Merge commit 'linux-pnfs/nfs41-for-2.6.31' into nfsv41-for-2.6.31
[linux-2.6] / net / rds / iw_rdma.c
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
35 #include "rds.h"
36 #include "rdma.h"
37 #include "iw.h"
38
39
40 /*
41  * This is stored as mr->r_trans_private.
42  */
43 struct rds_iw_mr {
44         struct rds_iw_device    *device;
45         struct rds_iw_mr_pool   *pool;
46         struct rdma_cm_id       *cm_id;
47
48         struct ib_mr    *mr;
49         struct ib_fast_reg_page_list *page_list;
50
51         struct rds_iw_mapping   mapping;
52         unsigned char           remap_count;
53 };
54
55 /*
56  * Our own little MR pool
57  */
58 struct rds_iw_mr_pool {
59         struct rds_iw_device    *device;                /* back ptr to the device that owns us */
60
61         struct mutex            flush_lock;             /* serialize fmr invalidate */
62         struct work_struct      flush_worker;           /* flush worker */
63
64         spinlock_t              list_lock;              /* protect variables below */
65         atomic_t                item_count;             /* total # of MRs */
66         atomic_t                dirty_count;            /* # dirty of MRs */
67         struct list_head        dirty_list;             /* dirty mappings */
68         struct list_head        clean_list;             /* unused & unamapped MRs */
69         atomic_t                free_pinned;            /* memory pinned by free MRs */
70         unsigned long           max_message_size;       /* in pages */
71         unsigned long           max_items;
72         unsigned long           max_items_soft;
73         unsigned long           max_free_pinned;
74         int                     max_pages;
75 };
76
77 static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all);
78 static void rds_iw_mr_pool_flush_worker(struct work_struct *work);
79 static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
80 static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
81                           struct rds_iw_mr *ibmr,
82                           struct scatterlist *sg, unsigned int nents);
83 static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
84 static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
85                         struct list_head *unmap_list,
86                         struct list_head *kill_list);
87 static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
88
89 static int rds_iw_get_device(struct rds_sock *rs, struct rds_iw_device **rds_iwdev, struct rdma_cm_id **cm_id)
90 {
91         struct rds_iw_device *iwdev;
92         struct rds_iw_cm_id *i_cm_id;
93
94         *rds_iwdev = NULL;
95         *cm_id = NULL;
96
97         list_for_each_entry(iwdev, &rds_iw_devices, list) {
98                 spin_lock_irq(&iwdev->spinlock);
99                 list_for_each_entry(i_cm_id, &iwdev->cm_id_list, list) {
100                         struct sockaddr_in *src_addr, *dst_addr;
101
102                         src_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.src_addr;
103                         dst_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.dst_addr;
104
105                         rdsdebug("local ipaddr = %x port %d, "
106                                  "remote ipaddr = %x port %d"
107                                  "..looking for %x port %d, "
108                                  "remote ipaddr = %x port %d\n",
109                                 src_addr->sin_addr.s_addr,
110                                 src_addr->sin_port,
111                                 dst_addr->sin_addr.s_addr,
112                                 dst_addr->sin_port,
113                                 rs->rs_bound_addr,
114                                 rs->rs_bound_port,
115                                 rs->rs_conn_addr,
116                                 rs->rs_conn_port);
117 #ifdef WORKING_TUPLE_DETECTION
118                         if (src_addr->sin_addr.s_addr == rs->rs_bound_addr &&
119                             src_addr->sin_port == rs->rs_bound_port &&
120                             dst_addr->sin_addr.s_addr == rs->rs_conn_addr &&
121                             dst_addr->sin_port == rs->rs_conn_port) {
122 #else
123                         /* FIXME - needs to compare the local and remote
124                          * ipaddr/port tuple, but the ipaddr is the only
125                          * available infomation in the rds_sock (as the rest are
126                          * zero'ed.  It doesn't appear to be properly populated
127                          * during connection setup...
128                          */
129                         if (src_addr->sin_addr.s_addr == rs->rs_bound_addr) {
130 #endif
131                                 spin_unlock_irq(&iwdev->spinlock);
132                                 *rds_iwdev = iwdev;
133                                 *cm_id = i_cm_id->cm_id;
134                                 return 0;
135                         }
136                 }
137                 spin_unlock_irq(&iwdev->spinlock);
138         }
139
140         return 1;
141 }
142
143 static int rds_iw_add_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
144 {
145         struct rds_iw_cm_id *i_cm_id;
146
147         i_cm_id = kmalloc(sizeof *i_cm_id, GFP_KERNEL);
148         if (!i_cm_id)
149                 return -ENOMEM;
150
151         i_cm_id->cm_id = cm_id;
152
153         spin_lock_irq(&rds_iwdev->spinlock);
154         list_add_tail(&i_cm_id->list, &rds_iwdev->cm_id_list);
155         spin_unlock_irq(&rds_iwdev->spinlock);
156
157         return 0;
158 }
159
160 void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
161 {
162         struct rds_iw_cm_id *i_cm_id;
163
164         spin_lock_irq(&rds_iwdev->spinlock);
165         list_for_each_entry(i_cm_id, &rds_iwdev->cm_id_list, list) {
166                 if (i_cm_id->cm_id == cm_id) {
167                         list_del(&i_cm_id->list);
168                         kfree(i_cm_id);
169                         break;
170                 }
171         }
172         spin_unlock_irq(&rds_iwdev->spinlock);
173 }
174
175
176 int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
177 {
178         struct sockaddr_in *src_addr, *dst_addr;
179         struct rds_iw_device *rds_iwdev_old;
180         struct rds_sock rs;
181         struct rdma_cm_id *pcm_id;
182         int rc;
183
184         src_addr = (struct sockaddr_in *)&cm_id->route.addr.src_addr;
185         dst_addr = (struct sockaddr_in *)&cm_id->route.addr.dst_addr;
186
187         rs.rs_bound_addr = src_addr->sin_addr.s_addr;
188         rs.rs_bound_port = src_addr->sin_port;
189         rs.rs_conn_addr = dst_addr->sin_addr.s_addr;
190         rs.rs_conn_port = dst_addr->sin_port;
191
192         rc = rds_iw_get_device(&rs, &rds_iwdev_old, &pcm_id);
193         if (rc)
194                 rds_iw_remove_cm_id(rds_iwdev, cm_id);
195
196         return rds_iw_add_cm_id(rds_iwdev, cm_id);
197 }
198
199 void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
200 {
201         struct rds_iw_connection *ic = conn->c_transport_data;
202
203         /* conn was previously on the nodev_conns_list */
204         spin_lock_irq(&iw_nodev_conns_lock);
205         BUG_ON(list_empty(&iw_nodev_conns));
206         BUG_ON(list_empty(&ic->iw_node));
207         list_del(&ic->iw_node);
208
209         spin_lock_irq(&rds_iwdev->spinlock);
210         list_add_tail(&ic->iw_node, &rds_iwdev->conn_list);
211         spin_unlock_irq(&rds_iwdev->spinlock);
212         spin_unlock_irq(&iw_nodev_conns_lock);
213
214         ic->rds_iwdev = rds_iwdev;
215 }
216
217 void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
218 {
219         struct rds_iw_connection *ic = conn->c_transport_data;
220
221         /* place conn on nodev_conns_list */
222         spin_lock(&iw_nodev_conns_lock);
223
224         spin_lock_irq(&rds_iwdev->spinlock);
225         BUG_ON(list_empty(&ic->iw_node));
226         list_del(&ic->iw_node);
227         spin_unlock_irq(&rds_iwdev->spinlock);
228
229         list_add_tail(&ic->iw_node, &iw_nodev_conns);
230
231         spin_unlock(&iw_nodev_conns_lock);
232
233         rds_iw_remove_cm_id(ic->rds_iwdev, ic->i_cm_id);
234         ic->rds_iwdev = NULL;
235 }
236
237 void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock)
238 {
239         struct rds_iw_connection *ic, *_ic;
240         LIST_HEAD(tmp_list);
241
242         /* avoid calling conn_destroy with irqs off */
243         spin_lock_irq(list_lock);
244         list_splice(list, &tmp_list);
245         INIT_LIST_HEAD(list);
246         spin_unlock_irq(list_lock);
247
248         list_for_each_entry_safe(ic, _ic, &tmp_list, iw_node) {
249                 if (ic->conn->c_passive)
250                         rds_conn_destroy(ic->conn->c_passive);
251                 rds_conn_destroy(ic->conn);
252         }
253 }
254
255 static 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
265 static u64 *rds_iw_map_scatterlist(struct rds_iw_device *rds_iwdev,
266                         struct rds_iw_scatterlist *sg,
267                         unsigned int dma_page_shift)
268 {
269         struct ib_device *dev = rds_iwdev->dev;
270         u64 *dma_pages = NULL;
271         u64 dma_mask;
272         unsigned int dma_page_size;
273         int i, j, ret;
274
275         dma_page_size = 1 << dma_page_shift;
276         dma_mask = dma_page_size - 1;
277
278         WARN_ON(sg->dma_len);
279
280         sg->dma_len = ib_dma_map_sg(dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
281         if (unlikely(!sg->dma_len)) {
282                 printk(KERN_WARNING "RDS/IW: dma_map_sg failed!\n");
283                 return ERR_PTR(-EBUSY);
284         }
285
286         sg->bytes = 0;
287         sg->dma_npages = 0;
288
289         ret = -EINVAL;
290         for (i = 0; i < sg->dma_len; ++i) {
291                 unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
292                 u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
293                 u64 end_addr;
294
295                 sg->bytes += dma_len;
296
297                 end_addr = dma_addr + dma_len;
298                 if (dma_addr & dma_mask) {
299                         if (i > 0)
300                                 goto out_unmap;
301                         dma_addr &= ~dma_mask;
302                 }
303                 if (end_addr & dma_mask) {
304                         if (i < sg->dma_len - 1)
305                                 goto out_unmap;
306                         end_addr = (end_addr + dma_mask) & ~dma_mask;
307                 }
308
309                 sg->dma_npages += (end_addr - dma_addr) >> dma_page_shift;
310         }
311
312         /* Now gather the dma addrs into one list */
313         if (sg->dma_npages > fastreg_message_size)
314                 goto out_unmap;
315
316         dma_pages = kmalloc(sizeof(u64) * sg->dma_npages, GFP_ATOMIC);
317         if (!dma_pages) {
318                 ret = -ENOMEM;
319                 goto out_unmap;
320         }
321
322         for (i = j = 0; i < sg->dma_len; ++i) {
323                 unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
324                 u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
325                 u64 end_addr;
326
327                 end_addr = dma_addr + dma_len;
328                 dma_addr &= ~dma_mask;
329                 for (; dma_addr < end_addr; dma_addr += dma_page_size)
330                         dma_pages[j++] = dma_addr;
331                 BUG_ON(j > sg->dma_npages);
332         }
333
334         return dma_pages;
335
336 out_unmap:
337         ib_dma_unmap_sg(rds_iwdev->dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
338         sg->dma_len = 0;
339         kfree(dma_pages);
340         return ERR_PTR(ret);
341 }
342
343
344 struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *rds_iwdev)
345 {
346         struct rds_iw_mr_pool *pool;
347
348         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
349         if (!pool) {
350                 printk(KERN_WARNING "RDS/IW: rds_iw_create_mr_pool alloc error\n");
351                 return ERR_PTR(-ENOMEM);
352         }
353
354         pool->device = rds_iwdev;
355         INIT_LIST_HEAD(&pool->dirty_list);
356         INIT_LIST_HEAD(&pool->clean_list);
357         mutex_init(&pool->flush_lock);
358         spin_lock_init(&pool->list_lock);
359         INIT_WORK(&pool->flush_worker, rds_iw_mr_pool_flush_worker);
360
361         pool->max_message_size = fastreg_message_size;
362         pool->max_items = fastreg_pool_size;
363         pool->max_free_pinned = pool->max_items * pool->max_message_size / 4;
364         pool->max_pages = fastreg_message_size;
365
366         /* We never allow more than max_items MRs to be allocated.
367          * When we exceed more than max_items_soft, we start freeing
368          * items more aggressively.
369          * Make sure that max_items > max_items_soft > max_items / 2
370          */
371         pool->max_items_soft = pool->max_items * 3 / 4;
372
373         return pool;
374 }
375
376 void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo)
377 {
378         struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
379
380         iinfo->rdma_mr_max = pool->max_items;
381         iinfo->rdma_mr_size = pool->max_pages;
382 }
383
384 void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *pool)
385 {
386         flush_workqueue(rds_wq);
387         rds_iw_flush_mr_pool(pool, 1);
388         BUG_ON(atomic_read(&pool->item_count));
389         BUG_ON(atomic_read(&pool->free_pinned));
390         kfree(pool);
391 }
392
393 static inline struct rds_iw_mr *rds_iw_reuse_fmr(struct rds_iw_mr_pool *pool)
394 {
395         struct rds_iw_mr *ibmr = NULL;
396         unsigned long flags;
397
398         spin_lock_irqsave(&pool->list_lock, flags);
399         if (!list_empty(&pool->clean_list)) {
400                 ibmr = list_entry(pool->clean_list.next, struct rds_iw_mr, mapping.m_list);
401                 list_del_init(&ibmr->mapping.m_list);
402         }
403         spin_unlock_irqrestore(&pool->list_lock, flags);
404
405         return ibmr;
406 }
407
408 static struct rds_iw_mr *rds_iw_alloc_mr(struct rds_iw_device *rds_iwdev)
409 {
410         struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
411         struct rds_iw_mr *ibmr = NULL;
412         int err = 0, iter = 0;
413
414         while (1) {
415                 ibmr = rds_iw_reuse_fmr(pool);
416                 if (ibmr)
417                         return ibmr;
418
419                 /* No clean MRs - now we have the choice of either
420                  * allocating a fresh MR up to the limit imposed by the
421                  * driver, or flush any dirty unused MRs.
422                  * We try to avoid stalling in the send path if possible,
423                  * so we allocate as long as we're allowed to.
424                  *
425                  * We're fussy with enforcing the FMR limit, though. If the driver
426                  * tells us we can't use more than N fmrs, we shouldn't start
427                  * arguing with it */
428                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
429                         break;
430
431                 atomic_dec(&pool->item_count);
432
433                 if (++iter > 2) {
434                         rds_iw_stats_inc(s_iw_rdma_mr_pool_depleted);
435                         return ERR_PTR(-EAGAIN);
436                 }
437
438                 /* We do have some empty MRs. Flush them out. */
439                 rds_iw_stats_inc(s_iw_rdma_mr_pool_wait);
440                 rds_iw_flush_mr_pool(pool, 0);
441         }
442
443         ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
444         if (!ibmr) {
445                 err = -ENOMEM;
446                 goto out_no_cigar;
447         }
448
449         spin_lock_init(&ibmr->mapping.m_lock);
450         INIT_LIST_HEAD(&ibmr->mapping.m_list);
451         ibmr->mapping.m_mr = ibmr;
452
453         err = rds_iw_init_fastreg(pool, ibmr);
454         if (err)
455                 goto out_no_cigar;
456
457         rds_iw_stats_inc(s_iw_rdma_mr_alloc);
458         return ibmr;
459
460 out_no_cigar:
461         if (ibmr) {
462                 rds_iw_destroy_fastreg(pool, ibmr);
463                 kfree(ibmr);
464         }
465         atomic_dec(&pool->item_count);
466         return ERR_PTR(err);
467 }
468
469 void rds_iw_sync_mr(void *trans_private, int direction)
470 {
471         struct rds_iw_mr *ibmr = trans_private;
472         struct rds_iw_device *rds_iwdev = ibmr->device;
473
474         switch (direction) {
475         case DMA_FROM_DEVICE:
476                 ib_dma_sync_sg_for_cpu(rds_iwdev->dev, ibmr->mapping.m_sg.list,
477                         ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
478                 break;
479         case DMA_TO_DEVICE:
480                 ib_dma_sync_sg_for_device(rds_iwdev->dev, ibmr->mapping.m_sg.list,
481                         ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
482                 break;
483         }
484 }
485
486 static inline unsigned int rds_iw_flush_goal(struct rds_iw_mr_pool *pool, int free_all)
487 {
488         unsigned int item_count;
489
490         item_count = atomic_read(&pool->item_count);
491         if (free_all)
492                 return item_count;
493
494         return 0;
495 }
496
497 /*
498  * Flush our pool of MRs.
499  * At a minimum, all currently unused MRs are unmapped.
500  * If the number of MRs allocated exceeds the limit, we also try
501  * to free as many MRs as needed to get back to this limit.
502  */
503 static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all)
504 {
505         struct rds_iw_mr *ibmr, *next;
506         LIST_HEAD(unmap_list);
507         LIST_HEAD(kill_list);
508         unsigned long flags;
509         unsigned int nfreed = 0, ncleaned = 0, free_goal;
510         int ret = 0;
511
512         rds_iw_stats_inc(s_iw_rdma_mr_pool_flush);
513
514         mutex_lock(&pool->flush_lock);
515
516         spin_lock_irqsave(&pool->list_lock, flags);
517         /* Get the list of all mappings to be destroyed */
518         list_splice_init(&pool->dirty_list, &unmap_list);
519         if (free_all)
520                 list_splice_init(&pool->clean_list, &kill_list);
521         spin_unlock_irqrestore(&pool->list_lock, flags);
522
523         free_goal = rds_iw_flush_goal(pool, free_all);
524
525         /* Batched invalidate of dirty MRs.
526          * For FMR based MRs, the mappings on the unmap list are
527          * actually members of an ibmr (ibmr->mapping). They either
528          * migrate to the kill_list, or have been cleaned and should be
529          * moved to the clean_list.
530          * For fastregs, they will be dynamically allocated, and
531          * will be destroyed by the unmap function.
532          */
533         if (!list_empty(&unmap_list)) {
534                 ncleaned = rds_iw_unmap_fastreg_list(pool, &unmap_list, &kill_list);
535                 /* If we've been asked to destroy all MRs, move those
536                  * that were simply cleaned to the kill list */
537                 if (free_all)
538                         list_splice_init(&unmap_list, &kill_list);
539         }
540
541         /* Destroy any MRs that are past their best before date */
542         list_for_each_entry_safe(ibmr, next, &kill_list, mapping.m_list) {
543                 rds_iw_stats_inc(s_iw_rdma_mr_free);
544                 list_del(&ibmr->mapping.m_list);
545                 rds_iw_destroy_fastreg(pool, ibmr);
546                 kfree(ibmr);
547                 nfreed++;
548         }
549
550         /* Anything that remains are laundered ibmrs, which we can add
551          * back to the clean list. */
552         if (!list_empty(&unmap_list)) {
553                 spin_lock_irqsave(&pool->list_lock, flags);
554                 list_splice(&unmap_list, &pool->clean_list);
555                 spin_unlock_irqrestore(&pool->list_lock, flags);
556         }
557
558         atomic_sub(ncleaned, &pool->dirty_count);
559         atomic_sub(nfreed, &pool->item_count);
560
561         mutex_unlock(&pool->flush_lock);
562         return ret;
563 }
564
565 static void rds_iw_mr_pool_flush_worker(struct work_struct *work)
566 {
567         struct rds_iw_mr_pool *pool = container_of(work, struct rds_iw_mr_pool, flush_worker);
568
569         rds_iw_flush_mr_pool(pool, 0);
570 }
571
572 void rds_iw_free_mr(void *trans_private, int invalidate)
573 {
574         struct rds_iw_mr *ibmr = trans_private;
575         struct rds_iw_mr_pool *pool = ibmr->device->mr_pool;
576
577         rdsdebug("RDS/IW: free_mr nents %u\n", ibmr->mapping.m_sg.len);
578         if (!pool)
579                 return;
580
581         /* Return it to the pool's free list */
582         rds_iw_free_fastreg(pool, ibmr);
583
584         /* If we've pinned too many pages, request a flush */
585         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned
586          || atomic_read(&pool->dirty_count) >= pool->max_items / 10)
587                 queue_work(rds_wq, &pool->flush_worker);
588
589         if (invalidate) {
590                 if (likely(!in_interrupt())) {
591                         rds_iw_flush_mr_pool(pool, 0);
592                 } else {
593                         /* We get here if the user created a MR marked
594                          * as use_once and invalidate at the same time. */
595                         queue_work(rds_wq, &pool->flush_worker);
596                 }
597         }
598 }
599
600 void rds_iw_flush_mrs(void)
601 {
602         struct rds_iw_device *rds_iwdev;
603
604         list_for_each_entry(rds_iwdev, &rds_iw_devices, list) {
605                 struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
606
607                 if (pool)
608                         rds_iw_flush_mr_pool(pool, 0);
609         }
610 }
611
612 void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
613                     struct rds_sock *rs, u32 *key_ret)
614 {
615         struct rds_iw_device *rds_iwdev;
616         struct rds_iw_mr *ibmr = NULL;
617         struct rdma_cm_id *cm_id;
618         int ret;
619
620         ret = rds_iw_get_device(rs, &rds_iwdev, &cm_id);
621         if (ret || !cm_id) {
622                 ret = -ENODEV;
623                 goto out;
624         }
625
626         if (!rds_iwdev->mr_pool) {
627                 ret = -ENODEV;
628                 goto out;
629         }
630
631         ibmr = rds_iw_alloc_mr(rds_iwdev);
632         if (IS_ERR(ibmr))
633                 return ibmr;
634
635         ibmr->cm_id = cm_id;
636         ibmr->device = rds_iwdev;
637
638         ret = rds_iw_map_fastreg(rds_iwdev->mr_pool, ibmr, sg, nents);
639         if (ret == 0)
640                 *key_ret = ibmr->mr->rkey;
641         else
642                 printk(KERN_WARNING "RDS/IW: failed to map mr (errno=%d)\n", ret);
643
644 out:
645         if (ret) {
646                 if (ibmr)
647                         rds_iw_free_mr(ibmr, 0);
648                 ibmr = ERR_PTR(ret);
649         }
650         return ibmr;
651 }
652
653 /*
654  * iWARP fastreg handling
655  *
656  * The life cycle of a fastreg registration is a bit different from
657  * FMRs.
658  * The idea behind fastreg is to have one MR, to which we bind different
659  * mappings over time. To avoid stalling on the expensive map and invalidate
660  * operations, these operations are pipelined on the same send queue on
661  * which we want to send the message containing the r_key.
662  *
663  * This creates a bit of a problem for us, as we do not have the destination
664  * IP in GET_MR, so the connection must be setup prior to the GET_MR call for
665  * RDMA to be correctly setup.  If a fastreg request is present, rds_iw_xmit
666  * will try to queue a LOCAL_INV (if needed) and a FAST_REG_MR work request
667  * before queuing the SEND. When completions for these arrive, they are
668  * dispatched to the MR has a bit set showing that RDMa can be performed.
669  *
670  * There is another interesting aspect that's related to invalidation.
671  * The application can request that a mapping is invalidated in FREE_MR.
672  * The expectation there is that this invalidation step includes ALL
673  * PREVIOUSLY FREED MRs.
674  */
675 static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool,
676                                 struct rds_iw_mr *ibmr)
677 {
678         struct rds_iw_device *rds_iwdev = pool->device;
679         struct ib_fast_reg_page_list *page_list = NULL;
680         struct ib_mr *mr;
681         int err;
682
683         mr = ib_alloc_fast_reg_mr(rds_iwdev->pd, pool->max_message_size);
684         if (IS_ERR(mr)) {
685                 err = PTR_ERR(mr);
686
687                 printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_mr failed (err=%d)\n", err);
688                 return err;
689         }
690
691         /* FIXME - this is overkill, but mapping->m_sg.dma_len/mapping->m_sg.dma_npages
692          * is not filled in.
693          */
694         page_list = ib_alloc_fast_reg_page_list(rds_iwdev->dev, pool->max_message_size);
695         if (IS_ERR(page_list)) {
696                 err = PTR_ERR(page_list);
697
698                 printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed (err=%d)\n", err);
699                 ib_dereg_mr(mr);
700                 return err;
701         }
702
703         ibmr->page_list = page_list;
704         ibmr->mr = mr;
705         return 0;
706 }
707
708 static int rds_iw_rdma_build_fastreg(struct rds_iw_mapping *mapping)
709 {
710         struct rds_iw_mr *ibmr = mapping->m_mr;
711         struct ib_send_wr f_wr, *failed_wr;
712         int ret;
713
714         /*
715          * Perform a WR for the fast_reg_mr. Each individual page
716          * in the sg list is added to the fast reg page list and placed
717          * inside the fast_reg_mr WR.  The key used is a rolling 8bit
718          * counter, which should guarantee uniqueness.
719          */
720         ib_update_fast_reg_key(ibmr->mr, ibmr->remap_count++);
721         mapping->m_rkey = ibmr->mr->rkey;
722
723         memset(&f_wr, 0, sizeof(f_wr));
724         f_wr.wr_id = RDS_IW_FAST_REG_WR_ID;
725         f_wr.opcode = IB_WR_FAST_REG_MR;
726         f_wr.wr.fast_reg.length = mapping->m_sg.bytes;
727         f_wr.wr.fast_reg.rkey = mapping->m_rkey;
728         f_wr.wr.fast_reg.page_list = ibmr->page_list;
729         f_wr.wr.fast_reg.page_list_len = mapping->m_sg.dma_len;
730         f_wr.wr.fast_reg.page_shift = ibmr->device->page_shift;
731         f_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE |
732                                 IB_ACCESS_REMOTE_READ |
733                                 IB_ACCESS_REMOTE_WRITE;
734         f_wr.wr.fast_reg.iova_start = 0;
735         f_wr.send_flags = IB_SEND_SIGNALED;
736
737         failed_wr = &f_wr;
738         ret = ib_post_send(ibmr->cm_id->qp, &f_wr, &failed_wr);
739         BUG_ON(failed_wr != &f_wr);
740         if (ret && printk_ratelimit())
741                 printk(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
742                         __func__, __LINE__, ret);
743         return ret;
744 }
745
746 static int rds_iw_rdma_fastreg_inv(struct rds_iw_mr *ibmr)
747 {
748         struct ib_send_wr s_wr, *failed_wr;
749         int ret = 0;
750
751         if (!ibmr->cm_id->qp || !ibmr->mr)
752                 goto out;
753
754         memset(&s_wr, 0, sizeof(s_wr));
755         s_wr.wr_id = RDS_IW_LOCAL_INV_WR_ID;
756         s_wr.opcode = IB_WR_LOCAL_INV;
757         s_wr.ex.invalidate_rkey = ibmr->mr->rkey;
758         s_wr.send_flags = IB_SEND_SIGNALED;
759
760         failed_wr = &s_wr;
761         ret = ib_post_send(ibmr->cm_id->qp, &s_wr, &failed_wr);
762         if (ret && printk_ratelimit()) {
763                 printk(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
764                         __func__, __LINE__, ret);
765                 goto out;
766         }
767 out:
768         return ret;
769 }
770
771 static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
772                         struct rds_iw_mr *ibmr,
773                         struct scatterlist *sg,
774                         unsigned int sg_len)
775 {
776         struct rds_iw_device *rds_iwdev = pool->device;
777         struct rds_iw_mapping *mapping = &ibmr->mapping;
778         u64 *dma_pages;
779         int i, ret = 0;
780
781         rds_iw_set_scatterlist(&mapping->m_sg, sg, sg_len);
782
783         dma_pages = rds_iw_map_scatterlist(rds_iwdev,
784                                 &mapping->m_sg,
785                                 rds_iwdev->page_shift);
786         if (IS_ERR(dma_pages)) {
787                 ret = PTR_ERR(dma_pages);
788                 dma_pages = NULL;
789                 goto out;
790         }
791
792         if (mapping->m_sg.dma_len > pool->max_message_size) {
793                 ret = -EMSGSIZE;
794                 goto out;
795         }
796
797         for (i = 0; i < mapping->m_sg.dma_npages; ++i)
798                 ibmr->page_list->page_list[i] = dma_pages[i];
799
800         ret = rds_iw_rdma_build_fastreg(mapping);
801         if (ret)
802                 goto out;
803
804         rds_iw_stats_inc(s_iw_rdma_mr_used);
805
806 out:
807         kfree(dma_pages);
808
809         return ret;
810 }
811
812 /*
813  * "Free" a fastreg MR.
814  */
815 static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool,
816                 struct rds_iw_mr *ibmr)
817 {
818         unsigned long flags;
819         int ret;
820
821         if (!ibmr->mapping.m_sg.dma_len)
822                 return;
823
824         ret = rds_iw_rdma_fastreg_inv(ibmr);
825         if (ret)
826                 return;
827
828         /* Try to post the LOCAL_INV WR to the queue. */
829         spin_lock_irqsave(&pool->list_lock, flags);
830
831         list_add_tail(&ibmr->mapping.m_list, &pool->dirty_list);
832         atomic_add(ibmr->mapping.m_sg.len, &pool->free_pinned);
833         atomic_inc(&pool->dirty_count);
834
835         spin_unlock_irqrestore(&pool->list_lock, flags);
836 }
837
838 static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
839                                 struct list_head *unmap_list,
840                                 struct list_head *kill_list)
841 {
842         struct rds_iw_mapping *mapping, *next;
843         unsigned int ncleaned = 0;
844         LIST_HEAD(laundered);
845
846         /* Batched invalidation of fastreg MRs.
847          * Why do we do it this way, even though we could pipeline unmap
848          * and remap? The reason is the application semantics - when the
849          * application requests an invalidation of MRs, it expects all
850          * previously released R_Keys to become invalid.
851          *
852          * If we implement MR reuse naively, we risk memory corruption
853          * (this has actually been observed). So the default behavior
854          * requires that a MR goes through an explicit unmap operation before
855          * we can reuse it again.
856          *
857          * We could probably improve on this a little, by allowing immediate
858          * reuse of a MR on the same socket (eg you could add small
859          * cache of unused MRs to strct rds_socket - GET_MR could grab one
860          * of these without requiring an explicit invalidate).
861          */
862         while (!list_empty(unmap_list)) {
863                 unsigned long flags;
864
865                 spin_lock_irqsave(&pool->list_lock, flags);
866                 list_for_each_entry_safe(mapping, next, unmap_list, m_list) {
867                         list_move(&mapping->m_list, &laundered);
868                         ncleaned++;
869                 }
870                 spin_unlock_irqrestore(&pool->list_lock, flags);
871         }
872
873         /* Move all laundered mappings back to the unmap list.
874          * We do not kill any WRs right now - it doesn't seem the
875          * fastreg API has a max_remap limit. */
876         list_splice_init(&laundered, unmap_list);
877
878         return ncleaned;
879 }
880
881 static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool,
882                 struct rds_iw_mr *ibmr)
883 {
884         if (ibmr->page_list)
885                 ib_free_fast_reg_page_list(ibmr->page_list);
886         if (ibmr->mr)
887                 ib_dereg_mr(ibmr->mr);
888 }