2 * Copyright (c) 2015 Oracle. All rights reserved.
3 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
6 /* Lightweight memory registration using Fast Memory Regions (FMR).
7 * Referred to sometimes as MTHCAFMR mode.
9 * FMR uses synchronous memory registration and deregistration.
10 * FMR registration is known to be fast, but FMR deregistration
11 * can take tens of usecs to complete.
16 * A Memory Region is prepared for RDMA READ or WRITE using the
17 * ib_map_phys_fmr verb (fmr_op_map). When the RDMA operation is
18 * finished, the Memory Region is unmapped using the ib_unmap_fmr
19 * verb (fmr_op_unmap).
24 * After a transport reconnect, fmr_op_map re-uses the MR already
25 * allocated for the RPC, but generates a fresh rkey then maps the
26 * MR again. This process is synchronous.
29 #include "xprt_rdma.h"
31 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
32 # define RPCDBG_FACILITY RPCDBG_TRANS
35 /* Maximum scatter/gather per FMR */
36 #define RPCRDMA_MAX_FMR_SGES (64)
38 /* Access mode of externally registered pages */
40 RPCRDMA_FMR_ACCESS_FLAGS = IB_ACCESS_REMOTE_WRITE |
41 IB_ACCESS_REMOTE_READ,
44 static struct workqueue_struct *fmr_recovery_wq;
46 #define FMR_RECOVERY_WQ_FLAGS (WQ_UNBOUND)
49 fmr_alloc_recovery_wq(void)
51 fmr_recovery_wq = alloc_workqueue("fmr_recovery", WQ_UNBOUND, 0);
52 return !fmr_recovery_wq ? -ENOMEM : 0;
56 fmr_destroy_recovery_wq(void)
58 struct workqueue_struct *wq;
64 fmr_recovery_wq = NULL;
65 destroy_workqueue(wq);
69 __fmr_init(struct rpcrdma_mw *mw, struct ib_pd *pd)
71 static struct ib_fmr_attr fmr_attr = {
72 .max_pages = RPCRDMA_MAX_FMR_SGES,
74 .page_shift = PAGE_SHIFT
77 mw->fmr.physaddrs = kcalloc(RPCRDMA_MAX_FMR_SGES,
78 sizeof(u64), GFP_KERNEL);
79 if (!mw->fmr.physaddrs)
82 mw->mw_sg = kcalloc(RPCRDMA_MAX_FMR_SGES,
83 sizeof(*mw->mw_sg), GFP_KERNEL);
87 sg_init_table(mw->mw_sg, RPCRDMA_MAX_FMR_SGES);
89 mw->fmr.fmr = ib_alloc_fmr(pd, RPCRDMA_FMR_ACCESS_FLAGS,
91 if (IS_ERR(mw->fmr.fmr))
97 dprintk("RPC: %s: ib_alloc_fmr returned %ld\n", __func__,
98 PTR_ERR(mw->fmr.fmr));
102 kfree(mw->fmr.physaddrs);
107 __fmr_unmap(struct rpcrdma_mw *mw)
112 list_add(&mw->fmr.fmr->list, &l);
113 rc = ib_unmap_fmr(&l);
114 list_del_init(&mw->fmr.fmr->list);
119 __fmr_dma_unmap(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg)
121 struct ib_device *device = r_xprt->rx_ia.ri_device;
122 int nsegs = seg->mr_nsegs;
125 rpcrdma_unmap_one(device, seg++);
129 __fmr_release(struct rpcrdma_mw *r)
133 kfree(r->fmr.physaddrs);
136 rc = ib_dealloc_fmr(r->fmr.fmr);
138 pr_err("rpcrdma: final ib_dealloc_fmr for %p returned %i\n",
142 /* Deferred reset of a single FMR. Generate a fresh rkey by
143 * replacing the MR. There's no recovery if this fails.
146 __fmr_recovery_worker(struct work_struct *work)
148 struct rpcrdma_mw *mw = container_of(work, struct rpcrdma_mw,
150 struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
153 rpcrdma_put_mw(r_xprt, mw);
157 /* A broken MR was discovered in a context that can't sleep.
158 * Defer recovery to the recovery worker.
161 __fmr_queue_recovery(struct rpcrdma_mw *mw)
163 INIT_WORK(&mw->mw_work, __fmr_recovery_worker);
164 queue_work(fmr_recovery_wq, &mw->mw_work);
168 fmr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
169 struct rpcrdma_create_data_internal *cdata)
171 rpcrdma_set_max_header_sizes(ia, cdata, max_t(unsigned int, 1,
172 RPCRDMA_MAX_DATA_SEGS /
173 RPCRDMA_MAX_FMR_SGES));
177 /* FMR mode conveys up to 64 pages of payload per chunk segment.
180 fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
182 return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
183 RPCRDMA_MAX_HDR_SEGS * RPCRDMA_MAX_FMR_SGES);
187 fmr_op_init(struct rpcrdma_xprt *r_xprt)
189 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
190 struct ib_pd *pd = r_xprt->rx_ia.ri_pd;
191 struct rpcrdma_mw *r;
194 spin_lock_init(&buf->rb_mwlock);
195 INIT_LIST_HEAD(&buf->rb_mws);
196 INIT_LIST_HEAD(&buf->rb_all);
198 i = max_t(int, RPCRDMA_MAX_DATA_SEGS / RPCRDMA_MAX_FMR_SGES, 1);
199 i += 2; /* head + tail */
200 i *= buf->rb_max_requests; /* one set for each RPC slot */
201 dprintk("RPC: %s: initalizing %d FMRs\n", __func__, i);
204 r = kzalloc(sizeof(*r), GFP_KERNEL);
208 rc = __fmr_init(r, pd);
215 list_add(&r->mw_list, &buf->rb_mws);
216 list_add(&r->mw_all, &buf->rb_all);
221 /* Use the ib_map_phys_fmr() verb to register a memory region
222 * for remote access via RDMA READ or RDMA WRITE.
225 fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
226 int nsegs, bool writing)
228 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
229 struct ib_device *device = ia->ri_device;
230 enum dma_data_direction direction = rpcrdma_data_dir(writing);
231 struct rpcrdma_mr_seg *seg1 = seg;
232 int len, pageoff, i, rc;
233 struct rpcrdma_mw *mw;
238 mw = rpcrdma_get_mw(r_xprt);
242 /* this is a retransmit; generate a fresh rkey */
243 rc = __fmr_unmap(mw);
248 pageoff = offset_in_page(seg1->mr_offset);
249 seg1->mr_offset -= pageoff; /* start of page */
250 seg1->mr_len += pageoff;
252 if (nsegs > RPCRDMA_MAX_FMR_SGES)
253 nsegs = RPCRDMA_MAX_FMR_SGES;
254 for (i = 0; i < nsegs;) {
255 rpcrdma_map_one(device, seg, direction);
256 mw->fmr.physaddrs[i] = seg->mr_dma;
260 /* Check for holes */
261 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
262 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
266 rc = ib_map_phys_fmr(mw->fmr.fmr, mw->fmr.physaddrs,
272 seg1->mr_rkey = mw->fmr.fmr->rkey;
273 seg1->mr_base = seg1->mr_dma + pageoff;
279 dprintk("RPC: %s: ib_map_phys_fmr %u@0x%llx+%i (%d) status %i\n",
280 __func__, len, (unsigned long long)seg1->mr_dma,
283 rpcrdma_unmap_one(device, --seg);
287 /* Invalidate all memory regions that were registered for "req".
289 * Sleeps until it is safe for the host CPU to access the
290 * previously mapped memory regions.
293 fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
295 struct rpcrdma_mr_seg *seg;
296 unsigned int i, nchunks;
297 struct rpcrdma_mw *mw;
298 LIST_HEAD(unmap_list);
301 dprintk("RPC: %s: req %p\n", __func__, req);
303 /* ORDER: Invalidate all of the req's MRs first
305 * ib_unmap_fmr() is slow, so use a single call instead
306 * of one call per mapped MR.
308 for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
309 seg = &req->rl_segments[i];
312 list_add_tail(&mw->fmr.fmr->list, &unmap_list);
316 rc = ib_unmap_fmr(&unmap_list);
318 pr_warn("%s: ib_unmap_fmr failed (%i)\n", __func__, rc);
320 /* ORDER: Now DMA unmap all of the req's MRs, and return
321 * them to the free MW list.
323 for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
324 seg = &req->rl_segments[i];
327 list_del_init(&mw->fmr.fmr->list);
328 __fmr_dma_unmap(r_xprt, seg);
329 rpcrdma_put_mw(r_xprt, seg->rl_mw);
339 /* Use a slow, safe mechanism to invalidate all memory regions
340 * that were registered for "req".
342 * In the asynchronous case, DMA unmapping occurs first here
343 * because the rpcrdma_mr_seg is released immediately after this
344 * call. It's contents won't be available in __fmr_dma_unmap later.
348 fmr_op_unmap_safe(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
351 struct rpcrdma_mr_seg *seg;
352 struct rpcrdma_mw *mw;
355 for (i = 0; req->rl_nchunks; req->rl_nchunks--) {
356 seg = &req->rl_segments[i];
362 __fmr_dma_unmap(r_xprt, seg);
363 rpcrdma_put_mw(r_xprt, mw);
365 __fmr_dma_unmap(r_xprt, seg);
366 __fmr_queue_recovery(mw);
376 fmr_op_destroy(struct rpcrdma_buffer *buf)
378 struct rpcrdma_mw *r;
380 while (!list_empty(&buf->rb_all)) {
381 r = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
382 list_del(&r->mw_all);
388 const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
389 .ro_map = fmr_op_map,
390 .ro_unmap_sync = fmr_op_unmap_sync,
391 .ro_unmap_safe = fmr_op_unmap_safe,
392 .ro_open = fmr_op_open,
393 .ro_maxpages = fmr_op_maxpages,
394 .ro_init = fmr_op_init,
395 .ro_destroy = fmr_op_destroy,
396 .ro_displayname = "fmr",