Root/net/rds/ib_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 "ib.h"
39
40
41/*
42 * This is stored as mr->r_trans_private.
43 */
44struct rds_ib_mr {
45    struct rds_ib_device *device;
46    struct rds_ib_mr_pool *pool;
47    struct ib_fmr *fmr;
48    struct list_head list;
49    unsigned int remap_count;
50
51    struct scatterlist *sg;
52    unsigned int sg_len;
53    u64 *dma;
54    int sg_dma_len;
55};
56
57/*
58 * Our own little FMR pool
59 */
60struct rds_ib_mr_pool {
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 drop_list; /* MRs that have reached their max_maps limit */
68    struct list_head free_list; /* unused MRs */
69    struct list_head clean_list; /* unused & unamapped MRs */
70    atomic_t free_pinned; /* memory pinned by free MRs */
71    unsigned long max_items;
72    unsigned long max_items_soft;
73    unsigned long max_free_pinned;
74    struct ib_fmr_attr fmr_attr;
75};
76
77static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all);
78static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
79static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
80
81static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
82{
83    struct rds_ib_device *rds_ibdev;
84    struct rds_ib_ipaddr *i_ipaddr;
85
86    list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
87        spin_lock_irq(&rds_ibdev->spinlock);
88        list_for_each_entry(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
89            if (i_ipaddr->ipaddr == ipaddr) {
90                spin_unlock_irq(&rds_ibdev->spinlock);
91                return rds_ibdev;
92            }
93        }
94        spin_unlock_irq(&rds_ibdev->spinlock);
95    }
96
97    return NULL;
98}
99
100static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
101{
102    struct rds_ib_ipaddr *i_ipaddr;
103
104    i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
105    if (!i_ipaddr)
106        return -ENOMEM;
107
108    i_ipaddr->ipaddr = ipaddr;
109
110    spin_lock_irq(&rds_ibdev->spinlock);
111    list_add_tail(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
112    spin_unlock_irq(&rds_ibdev->spinlock);
113
114    return 0;
115}
116
117static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
118{
119    struct rds_ib_ipaddr *i_ipaddr, *next;
120
121    spin_lock_irq(&rds_ibdev->spinlock);
122    list_for_each_entry_safe(i_ipaddr, next, &rds_ibdev->ipaddr_list, list) {
123        if (i_ipaddr->ipaddr == ipaddr) {
124            list_del(&i_ipaddr->list);
125            kfree(i_ipaddr);
126            break;
127        }
128    }
129    spin_unlock_irq(&rds_ibdev->spinlock);
130}
131
132int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
133{
134    struct rds_ib_device *rds_ibdev_old;
135
136    rds_ibdev_old = rds_ib_get_device(ipaddr);
137    if (rds_ibdev_old)
138        rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
139
140    return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
141}
142
143void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
144{
145    struct rds_ib_connection *ic = conn->c_transport_data;
146
147    /* conn was previously on the nodev_conns_list */
148    spin_lock_irq(&ib_nodev_conns_lock);
149    BUG_ON(list_empty(&ib_nodev_conns));
150    BUG_ON(list_empty(&ic->ib_node));
151    list_del(&ic->ib_node);
152
153    spin_lock_irq(&rds_ibdev->spinlock);
154    list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
155    spin_unlock_irq(&rds_ibdev->spinlock);
156    spin_unlock_irq(&ib_nodev_conns_lock);
157
158    ic->rds_ibdev = rds_ibdev;
159}
160
161void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
162{
163    struct rds_ib_connection *ic = conn->c_transport_data;
164
165    /* place conn on nodev_conns_list */
166    spin_lock(&ib_nodev_conns_lock);
167
168    spin_lock_irq(&rds_ibdev->spinlock);
169    BUG_ON(list_empty(&ic->ib_node));
170    list_del(&ic->ib_node);
171    spin_unlock_irq(&rds_ibdev->spinlock);
172
173    list_add_tail(&ic->ib_node, &ib_nodev_conns);
174
175    spin_unlock(&ib_nodev_conns_lock);
176
177    ic->rds_ibdev = NULL;
178}
179
180void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock)
181{
182    struct rds_ib_connection *ic, *_ic;
183    LIST_HEAD(tmp_list);
184
185    /* avoid calling conn_destroy with irqs off */
186    spin_lock_irq(list_lock);
187    list_splice(list, &tmp_list);
188    INIT_LIST_HEAD(list);
189    spin_unlock_irq(list_lock);
190
191    list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
192        rds_conn_destroy(ic->conn);
193}
194
195struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
196{
197    struct rds_ib_mr_pool *pool;
198
199    pool = kzalloc(sizeof(*pool), GFP_KERNEL);
200    if (!pool)
201        return ERR_PTR(-ENOMEM);
202
203    INIT_LIST_HEAD(&pool->free_list);
204    INIT_LIST_HEAD(&pool->drop_list);
205    INIT_LIST_HEAD(&pool->clean_list);
206    mutex_init(&pool->flush_lock);
207    spin_lock_init(&pool->list_lock);
208    INIT_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
209
210    pool->fmr_attr.max_pages = fmr_message_size;
211    pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
212    pool->fmr_attr.page_shift = PAGE_SHIFT;
213    pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
214
215    /* We never allow more than max_items MRs to be allocated.
216     * When we exceed more than max_items_soft, we start freeing
217     * items more aggressively.
218     * Make sure that max_items > max_items_soft > max_items / 2
219     */
220    pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
221    pool->max_items = rds_ibdev->max_fmrs;
222
223    return pool;
224}
225
226void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
227{
228    struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
229
230    iinfo->rdma_mr_max = pool->max_items;
231    iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
232}
233
234void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
235{
236    flush_workqueue(rds_wq);
237    rds_ib_flush_mr_pool(pool, 1);
238    BUG_ON(atomic_read(&pool->item_count));
239    BUG_ON(atomic_read(&pool->free_pinned));
240    kfree(pool);
241}
242
243static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
244{
245    struct rds_ib_mr *ibmr = NULL;
246    unsigned long flags;
247
248    spin_lock_irqsave(&pool->list_lock, flags);
249    if (!list_empty(&pool->clean_list)) {
250        ibmr = list_entry(pool->clean_list.next, struct rds_ib_mr, list);
251        list_del_init(&ibmr->list);
252    }
253    spin_unlock_irqrestore(&pool->list_lock, flags);
254
255    return ibmr;
256}
257
258static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
259{
260    struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
261    struct rds_ib_mr *ibmr = NULL;
262    int err = 0, iter = 0;
263
264    while (1) {
265        ibmr = rds_ib_reuse_fmr(pool);
266        if (ibmr)
267            return ibmr;
268
269        /* No clean MRs - now we have the choice of either
270         * allocating a fresh MR up to the limit imposed by the
271         * driver, or flush any dirty unused MRs.
272         * We try to avoid stalling in the send path if possible,
273         * so we allocate as long as we're allowed to.
274         *
275         * We're fussy with enforcing the FMR limit, though. If the driver
276         * tells us we can't use more than N fmrs, we shouldn't start
277         * arguing with it */
278        if (atomic_inc_return(&pool->item_count) <= pool->max_items)
279            break;
280
281        atomic_dec(&pool->item_count);
282
283        if (++iter > 2) {
284            rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
285            return ERR_PTR(-EAGAIN);
286        }
287
288        /* We do have some empty MRs. Flush them out. */
289        rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
290        rds_ib_flush_mr_pool(pool, 0);
291    }
292
293    ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
294    if (!ibmr) {
295        err = -ENOMEM;
296        goto out_no_cigar;
297    }
298
299    ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
300            (IB_ACCESS_LOCAL_WRITE |
301             IB_ACCESS_REMOTE_READ |
302             IB_ACCESS_REMOTE_WRITE),
303            &pool->fmr_attr);
304    if (IS_ERR(ibmr->fmr)) {
305        err = PTR_ERR(ibmr->fmr);
306        ibmr->fmr = NULL;
307        printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
308        goto out_no_cigar;
309    }
310
311    rds_ib_stats_inc(s_ib_rdma_mr_alloc);
312    return ibmr;
313
314out_no_cigar:
315    if (ibmr) {
316        if (ibmr->fmr)
317            ib_dealloc_fmr(ibmr->fmr);
318        kfree(ibmr);
319    }
320    atomic_dec(&pool->item_count);
321    return ERR_PTR(err);
322}
323
324static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
325           struct scatterlist *sg, unsigned int nents)
326{
327    struct ib_device *dev = rds_ibdev->dev;
328    struct scatterlist *scat = sg;
329    u64 io_addr = 0;
330    u64 *dma_pages;
331    u32 len;
332    int page_cnt, sg_dma_len;
333    int i, j;
334    int ret;
335
336    sg_dma_len = ib_dma_map_sg(dev, sg, nents,
337                 DMA_BIDIRECTIONAL);
338    if (unlikely(!sg_dma_len)) {
339        printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
340        return -EBUSY;
341    }
342
343    len = 0;
344    page_cnt = 0;
345
346    for (i = 0; i < sg_dma_len; ++i) {
347        unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
348        u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
349
350        if (dma_addr & ~PAGE_MASK) {
351            if (i > 0)
352                return -EINVAL;
353            else
354                ++page_cnt;
355        }
356        if ((dma_addr + dma_len) & ~PAGE_MASK) {
357            if (i < sg_dma_len - 1)
358                return -EINVAL;
359            else
360                ++page_cnt;
361        }
362
363        len += dma_len;
364    }
365
366    page_cnt += len >> PAGE_SHIFT;
367    if (page_cnt > fmr_message_size)
368        return -EINVAL;
369
370    dma_pages = kmalloc(sizeof(u64) * page_cnt, GFP_ATOMIC);
371    if (!dma_pages)
372        return -ENOMEM;
373
374    page_cnt = 0;
375    for (i = 0; i < sg_dma_len; ++i) {
376        unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
377        u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
378
379        for (j = 0; j < dma_len; j += PAGE_SIZE)
380            dma_pages[page_cnt++] =
381                (dma_addr & PAGE_MASK) + j;
382    }
383
384    ret = ib_map_phys_fmr(ibmr->fmr,
385                   dma_pages, page_cnt, io_addr);
386    if (ret)
387        goto out;
388
389    /* Success - we successfully remapped the MR, so we can
390     * safely tear down the old mapping. */
391    rds_ib_teardown_mr(ibmr);
392
393    ibmr->sg = scat;
394    ibmr->sg_len = nents;
395    ibmr->sg_dma_len = sg_dma_len;
396    ibmr->remap_count++;
397
398    rds_ib_stats_inc(s_ib_rdma_mr_used);
399    ret = 0;
400
401out:
402    kfree(dma_pages);
403
404    return ret;
405}
406
407void rds_ib_sync_mr(void *trans_private, int direction)
408{
409    struct rds_ib_mr *ibmr = trans_private;
410    struct rds_ib_device *rds_ibdev = ibmr->device;
411
412    switch (direction) {
413    case DMA_FROM_DEVICE:
414        ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
415            ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
416        break;
417    case DMA_TO_DEVICE:
418        ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
419            ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
420        break;
421    }
422}
423
424static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
425{
426    struct rds_ib_device *rds_ibdev = ibmr->device;
427
428    if (ibmr->sg_dma_len) {
429        ib_dma_unmap_sg(rds_ibdev->dev,
430                ibmr->sg, ibmr->sg_len,
431                DMA_BIDIRECTIONAL);
432        ibmr->sg_dma_len = 0;
433    }
434
435    /* Release the s/g list */
436    if (ibmr->sg_len) {
437        unsigned int i;
438
439        for (i = 0; i < ibmr->sg_len; ++i) {
440            struct page *page = sg_page(&ibmr->sg[i]);
441
442            /* FIXME we need a way to tell a r/w MR
443             * from a r/o MR */
444            set_page_dirty(page);
445            put_page(page);
446        }
447        kfree(ibmr->sg);
448
449        ibmr->sg = NULL;
450        ibmr->sg_len = 0;
451    }
452}
453
454static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
455{
456    unsigned int pinned = ibmr->sg_len;
457
458    __rds_ib_teardown_mr(ibmr);
459    if (pinned) {
460        struct rds_ib_device *rds_ibdev = ibmr->device;
461        struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
462
463        atomic_sub(pinned, &pool->free_pinned);
464    }
465}
466
467static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
468{
469    unsigned int item_count;
470
471    item_count = atomic_read(&pool->item_count);
472    if (free_all)
473        return item_count;
474
475    return 0;
476}
477
478/*
479 * Flush our pool of MRs.
480 * At a minimum, all currently unused MRs are unmapped.
481 * If the number of MRs allocated exceeds the limit, we also try
482 * to free as many MRs as needed to get back to this limit.
483 */
484static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all)
485{
486    struct rds_ib_mr *ibmr, *next;
487    LIST_HEAD(unmap_list);
488    LIST_HEAD(fmr_list);
489    unsigned long unpinned = 0;
490    unsigned long flags;
491    unsigned int nfreed = 0, ncleaned = 0, free_goal;
492    int ret = 0;
493
494    rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
495
496    mutex_lock(&pool->flush_lock);
497
498    spin_lock_irqsave(&pool->list_lock, flags);
499    /* Get the list of all MRs to be dropped. Ordering matters -
500     * we want to put drop_list ahead of free_list. */
501    list_splice_init(&pool->free_list, &unmap_list);
502    list_splice_init(&pool->drop_list, &unmap_list);
503    if (free_all)
504        list_splice_init(&pool->clean_list, &unmap_list);
505    spin_unlock_irqrestore(&pool->list_lock, flags);
506
507    free_goal = rds_ib_flush_goal(pool, free_all);
508
509    if (list_empty(&unmap_list))
510        goto out;
511
512    /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
513    list_for_each_entry(ibmr, &unmap_list, list)
514        list_add(&ibmr->fmr->list, &fmr_list);
515    ret = ib_unmap_fmr(&fmr_list);
516    if (ret)
517        printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
518
519    /* Now we can destroy the DMA mapping and unpin any pages */
520    list_for_each_entry_safe(ibmr, next, &unmap_list, list) {
521        unpinned += ibmr->sg_len;
522        __rds_ib_teardown_mr(ibmr);
523        if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
524            rds_ib_stats_inc(s_ib_rdma_mr_free);
525            list_del(&ibmr->list);
526            ib_dealloc_fmr(ibmr->fmr);
527            kfree(ibmr);
528            nfreed++;
529        }
530        ncleaned++;
531    }
532
533    spin_lock_irqsave(&pool->list_lock, flags);
534    list_splice(&unmap_list, &pool->clean_list);
535    spin_unlock_irqrestore(&pool->list_lock, flags);
536
537    atomic_sub(unpinned, &pool->free_pinned);
538    atomic_sub(ncleaned, &pool->dirty_count);
539    atomic_sub(nfreed, &pool->item_count);
540
541out:
542    mutex_unlock(&pool->flush_lock);
543    return ret;
544}
545
546static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
547{
548    struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker);
549
550    rds_ib_flush_mr_pool(pool, 0);
551}
552
553void rds_ib_free_mr(void *trans_private, int invalidate)
554{
555    struct rds_ib_mr *ibmr = trans_private;
556    struct rds_ib_device *rds_ibdev = ibmr->device;
557    struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
558    unsigned long flags;
559
560    rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
561
562    /* Return it to the pool's free list */
563    spin_lock_irqsave(&pool->list_lock, flags);
564    if (ibmr->remap_count >= pool->fmr_attr.max_maps)
565        list_add(&ibmr->list, &pool->drop_list);
566    else
567        list_add(&ibmr->list, &pool->free_list);
568
569    atomic_add(ibmr->sg_len, &pool->free_pinned);
570    atomic_inc(&pool->dirty_count);
571    spin_unlock_irqrestore(&pool->list_lock, flags);
572
573    /* If we've pinned too many pages, request a flush */
574    if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
575        atomic_read(&pool->dirty_count) >= pool->max_items / 10)
576        queue_work(rds_wq, &pool->flush_worker);
577
578    if (invalidate) {
579        if (likely(!in_interrupt())) {
580            rds_ib_flush_mr_pool(pool, 0);
581        } else {
582            /* We get here if the user created a MR marked
583             * as use_once and invalidate at the same time. */
584            queue_work(rds_wq, &pool->flush_worker);
585        }
586    }
587}
588
589void rds_ib_flush_mrs(void)
590{
591    struct rds_ib_device *rds_ibdev;
592
593    list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
594        struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
595
596        if (pool)
597            rds_ib_flush_mr_pool(pool, 0);
598    }
599}
600
601void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
602            struct rds_sock *rs, u32 *key_ret)
603{
604    struct rds_ib_device *rds_ibdev;
605    struct rds_ib_mr *ibmr = NULL;
606    int ret;
607
608    rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
609    if (!rds_ibdev) {
610        ret = -ENODEV;
611        goto out;
612    }
613
614    if (!rds_ibdev->mr_pool) {
615        ret = -ENODEV;
616        goto out;
617    }
618
619    ibmr = rds_ib_alloc_fmr(rds_ibdev);
620    if (IS_ERR(ibmr))
621        return ibmr;
622
623    ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
624    if (ret == 0)
625        *key_ret = ibmr->fmr->rkey;
626    else
627        printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
628
629    ibmr->device = rds_ibdev;
630
631 out:
632    if (ret) {
633        if (ibmr)
634            rds_ib_free_mr(ibmr, 0);
635        ibmr = ERR_PTR(ret);
636    }
637    return ibmr;
638}
639

Archive Download this file



interactive