xref: /linux/net/rds/ib_frmr.c (revision f79e4d5f92a129a1159c973735007d4ddc8541f3)
1 /*
2  * Copyright (c) 2016 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 "ib_mr.h"
34 
35 static struct rds_ib_mr *rds_ib_alloc_frmr(struct rds_ib_device *rds_ibdev,
36 					   int npages)
37 {
38 	struct rds_ib_mr_pool *pool;
39 	struct rds_ib_mr *ibmr = NULL;
40 	struct rds_ib_frmr *frmr;
41 	int err = 0;
42 
43 	if (npages <= RDS_MR_8K_MSG_SIZE)
44 		pool = rds_ibdev->mr_8k_pool;
45 	else
46 		pool = rds_ibdev->mr_1m_pool;
47 
48 	ibmr = rds_ib_try_reuse_ibmr(pool);
49 	if (ibmr)
50 		return ibmr;
51 
52 	ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL,
53 			    rdsibdev_to_node(rds_ibdev));
54 	if (!ibmr) {
55 		err = -ENOMEM;
56 		goto out_no_cigar;
57 	}
58 
59 	frmr = &ibmr->u.frmr;
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 		goto out_no_cigar;
65 	}
66 
67 	ibmr->pool = pool;
68 	if (pool->pool_type == RDS_IB_MR_8K_POOL)
69 		rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc);
70 	else
71 		rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc);
72 
73 	if (atomic_read(&pool->item_count) > pool->max_items_soft)
74 		pool->max_items_soft = pool->max_items;
75 
76 	frmr->fr_state = FRMR_IS_FREE;
77 	return ibmr;
78 
79 out_no_cigar:
80 	kfree(ibmr);
81 	atomic_dec(&pool->item_count);
82 	return ERR_PTR(err);
83 }
84 
85 static void rds_ib_free_frmr(struct rds_ib_mr *ibmr, bool drop)
86 {
87 	struct rds_ib_mr_pool *pool = ibmr->pool;
88 
89 	if (drop)
90 		llist_add(&ibmr->llnode, &pool->drop_list);
91 	else
92 		llist_add(&ibmr->llnode, &pool->free_list);
93 	atomic_add(ibmr->sg_len, &pool->free_pinned);
94 	atomic_inc(&pool->dirty_count);
95 
96 	/* If we've pinned too many pages, request a flush */
97 	if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
98 	    atomic_read(&pool->dirty_count) >= pool->max_items / 5)
99 		queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
100 }
101 
102 static int rds_ib_post_reg_frmr(struct rds_ib_mr *ibmr)
103 {
104 	struct rds_ib_frmr *frmr = &ibmr->u.frmr;
105 	struct ib_send_wr *failed_wr;
106 	struct ib_reg_wr reg_wr;
107 	int ret, off = 0;
108 
109 	while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) {
110 		atomic_inc(&ibmr->ic->i_fastreg_wrs);
111 		cpu_relax();
112 	}
113 
114 	ret = ib_map_mr_sg_zbva(frmr->mr, ibmr->sg, ibmr->sg_len,
115 				&off, PAGE_SIZE);
116 	if (unlikely(ret != ibmr->sg_len))
117 		return ret < 0 ? ret : -EINVAL;
118 
119 	/* Perform a WR for the fast_reg_mr. Each individual page
120 	 * in the sg list is added to the fast reg page list and placed
121 	 * inside the fast_reg_mr WR.  The key used is a rolling 8bit
122 	 * counter, which should guarantee uniqueness.
123 	 */
124 	ib_update_fast_reg_key(frmr->mr, ibmr->remap_count++);
125 	frmr->fr_state = FRMR_IS_INUSE;
126 
127 	memset(&reg_wr, 0, sizeof(reg_wr));
128 	reg_wr.wr.wr_id = (unsigned long)(void *)ibmr;
129 	reg_wr.wr.opcode = IB_WR_REG_MR;
130 	reg_wr.wr.num_sge = 0;
131 	reg_wr.mr = frmr->mr;
132 	reg_wr.key = frmr->mr->rkey;
133 	reg_wr.access = IB_ACCESS_LOCAL_WRITE |
134 			IB_ACCESS_REMOTE_READ |
135 			IB_ACCESS_REMOTE_WRITE;
136 	reg_wr.wr.send_flags = IB_SEND_SIGNALED;
137 
138 	failed_wr = &reg_wr.wr;
139 	ret = ib_post_send(ibmr->ic->i_cm_id->qp, &reg_wr.wr, &failed_wr);
140 	WARN_ON(failed_wr != &reg_wr.wr);
141 	if (unlikely(ret)) {
142 		/* Failure here can be because of -ENOMEM as well */
143 		frmr->fr_state = FRMR_IS_STALE;
144 		atomic_inc(&ibmr->ic->i_fastreg_wrs);
145 		if (printk_ratelimit())
146 			pr_warn("RDS/IB: %s returned error(%d)\n",
147 				__func__, ret);
148 	}
149 	return ret;
150 }
151 
152 static int rds_ib_map_frmr(struct rds_ib_device *rds_ibdev,
153 			   struct rds_ib_mr_pool *pool,
154 			   struct rds_ib_mr *ibmr,
155 			   struct scatterlist *sg, unsigned int sg_len)
156 {
157 	struct ib_device *dev = rds_ibdev->dev;
158 	struct rds_ib_frmr *frmr = &ibmr->u.frmr;
159 	int i;
160 	u32 len;
161 	int ret = 0;
162 
163 	/* We want to teardown old ibmr values here and fill it up with
164 	 * new sg values
165 	 */
166 	rds_ib_teardown_mr(ibmr);
167 
168 	ibmr->sg = sg;
169 	ibmr->sg_len = sg_len;
170 	ibmr->sg_dma_len = 0;
171 	frmr->sg_byte_len = 0;
172 	WARN_ON(ibmr->sg_dma_len);
173 	ibmr->sg_dma_len = ib_dma_map_sg(dev, ibmr->sg, ibmr->sg_len,
174 					 DMA_BIDIRECTIONAL);
175 	if (unlikely(!ibmr->sg_dma_len)) {
176 		pr_warn("RDS/IB: %s failed!\n", __func__);
177 		return -EBUSY;
178 	}
179 
180 	frmr->sg_byte_len = 0;
181 	frmr->dma_npages = 0;
182 	len = 0;
183 
184 	ret = -EINVAL;
185 	for (i = 0; i < ibmr->sg_dma_len; ++i) {
186 		unsigned int dma_len = ib_sg_dma_len(dev, &ibmr->sg[i]);
187 		u64 dma_addr = ib_sg_dma_address(dev, &ibmr->sg[i]);
188 
189 		frmr->sg_byte_len += dma_len;
190 		if (dma_addr & ~PAGE_MASK) {
191 			if (i > 0)
192 				goto out_unmap;
193 			else
194 				++frmr->dma_npages;
195 		}
196 
197 		if ((dma_addr + dma_len) & ~PAGE_MASK) {
198 			if (i < ibmr->sg_dma_len - 1)
199 				goto out_unmap;
200 			else
201 				++frmr->dma_npages;
202 		}
203 
204 		len += dma_len;
205 	}
206 	frmr->dma_npages += len >> PAGE_SHIFT;
207 
208 	if (frmr->dma_npages > ibmr->pool->fmr_attr.max_pages) {
209 		ret = -EMSGSIZE;
210 		goto out_unmap;
211 	}
212 
213 	ret = rds_ib_post_reg_frmr(ibmr);
214 	if (ret)
215 		goto out_unmap;
216 
217 	if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
218 		rds_ib_stats_inc(s_ib_rdma_mr_8k_used);
219 	else
220 		rds_ib_stats_inc(s_ib_rdma_mr_1m_used);
221 
222 	return ret;
223 
224 out_unmap:
225 	ib_dma_unmap_sg(rds_ibdev->dev, ibmr->sg, ibmr->sg_len,
226 			DMA_BIDIRECTIONAL);
227 	ibmr->sg_dma_len = 0;
228 	return ret;
229 }
230 
231 static int rds_ib_post_inv(struct rds_ib_mr *ibmr)
232 {
233 	struct ib_send_wr *s_wr, *failed_wr;
234 	struct rds_ib_frmr *frmr = &ibmr->u.frmr;
235 	struct rdma_cm_id *i_cm_id = ibmr->ic->i_cm_id;
236 	int ret = -EINVAL;
237 
238 	if (!i_cm_id || !i_cm_id->qp || !frmr->mr)
239 		goto out;
240 
241 	if (frmr->fr_state != FRMR_IS_INUSE)
242 		goto out;
243 
244 	while (atomic_dec_return(&ibmr->ic->i_fastunreg_wrs) <= 0) {
245 		atomic_inc(&ibmr->ic->i_fastunreg_wrs);
246 		cpu_relax();
247 	}
248 
249 	frmr->fr_inv = true;
250 	s_wr = &frmr->fr_wr;
251 
252 	memset(s_wr, 0, sizeof(*s_wr));
253 	s_wr->wr_id = (unsigned long)(void *)ibmr;
254 	s_wr->opcode = IB_WR_LOCAL_INV;
255 	s_wr->ex.invalidate_rkey = frmr->mr->rkey;
256 	s_wr->send_flags = IB_SEND_SIGNALED;
257 
258 	failed_wr = s_wr;
259 	ret = ib_post_send(i_cm_id->qp, s_wr, &failed_wr);
260 	WARN_ON(failed_wr != s_wr);
261 	if (unlikely(ret)) {
262 		frmr->fr_state = FRMR_IS_STALE;
263 		frmr->fr_inv = false;
264 		atomic_inc(&ibmr->ic->i_fastunreg_wrs);
265 		pr_err("RDS/IB: %s returned error(%d)\n", __func__, ret);
266 		goto out;
267 	}
268 out:
269 	return ret;
270 }
271 
272 void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
273 {
274 	struct rds_ib_mr *ibmr = (void *)(unsigned long)wc->wr_id;
275 	struct rds_ib_frmr *frmr = &ibmr->u.frmr;
276 
277 	if (wc->status != IB_WC_SUCCESS) {
278 		frmr->fr_state = FRMR_IS_STALE;
279 		if (rds_conn_up(ic->conn))
280 			rds_ib_conn_error(ic->conn,
281 					  "frmr completion <%pI4,%pI4> status %u(%s), vendor_err 0x%x, disconnecting and reconnecting\n",
282 					  &ic->conn->c_laddr,
283 					  &ic->conn->c_faddr,
284 					  wc->status,
285 					  ib_wc_status_msg(wc->status),
286 					  wc->vendor_err);
287 	}
288 
289 	if (frmr->fr_inv) {
290 		frmr->fr_state = FRMR_IS_FREE;
291 		frmr->fr_inv = false;
292 		atomic_inc(&ic->i_fastreg_wrs);
293 	} else {
294 		atomic_inc(&ic->i_fastunreg_wrs);
295 	}
296 }
297 
298 void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed,
299 		       unsigned long *unpinned, unsigned int goal)
300 {
301 	struct rds_ib_mr *ibmr, *next;
302 	struct rds_ib_frmr *frmr;
303 	int ret = 0;
304 	unsigned int freed = *nfreed;
305 
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);
310 	}
311 	if (ret)
312 		pr_warn("RDS/IB: %s failed (err=%d)\n", __func__, ret);
313 
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)
322 				continue;
323 
324 			if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
325 				rds_ib_stats_inc(s_ib_rdma_mr_8k_free);
326 			else
327 				rds_ib_stats_inc(s_ib_rdma_mr_1m_free);
328 			list_del(&ibmr->unmap_list);
329 			if (frmr->mr)
330 				ib_dereg_mr(frmr->mr);
331 			kfree(ibmr);
332 			freed++;
333 		}
334 	}
335 	*nfreed = freed;
336 }
337 
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)
342 {
343 	struct rds_ib_mr *ibmr = NULL;
344 	struct rds_ib_frmr *frmr;
345 	int ret;
346 
347 	if (!ic) {
348 		/* TODO: Add FRWR support for RDS_GET_MR using proxy qp*/
349 		return ERR_PTR(-EOPNOTSUPP);
350 	}
351 
352 	do {
353 		if (ibmr)
354 			rds_ib_free_frmr(ibmr, true);
355 		ibmr = rds_ib_alloc_frmr(rds_ibdev, nents);
356 		if (IS_ERR(ibmr))
357 			return ibmr;
358 		frmr = &ibmr->u.frmr;
359 	} while (frmr->fr_state != FRMR_IS_FREE);
360 
361 	ibmr->ic = ic;
362 	ibmr->device = rds_ibdev;
363 	ret = rds_ib_map_frmr(rds_ibdev, ibmr->pool, ibmr, sg, nents);
364 	if (ret == 0) {
365 		*key = frmr->mr->rkey;
366 	} else {
367 		rds_ib_free_frmr(ibmr, false);
368 		ibmr = ERR_PTR(ret);
369 	}
370 
371 	return ibmr;
372 }
373 
374 void rds_ib_free_frmr_list(struct rds_ib_mr *ibmr)
375 {
376 	struct rds_ib_mr_pool *pool = ibmr->pool;
377 	struct rds_ib_frmr *frmr = &ibmr->u.frmr;
378 
379 	if (frmr->fr_state == FRMR_IS_STALE)
380 		llist_add(&ibmr->llnode, &pool->drop_list);
381 	else
382 		llist_add(&ibmr->llnode, &pool->free_list);
383 }
384