Root/net/rds/rdma.c

1/*
2 * Copyright (c) 2007 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/pagemap.h>
34#include <linux/slab.h>
35#include <linux/rbtree.h>
36#include <linux/dma-mapping.h> /* for DMA_*_DEVICE */
37
38#include "rdma.h"
39
40/*
41 * XXX
42 * - build with sparse
43 * - should we limit the size of a mr region? let transport return failure?
44 * - should we detect duplicate keys on a socket? hmm.
45 * - an rdma is an mlock, apply rlimit?
46 */
47
48/*
49 * get the number of pages by looking at the page indices that the start and
50 * end addresses fall in.
51 *
52 * Returns 0 if the vec is invalid. It is invalid if the number of bytes
53 * causes the address to wrap or overflows an unsigned int. This comes
54 * from being stored in the 'length' member of 'struct scatterlist'.
55 */
56static unsigned int rds_pages_in_vec(struct rds_iovec *vec)
57{
58    if ((vec->addr + vec->bytes <= vec->addr) ||
59        (vec->bytes > (u64)UINT_MAX))
60        return 0;
61
62    return ((vec->addr + vec->bytes + PAGE_SIZE - 1) >> PAGE_SHIFT) -
63        (vec->addr >> PAGE_SHIFT);
64}
65
66static struct rds_mr *rds_mr_tree_walk(struct rb_root *root, u64 key,
67                       struct rds_mr *insert)
68{
69    struct rb_node **p = &root->rb_node;
70    struct rb_node *parent = NULL;
71    struct rds_mr *mr;
72
73    while (*p) {
74        parent = *p;
75        mr = rb_entry(parent, struct rds_mr, r_rb_node);
76
77        if (key < mr->r_key)
78            p = &(*p)->rb_left;
79        else if (key > mr->r_key)
80            p = &(*p)->rb_right;
81        else
82            return mr;
83    }
84
85    if (insert) {
86        rb_link_node(&insert->r_rb_node, parent, p);
87        rb_insert_color(&insert->r_rb_node, root);
88        atomic_inc(&insert->r_refcount);
89    }
90    return NULL;
91}
92
93/*
94 * Destroy the transport-specific part of a MR.
95 */
96static void rds_destroy_mr(struct rds_mr *mr)
97{
98    struct rds_sock *rs = mr->r_sock;
99    void *trans_private = NULL;
100    unsigned long flags;
101
102    rdsdebug("RDS: destroy mr key is %x refcnt %u\n",
103            mr->r_key, atomic_read(&mr->r_refcount));
104
105    if (test_and_set_bit(RDS_MR_DEAD, &mr->r_state))
106        return;
107
108    spin_lock_irqsave(&rs->rs_rdma_lock, flags);
109    if (!RB_EMPTY_NODE(&mr->r_rb_node))
110        rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
111    trans_private = mr->r_trans_private;
112    mr->r_trans_private = NULL;
113    spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
114
115    if (trans_private)
116        mr->r_trans->free_mr(trans_private, mr->r_invalidate);
117}
118
119void __rds_put_mr_final(struct rds_mr *mr)
120{
121    rds_destroy_mr(mr);
122    kfree(mr);
123}
124
125/*
126 * By the time this is called we can't have any more ioctls called on
127 * the socket so we don't need to worry about racing with others.
128 */
129void rds_rdma_drop_keys(struct rds_sock *rs)
130{
131    struct rds_mr *mr;
132    struct rb_node *node;
133
134    /* Release any MRs associated with this socket */
135    while ((node = rb_first(&rs->rs_rdma_keys))) {
136        mr = container_of(node, struct rds_mr, r_rb_node);
137        if (mr->r_trans == rs->rs_transport)
138            mr->r_invalidate = 0;
139        rds_mr_put(mr);
140    }
141
142    if (rs->rs_transport && rs->rs_transport->flush_mrs)
143        rs->rs_transport->flush_mrs();
144}
145
146/*
147 * Helper function to pin user pages.
148 */
149static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
150            struct page **pages, int write)
151{
152    int ret;
153
154    ret = get_user_pages_fast(user_addr, nr_pages, write, pages);
155
156    if (ret >= 0 && ret < nr_pages) {
157        while (ret--)
158            put_page(pages[ret]);
159        ret = -EFAULT;
160    }
161
162    return ret;
163}
164
165static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
166                u64 *cookie_ret, struct rds_mr **mr_ret)
167{
168    struct rds_mr *mr = NULL, *found;
169    unsigned int nr_pages;
170    struct page **pages = NULL;
171    struct scatterlist *sg;
172    void *trans_private;
173    unsigned long flags;
174    rds_rdma_cookie_t cookie;
175    unsigned int nents;
176    long i;
177    int ret;
178
179    if (rs->rs_bound_addr == 0) {
180        ret = -ENOTCONN; /* XXX not a great errno */
181        goto out;
182    }
183
184    if (rs->rs_transport->get_mr == NULL) {
185        ret = -EOPNOTSUPP;
186        goto out;
187    }
188
189    nr_pages = rds_pages_in_vec(&args->vec);
190    if (nr_pages == 0) {
191        ret = -EINVAL;
192        goto out;
193    }
194
195    rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
196        args->vec.addr, args->vec.bytes, nr_pages);
197
198    /* XXX clamp nr_pages to limit the size of this alloc? */
199    pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
200    if (pages == NULL) {
201        ret = -ENOMEM;
202        goto out;
203    }
204
205    mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
206    if (mr == NULL) {
207        ret = -ENOMEM;
208        goto out;
209    }
210
211    atomic_set(&mr->r_refcount, 1);
212    RB_CLEAR_NODE(&mr->r_rb_node);
213    mr->r_trans = rs->rs_transport;
214    mr->r_sock = rs;
215
216    if (args->flags & RDS_RDMA_USE_ONCE)
217        mr->r_use_once = 1;
218    if (args->flags & RDS_RDMA_INVALIDATE)
219        mr->r_invalidate = 1;
220    if (args->flags & RDS_RDMA_READWRITE)
221        mr->r_write = 1;
222
223    /*
224     * Pin the pages that make up the user buffer and transfer the page
225     * pointers to the mr's sg array. We check to see if we've mapped
226     * the whole region after transferring the partial page references
227     * to the sg array so that we can have one page ref cleanup path.
228     *
229     * For now we have no flag that tells us whether the mapping is
230     * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
231     * the zero page.
232     */
233    ret = rds_pin_pages(args->vec.addr & PAGE_MASK, nr_pages, pages, 1);
234    if (ret < 0)
235        goto out;
236
237    nents = ret;
238    sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
239    if (sg == NULL) {
240        ret = -ENOMEM;
241        goto out;
242    }
243    WARN_ON(!nents);
244    sg_init_table(sg, nents);
245
246    /* Stick all pages into the scatterlist */
247    for (i = 0 ; i < nents; i++)
248        sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
249
250    rdsdebug("RDS: trans_private nents is %u\n", nents);
251
252    /* Obtain a transport specific MR. If this succeeds, the
253     * s/g list is now owned by the MR.
254     * Note that dma_map() implies that pending writes are
255     * flushed to RAM, so no dma_sync is needed here. */
256    trans_private = rs->rs_transport->get_mr(sg, nents, rs,
257                         &mr->r_key);
258
259    if (IS_ERR(trans_private)) {
260        for (i = 0 ; i < nents; i++)
261            put_page(sg_page(&sg[i]));
262        kfree(sg);
263        ret = PTR_ERR(trans_private);
264        goto out;
265    }
266
267    mr->r_trans_private = trans_private;
268
269    rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
270           mr->r_key, (void *)(unsigned long) args->cookie_addr);
271
272    /* The user may pass us an unaligned address, but we can only
273     * map page aligned regions. So we keep the offset, and build
274     * a 64bit cookie containing <R_Key, offset> and pass that
275     * around. */
276    cookie = rds_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGE_MASK);
277    if (cookie_ret)
278        *cookie_ret = cookie;
279
280    if (args->cookie_addr && put_user(cookie, (u64 __user *)(unsigned long) args->cookie_addr)) {
281        ret = -EFAULT;
282        goto out;
283    }
284
285    /* Inserting the new MR into the rbtree bumps its
286     * reference count. */
287    spin_lock_irqsave(&rs->rs_rdma_lock, flags);
288    found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
289    spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
290
291    BUG_ON(found && found != mr);
292
293    rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
294    if (mr_ret) {
295        atomic_inc(&mr->r_refcount);
296        *mr_ret = mr;
297    }
298
299    ret = 0;
300out:
301    kfree(pages);
302    if (mr)
303        rds_mr_put(mr);
304    return ret;
305}
306
307int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen)
308{
309    struct rds_get_mr_args args;
310
311    if (optlen != sizeof(struct rds_get_mr_args))
312        return -EINVAL;
313
314    if (copy_from_user(&args, (struct rds_get_mr_args __user *)optval,
315               sizeof(struct rds_get_mr_args)))
316        return -EFAULT;
317
318    return __rds_rdma_map(rs, &args, NULL, NULL);
319}
320
321int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen)
322{
323    struct rds_get_mr_for_dest_args args;
324    struct rds_get_mr_args new_args;
325
326    if (optlen != sizeof(struct rds_get_mr_for_dest_args))
327        return -EINVAL;
328
329    if (copy_from_user(&args, (struct rds_get_mr_for_dest_args __user *)optval,
330               sizeof(struct rds_get_mr_for_dest_args)))
331        return -EFAULT;
332
333    /*
334     * Initially, just behave like get_mr().
335     * TODO: Implement get_mr as wrapper around this
336     * and deprecate it.
337     */
338    new_args.vec = args.vec;
339    new_args.cookie_addr = args.cookie_addr;
340    new_args.flags = args.flags;
341
342    return __rds_rdma_map(rs, &new_args, NULL, NULL);
343}
344
345/*
346 * Free the MR indicated by the given R_Key
347 */
348int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen)
349{
350    struct rds_free_mr_args args;
351    struct rds_mr *mr;
352    unsigned long flags;
353
354    if (optlen != sizeof(struct rds_free_mr_args))
355        return -EINVAL;
356
357    if (copy_from_user(&args, (struct rds_free_mr_args __user *)optval,
358               sizeof(struct rds_free_mr_args)))
359        return -EFAULT;
360
361    /* Special case - a null cookie means flush all unused MRs */
362    if (args.cookie == 0) {
363        if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
364            return -EINVAL;
365        rs->rs_transport->flush_mrs();
366        return 0;
367    }
368
369    /* Look up the MR given its R_key and remove it from the rbtree
370     * so nobody else finds it.
371     * This should also prevent races with rds_rdma_unuse.
372     */
373    spin_lock_irqsave(&rs->rs_rdma_lock, flags);
374    mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL);
375    if (mr) {
376        rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
377        RB_CLEAR_NODE(&mr->r_rb_node);
378        if (args.flags & RDS_RDMA_INVALIDATE)
379            mr->r_invalidate = 1;
380    }
381    spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
382
383    if (!mr)
384        return -EINVAL;
385
386    /*
387     * call rds_destroy_mr() ourselves so that we're sure it's done by the time
388     * we return. If we let rds_mr_put() do it it might not happen until
389     * someone else drops their ref.
390     */
391    rds_destroy_mr(mr);
392    rds_mr_put(mr);
393    return 0;
394}
395
396/*
397 * This is called when we receive an extension header that
398 * tells us this MR was used. It allows us to implement
399 * use_once semantics
400 */
401void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
402{
403    struct rds_mr *mr;
404    unsigned long flags;
405    int zot_me = 0;
406
407    spin_lock_irqsave(&rs->rs_rdma_lock, flags);
408    mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
409    if (mr && (mr->r_use_once || force)) {
410        rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
411        RB_CLEAR_NODE(&mr->r_rb_node);
412        zot_me = 1;
413    } else if (mr)
414        atomic_inc(&mr->r_refcount);
415    spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
416
417    /* May have to issue a dma_sync on this memory region.
418     * Note we could avoid this if the operation was a RDMA READ,
419     * but at this point we can't tell. */
420    if (mr != NULL) {
421        if (mr->r_trans->sync_mr)
422            mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
423
424        /* If the MR was marked as invalidate, this will
425         * trigger an async flush. */
426        if (zot_me)
427            rds_destroy_mr(mr);
428        rds_mr_put(mr);
429    }
430}
431
432void rds_rdma_free_op(struct rds_rdma_op *ro)
433{
434    unsigned int i;
435
436    for (i = 0; i < ro->r_nents; i++) {
437        struct page *page = sg_page(&ro->r_sg[i]);
438
439        /* Mark page dirty if it was possibly modified, which
440         * is the case for a RDMA_READ which copies from remote
441         * to local memory */
442        if (!ro->r_write)
443            set_page_dirty(page);
444        put_page(page);
445    }
446
447    kfree(ro->r_notifier);
448    kfree(ro);
449}
450
451/*
452 * args is a pointer to an in-kernel copy in the sendmsg cmsg.
453 */
454static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
455                        struct rds_rdma_args *args)
456{
457    struct rds_iovec vec;
458    struct rds_rdma_op *op = NULL;
459    unsigned int nr_pages;
460    unsigned int max_pages;
461    unsigned int nr_bytes;
462    struct page **pages = NULL;
463    struct rds_iovec __user *local_vec;
464    struct scatterlist *sg;
465    unsigned int nr;
466    unsigned int i, j;
467    int ret;
468
469
470    if (rs->rs_bound_addr == 0) {
471        ret = -ENOTCONN; /* XXX not a great errno */
472        goto out;
473    }
474
475    if (args->nr_local > (u64)UINT_MAX) {
476        ret = -EMSGSIZE;
477        goto out;
478    }
479
480    nr_pages = 0;
481    max_pages = 0;
482
483    local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
484
485    /* figure out the number of pages in the vector */
486    for (i = 0; i < args->nr_local; i++) {
487        if (copy_from_user(&vec, &local_vec[i],
488                   sizeof(struct rds_iovec))) {
489            ret = -EFAULT;
490            goto out;
491        }
492
493        nr = rds_pages_in_vec(&vec);
494        if (nr == 0) {
495            ret = -EINVAL;
496            goto out;
497        }
498
499        max_pages = max(nr, max_pages);
500        nr_pages += nr;
501    }
502
503    pages = kcalloc(max_pages, sizeof(struct page *), GFP_KERNEL);
504    if (pages == NULL) {
505        ret = -ENOMEM;
506        goto out;
507    }
508
509    op = kzalloc(offsetof(struct rds_rdma_op, r_sg[nr_pages]), GFP_KERNEL);
510    if (op == NULL) {
511        ret = -ENOMEM;
512        goto out;
513    }
514
515    op->r_write = !!(args->flags & RDS_RDMA_READWRITE);
516    op->r_fence = !!(args->flags & RDS_RDMA_FENCE);
517    op->r_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
518    op->r_recverr = rs->rs_recverr;
519    WARN_ON(!nr_pages);
520    sg_init_table(op->r_sg, nr_pages);
521
522    if (op->r_notify || op->r_recverr) {
523        /* We allocate an uninitialized notifier here, because
524         * we don't want to do that in the completion handler. We
525         * would have to use GFP_ATOMIC there, and don't want to deal
526         * with failed allocations.
527         */
528        op->r_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
529        if (!op->r_notifier) {
530            ret = -ENOMEM;
531            goto out;
532        }
533        op->r_notifier->n_user_token = args->user_token;
534        op->r_notifier->n_status = RDS_RDMA_SUCCESS;
535    }
536
537    /* The cookie contains the R_Key of the remote memory region, and
538     * optionally an offset into it. This is how we implement RDMA into
539     * unaligned memory.
540     * When setting up the RDMA, we need to add that offset to the
541     * destination address (which is really an offset into the MR)
542     * FIXME: We may want to move this into ib_rdma.c
543     */
544    op->r_key = rds_rdma_cookie_key(args->cookie);
545    op->r_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
546
547    nr_bytes = 0;
548
549    rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
550           (unsigned long long)args->nr_local,
551           (unsigned long long)args->remote_vec.addr,
552           op->r_key);
553
554    for (i = 0; i < args->nr_local; i++) {
555        if (copy_from_user(&vec, &local_vec[i],
556                   sizeof(struct rds_iovec))) {
557            ret = -EFAULT;
558            goto out;
559        }
560
561        nr = rds_pages_in_vec(&vec);
562        if (nr == 0) {
563            ret = -EINVAL;
564            goto out;
565        }
566
567        rs->rs_user_addr = vec.addr;
568        rs->rs_user_bytes = vec.bytes;
569
570        /* did the user change the vec under us? */
571        if (nr > max_pages || op->r_nents + nr > nr_pages) {
572            ret = -EINVAL;
573            goto out;
574        }
575        /* If it's a WRITE operation, we want to pin the pages for reading.
576         * If it's a READ operation, we need to pin the pages for writing.
577         */
578        ret = rds_pin_pages(vec.addr & PAGE_MASK, nr, pages, !op->r_write);
579        if (ret < 0)
580            goto out;
581
582        rdsdebug("RDS: nr_bytes %u nr %u vec.bytes %llu vec.addr %llx\n",
583               nr_bytes, nr, vec.bytes, vec.addr);
584
585        nr_bytes += vec.bytes;
586
587        for (j = 0; j < nr; j++) {
588            unsigned int offset = vec.addr & ~PAGE_MASK;
589
590            sg = &op->r_sg[op->r_nents + j];
591            sg_set_page(sg, pages[j],
592                    min_t(unsigned int, vec.bytes, PAGE_SIZE - offset),
593                    offset);
594
595            rdsdebug("RDS: sg->offset %x sg->len %x vec.addr %llx vec.bytes %llu\n",
596                   sg->offset, sg->length, vec.addr, vec.bytes);
597
598            vec.addr += sg->length;
599            vec.bytes -= sg->length;
600        }
601
602        op->r_nents += nr;
603    }
604
605
606    if (nr_bytes > args->remote_vec.bytes) {
607        rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
608                nr_bytes,
609                (unsigned int) args->remote_vec.bytes);
610        ret = -EINVAL;
611        goto out;
612    }
613    op->r_bytes = nr_bytes;
614
615    ret = 0;
616out:
617    kfree(pages);
618    if (ret) {
619        if (op)
620            rds_rdma_free_op(op);
621        op = ERR_PTR(ret);
622    }
623    return op;
624}
625
626/*
627 * The application asks for a RDMA transfer.
628 * Extract all arguments and set up the rdma_op
629 */
630int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
631              struct cmsghdr *cmsg)
632{
633    struct rds_rdma_op *op;
634
635    if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args)) ||
636        rm->m_rdma_op != NULL)
637        return -EINVAL;
638
639    op = rds_rdma_prepare(rs, CMSG_DATA(cmsg));
640    if (IS_ERR(op))
641        return PTR_ERR(op);
642    rds_stats_inc(s_send_rdma);
643    rm->m_rdma_op = op;
644    return 0;
645}
646
647/*
648 * The application wants us to pass an RDMA destination (aka MR)
649 * to the remote
650 */
651int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
652              struct cmsghdr *cmsg)
653{
654    unsigned long flags;
655    struct rds_mr *mr;
656    u32 r_key;
657    int err = 0;
658
659    if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
660        rm->m_rdma_cookie != 0)
661        return -EINVAL;
662
663    memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
664
665    /* We are reusing a previously mapped MR here. Most likely, the
666     * application has written to the buffer, so we need to explicitly
667     * flush those writes to RAM. Otherwise the HCA may not see them
668     * when doing a DMA from that buffer.
669     */
670    r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
671
672    spin_lock_irqsave(&rs->rs_rdma_lock, flags);
673    mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
674    if (mr == NULL)
675        err = -EINVAL; /* invalid r_key */
676    else
677        atomic_inc(&mr->r_refcount);
678    spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
679
680    if (mr) {
681        mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
682        rm->m_rdma_mr = mr;
683    }
684    return err;
685}
686
687/*
688 * The application passes us an address range it wants to enable RDMA
689 * to/from. We map the area, and save the <R_Key,offset> pair
690 * in rm->m_rdma_cookie. This causes it to be sent along to the peer
691 * in an extension header.
692 */
693int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
694              struct cmsghdr *cmsg)
695{
696    if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
697        rm->m_rdma_cookie != 0)
698        return -EINVAL;
699
700    return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->m_rdma_mr);
701}
702

Archive Download this file



interactive