1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2015, 2017 Oracle. All rights reserved.
4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5 */
6
7 /* Lightweight memory registration using Fast Memory Regions (FMR).
8 * Referred to sometimes as MTHCAFMR mode.
9 *
10 * FMR uses synchronous memory registration and deregistration.
11 * FMR registration is known to be fast, but FMR deregistration
12 * can take tens of usecs to complete.
13 */
14
15 /* Normal operation
16 *
17 * A Memory Region is prepared for RDMA READ or WRITE using the
18 * ib_map_phys_fmr verb (fmr_op_map). When the RDMA operation is
19 * finished, the Memory Region is unmapped using the ib_unmap_fmr
20 * verb (fmr_op_unmap).
21 */
22
23 #include <linux/sunrpc/svc_rdma.h>
24
25 #include "xprt_rdma.h"
26 #include <trace/events/rpcrdma.h>
27
28 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
29 # define RPCDBG_FACILITY RPCDBG_TRANS
30 #endif
31
32 /* Maximum scatter/gather per FMR */
33 #define RPCRDMA_MAX_FMR_SGES (64)
34
35 /* Access mode of externally registered pages */
36 enum {
37 RPCRDMA_FMR_ACCESS_FLAGS = IB_ACCESS_REMOTE_WRITE |
38 IB_ACCESS_REMOTE_READ,
39 };
40
41 bool
fmr_is_supported(struct rpcrdma_ia * ia)42 fmr_is_supported(struct rpcrdma_ia *ia)
43 {
44 if (!ia->ri_device->alloc_fmr) {
45 pr_info("rpcrdma: 'fmr' mode is not supported by device %s\n",
46 ia->ri_device->name);
47 return false;
48 }
49 return true;
50 }
51
52 static int
fmr_op_init_mr(struct rpcrdma_ia * ia,struct rpcrdma_mr * mr)53 fmr_op_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr)
54 {
55 static struct ib_fmr_attr fmr_attr = {
56 .max_pages = RPCRDMA_MAX_FMR_SGES,
57 .max_maps = 1,
58 .page_shift = PAGE_SHIFT
59 };
60
61 mr->fmr.fm_physaddrs = kcalloc(RPCRDMA_MAX_FMR_SGES,
62 sizeof(u64), GFP_KERNEL);
63 if (!mr->fmr.fm_physaddrs)
64 goto out_free;
65
66 mr->mr_sg = kcalloc(RPCRDMA_MAX_FMR_SGES,
67 sizeof(*mr->mr_sg), GFP_KERNEL);
68 if (!mr->mr_sg)
69 goto out_free;
70
71 sg_init_table(mr->mr_sg, RPCRDMA_MAX_FMR_SGES);
72
73 mr->fmr.fm_mr = ib_alloc_fmr(ia->ri_pd, RPCRDMA_FMR_ACCESS_FLAGS,
74 &fmr_attr);
75 if (IS_ERR(mr->fmr.fm_mr))
76 goto out_fmr_err;
77
78 INIT_LIST_HEAD(&mr->mr_list);
79 return 0;
80
81 out_fmr_err:
82 dprintk("RPC: %s: ib_alloc_fmr returned %ld\n", __func__,
83 PTR_ERR(mr->fmr.fm_mr));
84
85 out_free:
86 kfree(mr->mr_sg);
87 kfree(mr->fmr.fm_physaddrs);
88 return -ENOMEM;
89 }
90
91 static int
__fmr_unmap(struct rpcrdma_mr * mr)92 __fmr_unmap(struct rpcrdma_mr *mr)
93 {
94 LIST_HEAD(l);
95 int rc;
96
97 list_add(&mr->fmr.fm_mr->list, &l);
98 rc = ib_unmap_fmr(&l);
99 list_del(&mr->fmr.fm_mr->list);
100 return rc;
101 }
102
103 static void
fmr_op_release_mr(struct rpcrdma_mr * mr)104 fmr_op_release_mr(struct rpcrdma_mr *mr)
105 {
106 LIST_HEAD(unmap_list);
107 int rc;
108
109 kfree(mr->fmr.fm_physaddrs);
110 kfree(mr->mr_sg);
111
112 /* In case this one was left mapped, try to unmap it
113 * to prevent dealloc_fmr from failing with EBUSY
114 */
115 rc = __fmr_unmap(mr);
116 if (rc)
117 pr_err("rpcrdma: final ib_unmap_fmr for %p failed %i\n",
118 mr, rc);
119
120 rc = ib_dealloc_fmr(mr->fmr.fm_mr);
121 if (rc)
122 pr_err("rpcrdma: final ib_dealloc_fmr for %p returned %i\n",
123 mr, rc);
124
125 kfree(mr);
126 }
127
128 /* Reset of a single FMR.
129 */
130 static void
fmr_op_recover_mr(struct rpcrdma_mr * mr)131 fmr_op_recover_mr(struct rpcrdma_mr *mr)
132 {
133 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
134 int rc;
135
136 /* ORDER: invalidate first */
137 rc = __fmr_unmap(mr);
138 if (rc)
139 goto out_release;
140
141 /* ORDER: then DMA unmap */
142 rpcrdma_mr_unmap_and_put(mr);
143
144 r_xprt->rx_stats.mrs_recovered++;
145 return;
146
147 out_release:
148 pr_err("rpcrdma: FMR reset failed (%d), %p released\n", rc, mr);
149 r_xprt->rx_stats.mrs_orphaned++;
150
151 trace_xprtrdma_dma_unmap(mr);
152 ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
153 mr->mr_sg, mr->mr_nents, mr->mr_dir);
154
155 spin_lock(&r_xprt->rx_buf.rb_mrlock);
156 list_del(&mr->mr_all);
157 spin_unlock(&r_xprt->rx_buf.rb_mrlock);
158
159 fmr_op_release_mr(mr);
160 }
161
162 /* On success, sets:
163 * ep->rep_attr.cap.max_send_wr
164 * ep->rep_attr.cap.max_recv_wr
165 * cdata->max_requests
166 * ia->ri_max_segs
167 */
168 static int
fmr_op_open(struct rpcrdma_ia * ia,struct rpcrdma_ep * ep,struct rpcrdma_create_data_internal * cdata)169 fmr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
170 struct rpcrdma_create_data_internal *cdata)
171 {
172 int max_qp_wr;
173
174 max_qp_wr = ia->ri_device->attrs.max_qp_wr;
175 max_qp_wr -= RPCRDMA_BACKWARD_WRS;
176 max_qp_wr -= 1;
177 if (max_qp_wr < RPCRDMA_MIN_SLOT_TABLE)
178 return -ENOMEM;
179 if (cdata->max_requests > max_qp_wr)
180 cdata->max_requests = max_qp_wr;
181 ep->rep_attr.cap.max_send_wr = cdata->max_requests;
182 ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
183 ep->rep_attr.cap.max_send_wr += 1; /* for ib_drain_sq */
184 ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
185 ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
186 ep->rep_attr.cap.max_recv_wr += 1; /* for ib_drain_rq */
187
188 ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
189 RPCRDMA_MAX_FMR_SGES);
190 return 0;
191 }
192
193 /* FMR mode conveys up to 64 pages of payload per chunk segment.
194 */
195 static size_t
fmr_op_maxpages(struct rpcrdma_xprt * r_xprt)196 fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
197 {
198 return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
199 RPCRDMA_MAX_HDR_SEGS * RPCRDMA_MAX_FMR_SGES);
200 }
201
202 /* Use the ib_map_phys_fmr() verb to register a memory region
203 * for remote access via RDMA READ or RDMA WRITE.
204 */
205 static struct rpcrdma_mr_seg *
fmr_op_map(struct rpcrdma_xprt * r_xprt,struct rpcrdma_mr_seg * seg,int nsegs,bool writing,struct rpcrdma_mr ** out)206 fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
207 int nsegs, bool writing, struct rpcrdma_mr **out)
208 {
209 struct rpcrdma_mr_seg *seg1 = seg;
210 int len, pageoff, i, rc;
211 struct rpcrdma_mr *mr;
212 u64 *dma_pages;
213
214 mr = rpcrdma_mr_get(r_xprt);
215 if (!mr)
216 return ERR_PTR(-EAGAIN);
217
218 pageoff = offset_in_page(seg1->mr_offset);
219 seg1->mr_offset -= pageoff; /* start of page */
220 seg1->mr_len += pageoff;
221 len = -pageoff;
222 if (nsegs > RPCRDMA_MAX_FMR_SGES)
223 nsegs = RPCRDMA_MAX_FMR_SGES;
224 for (i = 0; i < nsegs;) {
225 if (seg->mr_page)
226 sg_set_page(&mr->mr_sg[i],
227 seg->mr_page,
228 seg->mr_len,
229 offset_in_page(seg->mr_offset));
230 else
231 sg_set_buf(&mr->mr_sg[i], seg->mr_offset,
232 seg->mr_len);
233 len += seg->mr_len;
234 ++seg;
235 ++i;
236 /* Check for holes */
237 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
238 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
239 break;
240 }
241 mr->mr_dir = rpcrdma_data_dir(writing);
242
243 mr->mr_nents = ib_dma_map_sg(r_xprt->rx_ia.ri_device,
244 mr->mr_sg, i, mr->mr_dir);
245 if (!mr->mr_nents)
246 goto out_dmamap_err;
247 trace_xprtrdma_dma_map(mr);
248
249 for (i = 0, dma_pages = mr->fmr.fm_physaddrs; i < mr->mr_nents; i++)
250 dma_pages[i] = sg_dma_address(&mr->mr_sg[i]);
251 rc = ib_map_phys_fmr(mr->fmr.fm_mr, dma_pages, mr->mr_nents,
252 dma_pages[0]);
253 if (rc)
254 goto out_maperr;
255
256 mr->mr_handle = mr->fmr.fm_mr->rkey;
257 mr->mr_length = len;
258 mr->mr_offset = dma_pages[0] + pageoff;
259
260 *out = mr;
261 return seg;
262
263 out_dmamap_err:
264 pr_err("rpcrdma: failed to DMA map sg %p sg_nents %d\n",
265 mr->mr_sg, i);
266 rpcrdma_mr_put(mr);
267 return ERR_PTR(-EIO);
268
269 out_maperr:
270 pr_err("rpcrdma: ib_map_phys_fmr %u@0x%llx+%i (%d) status %i\n",
271 len, (unsigned long long)dma_pages[0],
272 pageoff, mr->mr_nents, rc);
273 rpcrdma_mr_unmap_and_put(mr);
274 return ERR_PTR(-EIO);
275 }
276
277 /* Post Send WR containing the RPC Call message.
278 */
279 static int
fmr_op_send(struct rpcrdma_ia * ia,struct rpcrdma_req * req)280 fmr_op_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
281 {
282 return ib_post_send(ia->ri_id->qp, &req->rl_sendctx->sc_wr, NULL);
283 }
284
285 /* Invalidate all memory regions that were registered for "req".
286 *
287 * Sleeps until it is safe for the host CPU to access the
288 * previously mapped memory regions.
289 *
290 * Caller ensures that @mrs is not empty before the call. This
291 * function empties the list.
292 */
293 static void
fmr_op_unmap_sync(struct rpcrdma_xprt * r_xprt,struct list_head * mrs)294 fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct list_head *mrs)
295 {
296 struct rpcrdma_mr *mr;
297 LIST_HEAD(unmap_list);
298 int rc;
299
300 /* ORDER: Invalidate all of the req's MRs first
301 *
302 * ib_unmap_fmr() is slow, so use a single call instead
303 * of one call per mapped FMR.
304 */
305 list_for_each_entry(mr, mrs, mr_list) {
306 dprintk("RPC: %s: unmapping fmr %p\n",
307 __func__, &mr->fmr);
308 trace_xprtrdma_localinv(mr);
309 list_add_tail(&mr->fmr.fm_mr->list, &unmap_list);
310 }
311 r_xprt->rx_stats.local_inv_needed++;
312 rc = ib_unmap_fmr(&unmap_list);
313 if (rc)
314 goto out_reset;
315
316 /* ORDER: Now DMA unmap all of the req's MRs, and return
317 * them to the free MW list.
318 */
319 while (!list_empty(mrs)) {
320 mr = rpcrdma_mr_pop(mrs);
321 list_del(&mr->fmr.fm_mr->list);
322 rpcrdma_mr_unmap_and_put(mr);
323 }
324
325 return;
326
327 out_reset:
328 pr_err("rpcrdma: ib_unmap_fmr failed (%i)\n", rc);
329
330 while (!list_empty(mrs)) {
331 mr = rpcrdma_mr_pop(mrs);
332 list_del(&mr->fmr.fm_mr->list);
333 fmr_op_recover_mr(mr);
334 }
335 }
336
337 const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
338 .ro_map = fmr_op_map,
339 .ro_send = fmr_op_send,
340 .ro_unmap_sync = fmr_op_unmap_sync,
341 .ro_recover_mr = fmr_op_recover_mr,
342 .ro_open = fmr_op_open,
343 .ro_maxpages = fmr_op_maxpages,
344 .ro_init_mr = fmr_op_init_mr,
345 .ro_release_mr = fmr_op_release_mr,
346 .ro_displayname = "fmr",
347 .ro_send_w_inv_ok = 0,
348 };
349
