2 * Copyright (c) 2007 Oracle. 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 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?
49 * get the number of pages by looking at the page indices that the start and
50 * end addresses fall in.
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'.
56 static unsigned int rds_pages_in_vec(struct rds_iovec *vec)
58 if ((vec->addr + vec->bytes <= vec->addr) ||
59 (vec->bytes > (u64)UINT_MAX))
62 return ((vec->addr + vec->bytes + PAGE_SIZE - 1) >> PAGE_SHIFT) -
63 (vec->addr >> PAGE_SHIFT);
66 static struct rds_mr *rds_mr_tree_walk(struct rb_root *root, u64 key,
67 struct rds_mr *insert)
69 struct rb_node **p = &root->rb_node;
70 struct rb_node *parent = NULL;
75 mr = rb_entry(parent, struct rds_mr, r_rb_node);
79 else if (key > mr->r_key)
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);
94 * Destroy the transport-specific part of a MR.
96 static void rds_destroy_mr(struct rds_mr *mr)
98 struct rds_sock *rs = mr->r_sock;
99 void *trans_private = NULL;
102 rdsdebug("RDS: destroy mr key is %x refcnt %u\n",
103 mr->r_key, atomic_read(&mr->r_refcount));
105 if (test_and_set_bit(RDS_MR_DEAD, &mr->r_state))
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);
116 mr->r_trans->free_mr(trans_private, mr->r_invalidate);
119 void __rds_put_mr_final(struct rds_mr *mr)
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.
129 void rds_rdma_drop_keys(struct rds_sock *rs)
132 struct rb_node *node;
135 /* Release any MRs associated with this socket */
136 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
137 while ((node = rb_first(&rs->rs_rdma_keys))) {
138 mr = container_of(node, struct rds_mr, r_rb_node);
139 if (mr->r_trans == rs->rs_transport)
140 mr->r_invalidate = 0;
141 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
142 RB_CLEAR_NODE(&mr->r_rb_node);
143 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
146 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
148 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
150 if (rs->rs_transport && rs->rs_transport->flush_mrs)
151 rs->rs_transport->flush_mrs();
155 * Helper function to pin user pages.
157 static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
158 struct page **pages, int write)
162 ret = get_user_pages_fast(user_addr, nr_pages, write, pages);
164 if (ret >= 0 && ret < nr_pages) {
166 put_page(pages[ret]);
173 static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
174 u64 *cookie_ret, struct rds_mr **mr_ret)
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 (rs->rs_bound_addr == 0) {
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 rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
204 args->vec.addr, args->vec.bytes, nr_pages);
206 /* XXX clamp nr_pages to limit the size of this alloc? */
207 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
213 mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
219 atomic_set(&mr->r_refcount, 1);
220 RB_CLEAR_NODE(&mr->r_rb_node);
221 mr->r_trans = rs->rs_transport;
224 if (args->flags & RDS_RDMA_USE_ONCE)
226 if (args->flags & RDS_RDMA_INVALIDATE)
227 mr->r_invalidate = 1;
228 if (args->flags & RDS_RDMA_READWRITE)
232 * Pin the pages that make up the user buffer and transfer the page
233 * pointers to the mr's sg array. We check to see if we've mapped
234 * the whole region after transferring the partial page references
235 * to the sg array so that we can have one page ref cleanup path.
237 * For now we have no flag that tells us whether the mapping is
238 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
241 ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
246 sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
252 sg_init_table(sg, nents);
254 /* Stick all pages into the scatterlist */
255 for (i = 0 ; i < nents; i++)
256 sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
258 rdsdebug("RDS: trans_private nents is %u\n", nents);
260 /* Obtain a transport specific MR. If this succeeds, the
261 * s/g list is now owned by the MR.
262 * Note that dma_map() implies that pending writes are
263 * flushed to RAM, so no dma_sync is needed here. */
264 trans_private = rs->rs_transport->get_mr(sg, nents, rs,
267 if (IS_ERR(trans_private)) {
268 for (i = 0 ; i < nents; i++)
269 put_page(sg_page(&sg[i]));
271 ret = PTR_ERR(trans_private);
275 mr->r_trans_private = trans_private;
277 rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
278 mr->r_key, (void *)(unsigned long) args->cookie_addr);
280 /* The user may pass us an unaligned address, but we can only
281 * map page aligned regions. So we keep the offset, and build
282 * a 64bit cookie containing <R_Key, offset> and pass that
284 cookie = rds_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGE_MASK);
286 *cookie_ret = cookie;
288 if (args->cookie_addr && put_user(cookie, (u64 __user *)(unsigned long) args->cookie_addr)) {
293 /* Inserting the new MR into the rbtree bumps its
294 * reference count. */
295 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
296 found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
297 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
299 BUG_ON(found && found != mr);
301 rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
303 atomic_inc(&mr->r_refcount);
315 int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen)
317 struct rds_get_mr_args args;
319 if (optlen != sizeof(struct rds_get_mr_args))
322 if (copy_from_user(&args, (struct rds_get_mr_args __user *)optval,
323 sizeof(struct rds_get_mr_args)))
326 return __rds_rdma_map(rs, &args, NULL, NULL);
329 int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen)
331 struct rds_get_mr_for_dest_args args;
332 struct rds_get_mr_args new_args;
334 if (optlen != sizeof(struct rds_get_mr_for_dest_args))
337 if (copy_from_user(&args, (struct rds_get_mr_for_dest_args __user *)optval,
338 sizeof(struct rds_get_mr_for_dest_args)))
342 * Initially, just behave like get_mr().
343 * TODO: Implement get_mr as wrapper around this
346 new_args.vec = args.vec;
347 new_args.cookie_addr = args.cookie_addr;
348 new_args.flags = args.flags;
350 return __rds_rdma_map(rs, &new_args, NULL, NULL);
354 * Free the MR indicated by the given R_Key
356 int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen)
358 struct rds_free_mr_args args;
362 if (optlen != sizeof(struct rds_free_mr_args))
365 if (copy_from_user(&args, (struct rds_free_mr_args __user *)optval,
366 sizeof(struct rds_free_mr_args)))
369 /* Special case - a null cookie means flush all unused MRs */
370 if (args.cookie == 0) {
371 if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
373 rs->rs_transport->flush_mrs();
377 /* Look up the MR given its R_key and remove it from the rbtree
378 * so nobody else finds it.
379 * This should also prevent races with rds_rdma_unuse.
381 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
382 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL);
384 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
385 RB_CLEAR_NODE(&mr->r_rb_node);
386 if (args.flags & RDS_RDMA_INVALIDATE)
387 mr->r_invalidate = 1;
389 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
395 * call rds_destroy_mr() ourselves so that we're sure it's done by the time
396 * we return. If we let rds_mr_put() do it it might not happen until
397 * someone else drops their ref.
405 * This is called when we receive an extension header that
406 * tells us this MR was used. It allows us to implement
409 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
415 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
416 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
418 printk(KERN_ERR "rds: trying to unuse MR with unknown r_key %u!\n", r_key);
419 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
423 if (mr->r_use_once || force) {
424 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
425 RB_CLEAR_NODE(&mr->r_rb_node);
428 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
430 /* May have to issue a dma_sync on this memory region.
431 * Note we could avoid this if the operation was a RDMA READ,
432 * but at this point we can't tell. */
433 if (mr->r_trans->sync_mr)
434 mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
436 /* If the MR was marked as invalidate, this will
437 * trigger an async flush. */
443 void rds_rdma_free_op(struct rds_rdma_op *ro)
447 for (i = 0; i < ro->r_nents; i++) {
448 struct page *page = sg_page(&ro->r_sg[i]);
450 /* Mark page dirty if it was possibly modified, which
451 * is the case for a RDMA_READ which copies from remote
454 BUG_ON(irqs_disabled());
455 set_page_dirty(page);
460 kfree(ro->r_notifier);
461 ro->r_notifier = NULL;
466 * Count the number of pages needed to describe an incoming iovec.
468 static int rds_rdma_pages(struct rds_rdma_args *args)
470 struct rds_iovec vec;
471 struct rds_iovec __user *local_vec;
472 unsigned int tot_pages = 0;
473 unsigned int nr_pages;
476 local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
478 /* figure out the number of pages in the vector */
479 for (i = 0; i < args->nr_local; i++) {
480 if (copy_from_user(&vec, &local_vec[i],
481 sizeof(struct rds_iovec)))
484 nr_pages = rds_pages_in_vec(&vec);
488 tot_pages += nr_pages;
494 int rds_rdma_extra_size(struct rds_rdma_args *args)
496 return rds_rdma_pages(args) * sizeof(struct scatterlist);
500 * The application asks for a RDMA transfer.
501 * Extract all arguments and set up the rdma_op
503 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
504 struct cmsghdr *cmsg)
506 struct rds_rdma_args *args;
507 struct rds_iovec vec;
508 struct rds_rdma_op *op = &rm->rdma.m_rdma_op;
509 unsigned int nr_pages;
510 unsigned int nr_bytes;
511 struct page **pages = NULL;
512 struct rds_iovec __user *local_vec;
517 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
518 || rm->rdma.m_rdma_op.r_active)
521 args = CMSG_DATA(cmsg);
523 if (rs->rs_bound_addr == 0) {
524 ret = -ENOTCONN; /* XXX not a great errno */
528 if (args->nr_local > (u64)UINT_MAX) {
533 nr_pages = rds_rdma_pages(args);
537 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
543 op->r_write = !!(args->flags & RDS_RDMA_READWRITE);
544 op->r_fence = !!(args->flags & RDS_RDMA_FENCE);
545 op->r_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
547 op->r_recverr = rs->rs_recverr;
549 op->r_sg = rds_message_alloc_sgs(rm, nr_pages);
551 if (op->r_notify || op->r_recverr) {
552 /* We allocate an uninitialized notifier here, because
553 * we don't want to do that in the completion handler. We
554 * would have to use GFP_ATOMIC there, and don't want to deal
555 * with failed allocations.
557 op->r_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
558 if (!op->r_notifier) {
562 op->r_notifier->n_user_token = args->user_token;
563 op->r_notifier->n_status = RDS_RDMA_SUCCESS;
566 /* The cookie contains the R_Key of the remote memory region, and
567 * optionally an offset into it. This is how we implement RDMA into
569 * When setting up the RDMA, we need to add that offset to the
570 * destination address (which is really an offset into the MR)
571 * FIXME: We may want to move this into ib_rdma.c
573 op->r_key = rds_rdma_cookie_key(args->cookie);
574 op->r_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
578 rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
579 (unsigned long long)args->nr_local,
580 (unsigned long long)args->remote_vec.addr,
583 local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
585 for (i = 0; i < args->nr_local; i++) {
586 if (copy_from_user(&vec, &local_vec[i],
587 sizeof(struct rds_iovec))) {
592 nr = rds_pages_in_vec(&vec);
598 rs->rs_user_addr = vec.addr;
599 rs->rs_user_bytes = vec.bytes;
601 /* If it's a WRITE operation, we want to pin the pages for reading.
602 * If it's a READ operation, we need to pin the pages for writing.
604 ret = rds_pin_pages(vec.addr, nr, pages, !op->r_write);
608 rdsdebug("RDS: nr_bytes %u nr %u vec.bytes %llu vec.addr %llx\n",
609 nr_bytes, nr, vec.bytes, vec.addr);
611 nr_bytes += vec.bytes;
613 for (j = 0; j < nr; j++) {
614 unsigned int offset = vec.addr & ~PAGE_MASK;
615 struct scatterlist *sg;
617 sg = &op->r_sg[op->r_nents + j];
618 sg_set_page(sg, pages[j],
619 min_t(unsigned int, vec.bytes, PAGE_SIZE - offset),
622 rdsdebug("RDS: sg->offset %x sg->len %x vec.addr %llx vec.bytes %llu\n",
623 sg->offset, sg->length, vec.addr, vec.bytes);
625 vec.addr += sg->length;
626 vec.bytes -= sg->length;
632 if (nr_bytes > args->remote_vec.bytes) {
633 rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
635 (unsigned int) args->remote_vec.bytes);
639 op->r_bytes = nr_bytes;
645 rds_rdma_free_op(op);
647 rds_stats_inc(s_send_rdma);
653 * The application wants us to pass an RDMA destination (aka MR)
656 int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
657 struct cmsghdr *cmsg)
664 if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
665 rm->m_rdma_cookie != 0)
668 memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
670 /* We are reusing a previously mapped MR here. Most likely, the
671 * application has written to the buffer, so we need to explicitly
672 * flush those writes to RAM. Otherwise the HCA may not see them
673 * when doing a DMA from that buffer.
675 r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
677 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
678 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
680 err = -EINVAL; /* invalid r_key */
682 atomic_inc(&mr->r_refcount);
683 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
686 mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
687 rm->rdma.m_rdma_mr = mr;
693 * The application passes us an address range it wants to enable RDMA
694 * to/from. We map the area, and save the <R_Key,offset> pair
695 * in rm->m_rdma_cookie. This causes it to be sent along to the peer
696 * in an extension header.
698 int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
699 struct cmsghdr *cmsg)
701 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
702 rm->m_rdma_cookie != 0)
705 return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->rdma.m_rdma_mr);
709 * Fill in rds_message for an atomic request.
711 int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
712 struct cmsghdr *cmsg)
714 struct page *page = NULL;
715 struct rds_atomic_args *args;
718 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
719 || rm->atomic.op_active)
722 args = CMSG_DATA(cmsg);
724 if (cmsg->cmsg_type == RDS_CMSG_ATOMIC_CSWP) {
725 rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
726 rm->atomic.op_swap_add = args->cswp.swap;
727 rm->atomic.op_compare = args->cswp.compare;
729 rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
730 rm->atomic.op_swap_add = args->fadd.add;
733 rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
734 rm->atomic.op_recverr = rs->rs_recverr;
735 rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1);
737 /* verify 8 byte-aligned */
738 if (args->local_addr & 0x7) {
743 ret = rds_pin_pages(args->local_addr, 1, &page, 1);
748 sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
750 if (rm->atomic.op_notify || rm->atomic.op_recverr) {
751 /* We allocate an uninitialized notifier here, because
752 * we don't want to do that in the completion handler. We
753 * would have to use GFP_ATOMIC there, and don't want to deal
754 * with failed allocations.
756 rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL);
757 if (!rm->atomic.op_notifier) {
762 rm->atomic.op_notifier->n_user_token = args->user_token;
763 rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
766 rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
767 rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
769 rm->atomic.op_active = 1;
775 kfree(rm->atomic.op_notifier);