2 * Copyright (c) 2016 Oracle. 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
35 static struct rds_ib_mr
*rds_ib_alloc_frmr(struct rds_ib_device
*rds_ibdev
,
38 struct rds_ib_mr_pool
*pool
;
39 struct rds_ib_mr
*ibmr
= NULL
;
40 struct rds_ib_frmr
*frmr
;
43 if (npages
<= RDS_MR_8K_MSG_SIZE
)
44 pool
= rds_ibdev
->mr_8k_pool
;
46 pool
= rds_ibdev
->mr_1m_pool
;
48 ibmr
= rds_ib_try_reuse_ibmr(pool
);
52 ibmr
= kzalloc_node(sizeof(*ibmr
), GFP_KERNEL
,
53 rdsibdev_to_node(rds_ibdev
));
60 frmr
->mr
= ib_alloc_mr(rds_ibdev
->pd
, IB_MR_TYPE_MEM_REG
,
61 pool
->fmr_attr
.max_pages
);
62 if (IS_ERR(frmr
->mr
)) {
63 pr_warn("RDS/IB: %s failed to allocate MR", __func__
);
64 err
= PTR_ERR(frmr
->mr
);
69 if (pool
->pool_type
== RDS_IB_MR_8K_POOL
)
70 rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc
);
72 rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc
);
74 if (atomic_read(&pool
->item_count
) > pool
->max_items_soft
)
75 pool
->max_items_soft
= pool
->max_items
;
77 frmr
->fr_state
= FRMR_IS_FREE
;
82 atomic_dec(&pool
->item_count
);
86 static void rds_ib_free_frmr(struct rds_ib_mr
*ibmr
, bool drop
)
88 struct rds_ib_mr_pool
*pool
= ibmr
->pool
;
91 llist_add(&ibmr
->llnode
, &pool
->drop_list
);
93 llist_add(&ibmr
->llnode
, &pool
->free_list
);
94 atomic_add(ibmr
->sg_len
, &pool
->free_pinned
);
95 atomic_inc(&pool
->dirty_count
);
97 /* If we've pinned too many pages, request a flush */
98 if (atomic_read(&pool
->free_pinned
) >= pool
->max_free_pinned
||
99 atomic_read(&pool
->dirty_count
) >= pool
->max_items
/ 5)
100 queue_delayed_work(rds_ib_mr_wq
, &pool
->flush_worker
, 10);
103 static int rds_ib_post_reg_frmr(struct rds_ib_mr
*ibmr
)
105 struct rds_ib_frmr
*frmr
= &ibmr
->u
.frmr
;
106 struct ib_send_wr
*failed_wr
;
107 struct ib_reg_wr reg_wr
;
110 while (atomic_dec_return(&ibmr
->ic
->i_fastreg_wrs
) <= 0) {
111 atomic_inc(&ibmr
->ic
->i_fastreg_wrs
);
115 ret
= ib_map_mr_sg_zbva(frmr
->mr
, ibmr
->sg
, ibmr
->sg_len
,
117 if (unlikely(ret
!= ibmr
->sg_len
))
118 return ret
< 0 ? ret
: -EINVAL
;
120 /* Perform a WR for the fast_reg_mr. Each individual page
121 * in the sg list is added to the fast reg page list and placed
122 * inside the fast_reg_mr WR. The key used is a rolling 8bit
123 * counter, which should guarantee uniqueness.
125 ib_update_fast_reg_key(frmr
->mr
, ibmr
->remap_count
++);
126 frmr
->fr_state
= FRMR_IS_INUSE
;
128 memset(®_wr
, 0, sizeof(reg_wr
));
129 reg_wr
.wr
.wr_id
= (unsigned long)(void *)ibmr
;
130 reg_wr
.wr
.opcode
= IB_WR_REG_MR
;
131 reg_wr
.wr
.num_sge
= 0;
132 reg_wr
.mr
= frmr
->mr
;
133 reg_wr
.key
= frmr
->mr
->rkey
;
134 reg_wr
.access
= IB_ACCESS_LOCAL_WRITE
|
135 IB_ACCESS_REMOTE_READ
|
136 IB_ACCESS_REMOTE_WRITE
;
137 reg_wr
.wr
.send_flags
= IB_SEND_SIGNALED
;
139 failed_wr
= ®_wr
.wr
;
140 ret
= ib_post_send(ibmr
->ic
->i_cm_id
->qp
, ®_wr
.wr
, &failed_wr
);
141 WARN_ON(failed_wr
!= ®_wr
.wr
);
143 /* Failure here can be because of -ENOMEM as well */
144 frmr
->fr_state
= FRMR_IS_STALE
;
145 atomic_inc(&ibmr
->ic
->i_fastreg_wrs
);
146 if (printk_ratelimit())
147 pr_warn("RDS/IB: %s returned error(%d)\n",
153 static int rds_ib_map_frmr(struct rds_ib_device
*rds_ibdev
,
154 struct rds_ib_mr_pool
*pool
,
155 struct rds_ib_mr
*ibmr
,
156 struct scatterlist
*sg
, unsigned int sg_len
)
158 struct ib_device
*dev
= rds_ibdev
->dev
;
159 struct rds_ib_frmr
*frmr
= &ibmr
->u
.frmr
;
164 /* We want to teardown old ibmr values here and fill it up with
167 rds_ib_teardown_mr(ibmr
);
170 ibmr
->sg_len
= sg_len
;
171 ibmr
->sg_dma_len
= 0;
172 frmr
->sg_byte_len
= 0;
173 WARN_ON(ibmr
->sg_dma_len
);
174 ibmr
->sg_dma_len
= ib_dma_map_sg(dev
, ibmr
->sg
, ibmr
->sg_len
,
176 if (unlikely(!ibmr
->sg_dma_len
)) {
177 pr_warn("RDS/IB: %s failed!\n", __func__
);
181 frmr
->sg_byte_len
= 0;
182 frmr
->dma_npages
= 0;
186 for (i
= 0; i
< ibmr
->sg_dma_len
; ++i
) {
187 unsigned int dma_len
= ib_sg_dma_len(dev
, &ibmr
->sg
[i
]);
188 u64 dma_addr
= ib_sg_dma_address(dev
, &ibmr
->sg
[i
]);
190 frmr
->sg_byte_len
+= dma_len
;
191 if (dma_addr
& ~PAGE_MASK
) {
198 if ((dma_addr
+ dma_len
) & ~PAGE_MASK
) {
199 if (i
< ibmr
->sg_dma_len
- 1)
207 frmr
->dma_npages
+= len
>> PAGE_SHIFT
;
209 if (frmr
->dma_npages
> ibmr
->pool
->fmr_attr
.max_pages
) {
214 ret
= rds_ib_post_reg_frmr(ibmr
);
218 if (ibmr
->pool
->pool_type
== RDS_IB_MR_8K_POOL
)
219 rds_ib_stats_inc(s_ib_rdma_mr_8k_used
);
221 rds_ib_stats_inc(s_ib_rdma_mr_1m_used
);
226 ib_dma_unmap_sg(rds_ibdev
->dev
, ibmr
->sg
, ibmr
->sg_len
,
228 ibmr
->sg_dma_len
= 0;
232 static int rds_ib_post_inv(struct rds_ib_mr
*ibmr
)
234 struct ib_send_wr
*s_wr
, *failed_wr
;
235 struct rds_ib_frmr
*frmr
= &ibmr
->u
.frmr
;
236 struct rdma_cm_id
*i_cm_id
= ibmr
->ic
->i_cm_id
;
239 if (!i_cm_id
|| !i_cm_id
->qp
|| !frmr
->mr
)
242 if (frmr
->fr_state
!= FRMR_IS_INUSE
)
245 while (atomic_dec_return(&ibmr
->ic
->i_fastreg_wrs
) <= 0) {
246 atomic_inc(&ibmr
->ic
->i_fastreg_wrs
);
253 memset(s_wr
, 0, sizeof(*s_wr
));
254 s_wr
->wr_id
= (unsigned long)(void *)ibmr
;
255 s_wr
->opcode
= IB_WR_LOCAL_INV
;
256 s_wr
->ex
.invalidate_rkey
= frmr
->mr
->rkey
;
257 s_wr
->send_flags
= IB_SEND_SIGNALED
;
260 ret
= ib_post_send(i_cm_id
->qp
, s_wr
, &failed_wr
);
261 WARN_ON(failed_wr
!= s_wr
);
263 frmr
->fr_state
= FRMR_IS_STALE
;
264 frmr
->fr_inv
= false;
265 atomic_inc(&ibmr
->ic
->i_fastreg_wrs
);
266 pr_err("RDS/IB: %s returned error(%d)\n", __func__
, ret
);
273 void rds_ib_mr_cqe_handler(struct rds_ib_connection
*ic
, struct ib_wc
*wc
)
275 struct rds_ib_mr
*ibmr
= (void *)(unsigned long)wc
->wr_id
;
276 struct rds_ib_frmr
*frmr
= &ibmr
->u
.frmr
;
278 if (wc
->status
!= IB_WC_SUCCESS
) {
279 frmr
->fr_state
= FRMR_IS_STALE
;
280 if (rds_conn_up(ic
->conn
))
281 rds_ib_conn_error(ic
->conn
,
282 "frmr completion <%pI4,%pI4> status %u(%s), vendor_err 0x%x, disconnecting and reconnecting\n",
286 ib_wc_status_msg(wc
->status
),
291 frmr
->fr_state
= FRMR_IS_FREE
;
292 frmr
->fr_inv
= false;
295 atomic_inc(&ic
->i_fastreg_wrs
);
298 void rds_ib_unreg_frmr(struct list_head
*list
, unsigned int *nfreed
,
299 unsigned long *unpinned
, unsigned int goal
)
301 struct rds_ib_mr
*ibmr
, *next
;
302 struct rds_ib_frmr
*frmr
;
304 unsigned int freed
= *nfreed
;
306 /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
307 list_for_each_entry(ibmr
, list
, unmap_list
) {
308 if (ibmr
->sg_dma_len
)
309 ret
|= rds_ib_post_inv(ibmr
);
312 pr_warn("RDS/IB: %s failed (err=%d)\n", __func__
, ret
);
314 /* Now we can destroy the DMA mapping and unpin any pages */
315 list_for_each_entry_safe(ibmr
, next
, list
, unmap_list
) {
316 *unpinned
+= ibmr
->sg_len
;
317 frmr
= &ibmr
->u
.frmr
;
318 __rds_ib_teardown_mr(ibmr
);
319 if (freed
< goal
|| frmr
->fr_state
== FRMR_IS_STALE
) {
320 /* Don't de-allocate if the MR is not free yet */
321 if (frmr
->fr_state
== FRMR_IS_INUSE
)
324 if (ibmr
->pool
->pool_type
== RDS_IB_MR_8K_POOL
)
325 rds_ib_stats_inc(s_ib_rdma_mr_8k_free
);
327 rds_ib_stats_inc(s_ib_rdma_mr_1m_free
);
328 list_del(&ibmr
->unmap_list
);
330 ib_dereg_mr(frmr
->mr
);
338 struct rds_ib_mr
*rds_ib_reg_frmr(struct rds_ib_device
*rds_ibdev
,
339 struct rds_ib_connection
*ic
,
340 struct scatterlist
*sg
,
341 unsigned long nents
, u32
*key
)
343 struct rds_ib_mr
*ibmr
= NULL
;
344 struct rds_ib_frmr
*frmr
;
349 rds_ib_free_frmr(ibmr
, true);
350 ibmr
= rds_ib_alloc_frmr(rds_ibdev
, nents
);
353 frmr
= &ibmr
->u
.frmr
;
354 } while (frmr
->fr_state
!= FRMR_IS_FREE
);
357 ibmr
->device
= rds_ibdev
;
358 ret
= rds_ib_map_frmr(rds_ibdev
, ibmr
->pool
, ibmr
, sg
, nents
);
360 *key
= frmr
->mr
->rkey
;
362 rds_ib_free_frmr(ibmr
, false);
369 void rds_ib_free_frmr_list(struct rds_ib_mr
*ibmr
)
371 struct rds_ib_mr_pool
*pool
= ibmr
->pool
;
372 struct rds_ib_frmr
*frmr
= &ibmr
->u
.frmr
;
374 if (frmr
->fr_state
== FRMR_IS_STALE
)
375 llist_add(&ibmr
->llnode
, &pool
->drop_list
);
377 llist_add(&ibmr
->llnode
, &pool
->free_list
);