2 * Copyright (c) 2013, Mellanox Technologies inc. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
34 #include <linux/kref.h>
35 #include <linux/random.h>
36 #include <linux/debugfs.h>
37 #include <linux/export.h>
38 #include <linux/delay.h>
39 #include <rdma/ib_umem.h>
44 MAX_PENDING_REG_MR = 8,
51 static __be64 *mr_align(__be64 *ptr, int align)
53 unsigned long mask = align - 1;
55 return (__be64 *)(((unsigned long)ptr + mask) & ~mask);
58 static int order2idx(struct mlx5_ib_dev *dev, int order)
60 struct mlx5_mr_cache *cache = &dev->cache;
62 if (order < cache->ent[0].order)
65 return order - cache->ent[0].order;
68 static void reg_mr_callback(int status, void *context)
70 struct mlx5_ib_mr *mr = context;
71 struct mlx5_ib_dev *dev = mr->dev;
72 struct mlx5_mr_cache *cache = &dev->cache;
73 int c = order2idx(dev, mr->order);
74 struct mlx5_cache_ent *ent = &cache->ent[c];
76 unsigned long delta = jiffies - mr->start;
80 index = find_last_bit(&delta, 8 * sizeof(delta));
84 spin_lock_irqsave(&ent->lock, flags);
86 spin_unlock_irqrestore(&ent->lock, flags);
88 mlx5_ib_warn(dev, "async reg mr failed. status %d\n", status);
91 mod_timer(&dev->delay_timer, jiffies + HZ);
95 if (mr->out.hdr.status) {
96 mlx5_ib_warn(dev, "failed - status %d, syndorme 0x%x\n",
98 be32_to_cpu(mr->out.hdr.syndrome));
101 mod_timer(&dev->delay_timer, jiffies + HZ);
105 spin_lock_irqsave(&dev->mdev.priv.mkey_lock, flags);
106 key = dev->mdev.priv.mkey_key++;
107 spin_unlock_irqrestore(&dev->mdev.priv.mkey_lock, flags);
108 mr->mmr.key = mlx5_idx_to_mkey(be32_to_cpu(mr->out.mkey) & 0xffffff) | key;
110 cache->last_add = jiffies;
112 spin_lock_irqsave(&ent->lock, flags);
113 list_add_tail(&mr->list, &ent->head);
116 spin_unlock_irqrestore(&ent->lock, flags);
119 static int add_keys(struct mlx5_ib_dev *dev, int c, int num)
121 struct mlx5_mr_cache *cache = &dev->cache;
122 struct mlx5_cache_ent *ent = &cache->ent[c];
123 struct mlx5_create_mkey_mbox_in *in;
124 struct mlx5_ib_mr *mr;
125 int npages = 1 << ent->order;
129 in = kzalloc(sizeof(*in), GFP_KERNEL);
133 for (i = 0; i < num; i++) {
134 if (ent->pending >= MAX_PENDING_REG_MR) {
139 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
144 mr->order = ent->order;
147 in->seg.status = 1 << 6;
148 in->seg.xlt_oct_size = cpu_to_be32((npages + 1) / 2);
149 in->seg.qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
150 in->seg.flags = MLX5_ACCESS_MODE_MTT | MLX5_PERM_UMR_EN;
151 in->seg.log2_page_size = 12;
153 spin_lock_irq(&ent->lock);
155 spin_unlock_irq(&ent->lock);
157 err = mlx5_core_create_mkey(&dev->mdev, &mr->mmr, in,
158 sizeof(*in), reg_mr_callback,
161 mlx5_ib_warn(dev, "create mkey failed %d\n", err);
171 static void remove_keys(struct mlx5_ib_dev *dev, int c, int num)
173 struct mlx5_mr_cache *cache = &dev->cache;
174 struct mlx5_cache_ent *ent = &cache->ent[c];
175 struct mlx5_ib_mr *mr;
179 for (i = 0; i < num; i++) {
180 spin_lock_irq(&ent->lock);
181 if (list_empty(&ent->head)) {
182 spin_unlock_irq(&ent->lock);
185 mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
189 spin_unlock_irq(&ent->lock);
190 err = mlx5_core_destroy_mkey(&dev->mdev, &mr->mmr);
192 mlx5_ib_warn(dev, "failed destroy mkey\n");
198 static ssize_t size_write(struct file *filp, const char __user *buf,
199 size_t count, loff_t *pos)
201 struct mlx5_cache_ent *ent = filp->private_data;
202 struct mlx5_ib_dev *dev = ent->dev;
208 if (copy_from_user(lbuf, buf, sizeof(lbuf)))
211 c = order2idx(dev, ent->order);
212 lbuf[sizeof(lbuf) - 1] = 0;
214 if (sscanf(lbuf, "%u", &var) != 1)
217 if (var < ent->limit)
220 if (var > ent->size) {
222 err = add_keys(dev, c, var - ent->size);
223 if (err && err != -EAGAIN)
226 usleep_range(3000, 5000);
228 } else if (var < ent->size) {
229 remove_keys(dev, c, ent->size - var);
235 static ssize_t size_read(struct file *filp, char __user *buf, size_t count,
238 struct mlx5_cache_ent *ent = filp->private_data;
245 err = snprintf(lbuf, sizeof(lbuf), "%d\n", ent->size);
249 if (copy_to_user(buf, lbuf, err))
257 static const struct file_operations size_fops = {
258 .owner = THIS_MODULE,
264 static ssize_t limit_write(struct file *filp, const char __user *buf,
265 size_t count, loff_t *pos)
267 struct mlx5_cache_ent *ent = filp->private_data;
268 struct mlx5_ib_dev *dev = ent->dev;
274 if (copy_from_user(lbuf, buf, sizeof(lbuf)))
277 c = order2idx(dev, ent->order);
278 lbuf[sizeof(lbuf) - 1] = 0;
280 if (sscanf(lbuf, "%u", &var) != 1)
288 if (ent->cur < ent->limit) {
289 err = add_keys(dev, c, 2 * ent->limit - ent->cur);
297 static ssize_t limit_read(struct file *filp, char __user *buf, size_t count,
300 struct mlx5_cache_ent *ent = filp->private_data;
307 err = snprintf(lbuf, sizeof(lbuf), "%d\n", ent->limit);
311 if (copy_to_user(buf, lbuf, err))
319 static const struct file_operations limit_fops = {
320 .owner = THIS_MODULE,
322 .write = limit_write,
326 static int someone_adding(struct mlx5_mr_cache *cache)
330 for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
331 if (cache->ent[i].cur < cache->ent[i].limit)
338 static void __cache_work_func(struct mlx5_cache_ent *ent)
340 struct mlx5_ib_dev *dev = ent->dev;
341 struct mlx5_mr_cache *cache = &dev->cache;
342 int i = order2idx(dev, ent->order);
348 ent = &dev->cache.ent[i];
349 if (ent->cur < 2 * ent->limit && !dev->fill_delay) {
350 err = add_keys(dev, i, 1);
351 if (ent->cur < 2 * ent->limit) {
352 if (err == -EAGAIN) {
353 mlx5_ib_dbg(dev, "returned eagain, order %d\n",
355 queue_delayed_work(cache->wq, &ent->dwork,
356 msecs_to_jiffies(3));
358 mlx5_ib_warn(dev, "command failed order %d, err %d\n",
360 queue_delayed_work(cache->wq, &ent->dwork,
361 msecs_to_jiffies(1000));
363 queue_work(cache->wq, &ent->work);
366 } else if (ent->cur > 2 * ent->limit) {
367 if (!someone_adding(cache) &&
368 time_after(jiffies, cache->last_add + 300 * HZ)) {
369 remove_keys(dev, i, 1);
370 if (ent->cur > ent->limit)
371 queue_work(cache->wq, &ent->work);
373 queue_delayed_work(cache->wq, &ent->dwork, 300 * HZ);
378 static void delayed_cache_work_func(struct work_struct *work)
380 struct mlx5_cache_ent *ent;
382 ent = container_of(work, struct mlx5_cache_ent, dwork.work);
383 __cache_work_func(ent);
386 static void cache_work_func(struct work_struct *work)
388 struct mlx5_cache_ent *ent;
390 ent = container_of(work, struct mlx5_cache_ent, work);
391 __cache_work_func(ent);
394 static struct mlx5_ib_mr *alloc_cached_mr(struct mlx5_ib_dev *dev, int order)
396 struct mlx5_mr_cache *cache = &dev->cache;
397 struct mlx5_ib_mr *mr = NULL;
398 struct mlx5_cache_ent *ent;
402 c = order2idx(dev, order);
403 if (c < 0 || c >= MAX_MR_CACHE_ENTRIES) {
404 mlx5_ib_warn(dev, "order %d, cache index %d\n", order, c);
408 for (i = c; i < MAX_MR_CACHE_ENTRIES; i++) {
409 ent = &cache->ent[i];
411 mlx5_ib_dbg(dev, "order %d, cache index %d\n", ent->order, i);
413 spin_lock_irq(&ent->lock);
414 if (!list_empty(&ent->head)) {
415 mr = list_first_entry(&ent->head, struct mlx5_ib_mr,
419 spin_unlock_irq(&ent->lock);
420 if (ent->cur < ent->limit)
421 queue_work(cache->wq, &ent->work);
424 spin_unlock_irq(&ent->lock);
426 queue_work(cache->wq, &ent->work);
433 cache->ent[c].miss++;
438 static void free_cached_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
440 struct mlx5_mr_cache *cache = &dev->cache;
441 struct mlx5_cache_ent *ent;
445 c = order2idx(dev, mr->order);
446 if (c < 0 || c >= MAX_MR_CACHE_ENTRIES) {
447 mlx5_ib_warn(dev, "order %d, cache index %d\n", mr->order, c);
450 ent = &cache->ent[c];
451 spin_lock_irq(&ent->lock);
452 list_add_tail(&mr->list, &ent->head);
454 if (ent->cur > 2 * ent->limit)
456 spin_unlock_irq(&ent->lock);
459 queue_work(cache->wq, &ent->work);
462 static void clean_keys(struct mlx5_ib_dev *dev, int c)
464 struct mlx5_mr_cache *cache = &dev->cache;
465 struct mlx5_cache_ent *ent = &cache->ent[c];
466 struct mlx5_ib_mr *mr;
469 cancel_delayed_work(&ent->dwork);
471 spin_lock_irq(&ent->lock);
472 if (list_empty(&ent->head)) {
473 spin_unlock_irq(&ent->lock);
476 mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
480 spin_unlock_irq(&ent->lock);
481 err = mlx5_core_destroy_mkey(&dev->mdev, &mr->mmr);
483 mlx5_ib_warn(dev, "failed destroy mkey\n");
489 static int mlx5_mr_cache_debugfs_init(struct mlx5_ib_dev *dev)
491 struct mlx5_mr_cache *cache = &dev->cache;
492 struct mlx5_cache_ent *ent;
495 if (!mlx5_debugfs_root)
498 cache->root = debugfs_create_dir("mr_cache", dev->mdev.priv.dbg_root);
502 for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
503 ent = &cache->ent[i];
504 sprintf(ent->name, "%d", ent->order);
505 ent->dir = debugfs_create_dir(ent->name, cache->root);
509 ent->fsize = debugfs_create_file("size", 0600, ent->dir, ent,
514 ent->flimit = debugfs_create_file("limit", 0600, ent->dir, ent,
519 ent->fcur = debugfs_create_u32("cur", 0400, ent->dir,
524 ent->fmiss = debugfs_create_u32("miss", 0600, ent->dir,
533 static void mlx5_mr_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
535 if (!mlx5_debugfs_root)
538 debugfs_remove_recursive(dev->cache.root);
541 static void delay_time_func(unsigned long ctx)
543 struct mlx5_ib_dev *dev = (struct mlx5_ib_dev *)ctx;
548 int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
550 struct mlx5_mr_cache *cache = &dev->cache;
551 struct mlx5_cache_ent *ent;
557 cache->wq = create_singlethread_workqueue("mkey_cache");
559 mlx5_ib_warn(dev, "failed to create work queue\n");
563 setup_timer(&dev->delay_timer, delay_time_func, (unsigned long)dev);
564 for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
565 INIT_LIST_HEAD(&cache->ent[i].head);
566 spin_lock_init(&cache->ent[i].lock);
568 ent = &cache->ent[i];
569 INIT_LIST_HEAD(&ent->head);
570 spin_lock_init(&ent->lock);
574 if (dev->mdev.profile->mask & MLX5_PROF_MASK_MR_CACHE) {
575 size = dev->mdev.profile->mr_cache[i].size;
576 limit = dev->mdev.profile->mr_cache[i].limit;
578 size = DEF_CACHE_SIZE;
581 INIT_WORK(&ent->work, cache_work_func);
582 INIT_DELAYED_WORK(&ent->dwork, delayed_cache_work_func);
584 queue_work(cache->wq, &ent->work);
587 err = mlx5_mr_cache_debugfs_init(dev);
589 mlx5_ib_warn(dev, "cache debugfs failure\n");
594 int mlx5_mr_cache_cleanup(struct mlx5_ib_dev *dev)
598 dev->cache.stopped = 1;
599 flush_workqueue(dev->cache.wq);
601 mlx5_mr_cache_debugfs_cleanup(dev);
603 for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++)
606 destroy_workqueue(dev->cache.wq);
607 del_timer_sync(&dev->delay_timer);
612 struct ib_mr *mlx5_ib_get_dma_mr(struct ib_pd *pd, int acc)
614 struct mlx5_ib_dev *dev = to_mdev(pd->device);
615 struct mlx5_core_dev *mdev = &dev->mdev;
616 struct mlx5_create_mkey_mbox_in *in;
617 struct mlx5_mkey_seg *seg;
618 struct mlx5_ib_mr *mr;
621 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
623 return ERR_PTR(-ENOMEM);
625 in = kzalloc(sizeof(*in), GFP_KERNEL);
632 seg->flags = convert_access(acc) | MLX5_ACCESS_MODE_PA;
633 seg->flags_pd = cpu_to_be32(to_mpd(pd)->pdn | MLX5_MKEY_LEN64);
634 seg->qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
637 err = mlx5_core_create_mkey(mdev, &mr->mmr, in, sizeof(*in), NULL, NULL,
643 mr->ibmr.lkey = mr->mmr.key;
644 mr->ibmr.rkey = mr->mmr.key;
658 static int get_octo_len(u64 addr, u64 len, int page_size)
663 offset = addr & (page_size - 1);
664 npages = ALIGN(len + offset, page_size) >> ilog2(page_size);
665 return (npages + 1) / 2;
668 static int use_umr(int order)
673 static void prep_umr_reg_wqe(struct ib_pd *pd, struct ib_send_wr *wr,
674 struct ib_sge *sg, u64 dma, int n, u32 key,
675 int page_shift, u64 virt_addr, u64 len,
678 struct mlx5_ib_dev *dev = to_mdev(pd->device);
679 struct ib_mr *mr = dev->umrc.mr;
682 sg->length = ALIGN(sizeof(u64) * n, 64);
693 wr->opcode = MLX5_IB_WR_UMR;
694 wr->wr.fast_reg.page_list_len = n;
695 wr->wr.fast_reg.page_shift = page_shift;
696 wr->wr.fast_reg.rkey = key;
697 wr->wr.fast_reg.iova_start = virt_addr;
698 wr->wr.fast_reg.length = len;
699 wr->wr.fast_reg.access_flags = access_flags;
700 wr->wr.fast_reg.page_list = (struct ib_fast_reg_page_list *)pd;
703 static void prep_umr_unreg_wqe(struct mlx5_ib_dev *dev,
704 struct ib_send_wr *wr, u32 key)
706 wr->send_flags = MLX5_IB_SEND_UMR_UNREG;
707 wr->opcode = MLX5_IB_WR_UMR;
708 wr->wr.fast_reg.rkey = key;
711 void mlx5_umr_cq_handler(struct ib_cq *cq, void *cq_context)
713 struct mlx5_ib_mr *mr;
718 err = ib_poll_cq(cq, 1, &wc);
720 pr_warn("poll cq error %d\n", err);
726 mr = (struct mlx5_ib_mr *)(unsigned long)wc.wr_id;
727 mr->status = wc.status;
730 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
733 static struct mlx5_ib_mr *reg_umr(struct ib_pd *pd, struct ib_umem *umem,
734 u64 virt_addr, u64 len, int npages,
735 int page_shift, int order, int access_flags)
737 struct mlx5_ib_dev *dev = to_mdev(pd->device);
738 struct device *ddev = dev->ib_dev.dma_device;
739 struct umr_common *umrc = &dev->umrc;
740 struct ib_send_wr wr, *bad;
741 struct mlx5_ib_mr *mr;
743 int size = sizeof(u64) * npages;
747 for (i = 0; i < 1; i++) {
748 mr = alloc_cached_mr(dev, order);
752 err = add_keys(dev, order2idx(dev, order), 1);
753 if (err && err != -EAGAIN) {
754 mlx5_ib_warn(dev, "add_keys failed, err %d\n", err);
760 return ERR_PTR(-EAGAIN);
762 mr->pas = kmalloc(size + MLX5_UMR_ALIGN - 1, GFP_KERNEL);
768 mlx5_ib_populate_pas(dev, umem, page_shift,
769 mr_align(mr->pas, MLX5_UMR_ALIGN), 1);
771 mr->dma = dma_map_single(ddev, mr_align(mr->pas, MLX5_UMR_ALIGN), size,
773 if (dma_mapping_error(ddev, mr->dma)) {
779 memset(&wr, 0, sizeof(wr));
780 wr.wr_id = (u64)(unsigned long)mr;
781 prep_umr_reg_wqe(pd, &wr, &sg, mr->dma, npages, mr->mmr.key, page_shift, virt_addr, len, access_flags);
783 /* We serialize polls so one process does not kidnap another's
784 * completion. This is not a problem since wr is completed in
788 init_completion(&mr->done);
789 err = ib_post_send(umrc->qp, &wr, &bad);
791 mlx5_ib_warn(dev, "post send failed, err %d\n", err);
795 wait_for_completion(&mr->done);
798 dma_unmap_single(ddev, mr->dma, size, DMA_TO_DEVICE);
801 if (mr->status != IB_WC_SUCCESS) {
802 mlx5_ib_warn(dev, "reg umr failed\n");
810 free_cached_mr(dev, mr);
814 static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, u64 virt_addr,
815 u64 length, struct ib_umem *umem,
816 int npages, int page_shift,
819 struct mlx5_ib_dev *dev = to_mdev(pd->device);
820 struct mlx5_create_mkey_mbox_in *in;
821 struct mlx5_ib_mr *mr;
825 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
827 return ERR_PTR(-ENOMEM);
829 inlen = sizeof(*in) + sizeof(*in->pas) * ((npages + 1) / 2) * 2;
830 in = mlx5_vzalloc(inlen);
835 mlx5_ib_populate_pas(dev, umem, page_shift, in->pas, 0);
837 in->seg.flags = convert_access(access_flags) |
838 MLX5_ACCESS_MODE_MTT;
839 in->seg.flags_pd = cpu_to_be32(to_mpd(pd)->pdn);
840 in->seg.start_addr = cpu_to_be64(virt_addr);
841 in->seg.len = cpu_to_be64(length);
842 in->seg.bsfs_octo_size = 0;
843 in->seg.xlt_oct_size = cpu_to_be32(get_octo_len(virt_addr, length, 1 << page_shift));
844 in->seg.log2_page_size = page_shift;
845 in->seg.qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
846 in->xlat_oct_act_size = cpu_to_be32(get_octo_len(virt_addr, length,
848 err = mlx5_core_create_mkey(&dev->mdev, &mr->mmr, in, inlen, NULL,
851 mlx5_ib_warn(dev, "create mkey failed\n");
857 mlx5_ib_dbg(dev, "mkey = 0x%x\n", mr->mmr.key);
870 struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
871 u64 virt_addr, int access_flags,
872 struct ib_udata *udata)
874 struct mlx5_ib_dev *dev = to_mdev(pd->device);
875 struct mlx5_ib_mr *mr = NULL;
876 struct ib_umem *umem;
883 mlx5_ib_dbg(dev, "start 0x%llx, virt_addr 0x%llx, length 0x%llx\n",
884 start, virt_addr, length);
885 umem = ib_umem_get(pd->uobject->context, start, length, access_flags,
888 mlx5_ib_dbg(dev, "umem get failed\n");
892 mlx5_ib_cont_pages(umem, start, &npages, &page_shift, &ncont, &order);
894 mlx5_ib_warn(dev, "avoid zero region\n");
899 mlx5_ib_dbg(dev, "npages %d, ncont %d, order %d, page_shift %d\n",
900 npages, ncont, order, page_shift);
902 if (use_umr(order)) {
903 mr = reg_umr(pd, umem, virt_addr, length, ncont, page_shift,
904 order, access_flags);
905 if (PTR_ERR(mr) == -EAGAIN) {
906 mlx5_ib_dbg(dev, "cache empty for order %d", order);
912 mr = reg_create(pd, virt_addr, length, umem, ncont, page_shift,
920 mlx5_ib_dbg(dev, "mkey 0x%x\n", mr->mmr.key);
924 spin_lock(&dev->mr_lock);
925 dev->mdev.priv.reg_pages += npages;
926 spin_unlock(&dev->mr_lock);
927 mr->ibmr.lkey = mr->mmr.key;
928 mr->ibmr.rkey = mr->mmr.key;
933 ib_umem_release(umem);
937 static int unreg_umr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
939 struct umr_common *umrc = &dev->umrc;
940 struct ib_send_wr wr, *bad;
943 memset(&wr, 0, sizeof(wr));
944 wr.wr_id = (u64)(unsigned long)mr;
945 prep_umr_unreg_wqe(dev, &wr, mr->mmr.key);
948 init_completion(&mr->done);
949 err = ib_post_send(umrc->qp, &wr, &bad);
952 mlx5_ib_dbg(dev, "err %d\n", err);
955 wait_for_completion(&mr->done);
957 if (mr->status != IB_WC_SUCCESS) {
958 mlx5_ib_warn(dev, "unreg umr failed\n");
968 int mlx5_ib_dereg_mr(struct ib_mr *ibmr)
970 struct mlx5_ib_dev *dev = to_mdev(ibmr->device);
971 struct mlx5_ib_mr *mr = to_mmr(ibmr);
972 struct ib_umem *umem = mr->umem;
973 int npages = mr->npages;
974 int umred = mr->umred;
978 err = mlx5_core_destroy_mkey(&dev->mdev, &mr->mmr);
980 mlx5_ib_warn(dev, "failed to destroy mkey 0x%x (%d)\n",
985 err = unreg_umr(dev, mr);
987 mlx5_ib_warn(dev, "failed unregister\n");
990 free_cached_mr(dev, mr);
994 ib_umem_release(umem);
995 spin_lock(&dev->mr_lock);
996 dev->mdev.priv.reg_pages -= npages;
997 spin_unlock(&dev->mr_lock);
1006 struct ib_mr *mlx5_ib_alloc_fast_reg_mr(struct ib_pd *pd,
1007 int max_page_list_len)
1009 struct mlx5_ib_dev *dev = to_mdev(pd->device);
1010 struct mlx5_create_mkey_mbox_in *in;
1011 struct mlx5_ib_mr *mr;
1014 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1016 return ERR_PTR(-ENOMEM);
1018 in = kzalloc(sizeof(*in), GFP_KERNEL);
1024 in->seg.status = 1 << 6; /* free */
1025 in->seg.xlt_oct_size = cpu_to_be32((max_page_list_len + 1) / 2);
1026 in->seg.qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
1027 in->seg.flags = MLX5_PERM_UMR_EN | MLX5_ACCESS_MODE_MTT;
1028 in->seg.flags_pd = cpu_to_be32(to_mpd(pd)->pdn);
1030 * TBD not needed - issue 197292 */
1031 in->seg.log2_page_size = PAGE_SHIFT;
1033 err = mlx5_core_create_mkey(&dev->mdev, &mr->mmr, in, sizeof(*in), NULL,
1039 mr->ibmr.lkey = mr->mmr.key;
1040 mr->ibmr.rkey = mr->mmr.key;
1047 return ERR_PTR(err);
1050 struct ib_fast_reg_page_list *mlx5_ib_alloc_fast_reg_page_list(struct ib_device *ibdev,
1053 struct mlx5_ib_fast_reg_page_list *mfrpl;
1054 int size = page_list_len * sizeof(u64);
1056 mfrpl = kmalloc(sizeof(*mfrpl), GFP_KERNEL);
1058 return ERR_PTR(-ENOMEM);
1060 mfrpl->ibfrpl.page_list = kmalloc(size, GFP_KERNEL);
1061 if (!mfrpl->ibfrpl.page_list)
1064 mfrpl->mapped_page_list = dma_alloc_coherent(ibdev->dma_device,
1067 if (!mfrpl->mapped_page_list)
1070 WARN_ON(mfrpl->map & 0x3f);
1072 return &mfrpl->ibfrpl;
1075 kfree(mfrpl->ibfrpl.page_list);
1077 return ERR_PTR(-ENOMEM);
1080 void mlx5_ib_free_fast_reg_page_list(struct ib_fast_reg_page_list *page_list)
1082 struct mlx5_ib_fast_reg_page_list *mfrpl = to_mfrpl(page_list);
1083 struct mlx5_ib_dev *dev = to_mdev(page_list->device);
1084 int size = page_list->max_page_list_len * sizeof(u64);
1086 dma_free_coherent(&dev->mdev.pdev->dev, size, mfrpl->mapped_page_list,
1088 kfree(mfrpl->ibfrpl.page_list);