1/*
2 * Copyright (c) 2006 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/kernel.h>
34#include <linux/slab.h>
35#include <linux/ratelimit.h>
36
37#include "rds.h"
38#include "iw.h"
39
40
41/*
42 * This is stored as mr->r_trans_private.
43 */
44struct rds_iw_mr {
45	struct rds_iw_device	*device;
46	struct rds_iw_mr_pool	*pool;
47	struct rdma_cm_id	*cm_id;
48
49	struct ib_mr	*mr;
50	struct ib_fast_reg_page_list *page_list;
51
52	struct rds_iw_mapping	mapping;
53	unsigned char		remap_count;
54};
55
56/*
57 * Our own little MR pool
58 */
59struct rds_iw_mr_pool {
60	struct rds_iw_device	*device;		/* back ptr to the device that owns us */
61
62	struct mutex		flush_lock;		/* serialize fmr invalidate */
63	struct work_struct	flush_worker;		/* flush worker */
64
65	spinlock_t		list_lock;		/* protect variables below */
66	atomic_t		item_count;		/* total # of MRs */
67	atomic_t		dirty_count;		/* # dirty of MRs */
68	struct list_head	dirty_list;		/* dirty mappings */
69	struct list_head	clean_list;		/* unused & unamapped MRs */
70	atomic_t		free_pinned;		/* memory pinned by free MRs */
71	unsigned long		max_message_size;	/* in pages */
72	unsigned long		max_items;
73	unsigned long		max_items_soft;
74	unsigned long		max_free_pinned;
75	int			max_pages;
76};
77
78static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all);
79static void rds_iw_mr_pool_flush_worker(struct work_struct *work);
80static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
81static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
82			  struct rds_iw_mr *ibmr,
83			  struct scatterlist *sg, unsigned int nents);
84static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
85static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
86			struct list_head *unmap_list,
87			struct list_head *kill_list,
88			int *unpinned);
89static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
90
91static int rds_iw_get_device(struct rds_sock *rs, struct rds_iw_device **rds_iwdev, struct rdma_cm_id **cm_id)
92{
93	struct rds_iw_device *iwdev;
94	struct rds_iw_cm_id *i_cm_id;
95
96	*rds_iwdev = NULL;
97	*cm_id = NULL;
98
99	list_for_each_entry(iwdev, &rds_iw_devices, list) {
100		spin_lock_irq(&iwdev->spinlock);
101		list_for_each_entry(i_cm_id, &iwdev->cm_id_list, list) {
102			struct sockaddr_in *src_addr, *dst_addr;
103
104			src_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.src_addr;
105			dst_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.dst_addr;
106
107			rdsdebug("local ipaddr = %x port %d, "
108				 "remote ipaddr = %x port %d"
109				 "..looking for %x port %d, "
110				 "remote ipaddr = %x port %d\n",
111				src_addr->sin_addr.s_addr,
112				src_addr->sin_port,
113				dst_addr->sin_addr.s_addr,
114				dst_addr->sin_port,
115				rs->rs_bound_addr,
116				rs->rs_bound_port,
117				rs->rs_conn_addr,
118				rs->rs_conn_port);
119#ifdef WORKING_TUPLE_DETECTION
120			if (src_addr->sin_addr.s_addr == rs->rs_bound_addr &&
121			    src_addr->sin_port == rs->rs_bound_port &&
122			    dst_addr->sin_addr.s_addr == rs->rs_conn_addr &&
123			    dst_addr->sin_port == rs->rs_conn_port) {
124#else
125			/* FIXME - needs to compare the local and remote
126			 * ipaddr/port tuple, but the ipaddr is the only
127			 * available information in the rds_sock (as the rest are
128			 * zero'ed.  It doesn't appear to be properly populated
129			 * during connection setup...
130			 */
131			if (src_addr->sin_addr.s_addr == rs->rs_bound_addr) {
132#endif
133				spin_unlock_irq(&iwdev->spinlock);
134				*rds_iwdev = iwdev;
135				*cm_id = i_cm_id->cm_id;
136				return 0;
137			}
138		}
139		spin_unlock_irq(&iwdev->spinlock);
140	}
141
142	return 1;
143}
144
145static int rds_iw_add_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
146{
147	struct rds_iw_cm_id *i_cm_id;
148
149	i_cm_id = kmalloc(sizeof *i_cm_id, GFP_KERNEL);
150	if (!i_cm_id)
151		return -ENOMEM;
152
153	i_cm_id->cm_id = cm_id;
154
155	spin_lock_irq(&rds_iwdev->spinlock);
156	list_add_tail(&i_cm_id->list, &rds_iwdev->cm_id_list);
157	spin_unlock_irq(&rds_iwdev->spinlock);
158
159	return 0;
160}
161
162static void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev,
163				struct rdma_cm_id *cm_id)
164{
165	struct rds_iw_cm_id *i_cm_id;
166
167	spin_lock_irq(&rds_iwdev->spinlock);
168	list_for_each_entry(i_cm_id, &rds_iwdev->cm_id_list, list) {
169		if (i_cm_id->cm_id == cm_id) {
170			list_del(&i_cm_id->list);
171			kfree(i_cm_id);
172			break;
173		}
174	}
175	spin_unlock_irq(&rds_iwdev->spinlock);
176}
177
178
179int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
180{
181	struct sockaddr_in *src_addr, *dst_addr;
182	struct rds_iw_device *rds_iwdev_old;
183	struct rds_sock rs;
184	struct rdma_cm_id *pcm_id;
185	int rc;
186
187	src_addr = (struct sockaddr_in *)&cm_id->route.addr.src_addr;
188	dst_addr = (struct sockaddr_in *)&cm_id->route.addr.dst_addr;
189
190	rs.rs_bound_addr = src_addr->sin_addr.s_addr;
191	rs.rs_bound_port = src_addr->sin_port;
192	rs.rs_conn_addr = dst_addr->sin_addr.s_addr;
193	rs.rs_conn_port = dst_addr->sin_port;
194
195	rc = rds_iw_get_device(&rs, &rds_iwdev_old, &pcm_id);
196	if (rc)
197		rds_iw_remove_cm_id(rds_iwdev, cm_id);
198
199	return rds_iw_add_cm_id(rds_iwdev, cm_id);
200}
201
202void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
203{
204	struct rds_iw_connection *ic = conn->c_transport_data;
205
206	/* conn was previously on the nodev_conns_list */
207	spin_lock_irq(&iw_nodev_conns_lock);
208	BUG_ON(list_empty(&iw_nodev_conns));
209	BUG_ON(list_empty(&ic->iw_node));
210	list_del(&ic->iw_node);
211
212	spin_lock(&rds_iwdev->spinlock);
213	list_add_tail(&ic->iw_node, &rds_iwdev->conn_list);
214	spin_unlock(&rds_iwdev->spinlock);
215	spin_unlock_irq(&iw_nodev_conns_lock);
216
217	ic->rds_iwdev = rds_iwdev;
218}
219
220void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
221{
222	struct rds_iw_connection *ic = conn->c_transport_data;
223
224	/* place conn on nodev_conns_list */
225	spin_lock(&iw_nodev_conns_lock);
226
227	spin_lock_irq(&rds_iwdev->spinlock);
228	BUG_ON(list_empty(&ic->iw_node));
229	list_del(&ic->iw_node);
230	spin_unlock_irq(&rds_iwdev->spinlock);
231
232	list_add_tail(&ic->iw_node, &iw_nodev_conns);
233
234	spin_unlock(&iw_nodev_conns_lock);
235
236	rds_iw_remove_cm_id(ic->rds_iwdev, ic->i_cm_id);
237	ic->rds_iwdev = NULL;
238}
239
240void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock)
241{
242	struct rds_iw_connection *ic, *_ic;
243	LIST_HEAD(tmp_list);
244
245	/* avoid calling conn_destroy with irqs off */
246	spin_lock_irq(list_lock);
247	list_splice(list, &tmp_list);
248	INIT_LIST_HEAD(list);
249	spin_unlock_irq(list_lock);
250
251	list_for_each_entry_safe(ic, _ic, &tmp_list, iw_node)
252		rds_conn_destroy(ic->conn);
253}
254
255static void rds_iw_set_scatterlist(struct rds_iw_scatterlist *sg,
256		struct scatterlist *list, unsigned int sg_len)
257{
258	sg->list = list;
259	sg->len = sg_len;
260	sg->dma_len = 0;
261	sg->dma_npages = 0;
262	sg->bytes = 0;
263}
264
265static u64 *rds_iw_map_scatterlist(struct rds_iw_device *rds_iwdev,
266			struct rds_iw_scatterlist *sg)
267{
268	struct ib_device *dev = rds_iwdev->dev;
269	u64 *dma_pages = NULL;
270	int i, j, ret;
271
272	WARN_ON(sg->dma_len);
273
274	sg->dma_len = ib_dma_map_sg(dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
275	if (unlikely(!sg->dma_len)) {
276		printk(KERN_WARNING "RDS/IW: dma_map_sg failed!\n");
277		return ERR_PTR(-EBUSY);
278	}
279
280	sg->bytes = 0;
281	sg->dma_npages = 0;
282
283	ret = -EINVAL;
284	for (i = 0; i < sg->dma_len; ++i) {
285		unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
286		u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
287		u64 end_addr;
288
289		sg->bytes += dma_len;
290
291		end_addr = dma_addr + dma_len;
292		if (dma_addr & PAGE_MASK) {
293			if (i > 0)
294				goto out_unmap;
295			dma_addr &= ~PAGE_MASK;
296		}
297		if (end_addr & PAGE_MASK) {
298			if (i < sg->dma_len - 1)
299				goto out_unmap;
300			end_addr = (end_addr + PAGE_MASK) & ~PAGE_MASK;
301		}
302
303		sg->dma_npages += (end_addr - dma_addr) >> PAGE_SHIFT;
304	}
305
306	/* Now gather the dma addrs into one list */
307	if (sg->dma_npages > fastreg_message_size)
308		goto out_unmap;
309
310	dma_pages = kmalloc(sizeof(u64) * sg->dma_npages, GFP_ATOMIC);
311	if (!dma_pages) {
312		ret = -ENOMEM;
313		goto out_unmap;
314	}
315
316	for (i = j = 0; i < sg->dma_len; ++i) {
317		unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
318		u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
319		u64 end_addr;
320
321		end_addr = dma_addr + dma_len;
322		dma_addr &= ~PAGE_MASK;
323		for (; dma_addr < end_addr; dma_addr += PAGE_SIZE)
324			dma_pages[j++] = dma_addr;
325		BUG_ON(j > sg->dma_npages);
326	}
327
328	return dma_pages;
329
330out_unmap:
331	ib_dma_unmap_sg(rds_iwdev->dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
332	sg->dma_len = 0;
333	kfree(dma_pages);
334	return ERR_PTR(ret);
335}
336
337
338struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *rds_iwdev)
339{
340	struct rds_iw_mr_pool *pool;
341
342	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
343	if (!pool) {
344		printk(KERN_WARNING "RDS/IW: rds_iw_create_mr_pool alloc error\n");
345		return ERR_PTR(-ENOMEM);
346	}
347
348	pool->device = rds_iwdev;
349	INIT_LIST_HEAD(&pool->dirty_list);
350	INIT_LIST_HEAD(&pool->clean_list);
351	mutex_init(&pool->flush_lock);
352	spin_lock_init(&pool->list_lock);
353	INIT_WORK(&pool->flush_worker, rds_iw_mr_pool_flush_worker);
354
355	pool->max_message_size = fastreg_message_size;
356	pool->max_items = fastreg_pool_size;
357	pool->max_free_pinned = pool->max_items * pool->max_message_size / 4;
358	pool->max_pages = fastreg_message_size;
359
360	/* We never allow more than max_items MRs to be allocated.
361	 * When we exceed more than max_items_soft, we start freeing
362	 * items more aggressively.
363	 * Make sure that max_items > max_items_soft > max_items / 2
364	 */
365	pool->max_items_soft = pool->max_items * 3 / 4;
366
367	return pool;
368}
369
370void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo)
371{
372	struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
373
374	iinfo->rdma_mr_max = pool->max_items;
375	iinfo->rdma_mr_size = pool->max_pages;
376}
377
378void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *pool)
379{
380	flush_workqueue(rds_wq);
381	rds_iw_flush_mr_pool(pool, 1);
382	BUG_ON(atomic_read(&pool->item_count));
383	BUG_ON(atomic_read(&pool->free_pinned));
384	kfree(pool);
385}
386
387static inline struct rds_iw_mr *rds_iw_reuse_fmr(struct rds_iw_mr_pool *pool)
388{
389	struct rds_iw_mr *ibmr = NULL;
390	unsigned long flags;
391
392	spin_lock_irqsave(&pool->list_lock, flags);
393	if (!list_empty(&pool->clean_list)) {
394		ibmr = list_entry(pool->clean_list.next, struct rds_iw_mr, mapping.m_list);
395		list_del_init(&ibmr->mapping.m_list);
396	}
397	spin_unlock_irqrestore(&pool->list_lock, flags);
398
399	return ibmr;
400}
401
402static struct rds_iw_mr *rds_iw_alloc_mr(struct rds_iw_device *rds_iwdev)
403{
404	struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
405	struct rds_iw_mr *ibmr = NULL;
406	int err = 0, iter = 0;
407
408	while (1) {
409		ibmr = rds_iw_reuse_fmr(pool);
410		if (ibmr)
411			return ibmr;
412
413		/* No clean MRs - now we have the choice of either
414		 * allocating a fresh MR up to the limit imposed by the
415		 * driver, or flush any dirty unused MRs.
416		 * We try to avoid stalling in the send path if possible,
417		 * so we allocate as long as we're allowed to.
418		 *
419		 * We're fussy with enforcing the FMR limit, though. If the driver
420		 * tells us we can't use more than N fmrs, we shouldn't start
421		 * arguing with it */
422		if (atomic_inc_return(&pool->item_count) <= pool->max_items)
423			break;
424
425		atomic_dec(&pool->item_count);
426
427		if (++iter > 2) {
428			rds_iw_stats_inc(s_iw_rdma_mr_pool_depleted);
429			return ERR_PTR(-EAGAIN);
430		}
431
432		/* We do have some empty MRs. Flush them out. */
433		rds_iw_stats_inc(s_iw_rdma_mr_pool_wait);
434		rds_iw_flush_mr_pool(pool, 0);
435	}
436
437	ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
438	if (!ibmr) {
439		err = -ENOMEM;
440		goto out_no_cigar;
441	}
442
443	spin_lock_init(&ibmr->mapping.m_lock);
444	INIT_LIST_HEAD(&ibmr->mapping.m_list);
445	ibmr->mapping.m_mr = ibmr;
446
447	err = rds_iw_init_fastreg(pool, ibmr);
448	if (err)
449		goto out_no_cigar;
450
451	rds_iw_stats_inc(s_iw_rdma_mr_alloc);
452	return ibmr;
453
454out_no_cigar:
455	if (ibmr) {
456		rds_iw_destroy_fastreg(pool, ibmr);
457		kfree(ibmr);
458	}
459	atomic_dec(&pool->item_count);
460	return ERR_PTR(err);
461}
462
463void rds_iw_sync_mr(void *trans_private, int direction)
464{
465	struct rds_iw_mr *ibmr = trans_private;
466	struct rds_iw_device *rds_iwdev = ibmr->device;
467
468	switch (direction) {
469	case DMA_FROM_DEVICE:
470		ib_dma_sync_sg_for_cpu(rds_iwdev->dev, ibmr->mapping.m_sg.list,
471			ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
472		break;
473	case DMA_TO_DEVICE:
474		ib_dma_sync_sg_for_device(rds_iwdev->dev, ibmr->mapping.m_sg.list,
475			ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
476		break;
477	}
478}
479
480/*
481 * Flush our pool of MRs.
482 * At a minimum, all currently unused MRs are unmapped.
483 * If the number of MRs allocated exceeds the limit, we also try
484 * to free as many MRs as needed to get back to this limit.
485 */
486static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all)
487{
488	struct rds_iw_mr *ibmr, *next;
489	LIST_HEAD(unmap_list);
490	LIST_HEAD(kill_list);
491	unsigned long flags;
492	unsigned int nfreed = 0, ncleaned = 0, unpinned = 0;
493	int ret = 0;
494
495	rds_iw_stats_inc(s_iw_rdma_mr_pool_flush);
496
497	mutex_lock(&pool->flush_lock);
498
499	spin_lock_irqsave(&pool->list_lock, flags);
500	/* Get the list of all mappings to be destroyed */
501	list_splice_init(&pool->dirty_list, &unmap_list);
502	if (free_all)
503		list_splice_init(&pool->clean_list, &kill_list);
504	spin_unlock_irqrestore(&pool->list_lock, flags);
505
506	/* Batched invalidate of dirty MRs.
507	 * For FMR based MRs, the mappings on the unmap list are
508	 * actually members of an ibmr (ibmr->mapping). They either
509	 * migrate to the kill_list, or have been cleaned and should be
510	 * moved to the clean_list.
511	 * For fastregs, they will be dynamically allocated, and
512	 * will be destroyed by the unmap function.
513	 */
514	if (!list_empty(&unmap_list)) {
515		ncleaned = rds_iw_unmap_fastreg_list(pool, &unmap_list,
516						     &kill_list, &unpinned);
517		/* If we've been asked to destroy all MRs, move those
518		 * that were simply cleaned to the kill list */
519		if (free_all)
520			list_splice_init(&unmap_list, &kill_list);
521	}
522
523	/* Destroy any MRs that are past their best before date */
524	list_for_each_entry_safe(ibmr, next, &kill_list, mapping.m_list) {
525		rds_iw_stats_inc(s_iw_rdma_mr_free);
526		list_del(&ibmr->mapping.m_list);
527		rds_iw_destroy_fastreg(pool, ibmr);
528		kfree(ibmr);
529		nfreed++;
530	}
531
532	/* Anything that remains are laundered ibmrs, which we can add
533	 * back to the clean list. */
534	if (!list_empty(&unmap_list)) {
535		spin_lock_irqsave(&pool->list_lock, flags);
536		list_splice(&unmap_list, &pool->clean_list);
537		spin_unlock_irqrestore(&pool->list_lock, flags);
538	}
539
540	atomic_sub(unpinned, &pool->free_pinned);
541	atomic_sub(ncleaned, &pool->dirty_count);
542	atomic_sub(nfreed, &pool->item_count);
543
544	mutex_unlock(&pool->flush_lock);
545	return ret;
546}
547
548static void rds_iw_mr_pool_flush_worker(struct work_struct *work)
549{
550	struct rds_iw_mr_pool *pool = container_of(work, struct rds_iw_mr_pool, flush_worker);
551
552	rds_iw_flush_mr_pool(pool, 0);
553}
554
555void rds_iw_free_mr(void *trans_private, int invalidate)
556{
557	struct rds_iw_mr *ibmr = trans_private;
558	struct rds_iw_mr_pool *pool = ibmr->device->mr_pool;
559
560	rdsdebug("RDS/IW: free_mr nents %u\n", ibmr->mapping.m_sg.len);
561	if (!pool)
562		return;
563
564	/* Return it to the pool's free list */
565	rds_iw_free_fastreg(pool, ibmr);
566
567	/* If we've pinned too many pages, request a flush */
568	if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
569	    atomic_read(&pool->dirty_count) >= pool->max_items / 10)
570		queue_work(rds_wq, &pool->flush_worker);
571
572	if (invalidate) {
573		if (likely(!in_interrupt())) {
574			rds_iw_flush_mr_pool(pool, 0);
575		} else {
576			/* We get here if the user created a MR marked
577			 * as use_once and invalidate at the same time. */
578			queue_work(rds_wq, &pool->flush_worker);
579		}
580	}
581}
582
583void rds_iw_flush_mrs(void)
584{
585	struct rds_iw_device *rds_iwdev;
586
587	list_for_each_entry(rds_iwdev, &rds_iw_devices, list) {
588		struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
589
590		if (pool)
591			rds_iw_flush_mr_pool(pool, 0);
592	}
593}
594
595void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
596		    struct rds_sock *rs, u32 *key_ret)
597{
598	struct rds_iw_device *rds_iwdev;
599	struct rds_iw_mr *ibmr = NULL;
600	struct rdma_cm_id *cm_id;
601	int ret;
602
603	ret = rds_iw_get_device(rs, &rds_iwdev, &cm_id);
604	if (ret || !cm_id) {
605		ret = -ENODEV;
606		goto out;
607	}
608
609	if (!rds_iwdev->mr_pool) {
610		ret = -ENODEV;
611		goto out;
612	}
613
614	ibmr = rds_iw_alloc_mr(rds_iwdev);
615	if (IS_ERR(ibmr))
616		return ibmr;
617
618	ibmr->cm_id = cm_id;
619	ibmr->device = rds_iwdev;
620
621	ret = rds_iw_map_fastreg(rds_iwdev->mr_pool, ibmr, sg, nents);
622	if (ret == 0)
623		*key_ret = ibmr->mr->rkey;
624	else
625		printk(KERN_WARNING "RDS/IW: failed to map mr (errno=%d)\n", ret);
626
627out:
628	if (ret) {
629		if (ibmr)
630			rds_iw_free_mr(ibmr, 0);
631		ibmr = ERR_PTR(ret);
632	}
633	return ibmr;
634}
635
636/*
637 * iWARP fastreg handling
638 *
639 * The life cycle of a fastreg registration is a bit different from
640 * FMRs.
641 * The idea behind fastreg is to have one MR, to which we bind different
642 * mappings over time. To avoid stalling on the expensive map and invalidate
643 * operations, these operations are pipelined on the same send queue on
644 * which we want to send the message containing the r_key.
645 *
646 * This creates a bit of a problem for us, as we do not have the destination
647 * IP in GET_MR, so the connection must be setup prior to the GET_MR call for
648 * RDMA to be correctly setup.  If a fastreg request is present, rds_iw_xmit
649 * will try to queue a LOCAL_INV (if needed) and a FAST_REG_MR work request
650 * before queuing the SEND. When completions for these arrive, they are
651 * dispatched to the MR has a bit set showing that RDMa can be performed.
652 *
653 * There is another interesting aspect that's related to invalidation.
654 * The application can request that a mapping is invalidated in FREE_MR.
655 * The expectation there is that this invalidation step includes ALL
656 * PREVIOUSLY FREED MRs.
657 */
658static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool,
659				struct rds_iw_mr *ibmr)
660{
661	struct rds_iw_device *rds_iwdev = pool->device;
662	struct ib_fast_reg_page_list *page_list = NULL;
663	struct ib_mr *mr;
664	int err;
665
666	mr = ib_alloc_fast_reg_mr(rds_iwdev->pd, pool->max_message_size);
667	if (IS_ERR(mr)) {
668		err = PTR_ERR(mr);
669
670		printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_mr failed (err=%d)\n", err);
671		return err;
672	}
673
674	/* FIXME - this is overkill, but mapping->m_sg.dma_len/mapping->m_sg.dma_npages
675	 * is not filled in.
676	 */
677	page_list = ib_alloc_fast_reg_page_list(rds_iwdev->dev, pool->max_message_size);
678	if (IS_ERR(page_list)) {
679		err = PTR_ERR(page_list);
680
681		printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed (err=%d)\n", err);
682		ib_dereg_mr(mr);
683		return err;
684	}
685
686	ibmr->page_list = page_list;
687	ibmr->mr = mr;
688	return 0;
689}
690
691static int rds_iw_rdma_build_fastreg(struct rds_iw_mapping *mapping)
692{
693	struct rds_iw_mr *ibmr = mapping->m_mr;
694	struct ib_send_wr f_wr, *failed_wr;
695	int ret;
696
697	/*
698	 * Perform a WR for the fast_reg_mr. Each individual page
699	 * in the sg list is added to the fast reg page list and placed
700	 * inside the fast_reg_mr WR.  The key used is a rolling 8bit
701	 * counter, which should guarantee uniqueness.
702	 */
703	ib_update_fast_reg_key(ibmr->mr, ibmr->remap_count++);
704	mapping->m_rkey = ibmr->mr->rkey;
705
706	memset(&f_wr, 0, sizeof(f_wr));
707	f_wr.wr_id = RDS_IW_FAST_REG_WR_ID;
708	f_wr.opcode = IB_WR_FAST_REG_MR;
709	f_wr.wr.fast_reg.length = mapping->m_sg.bytes;
710	f_wr.wr.fast_reg.rkey = mapping->m_rkey;
711	f_wr.wr.fast_reg.page_list = ibmr->page_list;
712	f_wr.wr.fast_reg.page_list_len = mapping->m_sg.dma_len;
713	f_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
714	f_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE |
715				IB_ACCESS_REMOTE_READ |
716				IB_ACCESS_REMOTE_WRITE;
717	f_wr.wr.fast_reg.iova_start = 0;
718	f_wr.send_flags = IB_SEND_SIGNALED;
719
720	failed_wr = &f_wr;
721	ret = ib_post_send(ibmr->cm_id->qp, &f_wr, &failed_wr);
722	BUG_ON(failed_wr != &f_wr);
723	if (ret)
724		printk_ratelimited(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
725			__func__, __LINE__, ret);
726	return ret;
727}
728
729static int rds_iw_rdma_fastreg_inv(struct rds_iw_mr *ibmr)
730{
731	struct ib_send_wr s_wr, *failed_wr;
732	int ret = 0;
733
734	if (!ibmr->cm_id->qp || !ibmr->mr)
735		goto out;
736
737	memset(&s_wr, 0, sizeof(s_wr));
738	s_wr.wr_id = RDS_IW_LOCAL_INV_WR_ID;
739	s_wr.opcode = IB_WR_LOCAL_INV;
740	s_wr.ex.invalidate_rkey = ibmr->mr->rkey;
741	s_wr.send_flags = IB_SEND_SIGNALED;
742
743	failed_wr = &s_wr;
744	ret = ib_post_send(ibmr->cm_id->qp, &s_wr, &failed_wr);
745	if (ret) {
746		printk_ratelimited(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
747			__func__, __LINE__, ret);
748		goto out;
749	}
750out:
751	return ret;
752}
753
754static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
755			struct rds_iw_mr *ibmr,
756			struct scatterlist *sg,
757			unsigned int sg_len)
758{
759	struct rds_iw_device *rds_iwdev = pool->device;
760	struct rds_iw_mapping *mapping = &ibmr->mapping;
761	u64 *dma_pages;
762	int i, ret = 0;
763
764	rds_iw_set_scatterlist(&mapping->m_sg, sg, sg_len);
765
766	dma_pages = rds_iw_map_scatterlist(rds_iwdev, &mapping->m_sg);
767	if (IS_ERR(dma_pages)) {
768		ret = PTR_ERR(dma_pages);
769		dma_pages = NULL;
770		goto out;
771	}
772
773	if (mapping->m_sg.dma_len > pool->max_message_size) {
774		ret = -EMSGSIZE;
775		goto out;
776	}
777
778	for (i = 0; i < mapping->m_sg.dma_npages; ++i)
779		ibmr->page_list->page_list[i] = dma_pages[i];
780
781	ret = rds_iw_rdma_build_fastreg(mapping);
782	if (ret)
783		goto out;
784
785	rds_iw_stats_inc(s_iw_rdma_mr_used);
786
787out:
788	kfree(dma_pages);
789
790	return ret;
791}
792
793/*
794 * "Free" a fastreg MR.
795 */
796static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool,
797		struct rds_iw_mr *ibmr)
798{
799	unsigned long flags;
800	int ret;
801
802	if (!ibmr->mapping.m_sg.dma_len)
803		return;
804
805	ret = rds_iw_rdma_fastreg_inv(ibmr);
806	if (ret)
807		return;
808
809	/* Try to post the LOCAL_INV WR to the queue. */
810	spin_lock_irqsave(&pool->list_lock, flags);
811
812	list_add_tail(&ibmr->mapping.m_list, &pool->dirty_list);
813	atomic_add(ibmr->mapping.m_sg.len, &pool->free_pinned);
814	atomic_inc(&pool->dirty_count);
815
816	spin_unlock_irqrestore(&pool->list_lock, flags);
817}
818
819static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
820				struct list_head *unmap_list,
821				struct list_head *kill_list,
822				int *unpinned)
823{
824	struct rds_iw_mapping *mapping, *next;
825	unsigned int ncleaned = 0;
826	LIST_HEAD(laundered);
827
828	/* Batched invalidation of fastreg MRs.
829	 * Why do we do it this way, even though we could pipeline unmap
830	 * and remap? The reason is the application semantics - when the
831	 * application requests an invalidation of MRs, it expects all
832	 * previously released R_Keys to become invalid.
833	 *
834	 * If we implement MR reuse naively, we risk memory corruption
835	 * (this has actually been observed). So the default behavior
836	 * requires that a MR goes through an explicit unmap operation before
837	 * we can reuse it again.
838	 *
839	 * We could probably improve on this a little, by allowing immediate
840	 * reuse of a MR on the same socket (eg you could add small
841	 * cache of unused MRs to strct rds_socket - GET_MR could grab one
842	 * of these without requiring an explicit invalidate).
843	 */
844	while (!list_empty(unmap_list)) {
845		unsigned long flags;
846
847		spin_lock_irqsave(&pool->list_lock, flags);
848		list_for_each_entry_safe(mapping, next, unmap_list, m_list) {
849			*unpinned += mapping->m_sg.len;
850			list_move(&mapping->m_list, &laundered);
851			ncleaned++;
852		}
853		spin_unlock_irqrestore(&pool->list_lock, flags);
854	}
855
856	/* Move all laundered mappings back to the unmap list.
857	 * We do not kill any WRs right now - it doesn't seem the
858	 * fastreg API has a max_remap limit. */
859	list_splice_init(&laundered, unmap_list);
860
861	return ncleaned;
862}
863
864static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool,
865		struct rds_iw_mr *ibmr)
866{
867	if (ibmr->page_list)
868		ib_free_fast_reg_page_list(ibmr->page_list);
869	if (ibmr->mr)
870		ib_dereg_mr(ibmr->mr);
871}
872