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)
159 unsigned int gup_flags = FOLL_LONGTERM;
163 gup_flags |= FOLL_WRITE;
165 ret = get_user_pages_fast(user_addr, nr_pages, gup_flags, pages);
166 if (ret >= 0 && ret < nr_pages) {
168 put_page(pages[ret]);
175 static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
176 u64 *cookie_ret, struct rds_mr **mr_ret,
177 struct rds_conn_path *cp)
179 struct rds_mr *mr = NULL, *found;
180 struct scatterlist *sg = NULL;
181 unsigned int nr_pages;
182 struct page **pages = NULL;
185 rds_rdma_cookie_t cookie;
186 unsigned int nents = 0;
191 if (ipv6_addr_any(&rs->rs_bound_addr) || !rs->rs_transport) {
192 ret = -ENOTCONN; /* XXX not a great errno */
196 if (!rs->rs_transport->get_mr) {
201 /* If the combination of the addr and size requested for this memory
202 * region causes an integer overflow, return error.
204 if (((args->vec.addr + args->vec.bytes) < args->vec.addr) ||
205 PAGE_ALIGN(args->vec.addr + args->vec.bytes) <
206 (args->vec.addr + args->vec.bytes)) {
211 if (!can_do_mlock()) {
216 nr_pages = rds_pages_in_vec(&args->vec);
222 /* Restrict the size of mr irrespective of underlying transport
223 * To account for unaligned mr regions, subtract one from nr_pages
225 if ((nr_pages - 1) > (RDS_MAX_MSG_SIZE >> PAGE_SHIFT)) {
230 rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
231 args->vec.addr, args->vec.bytes, nr_pages);
233 /* XXX clamp nr_pages to limit the size of this alloc? */
234 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
240 mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
246 refcount_set(&mr->r_refcount, 1);
247 RB_CLEAR_NODE(&mr->r_rb_node);
248 mr->r_trans = rs->rs_transport;
251 if (args->flags & RDS_RDMA_USE_ONCE)
253 if (args->flags & RDS_RDMA_INVALIDATE)
254 mr->r_invalidate = 1;
255 if (args->flags & RDS_RDMA_READWRITE)
259 * Pin the pages that make up the user buffer and transfer the page
260 * pointers to the mr's sg array. We check to see if we've mapped
261 * the whole region after transferring the partial page references
262 * to the sg array so that we can have one page ref cleanup path.
264 * For now we have no flag that tells us whether the mapping is
265 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
268 ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
269 if (ret == -EOPNOTSUPP) {
271 } else if (ret <= 0) {
275 sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
281 sg_init_table(sg, nents);
283 /* Stick all pages into the scatterlist */
284 for (i = 0 ; i < nents; i++)
285 sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
287 rdsdebug("RDS: trans_private nents is %u\n", nents);
289 /* Obtain a transport specific MR. If this succeeds, the
290 * s/g list is now owned by the MR.
291 * Note that dma_map() implies that pending writes are
292 * flushed to RAM, so no dma_sync is needed here. */
293 trans_private = rs->rs_transport->get_mr(
294 sg, nents, rs, &mr->r_key, cp ? cp->cp_conn : NULL,
295 args->vec.addr, args->vec.bytes,
296 need_odp ? ODP_ZEROBASED : ODP_NOT_NEEDED);
298 if (IS_ERR(trans_private)) {
299 /* In ODP case, we don't GUP pages, so don't need
300 * to release anything.
303 for (i = 0 ; i < nents; i++)
304 put_page(sg_page(&sg[i]));
307 ret = PTR_ERR(trans_private);
311 mr->r_trans_private = trans_private;
313 rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
314 mr->r_key, (void *)(unsigned long) args->cookie_addr);
316 /* The user may pass us an unaligned address, but we can only
317 * map page aligned regions. So we keep the offset, and build
318 * a 64bit cookie containing <R_Key, offset> and pass that
321 cookie = rds_rdma_make_cookie(mr->r_key, 0);
323 cookie = rds_rdma_make_cookie(mr->r_key,
324 args->vec.addr & ~PAGE_MASK);
326 *cookie_ret = cookie;
328 if (args->cookie_addr && put_user(cookie, (u64 __user *)(unsigned long) args->cookie_addr)) {
333 /* Inserting the new MR into the rbtree bumps its
334 * reference count. */
335 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
336 found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
337 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
339 BUG_ON(found && found != mr);
341 rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
343 refcount_inc(&mr->r_refcount);
355 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen)
357 struct rds_get_mr_args args;
359 if (optlen != sizeof(struct rds_get_mr_args))
362 if (copy_from_user(&args, (struct rds_get_mr_args __user *)optval,
363 sizeof(struct rds_get_mr_args)))
366 return __rds_rdma_map(rs, &args, NULL, NULL, NULL);
369 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen)
371 struct rds_get_mr_for_dest_args args;
372 struct rds_get_mr_args new_args;
374 if (optlen != sizeof(struct rds_get_mr_for_dest_args))
377 if (copy_from_user(&args, (struct rds_get_mr_for_dest_args __user *)optval,
378 sizeof(struct rds_get_mr_for_dest_args)))
382 * Initially, just behave like get_mr().
383 * TODO: Implement get_mr as wrapper around this
386 new_args.vec = args.vec;
387 new_args.cookie_addr = args.cookie_addr;
388 new_args.flags = args.flags;
390 return __rds_rdma_map(rs, &new_args, NULL, NULL, NULL);
394 * Free the MR indicated by the given R_Key
396 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen)
398 struct rds_free_mr_args args;
402 if (optlen != sizeof(struct rds_free_mr_args))
405 if (copy_from_user(&args, (struct rds_free_mr_args __user *)optval,
406 sizeof(struct rds_free_mr_args)))
409 /* Special case - a null cookie means flush all unused MRs */
410 if (args.cookie == 0) {
411 if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
413 rs->rs_transport->flush_mrs();
417 /* Look up the MR given its R_key and remove it from the rbtree
418 * so nobody else finds it.
419 * This should also prevent races with rds_rdma_unuse.
421 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
422 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL);
424 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
425 RB_CLEAR_NODE(&mr->r_rb_node);
426 if (args.flags & RDS_RDMA_INVALIDATE)
427 mr->r_invalidate = 1;
429 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
435 * call rds_destroy_mr() ourselves so that we're sure it's done by the time
436 * we return. If we let rds_mr_put() do it it might not happen until
437 * someone else drops their ref.
445 * This is called when we receive an extension header that
446 * tells us this MR was used. It allows us to implement
449 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
455 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
456 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
458 pr_debug("rds: trying to unuse MR with unknown r_key %u!\n",
460 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
464 if (mr->r_use_once || force) {
465 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
466 RB_CLEAR_NODE(&mr->r_rb_node);
469 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
471 /* May have to issue a dma_sync on this memory region.
472 * Note we could avoid this if the operation was a RDMA READ,
473 * but at this point we can't tell. */
474 if (mr->r_trans->sync_mr)
475 mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
477 /* If the MR was marked as invalidate, this will
478 * trigger an async flush. */
485 void rds_rdma_free_op(struct rm_rdma_op *ro)
490 rds_mr_put(ro->op_odp_mr);
492 for (i = 0; i < ro->op_nents; i++) {
493 struct page *page = sg_page(&ro->op_sg[i]);
495 /* Mark page dirty if it was possibly modified, which
496 * is the case for a RDMA_READ which copies from remote
500 set_page_dirty(page);
505 kfree(ro->op_notifier);
506 ro->op_notifier = NULL;
508 ro->op_odp_mr = NULL;
511 void rds_atomic_free_op(struct rm_atomic_op *ao)
513 struct page *page = sg_page(ao->op_sg);
515 /* Mark page dirty if it was possibly modified, which
516 * is the case for a RDMA_READ which copies from remote
518 set_page_dirty(page);
521 kfree(ao->op_notifier);
522 ao->op_notifier = NULL;
528 * Count the number of pages needed to describe an incoming iovec array.
530 static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs)
533 unsigned int nr_pages;
536 /* figure out the number of pages in the vector */
537 for (i = 0; i < nr_iovecs; i++) {
538 nr_pages = rds_pages_in_vec(&iov[i]);
542 tot_pages += nr_pages;
545 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
546 * so tot_pages cannot overflow without first going negative.
555 int rds_rdma_extra_size(struct rds_rdma_args *args,
556 struct rds_iov_vector *iov)
558 struct rds_iovec *vec;
559 struct rds_iovec __user *local_vec;
561 unsigned int nr_pages;
564 local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
566 if (args->nr_local == 0)
569 iov->iov = kcalloc(args->nr_local,
570 sizeof(struct rds_iovec),
577 if (copy_from_user(vec, local_vec, args->nr_local *
578 sizeof(struct rds_iovec)))
580 iov->len = args->nr_local;
582 /* figure out the number of pages in the vector */
583 for (i = 0; i < args->nr_local; i++, vec++) {
585 nr_pages = rds_pages_in_vec(vec);
589 tot_pages += nr_pages;
592 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
593 * so tot_pages cannot overflow without first going negative.
599 return tot_pages * sizeof(struct scatterlist);
603 * The application asks for a RDMA transfer.
604 * Extract all arguments and set up the rdma_op
606 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
607 struct cmsghdr *cmsg,
608 struct rds_iov_vector *vec)
610 struct rds_rdma_args *args;
611 struct rm_rdma_op *op = &rm->rdma;
613 unsigned int nr_bytes;
614 struct page **pages = NULL;
615 struct rds_iovec *iovs;
618 bool odp_supported = true;
620 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
621 || rm->rdma.op_active)
624 args = CMSG_DATA(cmsg);
626 if (ipv6_addr_any(&rs->rs_bound_addr)) {
627 ret = -ENOTCONN; /* XXX not a great errno */
631 if (args->nr_local > UIO_MAXIOV) {
636 if (vec->len != args->nr_local) {
640 /* odp-mr is not supported for multiple requests within one message */
641 if (args->nr_local != 1)
642 odp_supported = false;
646 nr_pages = rds_rdma_pages(iovs, args->nr_local);
652 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
658 op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
659 op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
660 op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
661 op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
663 op->op_recverr = rs->rs_recverr;
664 op->op_odp_mr = NULL;
667 op->op_sg = rds_message_alloc_sgs(rm, nr_pages, &ret);
671 if (op->op_notify || op->op_recverr) {
672 /* We allocate an uninitialized notifier here, because
673 * we don't want to do that in the completion handler. We
674 * would have to use GFP_ATOMIC there, and don't want to deal
675 * with failed allocations.
677 op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
678 if (!op->op_notifier) {
682 op->op_notifier->n_user_token = args->user_token;
683 op->op_notifier->n_status = RDS_RDMA_SUCCESS;
686 /* The cookie contains the R_Key of the remote memory region, and
687 * optionally an offset into it. This is how we implement RDMA into
689 * When setting up the RDMA, we need to add that offset to the
690 * destination address (which is really an offset into the MR)
691 * FIXME: We may want to move this into ib_rdma.c
693 op->op_rkey = rds_rdma_cookie_key(args->cookie);
694 op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
698 rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
699 (unsigned long long)args->nr_local,
700 (unsigned long long)args->remote_vec.addr,
703 for (i = 0; i < args->nr_local; i++) {
704 struct rds_iovec *iov = &iovs[i];
705 /* don't need to check, rds_rdma_pages() verified nr will be +nonzero */
706 unsigned int nr = rds_pages_in_vec(iov);
708 rs->rs_user_addr = iov->addr;
709 rs->rs_user_bytes = iov->bytes;
711 /* If it's a WRITE operation, we want to pin the pages for reading.
712 * If it's a READ operation, we need to pin the pages for writing.
714 ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
715 if ((!odp_supported && ret <= 0) ||
716 (odp_supported && ret <= 0 && ret != -EOPNOTSUPP))
719 if (ret == -EOPNOTSUPP) {
720 struct rds_mr *local_odp_mr;
722 if (!rs->rs_transport->get_mr) {
727 kzalloc(sizeof(*local_odp_mr), GFP_KERNEL);
732 RB_CLEAR_NODE(&local_odp_mr->r_rb_node);
733 refcount_set(&local_odp_mr->r_refcount, 1);
734 local_odp_mr->r_trans = rs->rs_transport;
735 local_odp_mr->r_sock = rs;
736 local_odp_mr->r_trans_private =
737 rs->rs_transport->get_mr(
738 NULL, 0, rs, &local_odp_mr->r_key, NULL,
739 iov->addr, iov->bytes, ODP_VIRTUAL);
740 if (IS_ERR(local_odp_mr->r_trans_private)) {
741 ret = IS_ERR(local_odp_mr->r_trans_private);
742 rdsdebug("get_mr ret %d %p\"", ret,
743 local_odp_mr->r_trans_private);
748 rdsdebug("Need odp; local_odp_mr %p trans_private %p\n",
749 local_odp_mr, local_odp_mr->r_trans_private);
750 op->op_odp_mr = local_odp_mr;
751 op->op_odp_addr = iov->addr;
754 rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
755 nr_bytes, nr, iov->bytes, iov->addr);
757 nr_bytes += iov->bytes;
759 for (j = 0; j < nr; j++) {
760 unsigned int offset = iov->addr & ~PAGE_MASK;
761 struct scatterlist *sg;
763 sg = &op->op_sg[op->op_nents + j];
764 sg_set_page(sg, pages[j],
765 min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
768 sg_dma_len(sg) = sg->length;
769 rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n",
770 sg->offset, sg->length, iov->addr, iov->bytes);
772 iov->addr += sg->length;
773 iov->bytes -= sg->length;
779 if (nr_bytes > args->remote_vec.bytes) {
780 rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
782 (unsigned int) args->remote_vec.bytes);
786 op->op_bytes = nr_bytes;
793 rds_rdma_free_op(op);
795 rds_stats_inc(s_send_rdma);
801 * The application wants us to pass an RDMA destination (aka MR)
804 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
805 struct cmsghdr *cmsg)
812 if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
813 rm->m_rdma_cookie != 0)
816 memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
818 /* We are reusing a previously mapped MR here. Most likely, the
819 * application has written to the buffer, so we need to explicitly
820 * flush those writes to RAM. Otherwise the HCA may not see them
821 * when doing a DMA from that buffer.
823 r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
825 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
826 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
828 err = -EINVAL; /* invalid r_key */
830 refcount_inc(&mr->r_refcount);
831 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
834 mr->r_trans->sync_mr(mr->r_trans_private,
836 rm->rdma.op_rdma_mr = mr;
842 * The application passes us an address range it wants to enable RDMA
843 * to/from. We map the area, and save the <R_Key,offset> pair
844 * in rm->m_rdma_cookie. This causes it to be sent along to the peer
845 * in an extension header.
847 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
848 struct cmsghdr *cmsg)
850 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
851 rm->m_rdma_cookie != 0)
854 return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie,
855 &rm->rdma.op_rdma_mr, rm->m_conn_path);
859 * Fill in rds_message for an atomic request.
861 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
862 struct cmsghdr *cmsg)
864 struct page *page = NULL;
865 struct rds_atomic_args *args;
868 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
869 || rm->atomic.op_active)
872 args = CMSG_DATA(cmsg);
874 /* Nonmasked & masked cmsg ops converted to masked hw ops */
875 switch (cmsg->cmsg_type) {
876 case RDS_CMSG_ATOMIC_FADD:
877 rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
878 rm->atomic.op_m_fadd.add = args->fadd.add;
879 rm->atomic.op_m_fadd.nocarry_mask = 0;
881 case RDS_CMSG_MASKED_ATOMIC_FADD:
882 rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
883 rm->atomic.op_m_fadd.add = args->m_fadd.add;
884 rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
886 case RDS_CMSG_ATOMIC_CSWP:
887 rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
888 rm->atomic.op_m_cswp.compare = args->cswp.compare;
889 rm->atomic.op_m_cswp.swap = args->cswp.swap;
890 rm->atomic.op_m_cswp.compare_mask = ~0;
891 rm->atomic.op_m_cswp.swap_mask = ~0;
893 case RDS_CMSG_MASKED_ATOMIC_CSWP:
894 rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
895 rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
896 rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
897 rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
898 rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
901 BUG(); /* should never happen */
904 rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
905 rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
906 rm->atomic.op_active = 1;
907 rm->atomic.op_recverr = rs->rs_recverr;
908 rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1, &ret);
909 if (!rm->atomic.op_sg)
912 /* verify 8 byte-aligned */
913 if (args->local_addr & 0x7) {
918 ret = rds_pin_pages(args->local_addr, 1, &page, 1);
923 sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
925 if (rm->atomic.op_notify || rm->atomic.op_recverr) {
926 /* We allocate an uninitialized notifier here, because
927 * we don't want to do that in the completion handler. We
928 * would have to use GFP_ATOMIC there, and don't want to deal
929 * with failed allocations.
931 rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL);
932 if (!rm->atomic.op_notifier) {
937 rm->atomic.op_notifier->n_user_token = args->user_token;
938 rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
941 rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
942 rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
948 rm->atomic.op_active = 0;
949 kfree(rm->atomic.op_notifier);