1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2015 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_mw *mw)
52 static struct ib_fmr_attr fmr_attr = {
53 .max_pages = RPCRDMA_MAX_FMR_SGES,
55 .page_shift = PAGE_SHIFT
58 mw->fmr.fm_physaddrs = kcalloc(RPCRDMA_MAX_FMR_SGES,
59 sizeof(u64), GFP_KERNEL);
60 if (!mw->fmr.fm_physaddrs)
63 mw->mw_sg = kcalloc(RPCRDMA_MAX_FMR_SGES,
64 sizeof(*mw->mw_sg), GFP_KERNEL);
68 sg_init_table(mw->mw_sg, RPCRDMA_MAX_FMR_SGES);
70 mw->fmr.fm_mr = ib_alloc_fmr(ia->ri_pd, RPCRDMA_FMR_ACCESS_FLAGS,
72 if (IS_ERR(mw->fmr.fm_mr))
78 dprintk("RPC: %s: ib_alloc_fmr returned %ld\n", __func__,
79 PTR_ERR(mw->fmr.fm_mr));
83 kfree(mw->fmr.fm_physaddrs);
88 __fmr_unmap(struct rpcrdma_mw *mw)
93 list_add(&mw->fmr.fm_mr->list, &l);
94 rc = ib_unmap_fmr(&l);
95 list_del(&mw->fmr.fm_mr->list);
100 fmr_op_release_mr(struct rpcrdma_mw *r)
102 LIST_HEAD(unmap_list);
105 /* Ensure MW is not on any rl_registered list */
106 if (!list_empty(&r->mw_list))
107 list_del(&r->mw_list);
109 kfree(r->fmr.fm_physaddrs);
112 /* In case this one was left mapped, try to unmap it
113 * to prevent dealloc_fmr from failing with EBUSY
117 pr_err("rpcrdma: final ib_unmap_fmr for %p failed %i\n",
120 rc = ib_dealloc_fmr(r->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_mw *mw)
133 struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
136 /* ORDER: invalidate first */
137 rc = __fmr_unmap(mw);
139 /* ORDER: then DMA unmap */
140 ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
141 mw->mw_sg, mw->mw_nents, mw->mw_dir);
145 rpcrdma_put_mw(r_xprt, mw);
146 r_xprt->rx_stats.mrs_recovered++;
150 pr_err("rpcrdma: FMR reset failed (%d), %p released\n", rc, mw);
151 r_xprt->rx_stats.mrs_orphaned++;
153 spin_lock(&r_xprt->rx_buf.rb_mwlock);
154 list_del(&mw->mw_all);
155 spin_unlock(&r_xprt->rx_buf.rb_mwlock);
157 fmr_op_release_mr(mw);
161 fmr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
162 struct rpcrdma_create_data_internal *cdata)
164 ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
165 RPCRDMA_MAX_FMR_SGES);
169 /* FMR mode conveys up to 64 pages of payload per chunk segment.
172 fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
174 return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
175 RPCRDMA_MAX_HDR_SEGS * RPCRDMA_MAX_FMR_SGES);
178 /* Use the ib_map_phys_fmr() verb to register a memory region
179 * for remote access via RDMA READ or RDMA WRITE.
181 static struct rpcrdma_mr_seg *
182 fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
183 int nsegs, bool writing, struct rpcrdma_mw **out)
185 struct rpcrdma_mr_seg *seg1 = seg;
186 int len, pageoff, i, rc;
187 struct rpcrdma_mw *mw;
190 mw = rpcrdma_get_mw(r_xprt);
192 return ERR_PTR(-ENOBUFS);
194 pageoff = offset_in_page(seg1->mr_offset);
195 seg1->mr_offset -= pageoff; /* start of page */
196 seg1->mr_len += pageoff;
198 if (nsegs > RPCRDMA_MAX_FMR_SGES)
199 nsegs = RPCRDMA_MAX_FMR_SGES;
200 for (i = 0; i < nsegs;) {
202 sg_set_page(&mw->mw_sg[i],
205 offset_in_page(seg->mr_offset));
207 sg_set_buf(&mw->mw_sg[i], seg->mr_offset,
212 /* Check for holes */
213 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
214 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
217 mw->mw_dir = rpcrdma_data_dir(writing);
219 mw->mw_nents = ib_dma_map_sg(r_xprt->rx_ia.ri_device,
220 mw->mw_sg, i, mw->mw_dir);
224 for (i = 0, dma_pages = mw->fmr.fm_physaddrs; i < mw->mw_nents; i++)
225 dma_pages[i] = sg_dma_address(&mw->mw_sg[i]);
226 rc = ib_map_phys_fmr(mw->fmr.fm_mr, dma_pages, mw->mw_nents,
231 mw->mw_handle = mw->fmr.fm_mr->rkey;
233 mw->mw_offset = dma_pages[0] + pageoff;
239 pr_err("rpcrdma: failed to DMA map sg %p sg_nents %d\n",
241 rpcrdma_put_mw(r_xprt, mw);
242 return ERR_PTR(-EIO);
245 pr_err("rpcrdma: ib_map_phys_fmr %u@0x%llx+%i (%d) status %i\n",
246 len, (unsigned long long)dma_pages[0],
247 pageoff, mw->mw_nents, rc);
248 ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
249 mw->mw_sg, mw->mw_nents, mw->mw_dir);
250 rpcrdma_put_mw(r_xprt, mw);
251 return ERR_PTR(-EIO);
254 /* Invalidate all memory regions that were registered for "req".
256 * Sleeps until it is safe for the host CPU to access the
257 * previously mapped memory regions.
259 * Caller ensures that @mws is not empty before the call. This
260 * function empties the list.
263 fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct list_head *mws)
265 struct rpcrdma_mw *mw;
266 LIST_HEAD(unmap_list);
269 /* ORDER: Invalidate all of the req's MRs first
271 * ib_unmap_fmr() is slow, so use a single call instead
272 * of one call per mapped FMR.
274 list_for_each_entry(mw, mws, mw_list) {
275 dprintk("RPC: %s: unmapping fmr %p\n",
277 list_add_tail(&mw->fmr.fm_mr->list, &unmap_list);
279 r_xprt->rx_stats.local_inv_needed++;
280 rc = ib_unmap_fmr(&unmap_list);
284 /* ORDER: Now DMA unmap all of the req's MRs, and return
285 * them to the free MW list.
287 while (!list_empty(mws)) {
288 mw = rpcrdma_pop_mw(mws);
289 dprintk("RPC: %s: DMA unmapping fmr %p\n",
291 list_del(&mw->fmr.fm_mr->list);
292 ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
293 mw->mw_sg, mw->mw_nents, mw->mw_dir);
294 rpcrdma_put_mw(r_xprt, mw);
300 pr_err("rpcrdma: ib_unmap_fmr failed (%i)\n", rc);
302 while (!list_empty(mws)) {
303 mw = rpcrdma_pop_mw(mws);
304 list_del(&mw->fmr.fm_mr->list);
305 fmr_op_recover_mr(mw);
309 const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
310 .ro_map = fmr_op_map,
311 .ro_unmap_sync = fmr_op_unmap_sync,
312 .ro_recover_mr = fmr_op_recover_mr,
313 .ro_open = fmr_op_open,
314 .ro_maxpages = fmr_op_maxpages,
315 .ro_init_mr = fmr_op_init_mr,
316 .ro_release_mr = fmr_op_release_mr,
317 .ro_displayname = "fmr",
318 .ro_send_w_inv_ok = 0,