1/*
  2 * Copyright (c) 2007, 2020 Oracle and/or its affiliates.
  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 detect duplicate keys on a socket?  hmm.
 44 *  - an rdma is an mlock, apply rlimit?
 45 */
 46
 47/*
 48 * get the number of pages by looking at the page indices that the start and
 49 * end addresses fall in.
 50 *
 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'.
 54 */
 55static unsigned int rds_pages_in_vec(struct rds_iovec *vec)
 56{
 57	if ((vec->addr + vec->bytes <= vec->addr) ||
 58	    (vec->bytes > (u64)UINT_MAX))
 59		return 0;
 60
 61	return ((vec->addr + vec->bytes + PAGE_SIZE - 1) >> PAGE_SHIFT) -
 62		(vec->addr >> PAGE_SHIFT);
 63}
 64
 65static struct rds_mr *rds_mr_tree_walk(struct rb_root *root, u64 key,
 66				       struct rds_mr *insert)
 67{
 68	struct rb_node **p = &root->rb_node;
 69	struct rb_node *parent = NULL;
 70	struct rds_mr *mr;
 71
 72	while (*p) {
 73		parent = *p;
 74		mr = rb_entry(parent, struct rds_mr, r_rb_node);
 75
 76		if (key < mr->r_key)
 77			p = &(*p)->rb_left;
 78		else if (key > mr->r_key)
 79			p = &(*p)->rb_right;
 80		else
 81			return mr;
 82	}
 83
 84	if (insert) {
 85		rb_link_node(&insert->r_rb_node, parent, p);
 86		rb_insert_color(&insert->r_rb_node, root);
 87		kref_get(&insert->r_kref);
 88	}
 89	return NULL;
 90}
 91
 92/*
 93 * Destroy the transport-specific part of a MR.
 94 */
 95static void rds_destroy_mr(struct rds_mr *mr)
 96{
 97	struct rds_sock *rs = mr->r_sock;
 98	void *trans_private = NULL;
 99	unsigned long flags;
100
101	rdsdebug("RDS: destroy mr key is %x refcnt %u\n",
102		 mr->r_key, kref_read(&mr->r_kref));
103
104	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
105	if (!RB_EMPTY_NODE(&mr->r_rb_node))
106		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
107	trans_private = mr->r_trans_private;
108	mr->r_trans_private = NULL;
109	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
110
111	if (trans_private)
112		mr->r_trans->free_mr(trans_private, mr->r_invalidate);
113}
114
115void __rds_put_mr_final(struct kref *kref)
116{
117	struct rds_mr *mr = container_of(kref, struct rds_mr, r_kref);
118
119	rds_destroy_mr(mr);
120	kfree(mr);
121}
122
123/*
124 * By the time this is called we can't have any more ioctls called on
125 * the socket so we don't need to worry about racing with others.
126 */
127void rds_rdma_drop_keys(struct rds_sock *rs)
128{
129	struct rds_mr *mr;
130	struct rb_node *node;
131	unsigned long flags;
132
133	/* Release any MRs associated with this socket */
134	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
135	while ((node = rb_first(&rs->rs_rdma_keys))) {
136		mr = rb_entry(node, struct rds_mr, r_rb_node);
137		if (mr->r_trans == rs->rs_transport)
138			mr->r_invalidate = 0;
139		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
140		RB_CLEAR_NODE(&mr->r_rb_node);
141		spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
142		kref_put(&mr->r_kref, __rds_put_mr_final);
143		spin_lock_irqsave(&rs->rs_rdma_lock, flags);
144	}
145	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
146
147	if (rs->rs_transport && rs->rs_transport->flush_mrs)
148		rs->rs_transport->flush_mrs();
149}
150
151/*
152 * Helper function to pin user pages.
153 */
154static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
155			struct page **pages, int write)
156{
157	unsigned int gup_flags = FOLL_LONGTERM;
158	int ret;
159
160	if (write)
161		gup_flags |= FOLL_WRITE;
162
163	ret = pin_user_pages_fast(user_addr, nr_pages, gup_flags, pages);
164	if (ret >= 0 && ret < nr_pages) {
165		unpin_user_pages(pages, ret);
166		ret = -EFAULT;
167	}
168
169	return ret;
170}
171
172static 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)
175{
176	struct rds_mr *mr = NULL, *found;
177	struct scatterlist *sg = NULL;
178	unsigned int nr_pages;
179	struct page **pages = NULL;
180	void *trans_private;
181	unsigned long flags;
182	rds_rdma_cookie_t cookie;
183	unsigned int nents = 0;
184	int need_odp = 0;
185	long i;
186	int ret;
187
188	if (ipv6_addr_any(&rs->rs_bound_addr) || !rs->rs_transport) {
189		ret = -ENOTCONN; /* XXX not a great errno */
190		goto out;
191	}
192
193	if (!rs->rs_transport->get_mr) {
194		ret = -EOPNOTSUPP;
195		goto out;
196	}
197
198	/* If the combination of the addr and size requested for this memory
199	 * region causes an integer overflow, return error.
200	 */
201	if (((args->vec.addr + args->vec.bytes) < args->vec.addr) ||
202	    PAGE_ALIGN(args->vec.addr + args->vec.bytes) <
203		    (args->vec.addr + args->vec.bytes)) {
204		ret = -EINVAL;
205		goto out;
206	}
207
208	if (!can_do_mlock()) {
209		ret = -EPERM;
210		goto out;
211	}
212
213	nr_pages = rds_pages_in_vec(&args->vec);
214	if (nr_pages == 0) {
215		ret = -EINVAL;
216		goto out;
217	}
218
219	/* Restrict the size of mr irrespective of underlying transport
220	 * To account for unaligned mr regions, subtract one from nr_pages
221	 */
222	if ((nr_pages - 1) > (RDS_MAX_MSG_SIZE >> PAGE_SHIFT)) {
223		ret = -EMSGSIZE;
224		goto out;
225	}
226
227	rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
228		args->vec.addr, args->vec.bytes, nr_pages);
229
230	/* XXX clamp nr_pages to limit the size of this alloc? */
231	pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
232	if (!pages) {
233		ret = -ENOMEM;
234		goto out;
235	}
236
237	mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
238	if (!mr) {
239		ret = -ENOMEM;
240		goto out;
241	}
242
243	kref_init(&mr->r_kref);
244	RB_CLEAR_NODE(&mr->r_rb_node);
245	mr->r_trans = rs->rs_transport;
246	mr->r_sock = rs;
247
248	if (args->flags & RDS_RDMA_USE_ONCE)
249		mr->r_use_once = 1;
250	if (args->flags & RDS_RDMA_INVALIDATE)
251		mr->r_invalidate = 1;
252	if (args->flags & RDS_RDMA_READWRITE)
253		mr->r_write = 1;
254
255	/*
256	 * Pin the pages that make up the user buffer and transfer the page
257	 * pointers to the mr's sg array.  We check to see if we've mapped
258	 * the whole region after transferring the partial page references
259	 * to the sg array so that we can have one page ref cleanup path.
260	 *
261	 * For now we have no flag that tells us whether the mapping is
262	 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
263	 * the zero page.
264	 */
265	ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
266	if (ret == -EOPNOTSUPP) {
267		need_odp = 1;
268	} else if (ret <= 0) {
269		goto out;
270	} else {
271		nents = ret;
272		sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
273		if (!sg) {
274			ret = -ENOMEM;
275			goto out;
276		}
277		WARN_ON(!nents);
278		sg_init_table(sg, nents);
279
280		/* Stick all pages into the scatterlist */
281		for (i = 0 ; i < nents; i++)
282			sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
283
284		rdsdebug("RDS: trans_private nents is %u\n", nents);
285	}
286	/* Obtain a transport specific MR. If this succeeds, the
287	 * s/g list is now owned by the MR.
288	 * Note that dma_map() implies that pending writes are
289	 * flushed to RAM, so no dma_sync is needed here. */
290	trans_private = rs->rs_transport->get_mr(
291		sg, nents, rs, &mr->r_key, cp ? cp->cp_conn : NULL,
292		args->vec.addr, args->vec.bytes,
293		need_odp ? ODP_ZEROBASED : ODP_NOT_NEEDED);
294
295	if (IS_ERR(trans_private)) {
296		/* In ODP case, we don't GUP pages, so don't need
297		 * to release anything.
298		 */
299		if (!need_odp) {
300			unpin_user_pages(pages, nr_pages);
301			kfree(sg);
302		}
303		ret = PTR_ERR(trans_private);
 
 
 
304		goto out;
305	}
306
307	mr->r_trans_private = trans_private;
308
309	rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
310	       mr->r_key, (void *)(unsigned long) args->cookie_addr);
311
312	/* The user may pass us an unaligned address, but we can only
313	 * map page aligned regions. So we keep the offset, and build
314	 * a 64bit cookie containing <R_Key, offset> and pass that
315	 * around. */
316	if (need_odp)
317		cookie = rds_rdma_make_cookie(mr->r_key, 0);
318	else
319		cookie = rds_rdma_make_cookie(mr->r_key,
320					      args->vec.addr & ~PAGE_MASK);
321	if (cookie_ret)
322		*cookie_ret = cookie;
323
324	if (args->cookie_addr &&
325	    put_user(cookie, (u64 __user *)(unsigned long)args->cookie_addr)) {
326		if (!need_odp) {
327			unpin_user_pages(pages, nr_pages);
328			kfree(sg);
329		}
330		ret = -EFAULT;
331		goto out;
332	}
333
334	/* Inserting the new MR into the rbtree bumps its
335	 * reference count. */
336	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
337	found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
338	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
339
340	BUG_ON(found && found != mr);
341
342	rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
343	if (mr_ret) {
344		kref_get(&mr->r_kref);
345		*mr_ret = mr;
346	}
347
348	ret = 0;
349out:
350	kfree(pages);
351	if (mr)
352		kref_put(&mr->r_kref, __rds_put_mr_final);
353	return ret;
354}
355
356int rds_get_mr(struct rds_sock *rs, sockptr_t optval, int optlen)
357{
358	struct rds_get_mr_args args;
359
360	if (optlen != sizeof(struct rds_get_mr_args))
361		return -EINVAL;
362
363	if (copy_from_sockptr(&args, optval, sizeof(struct rds_get_mr_args)))
364		return -EFAULT;
365
366	return __rds_rdma_map(rs, &args, NULL, NULL, NULL);
367}
368
369int rds_get_mr_for_dest(struct rds_sock *rs, sockptr_t optval, int optlen)
370{
371	struct rds_get_mr_for_dest_args args;
372	struct rds_get_mr_args new_args;
373
374	if (optlen != sizeof(struct rds_get_mr_for_dest_args))
375		return -EINVAL;
376
377	if (copy_from_sockptr(&args, optval,
378			   sizeof(struct rds_get_mr_for_dest_args)))
379		return -EFAULT;
380
381	/*
382	 * Initially, just behave like get_mr().
383	 * TODO: Implement get_mr as wrapper around this
384	 *	 and deprecate it.
385	 */
386	new_args.vec = args.vec;
387	new_args.cookie_addr = args.cookie_addr;
388	new_args.flags = args.flags;
389
390	return __rds_rdma_map(rs, &new_args, NULL, NULL, NULL);
391}
392
393/*
394 * Free the MR indicated by the given R_Key
395 */
396int rds_free_mr(struct rds_sock *rs, sockptr_t optval, int optlen)
397{
398	struct rds_free_mr_args args;
399	struct rds_mr *mr;
400	unsigned long flags;
401
402	if (optlen != sizeof(struct rds_free_mr_args))
403		return -EINVAL;
404
405	if (copy_from_sockptr(&args, optval, sizeof(struct rds_free_mr_args)))
406		return -EFAULT;
407
408	/* Special case - a null cookie means flush all unused MRs */
409	if (args.cookie == 0) {
410		if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
411			return -EINVAL;
412		rs->rs_transport->flush_mrs();
413		return 0;
414	}
415
416	/* Look up the MR given its R_key and remove it from the rbtree
417	 * so nobody else finds it.
418	 * This should also prevent races with rds_rdma_unuse.
419	 */
420	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
421	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL);
422	if (mr) {
423		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
424		RB_CLEAR_NODE(&mr->r_rb_node);
425		if (args.flags & RDS_RDMA_INVALIDATE)
426			mr->r_invalidate = 1;
427	}
428	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
429
430	if (!mr)
431		return -EINVAL;
432
433	kref_put(&mr->r_kref, __rds_put_mr_final);
434	return 0;
435}
436
437/*
438 * This is called when we receive an extension header that
439 * tells us this MR was used. It allows us to implement
440 * use_once semantics
441 */
442void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
443{
444	struct rds_mr *mr;
445	unsigned long flags;
446	int zot_me = 0;
447
448	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
449	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
450	if (!mr) {
451		pr_debug("rds: trying to unuse MR with unknown r_key %u!\n",
452			 r_key);
453		spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
454		return;
455	}
456
457	/* Get a reference so that the MR won't go away before calling
458	 * sync_mr() below.
459	 */
460	kref_get(&mr->r_kref);
461
462	/* If it is going to be freed, remove it from the tree now so
463	 * that no other thread can find it and free it.
464	 */
465	if (mr->r_use_once || force) {
466		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
467		RB_CLEAR_NODE(&mr->r_rb_node);
468		zot_me = 1;
469	}
470	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
471
472	/* May have to issue a dma_sync on this memory region.
473	 * Note we could avoid this if the operation was a RDMA READ,
474	 * but at this point we can't tell. */
475	if (mr->r_trans->sync_mr)
476		mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
477
478	/* Release the reference held above. */
479	kref_put(&mr->r_kref, __rds_put_mr_final);
480
481	/* If the MR was marked as invalidate, this will
482	 * trigger an async flush. */
483	if (zot_me)
484		kref_put(&mr->r_kref, __rds_put_mr_final);
485}
486
487void rds_rdma_free_op(struct rm_rdma_op *ro)
488{
489	unsigned int i;
490
491	if (ro->op_odp_mr) {
492		kref_put(&ro->op_odp_mr->r_kref, __rds_put_mr_final);
493	} else {
494		for (i = 0; i < ro->op_nents; i++) {
495			struct page *page = sg_page(&ro->op_sg[i]);
496
497			/* Mark page dirty if it was possibly modified, which
498			 * is the case for a RDMA_READ which copies from remote
499			 * to local memory
500			 */
501			unpin_user_pages_dirty_lock(&page, 1, !ro->op_write);
502		}
503	}
504
505	kfree(ro->op_notifier);
506	ro->op_notifier = NULL;
507	ro->op_active = 0;
508	ro->op_odp_mr = NULL;
509}
510
511void rds_atomic_free_op(struct rm_atomic_op *ao)
512{
513	struct page *page = sg_page(ao->op_sg);
514
515	/* Mark page dirty if it was possibly modified, which
516	 * is the case for a RDMA_READ which copies from remote
517	 * to local memory */
518	unpin_user_pages_dirty_lock(&page, 1, true);
519
520	kfree(ao->op_notifier);
521	ao->op_notifier = NULL;
522	ao->op_active = 0;
523}
524
525
526/*
527 * Count the number of pages needed to describe an incoming iovec array.
528 */
529static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs)
530{
531	int tot_pages = 0;
532	unsigned int nr_pages;
533	unsigned int i;
534
535	/* figure out the number of pages in the vector */
536	for (i = 0; i < nr_iovecs; i++) {
537		nr_pages = rds_pages_in_vec(&iov[i]);
538		if (nr_pages == 0)
539			return -EINVAL;
540
541		tot_pages += nr_pages;
542
543		/*
544		 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
545		 * so tot_pages cannot overflow without first going negative.
546		 */
547		if (tot_pages < 0)
548			return -EINVAL;
549	}
550
551	return tot_pages;
552}
553
554int rds_rdma_extra_size(struct rds_rdma_args *args,
555			struct rds_iov_vector *iov)
556{
557	struct rds_iovec *vec;
558	struct rds_iovec __user *local_vec;
559	int tot_pages = 0;
560	unsigned int nr_pages;
561	unsigned int i;
562
563	local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
564
565	if (args->nr_local == 0)
566		return -EINVAL;
567
 
 
 
568	iov->iov = kcalloc(args->nr_local,
569			   sizeof(struct rds_iovec),
570			   GFP_KERNEL);
571	if (!iov->iov)
572		return -ENOMEM;
573
574	vec = &iov->iov[0];
575
576	if (copy_from_user(vec, local_vec, args->nr_local *
577			   sizeof(struct rds_iovec)))
578		return -EFAULT;
579	iov->len = args->nr_local;
580
581	/* figure out the number of pages in the vector */
582	for (i = 0; i < args->nr_local; i++, vec++) {
583
584		nr_pages = rds_pages_in_vec(vec);
585		if (nr_pages == 0)
586			return -EINVAL;
587
588		tot_pages += nr_pages;
589
590		/*
591		 * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
592		 * so tot_pages cannot overflow without first going negative.
593		 */
594		if (tot_pages < 0)
595			return -EINVAL;
596	}
597
598	return tot_pages * sizeof(struct scatterlist);
599}
600
601/*
602 * The application asks for a RDMA transfer.
603 * Extract all arguments and set up the rdma_op
604 */
605int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
606		       struct cmsghdr *cmsg,
607		       struct rds_iov_vector *vec)
608{
609	struct rds_rdma_args *args;
610	struct rm_rdma_op *op = &rm->rdma;
611	int nr_pages;
612	unsigned int nr_bytes;
613	struct page **pages = NULL;
614	struct rds_iovec *iovs;
615	unsigned int i, j;
616	int ret = 0;
617	bool odp_supported = true;
618
619	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
620	    || rm->rdma.op_active)
621		return -EINVAL;
622
623	args = CMSG_DATA(cmsg);
624
625	if (ipv6_addr_any(&rs->rs_bound_addr)) {
626		ret = -ENOTCONN; /* XXX not a great errno */
627		goto out_ret;
628	}
629
630	if (args->nr_local > UIO_MAXIOV) {
631		ret = -EMSGSIZE;
632		goto out_ret;
633	}
634
635	if (vec->len != args->nr_local) {
636		ret = -EINVAL;
637		goto out_ret;
638	}
639	/* odp-mr is not supported for multiple requests within one message */
640	if (args->nr_local != 1)
641		odp_supported = false;
642
643	iovs = vec->iov;
644
645	nr_pages = rds_rdma_pages(iovs, args->nr_local);
646	if (nr_pages < 0) {
647		ret = -EINVAL;
648		goto out_ret;
649	}
650
651	pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
652	if (!pages) {
653		ret = -ENOMEM;
654		goto out_ret;
655	}
656
657	op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
658	op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
659	op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
660	op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
661	op->op_active = 1;
662	op->op_recverr = rs->rs_recverr;
663	op->op_odp_mr = NULL;
664
665	WARN_ON(!nr_pages);
666	op->op_sg = rds_message_alloc_sgs(rm, nr_pages);
667	if (IS_ERR(op->op_sg)) {
668		ret = PTR_ERR(op->op_sg);
669		goto out_pages;
670	}
671
672	if (op->op_notify || op->op_recverr) {
673		/* We allocate an uninitialized notifier here, because
674		 * we don't want to do that in the completion handler. We
675		 * would have to use GFP_ATOMIC there, and don't want to deal
676		 * with failed allocations.
677		 */
678		op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
679		if (!op->op_notifier) {
680			ret = -ENOMEM;
681			goto out_pages;
682		}
683		op->op_notifier->n_user_token = args->user_token;
684		op->op_notifier->n_status = RDS_RDMA_SUCCESS;
685	}
686
687	/* The cookie contains the R_Key of the remote memory region, and
688	 * optionally an offset into it. This is how we implement RDMA into
689	 * unaligned memory.
690	 * When setting up the RDMA, we need to add that offset to the
691	 * destination address (which is really an offset into the MR)
692	 * FIXME: We may want to move this into ib_rdma.c
693	 */
694	op->op_rkey = rds_rdma_cookie_key(args->cookie);
695	op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
696
697	nr_bytes = 0;
698
699	rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
700	       (unsigned long long)args->nr_local,
701	       (unsigned long long)args->remote_vec.addr,
702	       op->op_rkey);
703
704	for (i = 0; i < args->nr_local; i++) {
705		struct rds_iovec *iov = &iovs[i];
706		/* don't need to check, rds_rdma_pages() verified nr will be +nonzero */
707		unsigned int nr = rds_pages_in_vec(iov);
708
709		rs->rs_user_addr = iov->addr;
710		rs->rs_user_bytes = iov->bytes;
711
712		/* If it's a WRITE operation, we want to pin the pages for reading.
713		 * If it's a READ operation, we need to pin the pages for writing.
714		 */
715		ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
716		if ((!odp_supported && ret <= 0) ||
717		    (odp_supported && ret <= 0 && ret != -EOPNOTSUPP))
718			goto out_pages;
719
720		if (ret == -EOPNOTSUPP) {
721			struct rds_mr *local_odp_mr;
722
723			if (!rs->rs_transport->get_mr) {
724				ret = -EOPNOTSUPP;
725				goto out_pages;
726			}
727			local_odp_mr =
728				kzalloc(sizeof(*local_odp_mr), GFP_KERNEL);
729			if (!local_odp_mr) {
730				ret = -ENOMEM;
731				goto out_pages;
732			}
733			RB_CLEAR_NODE(&local_odp_mr->r_rb_node);
734			kref_init(&local_odp_mr->r_kref);
735			local_odp_mr->r_trans = rs->rs_transport;
736			local_odp_mr->r_sock = rs;
737			local_odp_mr->r_trans_private =
738				rs->rs_transport->get_mr(
739					NULL, 0, rs, &local_odp_mr->r_key, NULL,
740					iov->addr, iov->bytes, ODP_VIRTUAL);
741			if (IS_ERR(local_odp_mr->r_trans_private)) {
742				ret = IS_ERR(local_odp_mr->r_trans_private);
743				rdsdebug("get_mr ret %d %p\"", ret,
744					 local_odp_mr->r_trans_private);
745				kfree(local_odp_mr);
746				ret = -EOPNOTSUPP;
747				goto out_pages;
748			}
749			rdsdebug("Need odp; local_odp_mr %p trans_private %p\n",
750				 local_odp_mr, local_odp_mr->r_trans_private);
751			op->op_odp_mr = local_odp_mr;
752			op->op_odp_addr = iov->addr;
753		}
754
755		rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
756			 nr_bytes, nr, iov->bytes, iov->addr);
757
758		nr_bytes += iov->bytes;
759
760		for (j = 0; j < nr; j++) {
761			unsigned int offset = iov->addr & ~PAGE_MASK;
762			struct scatterlist *sg;
763
764			sg = &op->op_sg[op->op_nents + j];
765			sg_set_page(sg, pages[j],
766					min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
767					offset);
768
769			sg_dma_len(sg) = sg->length;
770			rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n",
771			       sg->offset, sg->length, iov->addr, iov->bytes);
772
773			iov->addr += sg->length;
774			iov->bytes -= sg->length;
775		}
776
777		op->op_nents += nr;
778	}
779
780	if (nr_bytes > args->remote_vec.bytes) {
781		rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
782				nr_bytes,
783				(unsigned int) args->remote_vec.bytes);
784		ret = -EINVAL;
785		goto out_pages;
786	}
787	op->op_bytes = nr_bytes;
788	ret = 0;
789
790out_pages:
791	kfree(pages);
792out_ret:
793	if (ret)
794		rds_rdma_free_op(op);
795	else
796		rds_stats_inc(s_send_rdma);
797
798	return ret;
799}
800
801/*
802 * The application wants us to pass an RDMA destination (aka MR)
803 * to the remote
804 */
805int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
806			  struct cmsghdr *cmsg)
807{
808	unsigned long flags;
809	struct rds_mr *mr;
810	u32 r_key;
811	int err = 0;
812
813	if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
814	    rm->m_rdma_cookie != 0)
815		return -EINVAL;
816
817	memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
818
819	/* We are reusing a previously mapped MR here. Most likely, the
820	 * application has written to the buffer, so we need to explicitly
821	 * flush those writes to RAM. Otherwise the HCA may not see them
822	 * when doing a DMA from that buffer.
823	 */
824	r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
825
826	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
827	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
828	if (!mr)
829		err = -EINVAL;	/* invalid r_key */
830	else
831		kref_get(&mr->r_kref);
832	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
833
834	if (mr) {
835		mr->r_trans->sync_mr(mr->r_trans_private,
836				     DMA_TO_DEVICE);
837		rm->rdma.op_rdma_mr = mr;
838	}
839	return err;
840}
841
842/*
843 * The application passes us an address range it wants to enable RDMA
844 * to/from. We map the area, and save the <R_Key,offset> pair
845 * in rm->m_rdma_cookie. This causes it to be sent along to the peer
846 * in an extension header.
847 */
848int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
849			  struct cmsghdr *cmsg)
850{
851	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
852	    rm->m_rdma_cookie != 0)
853		return -EINVAL;
854
855	return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie,
856			      &rm->rdma.op_rdma_mr, rm->m_conn_path);
857}
858
859/*
860 * Fill in rds_message for an atomic request.
861 */
862int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
863		    struct cmsghdr *cmsg)
864{
865	struct page *page = NULL;
866	struct rds_atomic_args *args;
867	int ret = 0;
868
869	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
870	 || rm->atomic.op_active)
871		return -EINVAL;
872
873	args = CMSG_DATA(cmsg);
874
875	/* Nonmasked & masked cmsg ops converted to masked hw ops */
876	switch (cmsg->cmsg_type) {
877	case RDS_CMSG_ATOMIC_FADD:
878		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
879		rm->atomic.op_m_fadd.add = args->fadd.add;
880		rm->atomic.op_m_fadd.nocarry_mask = 0;
881		break;
882	case RDS_CMSG_MASKED_ATOMIC_FADD:
883		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
884		rm->atomic.op_m_fadd.add = args->m_fadd.add;
885		rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
886		break;
887	case RDS_CMSG_ATOMIC_CSWP:
888		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
889		rm->atomic.op_m_cswp.compare = args->cswp.compare;
890		rm->atomic.op_m_cswp.swap = args->cswp.swap;
891		rm->atomic.op_m_cswp.compare_mask = ~0;
892		rm->atomic.op_m_cswp.swap_mask = ~0;
893		break;
894	case RDS_CMSG_MASKED_ATOMIC_CSWP:
895		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
896		rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
897		rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
898		rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
899		rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
900		break;
901	default:
902		BUG(); /* should never happen */
903	}
904
905	rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
906	rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
907	rm->atomic.op_active = 1;
908	rm->atomic.op_recverr = rs->rs_recverr;
909	rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1);
910	if (IS_ERR(rm->atomic.op_sg)) {
911		ret = PTR_ERR(rm->atomic.op_sg);
912		goto err;
913	}
914
915	/* verify 8 byte-aligned */
916	if (args->local_addr & 0x7) {
917		ret = -EFAULT;
918		goto err;
919	}
920
921	ret = rds_pin_pages(args->local_addr, 1, &page, 1);
922	if (ret != 1)
923		goto err;
924	ret = 0;
925
926	sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
927
928	if (rm->atomic.op_notify || rm->atomic.op_recverr) {
929		/* We allocate an uninitialized notifier here, because
930		 * we don't want to do that in the completion handler. We
931		 * would have to use GFP_ATOMIC there, and don't want to deal
932		 * with failed allocations.
933		 */
934		rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL);
935		if (!rm->atomic.op_notifier) {
936			ret = -ENOMEM;
937			goto err;
938		}
939
940		rm->atomic.op_notifier->n_user_token = args->user_token;
941		rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
942	}
943
944	rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
945	rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
946
947	return ret;
948err:
949	if (page)
950		unpin_user_page(page);
951	rm->atomic.op_active = 0;
952	kfree(rm->atomic.op_notifier);
953
954	return ret;
955}