1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2015, 2017 Oracle. All rights reserved.
4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
7 /* Lightweight memory registration using Fast Memory Regions (FMR).
8 * Referred to sometimes as MTHCAFMR mode.
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.
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).
23 #include "xprt_rdma.h"
25 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
26 # define RPCDBG_FACILITY RPCDBG_TRANS
29 /* Maximum scatter/gather per FMR */
30 #define RPCRDMA_MAX_FMR_SGES (64)
32 /* Access mode of externally registered pages */
34 RPCRDMA_FMR_ACCESS_FLAGS = IB_ACCESS_REMOTE_WRITE |
35 IB_ACCESS_REMOTE_READ,
39 fmr_is_supported(struct rpcrdma_ia *ia)
41 if (!ia->ri_device->alloc_fmr) {
42 pr_info("rpcrdma: 'fmr' mode is not supported by device %s\n",
50 fmr_op_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr)
52 static struct ib_fmr_attr fmr_attr = {
53 .max_pages = RPCRDMA_MAX_FMR_SGES,
55 .page_shift = PAGE_SHIFT
58 mr->fmr.fm_physaddrs = kcalloc(RPCRDMA_MAX_FMR_SGES,
59 sizeof(u64), GFP_KERNEL);
60 if (!mr->fmr.fm_physaddrs)
63 mr->mr_sg = kcalloc(RPCRDMA_MAX_FMR_SGES,
64 sizeof(*mr->mr_sg), GFP_KERNEL);
68 sg_init_table(mr->mr_sg, RPCRDMA_MAX_FMR_SGES);
70 mr->fmr.fm_mr = ib_alloc_fmr(ia->ri_pd, RPCRDMA_FMR_ACCESS_FLAGS,
72 if (IS_ERR(mr->fmr.fm_mr))
78 dprintk("RPC: %s: ib_alloc_fmr returned %ld\n", __func__,
79 PTR_ERR(mr->fmr.fm_mr));
83 kfree(mr->fmr.fm_physaddrs);
88 __fmr_unmap(struct rpcrdma_mr *mr)
93 list_add(&mr->fmr.fm_mr->list, &l);
94 rc = ib_unmap_fmr(&l);
95 list_del(&mr->fmr.fm_mr->list);
100 fmr_op_release_mr(struct rpcrdma_mr *mr)
102 LIST_HEAD(unmap_list);
105 /* Ensure MW is not on any rl_registered list */
106 if (!list_empty(&mr->mr_list))
107 list_del(&mr->mr_list);
109 kfree(mr->fmr.fm_physaddrs);
112 /* In case this one was left mapped, try to unmap it
113 * to prevent dealloc_fmr from failing with EBUSY
115 rc = __fmr_unmap(mr);
117 pr_err("rpcrdma: final ib_unmap_fmr for %p failed %i\n",
120 rc = ib_dealloc_fmr(mr->fmr.fm_mr);
122 pr_err("rpcrdma: final ib_dealloc_fmr for %p returned %i\n",
128 /* Reset of a single FMR.
131 fmr_op_recover_mr(struct rpcrdma_mr *mr)
133 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
136 /* ORDER: invalidate first */
137 rc = __fmr_unmap(mr);
141 /* ORDER: then DMA unmap */
142 rpcrdma_mr_unmap_and_put(mr);
144 r_xprt->rx_stats.mrs_recovered++;
148 pr_err("rpcrdma: FMR reset failed (%d), %p released\n", rc, mr);
149 r_xprt->rx_stats.mrs_orphaned++;
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);
155 spin_lock(&r_xprt->rx_buf.rb_mrlock);
156 list_del(&mr->mr_all);
157 spin_unlock(&r_xprt->rx_buf.rb_mrlock);
159 fmr_op_release_mr(mr);
163 fmr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
164 struct rpcrdma_create_data_internal *cdata)
166 ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
167 RPCRDMA_MAX_FMR_SGES);
171 /* FMR mode conveys up to 64 pages of payload per chunk segment.
174 fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
176 return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
177 RPCRDMA_MAX_HDR_SEGS * RPCRDMA_MAX_FMR_SGES);
180 /* Use the ib_map_phys_fmr() verb to register a memory region
181 * for remote access via RDMA READ or RDMA WRITE.
183 static struct rpcrdma_mr_seg *
184 fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
185 int nsegs, bool writing, struct rpcrdma_mr **out)
187 struct rpcrdma_mr_seg *seg1 = seg;
188 int len, pageoff, i, rc;
189 struct rpcrdma_mr *mr;
192 mr = rpcrdma_mr_get(r_xprt);
194 return ERR_PTR(-EAGAIN);
196 pageoff = offset_in_page(seg1->mr_offset);
197 seg1->mr_offset -= pageoff; /* start of page */
198 seg1->mr_len += pageoff;
200 if (nsegs > RPCRDMA_MAX_FMR_SGES)
201 nsegs = RPCRDMA_MAX_FMR_SGES;
202 for (i = 0; i < nsegs;) {
204 sg_set_page(&mr->mr_sg[i],
207 offset_in_page(seg->mr_offset));
209 sg_set_buf(&mr->mr_sg[i], seg->mr_offset,
214 /* Check for holes */
215 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
216 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
219 mr->mr_dir = rpcrdma_data_dir(writing);
221 mr->mr_nents = ib_dma_map_sg(r_xprt->rx_ia.ri_device,
222 mr->mr_sg, i, mr->mr_dir);
226 for (i = 0, dma_pages = mr->fmr.fm_physaddrs; i < mr->mr_nents; i++)
227 dma_pages[i] = sg_dma_address(&mr->mr_sg[i]);
228 rc = ib_map_phys_fmr(mr->fmr.fm_mr, dma_pages, mr->mr_nents,
233 mr->mr_handle = mr->fmr.fm_mr->rkey;
235 mr->mr_offset = dma_pages[0] + pageoff;
241 pr_err("rpcrdma: failed to DMA map sg %p sg_nents %d\n",
244 return ERR_PTR(-EIO);
247 pr_err("rpcrdma: ib_map_phys_fmr %u@0x%llx+%i (%d) status %i\n",
248 len, (unsigned long long)dma_pages[0],
249 pageoff, mr->mr_nents, rc);
250 rpcrdma_mr_unmap_and_put(mr);
251 return ERR_PTR(-EIO);
254 /* Post Send WR containing the RPC Call message.
257 fmr_op_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
259 struct ib_send_wr *bad_wr;
261 return ib_post_send(ia->ri_id->qp, &req->rl_sendctx->sc_wr, &bad_wr);
264 /* Invalidate all memory regions that were registered for "req".
266 * Sleeps until it is safe for the host CPU to access the
267 * previously mapped memory regions.
269 * Caller ensures that @mrs is not empty before the call. This
270 * function empties the list.
273 fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct list_head *mrs)
275 struct rpcrdma_mr *mr;
276 LIST_HEAD(unmap_list);
279 /* ORDER: Invalidate all of the req's MRs first
281 * ib_unmap_fmr() is slow, so use a single call instead
282 * of one call per mapped FMR.
284 list_for_each_entry(mr, mrs, mr_list) {
285 dprintk("RPC: %s: unmapping fmr %p\n",
287 trace_xprtrdma_localinv(mr);
288 list_add_tail(&mr->fmr.fm_mr->list, &unmap_list);
290 r_xprt->rx_stats.local_inv_needed++;
291 rc = ib_unmap_fmr(&unmap_list);
295 /* ORDER: Now DMA unmap all of the req's MRs, and return
296 * them to the free MW list.
298 while (!list_empty(mrs)) {
299 mr = rpcrdma_mr_pop(mrs);
300 list_del(&mr->fmr.fm_mr->list);
301 rpcrdma_mr_unmap_and_put(mr);
307 pr_err("rpcrdma: ib_unmap_fmr failed (%i)\n", rc);
309 while (!list_empty(mrs)) {
310 mr = rpcrdma_mr_pop(mrs);
311 list_del(&mr->fmr.fm_mr->list);
312 fmr_op_recover_mr(mr);
316 const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
317 .ro_map = fmr_op_map,
318 .ro_send = fmr_op_send,
319 .ro_unmap_sync = fmr_op_unmap_sync,
320 .ro_recover_mr = fmr_op_recover_mr,
321 .ro_open = fmr_op_open,
322 .ro_maxpages = fmr_op_maxpages,
323 .ro_init_mr = fmr_op_init_mr,
324 .ro_release_mr = fmr_op_release_mr,
325 .ro_displayname = "fmr",
326 .ro_send_w_inv_ok = 0,