13dc1862d86264288eb0455056d289008d072234
[linux-3.10.git] / 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 #include <linux/slab.h>
35
36 #include "rds.h"
37 #include "rdma.h"
38 #include "iw.h"
39
40
41 /*
42  * This is stored as mr->r_trans_private.
43  */
44 struct 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  */
59 struct 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
78 static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all);
79 static void rds_iw_mr_pool_flush_worker(struct work_struct *work);
80 static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
81 static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
82                           struct rds_iw_mr *ibmr,
83                           struct scatterlist *sg, unsigned int nents);
84 static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
85 static 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 static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
89
90 static int rds_iw_get_device(struct rds_sock *rs, struct rds_iw_device **rds_iwdev, struct rdma_cm_id **cm_id)
91 {
92         struct rds_iw_device *iwdev;
93         struct rds_iw_cm_id *i_cm_id;
94
95         *rds_iwdev = NULL;
96         *cm_id = NULL;
97
98         list_for_each_entry(iwdev, &rds_iw_devices, list) {
99                 spin_lock_irq(&iwdev->spinlock);
100                 list_for_each_entry(i_cm_id, &iwdev->cm_id_list, list) {
101                         struct sockaddr_in *src_addr, *dst_addr;
102
103                         src_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.src_addr;
104                         dst_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.dst_addr;
105
106                         rdsdebug("local ipaddr = %x port %d, "
107                                  "remote ipaddr = %x port %d"
108                                  "..looking for %x port %d, "
109                                  "remote ipaddr = %x port %d\n",
110                                 src_addr->sin_addr.s_addr,
111                                 src_addr->sin_port,
112                                 dst_addr->sin_addr.s_addr,
113                                 dst_addr->sin_port,
114                                 rs->rs_bound_addr,
115                                 rs->rs_bound_port,
116                                 rs->rs_conn_addr,
117                                 rs->rs_conn_port);
118 #ifdef WORKING_TUPLE_DETECTION
119                         if (src_addr->sin_addr.s_addr == rs->rs_bound_addr &&
120                             src_addr->sin_port == rs->rs_bound_port &&
121                             dst_addr->sin_addr.s_addr == rs->rs_conn_addr &&
122                             dst_addr->sin_port == rs->rs_conn_port) {
123 #else
124                         /* FIXME - needs to compare the local and remote
125                          * ipaddr/port tuple, but the ipaddr is the only
126                          * available infomation in the rds_sock (as the rest are
127                          * zero'ed.  It doesn't appear to be properly populated
128                          * during connection setup...
129                          */
130                         if (src_addr->sin_addr.s_addr == rs->rs_bound_addr) {
131 #endif
132                                 spin_unlock_irq(&iwdev->spinlock);
133                                 *rds_iwdev = iwdev;
134                                 *cm_id = i_cm_id->cm_id;
135                                 return 0;
136                         }
137                 }
138                 spin_unlock_irq(&iwdev->spinlock);
139         }
140
141         return 1;
142 }
143
144 static int rds_iw_add_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
145 {
146         struct rds_iw_cm_id *i_cm_id;
147
148         i_cm_id = kmalloc(sizeof *i_cm_id, GFP_KERNEL);
149         if (!i_cm_id)
150                 return -ENOMEM;
151
152         i_cm_id->cm_id = cm_id;
153
154         spin_lock_irq(&rds_iwdev->spinlock);
155         list_add_tail(&i_cm_id->list, &rds_iwdev->cm_id_list);
156         spin_unlock_irq(&rds_iwdev->spinlock);
157
158         return 0;
159 }
160
161 void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
162 {
163         struct rds_iw_cm_id *i_cm_id;
164
165         spin_lock_irq(&rds_iwdev->spinlock);
166         list_for_each_entry(i_cm_id, &rds_iwdev->cm_id_list, list) {
167                 if (i_cm_id->cm_id == cm_id) {
168                         list_del(&i_cm_id->list);
169                         kfree(i_cm_id);
170                         break;
171                 }
172         }
173         spin_unlock_irq(&rds_iwdev->spinlock);
174 }
175
176
177 int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
178 {
179         struct sockaddr_in *src_addr, *dst_addr;
180         struct rds_iw_device *rds_iwdev_old;
181         struct rds_sock rs;
182         struct rdma_cm_id *pcm_id;
183         int rc;
184
185         src_addr = (struct sockaddr_in *)&cm_id->route.addr.src_addr;
186         dst_addr = (struct sockaddr_in *)&cm_id->route.addr.dst_addr;
187
188         rs.rs_bound_addr = src_addr->sin_addr.s_addr;
189         rs.rs_bound_port = src_addr->sin_port;
190         rs.rs_conn_addr = dst_addr->sin_addr.s_addr;
191         rs.rs_conn_port = dst_addr->sin_port;
192
193         rc = rds_iw_get_device(&rs, &rds_iwdev_old, &pcm_id);
194         if (rc)
195                 rds_iw_remove_cm_id(rds_iwdev, cm_id);
196
197         return rds_iw_add_cm_id(rds_iwdev, cm_id);
198 }
199
200 void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
201 {
202         struct rds_iw_connection *ic = conn->c_transport_data;
203
204         /* conn was previously on the nodev_conns_list */
205         spin_lock_irq(&iw_nodev_conns_lock);
206         BUG_ON(list_empty(&iw_nodev_conns));
207         BUG_ON(list_empty(&ic->iw_node));
208         list_del(&ic->iw_node);
209
210         spin_lock_irq(&rds_iwdev->spinlock);
211         list_add_tail(&ic->iw_node, &rds_iwdev->conn_list);
212         spin_unlock_irq(&rds_iwdev->spinlock);
213         spin_unlock_irq(&iw_nodev_conns_lock);
214
215         ic->rds_iwdev = rds_iwdev;
216 }
217
218 void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
219 {
220         struct rds_iw_connection *ic = conn->c_transport_data;
221
222         /* place conn on nodev_conns_list */
223         spin_lock(&iw_nodev_conns_lock);
224
225         spin_lock_irq(&rds_iwdev->spinlock);
226         BUG_ON(list_empty(&ic->iw_node));
227         list_del(&ic->iw_node);
228         spin_unlock_irq(&rds_iwdev->spinlock);
229
230         list_add_tail(&ic->iw_node, &iw_nodev_conns);
231
232         spin_unlock(&iw_nodev_conns_lock);
233
234         rds_iw_remove_cm_id(ic->rds_iwdev, ic->i_cm_id);
235         ic->rds_iwdev = NULL;
236 }
237
238 void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock)
239 {
240         struct rds_iw_connection *ic, *_ic;
241         LIST_HEAD(tmp_list);
242
243         /* avoid calling conn_destroy with irqs off */
244         spin_lock_irq(list_lock);
245         list_splice(list, &tmp_list);
246         INIT_LIST_HEAD(list);
247         spin_unlock_irq(list_lock);
248
249         list_for_each_entry_safe(ic, _ic, &tmp_list, iw_node)
250                 rds_conn_destroy(ic->conn);
251 }
252
253 static void rds_iw_set_scatterlist(struct rds_iw_scatterlist *sg,
254                 struct scatterlist *list, unsigned int sg_len)
255 {
256         sg->list = list;
257         sg->len = sg_len;
258         sg->dma_len = 0;
259         sg->dma_npages = 0;
260         sg->bytes = 0;
261 }
262
263 static u64 *rds_iw_map_scatterlist(struct rds_iw_device *rds_iwdev,
264                         struct rds_iw_scatterlist *sg)
265 {
266         struct ib_device *dev = rds_iwdev->dev;
267         u64 *dma_pages = NULL;
268         int i, j, ret;
269
270         WARN_ON(sg->dma_len);
271
272         sg->dma_len = ib_dma_map_sg(dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
273         if (unlikely(!sg->dma_len)) {
274                 printk(KERN_WARNING "RDS/IW: dma_map_sg failed!\n");
275                 return ERR_PTR(-EBUSY);
276         }
277
278         sg->bytes = 0;
279         sg->dma_npages = 0;
280
281         ret = -EINVAL;
282         for (i = 0; i < sg->dma_len; ++i) {
283                 unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
284                 u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
285                 u64 end_addr;
286
287                 sg->bytes += dma_len;
288
289                 end_addr = dma_addr + dma_len;
290                 if (dma_addr & PAGE_MASK) {
291                         if (i > 0)
292                                 goto out_unmap;
293                         dma_addr &= ~PAGE_MASK;
294                 }
295                 if (end_addr & PAGE_MASK) {
296                         if (i < sg->dma_len - 1)
297                                 goto out_unmap;
298                         end_addr = (end_addr + PAGE_MASK) & ~PAGE_MASK;
299                 }
300
301                 sg->dma_npages += (end_addr - dma_addr) >> PAGE_SHIFT;
302         }
303
304         /* Now gather the dma addrs into one list */
305         if (sg->dma_npages > fastreg_message_size)
306                 goto out_unmap;
307
308         dma_pages = kmalloc(sizeof(u64) * sg->dma_npages, GFP_ATOMIC);
309         if (!dma_pages) {
310                 ret = -ENOMEM;
311                 goto out_unmap;
312         }
313
314         for (i = j = 0; i < sg->dma_len; ++i) {
315                 unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
316                 u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
317                 u64 end_addr;
318
319                 end_addr = dma_addr + dma_len;
320                 dma_addr &= ~PAGE_MASK;
321                 for (; dma_addr < end_addr; dma_addr += PAGE_SIZE)
322                         dma_pages[j++] = dma_addr;
323                 BUG_ON(j > sg->dma_npages);
324         }
325
326         return dma_pages;
327
328 out_unmap:
329         ib_dma_unmap_sg(rds_iwdev->dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
330         sg->dma_len = 0;
331         kfree(dma_pages);
332         return ERR_PTR(ret);
333 }
334
335
336 struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *rds_iwdev)
337 {
338         struct rds_iw_mr_pool *pool;
339
340         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
341         if (!pool) {
342                 printk(KERN_WARNING "RDS/IW: rds_iw_create_mr_pool alloc error\n");
343                 return ERR_PTR(-ENOMEM);
344         }
345
346         pool->device = rds_iwdev;
347         INIT_LIST_HEAD(&pool->dirty_list);
348         INIT_LIST_HEAD(&pool->clean_list);
349         mutex_init(&pool->flush_lock);
350         spin_lock_init(&pool->list_lock);
351         INIT_WORK(&pool->flush_worker, rds_iw_mr_pool_flush_worker);
352
353         pool->max_message_size = fastreg_message_size;
354         pool->max_items = fastreg_pool_size;
355         pool->max_free_pinned = pool->max_items * pool->max_message_size / 4;
356         pool->max_pages = fastreg_message_size;
357
358         /* We never allow more than max_items MRs to be allocated.
359          * When we exceed more than max_items_soft, we start freeing
360          * items more aggressively.
361          * Make sure that max_items > max_items_soft > max_items / 2
362          */
363         pool->max_items_soft = pool->max_items * 3 / 4;
364
365         return pool;
366 }
367
368 void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo)
369 {
370         struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
371
372         iinfo->rdma_mr_max = pool->max_items;
373         iinfo->rdma_mr_size = pool->max_pages;
374 }
375
376 void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *pool)
377 {
378         flush_workqueue(rds_wq);
379         rds_iw_flush_mr_pool(pool, 1);
380         BUG_ON(atomic_read(&pool->item_count));
381         BUG_ON(atomic_read(&pool->free_pinned));
382         kfree(pool);
383 }
384
385 static inline struct rds_iw_mr *rds_iw_reuse_fmr(struct rds_iw_mr_pool *pool)
386 {
387         struct rds_iw_mr *ibmr = NULL;
388         unsigned long flags;
389
390         spin_lock_irqsave(&pool->list_lock, flags);
391         if (!list_empty(&pool->clean_list)) {
392                 ibmr = list_entry(pool->clean_list.next, struct rds_iw_mr, mapping.m_list);
393                 list_del_init(&ibmr->mapping.m_list);
394         }
395         spin_unlock_irqrestore(&pool->list_lock, flags);
396
397         return ibmr;
398 }
399
400 static struct rds_iw_mr *rds_iw_alloc_mr(struct rds_iw_device *rds_iwdev)
401 {
402         struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
403         struct rds_iw_mr *ibmr = NULL;
404         int err = 0, iter = 0;
405
406         while (1) {
407                 ibmr = rds_iw_reuse_fmr(pool);
408                 if (ibmr)
409                         return ibmr;
410
411                 /* No clean MRs - now we have the choice of either
412                  * allocating a fresh MR up to the limit imposed by the
413                  * driver, or flush any dirty unused MRs.
414                  * We try to avoid stalling in the send path if possible,
415                  * so we allocate as long as we're allowed to.
416                  *
417                  * We're fussy with enforcing the FMR limit, though. If the driver
418                  * tells us we can't use more than N fmrs, we shouldn't start
419                  * arguing with it */
420                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
421                         break;
422
423                 atomic_dec(&pool->item_count);
424
425                 if (++iter > 2) {
426                         rds_iw_stats_inc(s_iw_rdma_mr_pool_depleted);
427                         return ERR_PTR(-EAGAIN);
428                 }
429
430                 /* We do have some empty MRs. Flush them out. */
431                 rds_iw_stats_inc(s_iw_rdma_mr_pool_wait);
432                 rds_iw_flush_mr_pool(pool, 0);
433         }
434
435         ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
436         if (!ibmr) {
437                 err = -ENOMEM;
438                 goto out_no_cigar;
439         }
440
441         spin_lock_init(&ibmr->mapping.m_lock);
442         INIT_LIST_HEAD(&ibmr->mapping.m_list);
443         ibmr->mapping.m_mr = ibmr;
444
445         err = rds_iw_init_fastreg(pool, ibmr);
446         if (err)
447                 goto out_no_cigar;
448
449         rds_iw_stats_inc(s_iw_rdma_mr_alloc);
450         return ibmr;
451
452 out_no_cigar:
453         if (ibmr) {
454                 rds_iw_destroy_fastreg(pool, ibmr);
455                 kfree(ibmr);
456         }
457         atomic_dec(&pool->item_count);
458         return ERR_PTR(err);
459 }
460
461 void rds_iw_sync_mr(void *trans_private, int direction)
462 {
463         struct rds_iw_mr *ibmr = trans_private;
464         struct rds_iw_device *rds_iwdev = ibmr->device;
465
466         switch (direction) {
467         case DMA_FROM_DEVICE:
468                 ib_dma_sync_sg_for_cpu(rds_iwdev->dev, ibmr->mapping.m_sg.list,
469                         ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
470                 break;
471         case DMA_TO_DEVICE:
472                 ib_dma_sync_sg_for_device(rds_iwdev->dev, ibmr->mapping.m_sg.list,
473                         ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
474                 break;
475         }
476 }
477
478 static inline unsigned int rds_iw_flush_goal(struct rds_iw_mr_pool *pool, int free_all)
479 {
480         unsigned int item_count;
481
482         item_count = atomic_read(&pool->item_count);
483         if (free_all)
484                 return item_count;
485
486         return 0;
487 }
488
489 /*
490  * Flush our pool of MRs.
491  * At a minimum, all currently unused MRs are unmapped.
492  * If the number of MRs allocated exceeds the limit, we also try
493  * to free as many MRs as needed to get back to this limit.
494  */
495 static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all)
496 {
497         struct rds_iw_mr *ibmr, *next;
498         LIST_HEAD(unmap_list);
499         LIST_HEAD(kill_list);
500         unsigned long flags;
501         unsigned int nfreed = 0, ncleaned = 0, free_goal;
502         int ret = 0;
503
504         rds_iw_stats_inc(s_iw_rdma_mr_pool_flush);
505
506         mutex_lock(&pool->flush_lock);
507
508         spin_lock_irqsave(&pool->list_lock, flags);
509         /* Get the list of all mappings to be destroyed */
510         list_splice_init(&pool->dirty_list, &unmap_list);
511         if (free_all)
512                 list_splice_init(&pool->clean_list, &kill_list);
513         spin_unlock_irqrestore(&pool->list_lock, flags);
514
515         free_goal = rds_iw_flush_goal(pool, free_all);
516
517         /* Batched invalidate of dirty MRs.
518          * For FMR based MRs, the mappings on the unmap list are
519          * actually members of an ibmr (ibmr->mapping). They either
520          * migrate to the kill_list, or have been cleaned and should be
521          * moved to the clean_list.
522          * For fastregs, they will be dynamically allocated, and
523          * will be destroyed by the unmap function.
524          */
525         if (!list_empty(&unmap_list)) {
526                 ncleaned = rds_iw_unmap_fastreg_list(pool, &unmap_list, &kill_list);
527                 /* If we've been asked to destroy all MRs, move those
528                  * that were simply cleaned to the kill list */
529                 if (free_all)
530                         list_splice_init(&unmap_list, &kill_list);
531         }
532
533         /* Destroy any MRs that are past their best before date */
534         list_for_each_entry_safe(ibmr, next, &kill_list, mapping.m_list) {
535                 rds_iw_stats_inc(s_iw_rdma_mr_free);
536                 list_del(&ibmr->mapping.m_list);
537                 rds_iw_destroy_fastreg(pool, ibmr);
538                 kfree(ibmr);
539                 nfreed++;
540         }
541
542         /* Anything that remains are laundered ibmrs, which we can add
543          * back to the clean list. */
544         if (!list_empty(&unmap_list)) {
545                 spin_lock_irqsave(&pool->list_lock, flags);
546                 list_splice(&unmap_list, &pool->clean_list);
547                 spin_unlock_irqrestore(&pool->list_lock, flags);
548         }
549
550         atomic_sub(ncleaned, &pool->dirty_count);
551         atomic_sub(nfreed, &pool->item_count);
552
553         mutex_unlock(&pool->flush_lock);
554         return ret;
555 }
556
557 static void rds_iw_mr_pool_flush_worker(struct work_struct *work)
558 {
559         struct rds_iw_mr_pool *pool = container_of(work, struct rds_iw_mr_pool, flush_worker);
560
561         rds_iw_flush_mr_pool(pool, 0);
562 }
563
564 void rds_iw_free_mr(void *trans_private, int invalidate)
565 {
566         struct rds_iw_mr *ibmr = trans_private;
567         struct rds_iw_mr_pool *pool = ibmr->device->mr_pool;
568
569         rdsdebug("RDS/IW: free_mr nents %u\n", ibmr->mapping.m_sg.len);
570         if (!pool)
571                 return;
572
573         /* Return it to the pool's free list */
574         rds_iw_free_fastreg(pool, ibmr);
575
576         /* If we've pinned too many pages, request a flush */
577         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
578             atomic_read(&pool->dirty_count) >= pool->max_items / 10)
579                 queue_work(rds_wq, &pool->flush_worker);
580
581         if (invalidate) {
582                 if (likely(!in_interrupt())) {
583                         rds_iw_flush_mr_pool(pool, 0);
584                 } else {
585                         /* We get here if the user created a MR marked
586                          * as use_once and invalidate at the same time. */
587                         queue_work(rds_wq, &pool->flush_worker);
588                 }
589         }
590 }
591
592 void rds_iw_flush_mrs(void)
593 {
594         struct rds_iw_device *rds_iwdev;
595
596         list_for_each_entry(rds_iwdev, &rds_iw_devices, list) {
597                 struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
598
599                 if (pool)
600                         rds_iw_flush_mr_pool(pool, 0);
601         }
602 }
603
604 void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
605                     struct rds_sock *rs, u32 *key_ret)
606 {
607         struct rds_iw_device *rds_iwdev;
608         struct rds_iw_mr *ibmr = NULL;
609         struct rdma_cm_id *cm_id;
610         int ret;
611
612         ret = rds_iw_get_device(rs, &rds_iwdev, &cm_id);
613         if (ret || !cm_id) {
614                 ret = -ENODEV;
615                 goto out;
616         }
617
618         if (!rds_iwdev->mr_pool) {
619                 ret = -ENODEV;
620                 goto out;
621         }
622
623         ibmr = rds_iw_alloc_mr(rds_iwdev);
624         if (IS_ERR(ibmr))
625                 return ibmr;
626
627         ibmr->cm_id = cm_id;
628         ibmr->device = rds_iwdev;
629
630         ret = rds_iw_map_fastreg(rds_iwdev->mr_pool, ibmr, sg, nents);
631         if (ret == 0)
632                 *key_ret = ibmr->mr->rkey;
633         else
634                 printk(KERN_WARNING "RDS/IW: failed to map mr (errno=%d)\n", ret);
635
636 out:
637         if (ret) {
638                 if (ibmr)
639                         rds_iw_free_mr(ibmr, 0);
640                 ibmr = ERR_PTR(ret);
641         }
642         return ibmr;
643 }
644
645 /*
646  * iWARP fastreg handling
647  *
648  * The life cycle of a fastreg registration is a bit different from
649  * FMRs.
650  * The idea behind fastreg is to have one MR, to which we bind different
651  * mappings over time. To avoid stalling on the expensive map and invalidate
652  * operations, these operations are pipelined on the same send queue on
653  * which we want to send the message containing the r_key.
654  *
655  * This creates a bit of a problem for us, as we do not have the destination
656  * IP in GET_MR, so the connection must be setup prior to the GET_MR call for
657  * RDMA to be correctly setup.  If a fastreg request is present, rds_iw_xmit
658  * will try to queue a LOCAL_INV (if needed) and a FAST_REG_MR work request
659  * before queuing the SEND. When completions for these arrive, they are
660  * dispatched to the MR has a bit set showing that RDMa can be performed.
661  *
662  * There is another interesting aspect that's related to invalidation.
663  * The application can request that a mapping is invalidated in FREE_MR.
664  * The expectation there is that this invalidation step includes ALL
665  * PREVIOUSLY FREED MRs.
666  */
667 static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool,
668                                 struct rds_iw_mr *ibmr)
669 {
670         struct rds_iw_device *rds_iwdev = pool->device;
671         struct ib_fast_reg_page_list *page_list = NULL;
672         struct ib_mr *mr;
673         int err;
674
675         mr = ib_alloc_fast_reg_mr(rds_iwdev->pd, pool->max_message_size);
676         if (IS_ERR(mr)) {
677                 err = PTR_ERR(mr);
678
679                 printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_mr failed (err=%d)\n", err);
680                 return err;
681         }
682
683         /* FIXME - this is overkill, but mapping->m_sg.dma_len/mapping->m_sg.dma_npages
684          * is not filled in.
685          */
686         page_list = ib_alloc_fast_reg_page_list(rds_iwdev->dev, pool->max_message_size);
687         if (IS_ERR(page_list)) {
688                 err = PTR_ERR(page_list);
689
690                 printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed (err=%d)\n", err);
691                 ib_dereg_mr(mr);
692                 return err;
693         }
694
695         ibmr->page_list = page_list;
696         ibmr->mr = mr;
697         return 0;
698 }
699
700 static int rds_iw_rdma_build_fastreg(struct rds_iw_mapping *mapping)
701 {
702         struct rds_iw_mr *ibmr = mapping->m_mr;
703         struct ib_send_wr f_wr, *failed_wr;
704         int ret;
705
706         /*
707          * Perform a WR for the fast_reg_mr. Each individual page
708          * in the sg list is added to the fast reg page list and placed
709          * inside the fast_reg_mr WR.  The key used is a rolling 8bit
710          * counter, which should guarantee uniqueness.
711          */
712         ib_update_fast_reg_key(ibmr->mr, ibmr->remap_count++);
713         mapping->m_rkey = ibmr->mr->rkey;
714
715         memset(&f_wr, 0, sizeof(f_wr));
716         f_wr.wr_id = RDS_IW_FAST_REG_WR_ID;
717         f_wr.opcode = IB_WR_FAST_REG_MR;
718         f_wr.wr.fast_reg.length = mapping->m_sg.bytes;
719         f_wr.wr.fast_reg.rkey = mapping->m_rkey;
720         f_wr.wr.fast_reg.page_list = ibmr->page_list;
721         f_wr.wr.fast_reg.page_list_len = mapping->m_sg.dma_len;
722         f_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
723         f_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE |
724                                 IB_ACCESS_REMOTE_READ |
725                                 IB_ACCESS_REMOTE_WRITE;
726         f_wr.wr.fast_reg.iova_start = 0;
727         f_wr.send_flags = IB_SEND_SIGNALED;
728
729         failed_wr = &f_wr;
730         ret = ib_post_send(ibmr->cm_id->qp, &f_wr, &failed_wr);
731         BUG_ON(failed_wr != &f_wr);
732         if (ret && printk_ratelimit())
733                 printk(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
734                         __func__, __LINE__, ret);
735         return ret;
736 }
737
738 static int rds_iw_rdma_fastreg_inv(struct rds_iw_mr *ibmr)
739 {
740         struct ib_send_wr s_wr, *failed_wr;
741         int ret = 0;
742
743         if (!ibmr->cm_id->qp || !ibmr->mr)
744                 goto out;
745
746         memset(&s_wr, 0, sizeof(s_wr));
747         s_wr.wr_id = RDS_IW_LOCAL_INV_WR_ID;
748         s_wr.opcode = IB_WR_LOCAL_INV;
749         s_wr.ex.invalidate_rkey = ibmr->mr->rkey;
750         s_wr.send_flags = IB_SEND_SIGNALED;
751
752         failed_wr = &s_wr;
753         ret = ib_post_send(ibmr->cm_id->qp, &s_wr, &failed_wr);
754         if (ret && printk_ratelimit()) {
755                 printk(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
756                         __func__, __LINE__, ret);
757                 goto out;
758         }
759 out:
760         return ret;
761 }
762
763 static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
764                         struct rds_iw_mr *ibmr,
765                         struct scatterlist *sg,
766                         unsigned int sg_len)
767 {
768         struct rds_iw_device *rds_iwdev = pool->device;
769         struct rds_iw_mapping *mapping = &ibmr->mapping;
770         u64 *dma_pages;
771         int i, ret = 0;
772
773         rds_iw_set_scatterlist(&mapping->m_sg, sg, sg_len);
774
775         dma_pages = rds_iw_map_scatterlist(rds_iwdev, &mapping->m_sg);
776         if (IS_ERR(dma_pages)) {
777                 ret = PTR_ERR(dma_pages);
778                 dma_pages = NULL;
779                 goto out;
780         }
781
782         if (mapping->m_sg.dma_len > pool->max_message_size) {
783                 ret = -EMSGSIZE;
784                 goto out;
785         }
786
787         for (i = 0; i < mapping->m_sg.dma_npages; ++i)
788                 ibmr->page_list->page_list[i] = dma_pages[i];
789
790         ret = rds_iw_rdma_build_fastreg(mapping);
791         if (ret)
792                 goto out;
793
794         rds_iw_stats_inc(s_iw_rdma_mr_used);
795
796 out:
797         kfree(dma_pages);
798
799         return ret;
800 }
801
802 /*
803  * "Free" a fastreg MR.
804  */
805 static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool,
806                 struct rds_iw_mr *ibmr)
807 {
808         unsigned long flags;
809         int ret;
810
811         if (!ibmr->mapping.m_sg.dma_len)
812                 return;
813
814         ret = rds_iw_rdma_fastreg_inv(ibmr);
815         if (ret)
816                 return;
817
818         /* Try to post the LOCAL_INV WR to the queue. */
819         spin_lock_irqsave(&pool->list_lock, flags);
820
821         list_add_tail(&ibmr->mapping.m_list, &pool->dirty_list);
822         atomic_add(ibmr->mapping.m_sg.len, &pool->free_pinned);
823         atomic_inc(&pool->dirty_count);
824
825         spin_unlock_irqrestore(&pool->list_lock, flags);
826 }
827
828 static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
829                                 struct list_head *unmap_list,
830                                 struct list_head *kill_list)
831 {
832         struct rds_iw_mapping *mapping, *next;
833         unsigned int ncleaned = 0;
834         LIST_HEAD(laundered);
835
836         /* Batched invalidation of fastreg MRs.
837          * Why do we do it this way, even though we could pipeline unmap
838          * and remap? The reason is the application semantics - when the
839          * application requests an invalidation of MRs, it expects all
840          * previously released R_Keys to become invalid.
841          *
842          * If we implement MR reuse naively, we risk memory corruption
843          * (this has actually been observed). So the default behavior
844          * requires that a MR goes through an explicit unmap operation before
845          * we can reuse it again.
846          *
847          * We could probably improve on this a little, by allowing immediate
848          * reuse of a MR on the same socket (eg you could add small
849          * cache of unused MRs to strct rds_socket - GET_MR could grab one
850          * of these without requiring an explicit invalidate).
851          */
852         while (!list_empty(unmap_list)) {
853                 unsigned long flags;
854
855                 spin_lock_irqsave(&pool->list_lock, flags);
856                 list_for_each_entry_safe(mapping, next, unmap_list, m_list) {
857                         list_move(&mapping->m_list, &laundered);
858                         ncleaned++;
859                 }
860                 spin_unlock_irqrestore(&pool->list_lock, flags);
861         }
862
863         /* Move all laundered mappings back to the unmap list.
864          * We do not kill any WRs right now - it doesn't seem the
865          * fastreg API has a max_remap limit. */
866         list_splice_init(&laundered, unmap_list);
867
868         return ncleaned;
869 }
870
871 static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool,
872                 struct rds_iw_mr *ibmr)
873 {
874         if (ibmr->page_list)
875                 ib_free_fast_reg_page_list(ibmr->page_list);
876         if (ibmr->mr)
877                 ib_dereg_mr(ibmr->mr);
878 }