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 <linux/sunrpc/svc_rdma.h>
25 #include "xprt_rdma.h"
26 #include <trace/events/rpcrdma.h>
28 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
29 # define RPCDBG_FACILITY RPCDBG_TRANS
32 /* Maximum scatter/gather per FMR */
33 #define RPCRDMA_MAX_FMR_SGES (64)
35 /* Access mode of externally registered pages */
37 RPCRDMA_FMR_ACCESS_FLAGS = IB_ACCESS_REMOTE_WRITE |
38 IB_ACCESS_REMOTE_READ,
42 fmr_is_supported(struct rpcrdma_ia *ia)
44 if (!ia->ri_device->ops.alloc_fmr) {
45 pr_info("rpcrdma: 'fmr' mode is not supported by device %s\n",
53 __fmr_unmap(struct rpcrdma_mr *mr)
58 list_add(&mr->fmr.fm_mr->list, &l);
59 rc = ib_unmap_fmr(&l);
60 list_del(&mr->fmr.fm_mr->list);
62 pr_err("rpcrdma: final ib_unmap_fmr for %p failed %i\n",
69 fmr_op_release_mr(struct rpcrdma_mr *mr)
73 kfree(mr->fmr.fm_physaddrs);
76 /* In case this one was left mapped, try to unmap it
77 * to prevent dealloc_fmr from failing with EBUSY
81 rc = ib_dealloc_fmr(mr->fmr.fm_mr);
83 pr_err("rpcrdma: final ib_dealloc_fmr for %p returned %i\n",
89 /* MRs are dynamically allocated, so simply clean up and release the MR.
90 * A replacement MR will subsequently be allocated on demand.
93 fmr_mr_recycle_worker(struct work_struct *work)
95 struct rpcrdma_mr *mr = container_of(work, struct rpcrdma_mr, mr_recycle);
96 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
98 trace_xprtrdma_mr_recycle(mr);
100 trace_xprtrdma_mr_unmap(mr);
101 ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
102 mr->mr_sg, mr->mr_nents, mr->mr_dir);
104 spin_lock(&r_xprt->rx_buf.rb_mrlock);
105 list_del(&mr->mr_all);
106 r_xprt->rx_stats.mrs_recycled++;
107 spin_unlock(&r_xprt->rx_buf.rb_mrlock);
108 fmr_op_release_mr(mr);
112 fmr_op_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr)
114 static struct ib_fmr_attr fmr_attr = {
115 .max_pages = RPCRDMA_MAX_FMR_SGES,
117 .page_shift = PAGE_SHIFT
120 mr->fmr.fm_physaddrs = kcalloc(RPCRDMA_MAX_FMR_SGES,
121 sizeof(u64), GFP_KERNEL);
122 if (!mr->fmr.fm_physaddrs)
125 mr->mr_sg = kcalloc(RPCRDMA_MAX_FMR_SGES,
126 sizeof(*mr->mr_sg), GFP_KERNEL);
130 sg_init_table(mr->mr_sg, RPCRDMA_MAX_FMR_SGES);
132 mr->fmr.fm_mr = ib_alloc_fmr(ia->ri_pd, RPCRDMA_FMR_ACCESS_FLAGS,
134 if (IS_ERR(mr->fmr.fm_mr))
137 INIT_LIST_HEAD(&mr->mr_list);
138 INIT_WORK(&mr->mr_recycle, fmr_mr_recycle_worker);
142 dprintk("RPC: %s: ib_alloc_fmr returned %ld\n", __func__,
143 PTR_ERR(mr->fmr.fm_mr));
147 kfree(mr->fmr.fm_physaddrs);
152 * ep->rep_attr.cap.max_send_wr
153 * ep->rep_attr.cap.max_recv_wr
154 * cdata->max_requests
158 fmr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
159 struct rpcrdma_create_data_internal *cdata)
163 max_qp_wr = ia->ri_device->attrs.max_qp_wr;
164 max_qp_wr -= RPCRDMA_BACKWARD_WRS;
166 if (max_qp_wr < RPCRDMA_MIN_SLOT_TABLE)
168 if (cdata->max_requests > max_qp_wr)
169 cdata->max_requests = max_qp_wr;
170 ep->rep_attr.cap.max_send_wr = cdata->max_requests;
171 ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
172 ep->rep_attr.cap.max_send_wr += 1; /* for ib_drain_sq */
173 ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
174 ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
175 ep->rep_attr.cap.max_recv_wr += 1; /* for ib_drain_rq */
177 ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
178 RPCRDMA_MAX_FMR_SGES);
179 ia->ri_max_segs += 2; /* segments for head and tail buffers */
183 /* FMR mode conveys up to 64 pages of payload per chunk segment.
186 fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
188 return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
189 RPCRDMA_MAX_HDR_SEGS * RPCRDMA_MAX_FMR_SGES);
192 /* Use the ib_map_phys_fmr() verb to register a memory region
193 * for remote access via RDMA READ or RDMA WRITE.
195 static struct rpcrdma_mr_seg *
196 fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
197 int nsegs, bool writing, struct rpcrdma_mr **out)
199 struct rpcrdma_mr_seg *seg1 = seg;
200 int len, pageoff, i, rc;
201 struct rpcrdma_mr *mr;
204 mr = rpcrdma_mr_get(r_xprt);
206 return ERR_PTR(-EAGAIN);
208 pageoff = offset_in_page(seg1->mr_offset);
209 seg1->mr_offset -= pageoff; /* start of page */
210 seg1->mr_len += pageoff;
212 if (nsegs > RPCRDMA_MAX_FMR_SGES)
213 nsegs = RPCRDMA_MAX_FMR_SGES;
214 for (i = 0; i < nsegs;) {
216 sg_set_page(&mr->mr_sg[i],
219 offset_in_page(seg->mr_offset));
221 sg_set_buf(&mr->mr_sg[i], seg->mr_offset,
226 /* Check for holes */
227 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
228 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
231 mr->mr_dir = rpcrdma_data_dir(writing);
233 mr->mr_nents = ib_dma_map_sg(r_xprt->rx_ia.ri_device,
234 mr->mr_sg, i, mr->mr_dir);
237 trace_xprtrdma_mr_map(mr);
239 for (i = 0, dma_pages = mr->fmr.fm_physaddrs; i < mr->mr_nents; i++)
240 dma_pages[i] = sg_dma_address(&mr->mr_sg[i]);
241 rc = ib_map_phys_fmr(mr->fmr.fm_mr, dma_pages, mr->mr_nents,
246 mr->mr_handle = mr->fmr.fm_mr->rkey;
248 mr->mr_offset = dma_pages[0] + pageoff;
254 pr_err("rpcrdma: failed to DMA map sg %p sg_nents %d\n",
257 return ERR_PTR(-EIO);
260 pr_err("rpcrdma: ib_map_phys_fmr %u@0x%llx+%i (%d) status %i\n",
261 len, (unsigned long long)dma_pages[0],
262 pageoff, mr->mr_nents, rc);
263 rpcrdma_mr_unmap_and_put(mr);
264 return ERR_PTR(-EIO);
267 /* Post Send WR containing the RPC Call message.
270 fmr_op_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
272 return ib_post_send(ia->ri_id->qp, &req->rl_sendctx->sc_wr, NULL);
275 /* Invalidate all memory regions that were registered for "req".
277 * Sleeps until it is safe for the host CPU to access the
278 * previously mapped memory regions.
280 * Caller ensures that @mrs is not empty before the call. This
281 * function empties the list.
284 fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct list_head *mrs)
286 struct rpcrdma_mr *mr;
287 LIST_HEAD(unmap_list);
290 /* ORDER: Invalidate all of the req's MRs first
292 * ib_unmap_fmr() is slow, so use a single call instead
293 * of one call per mapped FMR.
295 list_for_each_entry(mr, mrs, mr_list) {
296 dprintk("RPC: %s: unmapping fmr %p\n",
298 trace_xprtrdma_mr_localinv(mr);
299 list_add_tail(&mr->fmr.fm_mr->list, &unmap_list);
301 r_xprt->rx_stats.local_inv_needed++;
302 rc = ib_unmap_fmr(&unmap_list);
306 /* ORDER: Now DMA unmap all of the req's MRs, and return
307 * them to the free MW list.
309 while (!list_empty(mrs)) {
310 mr = rpcrdma_mr_pop(mrs);
311 list_del(&mr->fmr.fm_mr->list);
312 rpcrdma_mr_unmap_and_put(mr);
318 pr_err("rpcrdma: ib_unmap_fmr failed (%i)\n", rc);
320 while (!list_empty(mrs)) {
321 mr = rpcrdma_mr_pop(mrs);
322 list_del(&mr->fmr.fm_mr->list);
323 rpcrdma_mr_recycle(mr);
327 const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
328 .ro_map = fmr_op_map,
329 .ro_send = fmr_op_send,
330 .ro_unmap_sync = fmr_op_unmap_sync,
331 .ro_open = fmr_op_open,
332 .ro_maxpages = fmr_op_maxpages,
333 .ro_init_mr = fmr_op_init_mr,
334 .ro_release_mr = fmr_op_release_mr,
335 .ro_displayname = "fmr",
336 .ro_send_w_inv_ok = 0,