2 * Copyright (c) 2007, 2017 Oracle and/or its affiliates. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
33 #include <linux/pagemap.h>
34 #include <linux/slab.h>
35 #include <linux/rbtree.h>
36 #include <linux/dma-mapping.h> /* for DMA_*_DEVICE */
43 * - should we detect duplicate keys on a socket? hmm.
44 * - an rdma is an mlock, apply rlimit?
48 * get the number of pages by looking at the page indices that the start and
49 * end addresses fall in.
51 * Returns 0 if the vec is invalid. It is invalid if the number of bytes
52 * causes the address to wrap or overflows an unsigned int. This comes
53 * from being stored in the 'length' member of 'struct scatterlist'.
55 static unsigned int rds_pages_in_vec(struct rds_iovec *vec)
57 if ((vec->addr + vec->bytes <= vec->addr) ||
58 (vec->bytes > (u64)UINT_MAX))
61 return ((vec->addr + vec->bytes + PAGE_SIZE - 1) >> PAGE_SHIFT) -
62 (vec->addr >> PAGE_SHIFT);
65 static struct rds_mr *rds_mr_tree_walk(struct rb_root *root, u64 key,
66 struct rds_mr *insert)
68 struct rb_node **p = &root->rb_node;
69 struct rb_node *parent = NULL;
74 mr = rb_entry(parent, struct rds_mr, r_rb_node);
78 else if (key > mr->r_key)
85 rb_link_node(&insert->r_rb_node, parent, p);
86 rb_insert_color(&insert->r_rb_node, root);
87 refcount_inc(&insert->r_refcount);
93 * Destroy the transport-specific part of a MR.
95 static void rds_destroy_mr(struct rds_mr *mr)
97 struct rds_sock *rs = mr->r_sock;
98 void *trans_private = NULL;
101 rdsdebug("RDS: destroy mr key is %x refcnt %u\n",
102 mr->r_key, refcount_read(&mr->r_refcount));
104 if (test_and_set_bit(RDS_MR_DEAD, &mr->r_state))
107 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
108 if (!RB_EMPTY_NODE(&mr->r_rb_node))
109 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
110 trans_private = mr->r_trans_private;
111 mr->r_trans_private = NULL;
112 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
115 mr->r_trans->free_mr(trans_private, mr->r_invalidate);
118 void __rds_put_mr_final(struct rds_mr *mr)
125 * By the time this is called we can't have any more ioctls called on
126 * the socket so we don't need to worry about racing with others.
128 void rds_rdma_drop_keys(struct rds_sock *rs)
131 struct rb_node *node;
134 /* Release any MRs associated with this socket */
135 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
136 while ((node = rb_first(&rs->rs_rdma_keys))) {
137 mr = rb_entry(node, struct rds_mr, r_rb_node);
138 if (mr->r_trans == rs->rs_transport)
139 mr->r_invalidate = 0;
140 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
141 RB_CLEAR_NODE(&mr->r_rb_node);
142 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
145 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
147 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
149 if (rs->rs_transport && rs->rs_transport->flush_mrs)
150 rs->rs_transport->flush_mrs();
154 * Helper function to pin user pages.
156 static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
157 struct page **pages, int write)
161 ret = get_user_pages_fast(user_addr, nr_pages, write, pages);
163 if (ret >= 0 && ret < nr_pages) {
165 put_page(pages[ret]);
172 static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
173 u64 *cookie_ret, struct rds_mr **mr_ret,
174 struct rds_conn_path *cp)
176 struct rds_mr *mr = NULL, *found;
177 unsigned int nr_pages;
178 struct page **pages = NULL;
179 struct scatterlist *sg;
182 rds_rdma_cookie_t cookie;
187 if (ipv6_addr_any(&rs->rs_bound_addr) || !rs->rs_transport) {
188 ret = -ENOTCONN; /* XXX not a great errno */
192 if (!rs->rs_transport->get_mr) {
197 nr_pages = rds_pages_in_vec(&args->vec);
203 /* Restrict the size of mr irrespective of underlying transport
204 * To account for unaligned mr regions, subtract one from nr_pages
206 if ((nr_pages - 1) > (RDS_MAX_MSG_SIZE >> PAGE_SHIFT)) {
211 rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
212 args->vec.addr, args->vec.bytes, nr_pages);
214 /* XXX clamp nr_pages to limit the size of this alloc? */
215 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
221 mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
227 refcount_set(&mr->r_refcount, 1);
228 RB_CLEAR_NODE(&mr->r_rb_node);
229 mr->r_trans = rs->rs_transport;
232 if (args->flags & RDS_RDMA_USE_ONCE)
234 if (args->flags & RDS_RDMA_INVALIDATE)
235 mr->r_invalidate = 1;
236 if (args->flags & RDS_RDMA_READWRITE)
240 * Pin the pages that make up the user buffer and transfer the page
241 * pointers to the mr's sg array. We check to see if we've mapped
242 * the whole region after transferring the partial page references
243 * to the sg array so that we can have one page ref cleanup path.
245 * For now we have no flag that tells us whether the mapping is
246 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
249 ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
254 sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
260 sg_init_table(sg, nents);
262 /* Stick all pages into the scatterlist */
263 for (i = 0 ; i < nents; i++)
264 sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
266 rdsdebug("RDS: trans_private nents is %u\n", nents);
268 /* Obtain a transport specific MR. If this succeeds, the
269 * s/g list is now owned by the MR.
270 * Note that dma_map() implies that pending writes are
271 * flushed to RAM, so no dma_sync is needed here. */
272 trans_private = rs->rs_transport->get_mr(sg, nents, rs,
274 cp ? cp->cp_conn : NULL);
276 if (IS_ERR(trans_private)) {
277 for (i = 0 ; i < nents; i++)
278 put_page(sg_page(&sg[i]));
280 ret = PTR_ERR(trans_private);
284 mr->r_trans_private = trans_private;
286 rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
287 mr->r_key, (void *)(unsigned long) args->cookie_addr);
289 /* The user may pass us an unaligned address, but we can only
290 * map page aligned regions. So we keep the offset, and build
291 * a 64bit cookie containing <R_Key, offset> and pass that
293 cookie = rds_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGE_MASK);
295 *cookie_ret = cookie;
297 if (args->cookie_addr && put_user(cookie, (u64 __user *)(unsigned long) args->cookie_addr)) {
302 /* Inserting the new MR into the rbtree bumps its
303 * reference count. */
304 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
305 found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
306 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
308 BUG_ON(found && found != mr);
310 rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
312 refcount_inc(&mr->r_refcount);
324 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen)
326 struct rds_get_mr_args args;
328 if (optlen != sizeof(struct rds_get_mr_args))
331 if (copy_from_user(&args, (struct rds_get_mr_args __user *)optval,
332 sizeof(struct rds_get_mr_args)))
335 return __rds_rdma_map(rs, &args, NULL, NULL, NULL);
338 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen)
340 struct rds_get_mr_for_dest_args args;
341 struct rds_get_mr_args new_args;
343 if (optlen != sizeof(struct rds_get_mr_for_dest_args))
346 if (copy_from_user(&args, (struct rds_get_mr_for_dest_args __user *)optval,
347 sizeof(struct rds_get_mr_for_dest_args)))
351 * Initially, just behave like get_mr().
352 * TODO: Implement get_mr as wrapper around this
355 new_args.vec = args.vec;
356 new_args.cookie_addr = args.cookie_addr;
357 new_args.flags = args.flags;
359 return __rds_rdma_map(rs, &new_args, NULL, NULL, NULL);
363 * Free the MR indicated by the given R_Key
365 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen)
367 struct rds_free_mr_args args;
371 if (optlen != sizeof(struct rds_free_mr_args))
374 if (copy_from_user(&args, (struct rds_free_mr_args __user *)optval,
375 sizeof(struct rds_free_mr_args)))
378 /* Special case - a null cookie means flush all unused MRs */
379 if (args.cookie == 0) {
380 if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
382 rs->rs_transport->flush_mrs();
386 /* Look up the MR given its R_key and remove it from the rbtree
387 * so nobody else finds it.
388 * This should also prevent races with rds_rdma_unuse.
390 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
391 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL);
393 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
394 RB_CLEAR_NODE(&mr->r_rb_node);
395 if (args.flags & RDS_RDMA_INVALIDATE)
396 mr->r_invalidate = 1;
398 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
404 * call rds_destroy_mr() ourselves so that we're sure it's done by the time
405 * we return. If we let rds_mr_put() do it it might not happen until
406 * someone else drops their ref.
414 * This is called when we receive an extension header that
415 * tells us this MR was used. It allows us to implement
418 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
424 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
425 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
427 pr_debug("rds: trying to unuse MR with unknown r_key %u!\n",
429 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
433 if (mr->r_use_once || force) {
434 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
435 RB_CLEAR_NODE(&mr->r_rb_node);
438 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
440 /* May have to issue a dma_sync on this memory region.
441 * Note we could avoid this if the operation was a RDMA READ,
442 * but at this point we can't tell. */
443 if (mr->r_trans->sync_mr)
444 mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
446 /* If the MR was marked as invalidate, this will
447 * trigger an async flush. */
454 void rds_rdma_free_op(struct rm_rdma_op *ro)
458 for (i = 0; i < ro->op_nents; i++) {
459 struct page *page = sg_page(&ro->op_sg[i]);
461 /* Mark page dirty if it was possibly modified, which
462 * is the case for a RDMA_READ which copies from remote
465 WARN_ON(!page->mapping && irqs_disabled());
466 set_page_dirty(page);
471 kfree(ro->op_notifier);
472 ro->op_notifier = NULL;
476 void rds_atomic_free_op(struct rm_atomic_op *ao)
478 struct page *page = sg_page(ao->op_sg);
480 /* Mark page dirty if it was possibly modified, which
481 * is the case for a RDMA_READ which copies from remote
483 set_page_dirty(page);
486 kfree(ao->op_notifier);
487 ao->op_notifier = NULL;
493 * Count the number of pages needed to describe an incoming iovec array.
495 static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs)
498 unsigned int nr_pages;
501 /* figure out the number of pages in the vector */
502 for (i = 0; i < nr_iovecs; i++) {
503 nr_pages = rds_pages_in_vec(&iov[i]);
507 tot_pages += nr_pages;
510 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
511 * so tot_pages cannot overflow without first going negative.
520 int rds_rdma_extra_size(struct rds_rdma_args *args,
521 struct rds_iov_vector *iov)
523 struct rds_iovec *vec;
524 struct rds_iovec __user *local_vec;
526 unsigned int nr_pages;
529 local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
531 if (args->nr_local == 0)
534 if (args->nr_local > UIO_MAXIOV)
537 iov->iov = kcalloc(args->nr_local,
538 sizeof(struct rds_iovec),
545 if (copy_from_user(vec, local_vec, args->nr_local *
546 sizeof(struct rds_iovec)))
548 iov->len = args->nr_local;
550 /* figure out the number of pages in the vector */
551 for (i = 0; i < args->nr_local; i++, vec++) {
553 nr_pages = rds_pages_in_vec(vec);
557 tot_pages += nr_pages;
560 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
561 * so tot_pages cannot overflow without first going negative.
567 return tot_pages * sizeof(struct scatterlist);
571 * The application asks for a RDMA transfer.
572 * Extract all arguments and set up the rdma_op
574 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
575 struct cmsghdr *cmsg,
576 struct rds_iov_vector *vec)
578 struct rds_rdma_args *args;
579 struct rm_rdma_op *op = &rm->rdma;
581 unsigned int nr_bytes;
582 struct page **pages = NULL;
583 struct rds_iovec *iovs;
587 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
588 || rm->rdma.op_active)
591 args = CMSG_DATA(cmsg);
593 if (ipv6_addr_any(&rs->rs_bound_addr)) {
594 ret = -ENOTCONN; /* XXX not a great errno */
598 if (args->nr_local > UIO_MAXIOV) {
603 if (vec->len != args->nr_local) {
610 nr_pages = rds_rdma_pages(iovs, args->nr_local);
616 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
622 op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
623 op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
624 op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
625 op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
627 op->op_recverr = rs->rs_recverr;
629 op->op_sg = rds_message_alloc_sgs(rm, nr_pages);
635 if (op->op_notify || op->op_recverr) {
636 /* We allocate an uninitialized notifier here, because
637 * we don't want to do that in the completion handler. We
638 * would have to use GFP_ATOMIC there, and don't want to deal
639 * with failed allocations.
641 op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
642 if (!op->op_notifier) {
646 op->op_notifier->n_user_token = args->user_token;
647 op->op_notifier->n_status = RDS_RDMA_SUCCESS;
649 /* Enable rmda notification on data operation for composite
650 * rds messages and make sure notification is enabled only
651 * for the data operation which follows it so that application
652 * gets notified only after full message gets delivered.
654 if (rm->data.op_sg) {
655 rm->rdma.op_notify = 0;
656 rm->data.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
660 /* The cookie contains the R_Key of the remote memory region, and
661 * optionally an offset into it. This is how we implement RDMA into
663 * When setting up the RDMA, we need to add that offset to the
664 * destination address (which is really an offset into the MR)
665 * FIXME: We may want to move this into ib_rdma.c
667 op->op_rkey = rds_rdma_cookie_key(args->cookie);
668 op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
672 rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
673 (unsigned long long)args->nr_local,
674 (unsigned long long)args->remote_vec.addr,
677 for (i = 0; i < args->nr_local; i++) {
678 struct rds_iovec *iov = &iovs[i];
679 /* don't need to check, rds_rdma_pages() verified nr will be +nonzero */
680 unsigned int nr = rds_pages_in_vec(iov);
682 rs->rs_user_addr = iov->addr;
683 rs->rs_user_bytes = iov->bytes;
685 /* If it's a WRITE operation, we want to pin the pages for reading.
686 * If it's a READ operation, we need to pin the pages for writing.
688 ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
694 rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
695 nr_bytes, nr, iov->bytes, iov->addr);
697 nr_bytes += iov->bytes;
699 for (j = 0; j < nr; j++) {
700 unsigned int offset = iov->addr & ~PAGE_MASK;
701 struct scatterlist *sg;
703 sg = &op->op_sg[op->op_nents + j];
704 sg_set_page(sg, pages[j],
705 min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
708 rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n",
709 sg->offset, sg->length, iov->addr, iov->bytes);
711 iov->addr += sg->length;
712 iov->bytes -= sg->length;
718 if (nr_bytes > args->remote_vec.bytes) {
719 rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
721 (unsigned int) args->remote_vec.bytes);
725 op->op_bytes = nr_bytes;
731 rds_rdma_free_op(op);
733 rds_stats_inc(s_send_rdma);
739 * The application wants us to pass an RDMA destination (aka MR)
742 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
743 struct cmsghdr *cmsg)
750 if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
751 rm->m_rdma_cookie != 0)
754 memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
756 /* We are reusing a previously mapped MR here. Most likely, the
757 * application has written to the buffer, so we need to explicitly
758 * flush those writes to RAM. Otherwise the HCA may not see them
759 * when doing a DMA from that buffer.
761 r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
763 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
764 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
766 err = -EINVAL; /* invalid r_key */
768 refcount_inc(&mr->r_refcount);
769 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
772 mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
773 rm->rdma.op_rdma_mr = mr;
779 * The application passes us an address range it wants to enable RDMA
780 * to/from. We map the area, and save the <R_Key,offset> pair
781 * in rm->m_rdma_cookie. This causes it to be sent along to the peer
782 * in an extension header.
784 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
785 struct cmsghdr *cmsg)
787 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
788 rm->m_rdma_cookie != 0)
791 return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie,
792 &rm->rdma.op_rdma_mr, rm->m_conn_path);
796 * Fill in rds_message for an atomic request.
798 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
799 struct cmsghdr *cmsg)
801 struct page *page = NULL;
802 struct rds_atomic_args *args;
805 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
806 || rm->atomic.op_active)
809 args = CMSG_DATA(cmsg);
811 /* Nonmasked & masked cmsg ops converted to masked hw ops */
812 switch (cmsg->cmsg_type) {
813 case RDS_CMSG_ATOMIC_FADD:
814 rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
815 rm->atomic.op_m_fadd.add = args->fadd.add;
816 rm->atomic.op_m_fadd.nocarry_mask = 0;
818 case RDS_CMSG_MASKED_ATOMIC_FADD:
819 rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
820 rm->atomic.op_m_fadd.add = args->m_fadd.add;
821 rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
823 case RDS_CMSG_ATOMIC_CSWP:
824 rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
825 rm->atomic.op_m_cswp.compare = args->cswp.compare;
826 rm->atomic.op_m_cswp.swap = args->cswp.swap;
827 rm->atomic.op_m_cswp.compare_mask = ~0;
828 rm->atomic.op_m_cswp.swap_mask = ~0;
830 case RDS_CMSG_MASKED_ATOMIC_CSWP:
831 rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
832 rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
833 rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
834 rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
835 rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
838 BUG(); /* should never happen */
841 rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
842 rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
843 rm->atomic.op_active = 1;
844 rm->atomic.op_recverr = rs->rs_recverr;
845 rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1);
846 if (!rm->atomic.op_sg) {
851 /* verify 8 byte-aligned */
852 if (args->local_addr & 0x7) {
857 ret = rds_pin_pages(args->local_addr, 1, &page, 1);
862 sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
864 if (rm->atomic.op_notify || rm->atomic.op_recverr) {
865 /* We allocate an uninitialized notifier here, because
866 * we don't want to do that in the completion handler. We
867 * would have to use GFP_ATOMIC there, and don't want to deal
868 * with failed allocations.
870 rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL);
871 if (!rm->atomic.op_notifier) {
876 rm->atomic.op_notifier->n_user_token = args->user_token;
877 rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
880 rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
881 rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
887 rm->atomic.op_active = 0;
888 kfree(rm->atomic.op_notifier);