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 "rds.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	unsigned long flags;
134
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);
144		rds_destroy_mr(mr);
145		rds_mr_put(mr);
146		spin_lock_irqsave(&rs->rs_rdma_lock, flags);
147	}
148	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
149
150	if (rs->rs_transport && rs->rs_transport->flush_mrs)
151		rs->rs_transport->flush_mrs();
152}
153
154/*
155 * Helper function to pin user pages.
156 */
157static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
158			struct page **pages, int write)
159{
160	int ret;
161
162	ret = get_user_pages_fast(user_addr, nr_pages, write, pages);
 
163
164	if (ret >= 0 && ret < nr_pages) {
165		while (ret--)
166			put_page(pages[ret]);
167		ret = -EFAULT;
168	}
169
170	return ret;
171}
172
173static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
174				u64 *cookie_ret, struct rds_mr **mr_ret)
 
175{
176	struct rds_mr *mr = NULL, *found;
177	unsigned int nr_pages;
178	struct page **pages = NULL;
179	struct scatterlist *sg;
180	void *trans_private;
181	unsigned long flags;
182	rds_rdma_cookie_t cookie;
183	unsigned int nents;
184	long i;
185	int ret;
186
187	if (rs->rs_bound_addr == 0) {
188		ret = -ENOTCONN; /* XXX not a great errno */
189		goto out;
190	}
191
192	if (!rs->rs_transport->get_mr) {
193		ret = -EOPNOTSUPP;
194		goto out;
195	}
196
197	nr_pages = rds_pages_in_vec(&args->vec);
198	if (nr_pages == 0) {
199		ret = -EINVAL;
200		goto out;
201	}
202
 
 
 
 
 
 
 
 
203	rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
204		args->vec.addr, args->vec.bytes, nr_pages);
205
206	/* XXX clamp nr_pages to limit the size of this alloc? */
207	pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
208	if (!pages) {
209		ret = -ENOMEM;
210		goto out;
211	}
212
213	mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
214	if (!mr) {
215		ret = -ENOMEM;
216		goto out;
217	}
218
219	atomic_set(&mr->r_refcount, 1);
220	RB_CLEAR_NODE(&mr->r_rb_node);
221	mr->r_trans = rs->rs_transport;
222	mr->r_sock = rs;
223
224	if (args->flags & RDS_RDMA_USE_ONCE)
225		mr->r_use_once = 1;
226	if (args->flags & RDS_RDMA_INVALIDATE)
227		mr->r_invalidate = 1;
228	if (args->flags & RDS_RDMA_READWRITE)
229		mr->r_write = 1;
230
231	/*
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.
236	 *
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
239	 * the zero page.
240	 */
241	ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
242	if (ret < 0)
243		goto out;
244
245	nents = ret;
246	sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
247	if (!sg) {
248		ret = -ENOMEM;
249		goto out;
250	}
251	WARN_ON(!nents);
252	sg_init_table(sg, nents);
253
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);
257
258	rdsdebug("RDS: trans_private nents is %u\n", nents);
259
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,
265						 &mr->r_key);
 
266
267	if (IS_ERR(trans_private)) {
268		for (i = 0 ; i < nents; i++)
269			put_page(sg_page(&sg[i]));
270		kfree(sg);
271		ret = PTR_ERR(trans_private);
272		goto out;
273	}
274
275	mr->r_trans_private = trans_private;
276
277	rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
278	       mr->r_key, (void *)(unsigned long) args->cookie_addr);
279
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
283	 * around. */
284	cookie = rds_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGE_MASK);
285	if (cookie_ret)
286		*cookie_ret = cookie;
287
288	if (args->cookie_addr && put_user(cookie, (u64 __user *)(unsigned long) args->cookie_addr)) {
289		ret = -EFAULT;
290		goto out;
291	}
292
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);
298
299	BUG_ON(found && found != mr);
300
301	rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
302	if (mr_ret) {
303		atomic_inc(&mr->r_refcount);
304		*mr_ret = mr;
305	}
306
307	ret = 0;
308out:
309	kfree(pages);
310	if (mr)
311		rds_mr_put(mr);
312	return ret;
313}
314
315int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen)
316{
317	struct rds_get_mr_args args;
318
319	if (optlen != sizeof(struct rds_get_mr_args))
320		return -EINVAL;
321
322	if (copy_from_user(&args, (struct rds_get_mr_args __user *)optval,
323			   sizeof(struct rds_get_mr_args)))
324		return -EFAULT;
325
326	return __rds_rdma_map(rs, &args, NULL, NULL);
327}
328
329int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen)
330{
331	struct rds_get_mr_for_dest_args args;
332	struct rds_get_mr_args new_args;
333
334	if (optlen != sizeof(struct rds_get_mr_for_dest_args))
335		return -EINVAL;
336
337	if (copy_from_user(&args, (struct rds_get_mr_for_dest_args __user *)optval,
338			   sizeof(struct rds_get_mr_for_dest_args)))
339		return -EFAULT;
340
341	/*
342	 * Initially, just behave like get_mr().
343	 * TODO: Implement get_mr as wrapper around this
344	 *	 and deprecate it.
345	 */
346	new_args.vec = args.vec;
347	new_args.cookie_addr = args.cookie_addr;
348	new_args.flags = args.flags;
349
350	return __rds_rdma_map(rs, &new_args, NULL, NULL);
351}
352
353/*
354 * Free the MR indicated by the given R_Key
355 */
356int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen)
357{
358	struct rds_free_mr_args args;
359	struct rds_mr *mr;
360	unsigned long flags;
361
362	if (optlen != sizeof(struct rds_free_mr_args))
363		return -EINVAL;
364
365	if (copy_from_user(&args, (struct rds_free_mr_args __user *)optval,
366			   sizeof(struct rds_free_mr_args)))
367		return -EFAULT;
368
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)
372			return -EINVAL;
373		rs->rs_transport->flush_mrs();
374		return 0;
375	}
376
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.
380	 */
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);
383	if (mr) {
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;
388	}
389	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
390
391	if (!mr)
392		return -EINVAL;
393
394	/*
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.
398	 */
399	rds_destroy_mr(mr);
400	rds_mr_put(mr);
401	return 0;
402}
403
404/*
405 * This is called when we receive an extension header that
406 * tells us this MR was used. It allows us to implement
407 * use_once semantics
408 */
409void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
410{
411	struct rds_mr *mr;
412	unsigned long flags;
413	int zot_me = 0;
414
415	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
416	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
417	if (!mr) {
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);
420		return;
421	}
422
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);
426		zot_me = 1;
427	}
428	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
429
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);
435
436	/* If the MR was marked as invalidate, this will
437	 * trigger an async flush. */
438	if (zot_me) {
439		rds_destroy_mr(mr);
440		rds_mr_put(mr);
441	}
442}
443
444void rds_rdma_free_op(struct rm_rdma_op *ro)
445{
446	unsigned int i;
447
448	for (i = 0; i < ro->op_nents; i++) {
449		struct page *page = sg_page(&ro->op_sg[i]);
450
451		/* Mark page dirty if it was possibly modified, which
452		 * is the case for a RDMA_READ which copies from remote
453		 * to local memory */
454		if (!ro->op_write) {
455			WARN_ON(!page->mapping && irqs_disabled());
456			set_page_dirty(page);
457		}
458		put_page(page);
459	}
460
461	kfree(ro->op_notifier);
462	ro->op_notifier = NULL;
463	ro->op_active = 0;
464}
465
466void rds_atomic_free_op(struct rm_atomic_op *ao)
467{
468	struct page *page = sg_page(ao->op_sg);
469
470	/* Mark page dirty if it was possibly modified, which
471	 * is the case for a RDMA_READ which copies from remote
472	 * to local memory */
473	set_page_dirty(page);
474	put_page(page);
475
476	kfree(ao->op_notifier);
477	ao->op_notifier = NULL;
478	ao->op_active = 0;
479}
480
481
482/*
483 * Count the number of pages needed to describe an incoming iovec array.
484 */
485static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs)
486{
487	int tot_pages = 0;
488	unsigned int nr_pages;
489	unsigned int i;
490
491	/* figure out the number of pages in the vector */
492	for (i = 0; i < nr_iovecs; i++) {
493		nr_pages = rds_pages_in_vec(&iov[i]);
494		if (nr_pages == 0)
495			return -EINVAL;
496
497		tot_pages += nr_pages;
498
499		/*
500		 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
501		 * so tot_pages cannot overflow without first going negative.
502		 */
503		if (tot_pages < 0)
504			return -EINVAL;
505	}
506
507	return tot_pages;
508}
509
510int rds_rdma_extra_size(struct rds_rdma_args *args)
 
511{
512	struct rds_iovec vec;
513	struct rds_iovec __user *local_vec;
514	int tot_pages = 0;
515	unsigned int nr_pages;
516	unsigned int i;
517
518	local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
519
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
520	/* figure out the number of pages in the vector */
521	for (i = 0; i < args->nr_local; i++) {
522		if (copy_from_user(&vec, &local_vec[i],
523				   sizeof(struct rds_iovec)))
524			return -EFAULT;
525
526		nr_pages = rds_pages_in_vec(&vec);
527		if (nr_pages == 0)
528			return -EINVAL;
529
530		tot_pages += nr_pages;
531
532		/*
533		 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
534		 * so tot_pages cannot overflow without first going negative.
535		 */
536		if (tot_pages < 0)
537			return -EINVAL;
538	}
539
540	return tot_pages * sizeof(struct scatterlist);
541}
542
543/*
544 * The application asks for a RDMA transfer.
545 * Extract all arguments and set up the rdma_op
546 */
547int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
548			  struct cmsghdr *cmsg)
 
549{
550	struct rds_rdma_args *args;
551	struct rm_rdma_op *op = &rm->rdma;
552	int nr_pages;
553	unsigned int nr_bytes;
554	struct page **pages = NULL;
555	struct rds_iovec iovstack[UIO_FASTIOV], *iovs = iovstack;
556	int iov_size;
557	unsigned int i, j;
558	int ret = 0;
559
560	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
561	    || rm->rdma.op_active)
562		return -EINVAL;
563
564	args = CMSG_DATA(cmsg);
565
566	if (rs->rs_bound_addr == 0) {
567		ret = -ENOTCONN; /* XXX not a great errno */
568		goto out_ret;
569	}
570
571	if (args->nr_local > UIO_MAXIOV) {
572		ret = -EMSGSIZE;
573		goto out_ret;
574	}
575
576	/* Check whether to allocate the iovec area */
577	iov_size = args->nr_local * sizeof(struct rds_iovec);
578	if (args->nr_local > UIO_FASTIOV) {
579		iovs = sock_kmalloc(rds_rs_to_sk(rs), iov_size, GFP_KERNEL);
580		if (!iovs) {
581			ret = -ENOMEM;
582			goto out_ret;
583		}
584	}
585
586	if (copy_from_user(iovs, (struct rds_iovec __user *)(unsigned long) args->local_vec_addr, iov_size)) {
587		ret = -EFAULT;
588		goto out;
589	}
590
591	nr_pages = rds_rdma_pages(iovs, args->nr_local);
592	if (nr_pages < 0) {
593		ret = -EINVAL;
594		goto out;
595	}
596
597	pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
598	if (!pages) {
599		ret = -ENOMEM;
600		goto out;
601	}
602
603	op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
604	op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
605	op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
606	op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
607	op->op_active = 1;
608	op->op_recverr = rs->rs_recverr;
609	WARN_ON(!nr_pages);
610	op->op_sg = rds_message_alloc_sgs(rm, nr_pages);
611	if (!op->op_sg) {
612		ret = -ENOMEM;
613		goto out;
614	}
615
616	if (op->op_notify || op->op_recverr) {
617		/* We allocate an uninitialized notifier here, because
618		 * we don't want to do that in the completion handler. We
619		 * would have to use GFP_ATOMIC there, and don't want to deal
620		 * with failed allocations.
621		 */
622		op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
623		if (!op->op_notifier) {
624			ret = -ENOMEM;
625			goto out;
626		}
627		op->op_notifier->n_user_token = args->user_token;
628		op->op_notifier->n_status = RDS_RDMA_SUCCESS;
629	}
630
631	/* The cookie contains the R_Key of the remote memory region, and
632	 * optionally an offset into it. This is how we implement RDMA into
633	 * unaligned memory.
634	 * When setting up the RDMA, we need to add that offset to the
635	 * destination address (which is really an offset into the MR)
636	 * FIXME: We may want to move this into ib_rdma.c
637	 */
638	op->op_rkey = rds_rdma_cookie_key(args->cookie);
639	op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
640
641	nr_bytes = 0;
642
643	rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
644	       (unsigned long long)args->nr_local,
645	       (unsigned long long)args->remote_vec.addr,
646	       op->op_rkey);
647
648	for (i = 0; i < args->nr_local; i++) {
649		struct rds_iovec *iov = &iovs[i];
650		/* don't need to check, rds_rdma_pages() verified nr will be +nonzero */
651		unsigned int nr = rds_pages_in_vec(iov);
652
653		rs->rs_user_addr = iov->addr;
654		rs->rs_user_bytes = iov->bytes;
655
656		/* If it's a WRITE operation, we want to pin the pages for reading.
657		 * If it's a READ operation, we need to pin the pages for writing.
658		 */
659		ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
660		if (ret < 0)
661			goto out;
662		else
663			ret = 0;
664
665		rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
666			 nr_bytes, nr, iov->bytes, iov->addr);
667
668		nr_bytes += iov->bytes;
669
670		for (j = 0; j < nr; j++) {
671			unsigned int offset = iov->addr & ~PAGE_MASK;
672			struct scatterlist *sg;
673
674			sg = &op->op_sg[op->op_nents + j];
675			sg_set_page(sg, pages[j],
676					min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
677					offset);
678
679			rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n",
680			       sg->offset, sg->length, iov->addr, iov->bytes);
681
682			iov->addr += sg->length;
683			iov->bytes -= sg->length;
684		}
685
686		op->op_nents += nr;
687	}
688
689	if (nr_bytes > args->remote_vec.bytes) {
690		rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
691				nr_bytes,
692				(unsigned int) args->remote_vec.bytes);
693		ret = -EINVAL;
694		goto out;
695	}
696	op->op_bytes = nr_bytes;
697
698out:
699	if (iovs != iovstack)
700		sock_kfree_s(rds_rs_to_sk(rs), iovs, iov_size);
701	kfree(pages);
702out_ret:
703	if (ret)
704		rds_rdma_free_op(op);
705	else
706		rds_stats_inc(s_send_rdma);
707
708	return ret;
709}
710
711/*
712 * The application wants us to pass an RDMA destination (aka MR)
713 * to the remote
714 */
715int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
716			  struct cmsghdr *cmsg)
717{
718	unsigned long flags;
719	struct rds_mr *mr;
720	u32 r_key;
721	int err = 0;
722
723	if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
724	    rm->m_rdma_cookie != 0)
725		return -EINVAL;
726
727	memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
728
729	/* We are reusing a previously mapped MR here. Most likely, the
730	 * application has written to the buffer, so we need to explicitly
731	 * flush those writes to RAM. Otherwise the HCA may not see them
732	 * when doing a DMA from that buffer.
733	 */
734	r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
735
736	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
737	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
738	if (!mr)
739		err = -EINVAL;	/* invalid r_key */
740	else
741		atomic_inc(&mr->r_refcount);
742	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
743
744	if (mr) {
745		mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
746		rm->rdma.op_rdma_mr = mr;
747	}
748	return err;
749}
750
751/*
752 * The application passes us an address range it wants to enable RDMA
753 * to/from. We map the area, and save the <R_Key,offset> pair
754 * in rm->m_rdma_cookie. This causes it to be sent along to the peer
755 * in an extension header.
756 */
757int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
758			  struct cmsghdr *cmsg)
759{
760	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
761	    rm->m_rdma_cookie != 0)
762		return -EINVAL;
763
764	return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->rdma.op_rdma_mr);
 
765}
766
767/*
768 * Fill in rds_message for an atomic request.
769 */
770int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
771		    struct cmsghdr *cmsg)
772{
773	struct page *page = NULL;
774	struct rds_atomic_args *args;
775	int ret = 0;
776
777	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
778	 || rm->atomic.op_active)
779		return -EINVAL;
780
781	args = CMSG_DATA(cmsg);
782
783	/* Nonmasked & masked cmsg ops converted to masked hw ops */
784	switch (cmsg->cmsg_type) {
785	case RDS_CMSG_ATOMIC_FADD:
786		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
787		rm->atomic.op_m_fadd.add = args->fadd.add;
788		rm->atomic.op_m_fadd.nocarry_mask = 0;
789		break;
790	case RDS_CMSG_MASKED_ATOMIC_FADD:
791		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
792		rm->atomic.op_m_fadd.add = args->m_fadd.add;
793		rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
794		break;
795	case RDS_CMSG_ATOMIC_CSWP:
796		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
797		rm->atomic.op_m_cswp.compare = args->cswp.compare;
798		rm->atomic.op_m_cswp.swap = args->cswp.swap;
799		rm->atomic.op_m_cswp.compare_mask = ~0;
800		rm->atomic.op_m_cswp.swap_mask = ~0;
801		break;
802	case RDS_CMSG_MASKED_ATOMIC_CSWP:
803		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
804		rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
805		rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
806		rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
807		rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
808		break;
809	default:
810		BUG(); /* should never happen */
811	}
812
813	rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
814	rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
815	rm->atomic.op_active = 1;
816	rm->atomic.op_recverr = rs->rs_recverr;
817	rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1);
818	if (!rm->atomic.op_sg) {
819		ret = -ENOMEM;
820		goto err;
821	}
822
823	/* verify 8 byte-aligned */
824	if (args->local_addr & 0x7) {
825		ret = -EFAULT;
826		goto err;
827	}
828
829	ret = rds_pin_pages(args->local_addr, 1, &page, 1);
830	if (ret != 1)
831		goto err;
832	ret = 0;
833
834	sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
835
836	if (rm->atomic.op_notify || rm->atomic.op_recverr) {
837		/* We allocate an uninitialized notifier here, because
838		 * we don't want to do that in the completion handler. We
839		 * would have to use GFP_ATOMIC there, and don't want to deal
840		 * with failed allocations.
841		 */
842		rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL);
843		if (!rm->atomic.op_notifier) {
844			ret = -ENOMEM;
845			goto err;
846		}
847
848		rm->atomic.op_notifier->n_user_token = args->user_token;
849		rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
850	}
851
852	rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
853	rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
854
855	return ret;
856err:
857	if (page)
858		put_page(page);
 
859	kfree(rm->atomic.op_notifier);
860
861	return ret;
862}