info;
448
449 rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from));
450
451 /*
452 * now allocate and copy in the data payload.
453 */
454 sg = rm->data.op_sg;
455
456 info = kzalloc_obj(*info);
457 if (!info)
458 return -ENOMEM;
459 INIT_LIST_HEAD(&info->rs_zcookie_next);
460 rm->data.op_mmp_znotifier = &info->znotif;
461 if (mm_account_pinned_pages(&rm->data.op_mmp_znotifier->z_mmp,
462 length)) {
463 ret = -ENOMEM;
464 goto err;
465 }
466 while (iov_iter_count(from)) {
467 struct page *pages;
468 size_t start;
469 ssize_t copied;
470
471 copied = iov_iter_get_pages2(from, &pages, PAGE_SIZE,
472 1, &start);
473 if (copied < 0) {
474 struct mmpin *mmp;
475 int i;
476
477 for (i = 0; i < rm->data.op_nents; i++)
478 put_page(sg_page(&rm->data.op_sg[i]));
479 rm->data.op_nents = 0;
480 mmp = &rm->data.op_mmp_znotifier->z_mmp;
481 mm_unaccount_pinned_pages(mmp);
482 ret = -EFAULT;
483 goto err;
484 }
485 length -= copied;
486 sg_set_page(sg, pages, copied, start);
487 rm->data.op_nents++;
488 sg++;
489 }
490 WARN_ON_ONCE(length != 0);
491 return ret;
492err:
493 kfree(info);
494 rm->data.op_mmp_znotifier = NULL;
495 return ret;
496}
497
498int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
499 bool zcopy)
500{
501 unsigned long to_copy, nbytes;
502 unsigned long sg_off;
503 struct scatterlist *sg;
504 int ret = 0;
505
506 rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from));
507
508 /* now allocate and copy in the data payload. */
509 sg = rm->data.op_sg;
510 sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
511
512 if (zcopy)
513 return rds_message_zcopy_from_user(rm, from);
514
515 while (iov_iter_count(from)) {
516 if (!sg_page(sg)) {
517 ret = rds_page_remainder_alloc(sg, iov_iter_count(from),
518 GFP_HIGHUSER);
519 if (ret)
520 return ret;
521 rm->data.op_nents++;
522 sg_off = 0;
523 }
524
525 to_copy = min_t(unsigned long, iov_iter_count(from),
526 sg->length - sg_off);
527
528 rds_stats_add(s_copy_from_user, to_copy);
529 nbytes = copy_page_from_iter(sg_page(sg), sg->offset + sg_off,
530 to_copy, from);
531 if (nbytes != to_copy)
532 return -EFAULT;
533
534 sg_off += to_copy;
535
536 if (sg_off == sg->length)
537 sg++;
538 }
539
540 return ret;
541}
542
543int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
544{
545 struct rds_message *rm;
546 struct scatterlist *sg;
547 unsigned long to_copy;
548 unsigned long vec_off;
549 int copied;
550 int ret;
551 u32 len;
552
553 rm = container_of(inc, struct rds_message, m_inc);
554 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
555
556 sg = rm->data.op_sg;
557 vec_off = 0;
558 copied = 0;
559
560 while (iov_iter_count(to) && copied < len) {
561 to_copy = min_t(unsigned long, iov_iter_count(to),
562 sg->length - vec_off);
563 to_copy = min_t(unsigned long, to_copy, len - copied);
564
565 rds_stats_add(s_copy_to_user, to_copy);
566 ret = copy_page_to_iter(sg_page(sg), sg->offset + vec_off,
567 to_copy, to);
568 if (ret != to_copy)
569 return -EFAULT;
570
571 vec_off += to_copy;
572 copied += to_copy;
573
574 if (vec_off == sg->length) {
575 vec_off = 0;
576 sg++;
577 }
578 }
579
580 return copied;
581}
582
583/*
584 * If the message is still on the send queue, wait until the transport
585 * is done with it. This is particularly important for RDMA operations.
586 */
587void rds_message_wait(struct rds_message *rm)
588{
589 wait_event_interruptible(rm->m_flush_wait,
590 !test_bit(RDS_MSG_MAPPED, &rm->m_flags));
591}
592
593void rds_message_unmapped(struct rds_message *rm)
594{
595 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
596 wake_up_interruptible(&rm->m_flush_wait);
597}
598EXPORT_SYMBOL_GPL(rds_message_unmapped);