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 Registration Work
10 * FRWR features ordered asynchronous registration and invalidation
11 * of arbitrarily-sized memory regions. This is the fastest and safest
12 * but most complex memory registration mode.
17 * A Memory Region is prepared for RDMA Read or Write using a FAST_REG
18 * Work Request (frwr_map). When the RDMA operation is finished, this
19 * Memory Region is invalidated using a LOCAL_INV Work Request
20 * (frwr_unmap_async and frwr_unmap_sync).
22 * Typically FAST_REG Work Requests are not signaled, and neither are
23 * RDMA Send Work Requests (with the exception of signaling occasionally
24 * to prevent provider work queue overflows). This greatly reduces HCA
30 * frwr_map and frwr_unmap_* cannot run at the same time the transport
31 * connect worker is running. The connect worker holds the transport
32 * send lock, just as ->send_request does. This prevents frwr_map and
33 * the connect worker from running concurrently. When a connection is
34 * closed, the Receive completion queue is drained before the allowing
35 * the connect worker to get control. This prevents frwr_unmap and the
36 * connect worker from running concurrently.
38 * When the underlying transport disconnects, MRs that are in flight
39 * are flushed and are likely unusable. Thus all MRs are destroyed.
40 * New MRs are created on demand.
43 #include <linux/sunrpc/rpc_rdma.h>
44 #include <linux/sunrpc/svc_rdma.h>
46 #include "xprt_rdma.h"
47 #include <trace/events/rpcrdma.h>
49 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
50 # define RPCDBG_FACILITY RPCDBG_TRANS
54 * frwr_is_supported - Check if device supports FRWR
55 * @device: interface adapter to check
57 * Returns true if device supports FRWR, otherwise false
59 bool frwr_is_supported(struct ib_device *device)
61 struct ib_device_attr *attrs = &device->attrs;
63 if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
64 goto out_not_supported;
65 if (attrs->max_fast_reg_page_list_len == 0)
66 goto out_not_supported;
70 pr_info("rpcrdma: 'frwr' mode is not supported by device %s\n",
76 * frwr_release_mr - Destroy one MR
77 * @mr: MR allocated by frwr_init_mr
80 void frwr_release_mr(struct rpcrdma_mr *mr)
84 rc = ib_dereg_mr(mr->frwr.fr_mr);
86 trace_xprtrdma_frwr_dereg(mr, rc);
91 static void frwr_mr_recycle(struct rpcrdma_mr *mr)
93 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
95 trace_xprtrdma_mr_recycle(mr);
97 if (mr->mr_dir != DMA_NONE) {
98 trace_xprtrdma_mr_unmap(mr);
99 ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device,
100 mr->mr_sg, mr->mr_nents, mr->mr_dir);
101 mr->mr_dir = DMA_NONE;
104 spin_lock(&r_xprt->rx_buf.rb_lock);
105 list_del(&mr->mr_all);
106 r_xprt->rx_stats.mrs_recycled++;
107 spin_unlock(&r_xprt->rx_buf.rb_lock);
112 /* frwr_reset - Place MRs back on the free list
113 * @req: request to reset
115 * Used after a failed marshal. For FRWR, this means the MRs
116 * don't have to be fully released and recreated.
118 * NB: This is safe only as long as none of @req's MRs are
119 * involved with an ongoing asynchronous FAST_REG or LOCAL_INV
122 void frwr_reset(struct rpcrdma_req *req)
124 struct rpcrdma_mr *mr;
126 while ((mr = rpcrdma_mr_pop(&req->rl_registered)))
131 * frwr_init_mr - Initialize one MR
132 * @ia: interface adapter
133 * @mr: generic MR to prepare for FRWR
135 * Returns zero if successful. Otherwise a negative errno
138 int frwr_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr)
140 unsigned int depth = ia->ri_max_frwr_depth;
141 struct scatterlist *sg;
145 frmr = ib_alloc_mr(ia->ri_pd, ia->ri_mrtype, depth);
149 sg = kcalloc(depth, sizeof(*sg), GFP_NOFS);
153 mr->frwr.fr_mr = frmr;
154 mr->mr_dir = DMA_NONE;
155 INIT_LIST_HEAD(&mr->mr_list);
156 init_completion(&mr->frwr.fr_linv_done);
158 sg_init_table(sg, depth);
164 trace_xprtrdma_frwr_alloc(mr, rc);
173 * frwr_open - Prepare an endpoint for use with FRWR
174 * @ia: interface adapter this endpoint will use
175 * @ep: endpoint to prepare
178 * ep->rep_attr.cap.max_send_wr
179 * ep->rep_attr.cap.max_recv_wr
180 * ep->rep_max_requests
183 * And these FRWR-related fields:
184 * ia->ri_max_frwr_depth
187 * On failure, a negative errno is returned.
189 int frwr_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep)
191 struct ib_device_attr *attrs = &ia->ri_id->device->attrs;
192 int max_qp_wr, depth, delta;
194 ia->ri_mrtype = IB_MR_TYPE_MEM_REG;
195 if (attrs->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
196 ia->ri_mrtype = IB_MR_TYPE_SG_GAPS;
198 /* Quirk: Some devices advertise a large max_fast_reg_page_list_len
199 * capability, but perform optimally when the MRs are not larger
202 if (attrs->max_sge_rd > 1)
203 ia->ri_max_frwr_depth = attrs->max_sge_rd;
205 ia->ri_max_frwr_depth = attrs->max_fast_reg_page_list_len;
206 if (ia->ri_max_frwr_depth > RPCRDMA_MAX_DATA_SEGS)
207 ia->ri_max_frwr_depth = RPCRDMA_MAX_DATA_SEGS;
208 dprintk("RPC: %s: max FR page list depth = %u\n",
209 __func__, ia->ri_max_frwr_depth);
211 /* Add room for frwr register and invalidate WRs.
212 * 1. FRWR reg WR for head
213 * 2. FRWR invalidate WR for head
214 * 3. N FRWR reg WRs for pagelist
215 * 4. N FRWR invalidate WRs for pagelist
216 * 5. FRWR reg WR for tail
217 * 6. FRWR invalidate WR for tail
218 * 7. The RDMA_SEND WR
222 /* Calculate N if the device max FRWR depth is smaller than
223 * RPCRDMA_MAX_DATA_SEGS.
225 if (ia->ri_max_frwr_depth < RPCRDMA_MAX_DATA_SEGS) {
226 delta = RPCRDMA_MAX_DATA_SEGS - ia->ri_max_frwr_depth;
228 depth += 2; /* FRWR reg + invalidate */
229 delta -= ia->ri_max_frwr_depth;
233 max_qp_wr = ia->ri_id->device->attrs.max_qp_wr;
234 max_qp_wr -= RPCRDMA_BACKWARD_WRS;
236 if (max_qp_wr < RPCRDMA_MIN_SLOT_TABLE)
238 if (ep->rep_max_requests > max_qp_wr)
239 ep->rep_max_requests = max_qp_wr;
240 ep->rep_attr.cap.max_send_wr = ep->rep_max_requests * depth;
241 if (ep->rep_attr.cap.max_send_wr > max_qp_wr) {
242 ep->rep_max_requests = max_qp_wr / depth;
243 if (!ep->rep_max_requests)
245 ep->rep_attr.cap.max_send_wr = ep->rep_max_requests * depth;
247 ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
248 ep->rep_attr.cap.max_send_wr += 1; /* for ib_drain_sq */
249 ep->rep_attr.cap.max_recv_wr = ep->rep_max_requests;
250 ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
251 ep->rep_attr.cap.max_recv_wr += 1; /* for ib_drain_rq */
254 DIV_ROUND_UP(RPCRDMA_MAX_DATA_SEGS, ia->ri_max_frwr_depth);
255 /* Reply chunks require segments for head and tail buffers */
256 ia->ri_max_segs += 2;
257 if (ia->ri_max_segs > RPCRDMA_MAX_HDR_SEGS)
258 ia->ri_max_segs = RPCRDMA_MAX_HDR_SEGS;
263 * frwr_maxpages - Compute size of largest payload
266 * Returns maximum size of an RPC message, in pages.
268 * FRWR mode conveys a list of pages per chunk segment. The
269 * maximum length of that list is the FRWR page list depth.
271 size_t frwr_maxpages(struct rpcrdma_xprt *r_xprt)
273 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
275 return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
276 (ia->ri_max_segs - 2) * ia->ri_max_frwr_depth);
280 * frwr_map - Register a memory region
281 * @r_xprt: controlling transport
282 * @seg: memory region co-ordinates
283 * @nsegs: number of segments remaining
284 * @writing: true when RDMA Write will be used
285 * @xid: XID of RPC using the registered memory
288 * Prepare a REG_MR Work Request to register a memory region
289 * for remote access via RDMA READ or RDMA WRITE.
291 * Returns the next segment or a negative errno pointer.
292 * On success, @mr is filled in.
294 struct rpcrdma_mr_seg *frwr_map(struct rpcrdma_xprt *r_xprt,
295 struct rpcrdma_mr_seg *seg,
296 int nsegs, bool writing, __be32 xid,
297 struct rpcrdma_mr *mr)
299 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
300 struct ib_reg_wr *reg_wr;
305 if (nsegs > ia->ri_max_frwr_depth)
306 nsegs = ia->ri_max_frwr_depth;
307 for (i = 0; i < nsegs;) {
309 sg_set_page(&mr->mr_sg[i],
312 offset_in_page(seg->mr_offset));
314 sg_set_buf(&mr->mr_sg[i], seg->mr_offset,
319 if (ia->ri_mrtype == IB_MR_TYPE_SG_GAPS)
321 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
322 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
325 mr->mr_dir = rpcrdma_data_dir(writing);
328 ib_dma_map_sg(ia->ri_id->device, mr->mr_sg, i, mr->mr_dir);
332 ibmr = mr->frwr.fr_mr;
333 n = ib_map_mr_sg(ibmr, mr->mr_sg, mr->mr_nents, NULL, PAGE_SIZE);
334 if (unlikely(n != mr->mr_nents))
337 ibmr->iova &= 0x00000000ffffffff;
338 ibmr->iova |= ((u64)be32_to_cpu(xid)) << 32;
339 key = (u8)(ibmr->rkey & 0x000000FF);
340 ib_update_fast_reg_key(ibmr, ++key);
342 reg_wr = &mr->frwr.fr_regwr;
344 reg_wr->key = ibmr->rkey;
345 reg_wr->access = writing ?
346 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
347 IB_ACCESS_REMOTE_READ;
349 mr->mr_handle = ibmr->rkey;
350 mr->mr_length = ibmr->length;
351 mr->mr_offset = ibmr->iova;
352 trace_xprtrdma_mr_map(mr);
357 mr->mr_dir = DMA_NONE;
358 trace_xprtrdma_frwr_sgerr(mr, i);
359 return ERR_PTR(-EIO);
362 trace_xprtrdma_frwr_maperr(mr, n);
363 return ERR_PTR(-EIO);
367 * frwr_wc_fastreg - Invoked by RDMA provider for a flushed FastReg WC
368 * @cq: completion queue (ignored)
372 static void frwr_wc_fastreg(struct ib_cq *cq, struct ib_wc *wc)
374 struct ib_cqe *cqe = wc->wr_cqe;
375 struct rpcrdma_frwr *frwr =
376 container_of(cqe, struct rpcrdma_frwr, fr_cqe);
378 /* WARNING: Only wr_cqe and status are reliable at this point */
379 trace_xprtrdma_wc_fastreg(wc, frwr);
380 /* The MR will get recycled when the associated req is retransmitted */
384 * frwr_send - post Send WR containing the RPC Call message
385 * @ia: interface adapter
386 * @req: Prepared RPC Call
388 * For FRWR, chain any FastReg WRs to the Send WR. Only a
389 * single ib_post_send call is needed to register memory
390 * and then post the Send WR.
392 * Returns the result of ib_post_send.
394 int frwr_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
396 struct ib_send_wr *post_wr;
397 struct rpcrdma_mr *mr;
399 post_wr = &req->rl_wr;
400 list_for_each_entry(mr, &req->rl_registered, mr_list) {
401 struct rpcrdma_frwr *frwr;
405 frwr->fr_cqe.done = frwr_wc_fastreg;
406 frwr->fr_regwr.wr.next = post_wr;
407 frwr->fr_regwr.wr.wr_cqe = &frwr->fr_cqe;
408 frwr->fr_regwr.wr.num_sge = 0;
409 frwr->fr_regwr.wr.opcode = IB_WR_REG_MR;
410 frwr->fr_regwr.wr.send_flags = 0;
412 post_wr = &frwr->fr_regwr.wr;
415 return ib_post_send(ia->ri_id->qp, post_wr, NULL);
419 * frwr_reminv - handle a remotely invalidated mr on the @mrs list
420 * @rep: Received reply
421 * @mrs: list of MRs to check
424 void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs)
426 struct rpcrdma_mr *mr;
428 list_for_each_entry(mr, mrs, mr_list)
429 if (mr->mr_handle == rep->rr_inv_rkey) {
430 list_del_init(&mr->mr_list);
431 trace_xprtrdma_mr_remoteinv(mr);
433 break; /* only one invalidated MR per RPC */
437 static void __frwr_release_mr(struct ib_wc *wc, struct rpcrdma_mr *mr)
439 if (wc->status != IB_WC_SUCCESS)
446 * frwr_wc_localinv - Invoked by RDMA provider for a LOCAL_INV WC
447 * @cq: completion queue (ignored)
451 static void frwr_wc_localinv(struct ib_cq *cq, struct ib_wc *wc)
453 struct ib_cqe *cqe = wc->wr_cqe;
454 struct rpcrdma_frwr *frwr =
455 container_of(cqe, struct rpcrdma_frwr, fr_cqe);
456 struct rpcrdma_mr *mr = container_of(frwr, struct rpcrdma_mr, frwr);
458 /* WARNING: Only wr_cqe and status are reliable at this point */
459 trace_xprtrdma_wc_li(wc, frwr);
460 __frwr_release_mr(wc, mr);
464 * frwr_wc_localinv_wake - Invoked by RDMA provider for a LOCAL_INV WC
465 * @cq: completion queue (ignored)
468 * Awaken anyone waiting for an MR to finish being fenced.
470 static void frwr_wc_localinv_wake(struct ib_cq *cq, struct ib_wc *wc)
472 struct ib_cqe *cqe = wc->wr_cqe;
473 struct rpcrdma_frwr *frwr =
474 container_of(cqe, struct rpcrdma_frwr, fr_cqe);
475 struct rpcrdma_mr *mr = container_of(frwr, struct rpcrdma_mr, frwr);
477 /* WARNING: Only wr_cqe and status are reliable at this point */
478 trace_xprtrdma_wc_li_wake(wc, frwr);
479 __frwr_release_mr(wc, mr);
480 complete(&frwr->fr_linv_done);
484 * frwr_unmap_sync - invalidate memory regions that were registered for @req
485 * @r_xprt: controlling transport instance
486 * @req: rpcrdma_req with a non-empty list of MRs to process
488 * Sleeps until it is safe for the host CPU to access the previously mapped
489 * memory regions. This guarantees that registered MRs are properly fenced
490 * from the server before the RPC consumer accesses the data in them. It
491 * also ensures proper Send flow control: waking the next RPC waits until
492 * this RPC has relinquished all its Send Queue entries.
494 void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
496 struct ib_send_wr *first, **prev, *last;
497 const struct ib_send_wr *bad_wr;
498 struct rpcrdma_frwr *frwr;
499 struct rpcrdma_mr *mr;
502 /* ORDER: Invalidate all of the MRs first
504 * Chain the LOCAL_INV Work Requests and post them with
505 * a single ib_post_send() call.
509 while ((mr = rpcrdma_mr_pop(&req->rl_registered))) {
511 trace_xprtrdma_mr_localinv(mr);
512 r_xprt->rx_stats.local_inv_needed++;
515 frwr->fr_cqe.done = frwr_wc_localinv;
516 last = &frwr->fr_invwr;
518 last->wr_cqe = &frwr->fr_cqe;
519 last->sg_list = NULL;
521 last->opcode = IB_WR_LOCAL_INV;
522 last->send_flags = IB_SEND_SIGNALED;
523 last->ex.invalidate_rkey = mr->mr_handle;
529 /* Strong send queue ordering guarantees that when the
530 * last WR in the chain completes, all WRs in the chain
533 frwr->fr_cqe.done = frwr_wc_localinv_wake;
534 reinit_completion(&frwr->fr_linv_done);
536 /* Transport disconnect drains the receive CQ before it
537 * replaces the QP. The RPC reply handler won't call us
538 * unless ri_id->qp is a valid pointer.
541 rc = ib_post_send(r_xprt->rx_ia.ri_id->qp, first, &bad_wr);
543 /* The final LOCAL_INV WR in the chain is supposed to
544 * do the wake. If it was never posted, the wake will
545 * not happen, so don't wait in that case.
548 wait_for_completion(&frwr->fr_linv_done);
552 /* Recycle MRs in the LOCAL_INV chain that did not get posted.
554 trace_xprtrdma_post_linv(req, rc);
556 frwr = container_of(bad_wr, struct rpcrdma_frwr,
558 mr = container_of(frwr, struct rpcrdma_mr, frwr);
559 bad_wr = bad_wr->next;
561 list_del_init(&mr->mr_list);
567 * frwr_wc_localinv_done - Invoked by RDMA provider for a signaled LOCAL_INV WC
568 * @cq: completion queue (ignored)
572 static void frwr_wc_localinv_done(struct ib_cq *cq, struct ib_wc *wc)
574 struct ib_cqe *cqe = wc->wr_cqe;
575 struct rpcrdma_frwr *frwr =
576 container_of(cqe, struct rpcrdma_frwr, fr_cqe);
577 struct rpcrdma_mr *mr = container_of(frwr, struct rpcrdma_mr, frwr);
578 struct rpcrdma_rep *rep = mr->mr_req->rl_reply;
580 /* WARNING: Only wr_cqe and status are reliable at this point */
581 trace_xprtrdma_wc_li_done(wc, frwr);
582 __frwr_release_mr(wc, mr);
584 /* Ensure @rep is generated before __frwr_release_mr */
586 rpcrdma_complete_rqst(rep);
590 * frwr_unmap_async - invalidate memory regions that were registered for @req
591 * @r_xprt: controlling transport instance
592 * @req: rpcrdma_req with a non-empty list of MRs to process
594 * This guarantees that registered MRs are properly fenced from the
595 * server before the RPC consumer accesses the data in them. It also
596 * ensures proper Send flow control: waking the next RPC waits until
597 * this RPC has relinquished all its Send Queue entries.
599 void frwr_unmap_async(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
601 struct ib_send_wr *first, *last, **prev;
602 const struct ib_send_wr *bad_wr;
603 struct rpcrdma_frwr *frwr;
604 struct rpcrdma_mr *mr;
607 /* Chain the LOCAL_INV Work Requests and post them with
608 * a single ib_post_send() call.
612 while ((mr = rpcrdma_mr_pop(&req->rl_registered))) {
614 trace_xprtrdma_mr_localinv(mr);
615 r_xprt->rx_stats.local_inv_needed++;
618 frwr->fr_cqe.done = frwr_wc_localinv;
619 last = &frwr->fr_invwr;
621 last->wr_cqe = &frwr->fr_cqe;
622 last->sg_list = NULL;
624 last->opcode = IB_WR_LOCAL_INV;
625 last->send_flags = IB_SEND_SIGNALED;
626 last->ex.invalidate_rkey = mr->mr_handle;
632 /* Strong send queue ordering guarantees that when the
633 * last WR in the chain completes, all WRs in the chain
634 * are complete. The last completion will wake up the
637 frwr->fr_cqe.done = frwr_wc_localinv_done;
639 /* Transport disconnect drains the receive CQ before it
640 * replaces the QP. The RPC reply handler won't call us
641 * unless ri_id->qp is a valid pointer.
644 rc = ib_post_send(r_xprt->rx_ia.ri_id->qp, first, &bad_wr);
648 /* Recycle MRs in the LOCAL_INV chain that did not get posted.
650 trace_xprtrdma_post_linv(req, rc);
652 frwr = container_of(bad_wr, struct rpcrdma_frwr, fr_invwr);
653 mr = container_of(frwr, struct rpcrdma_mr, frwr);
654 bad_wr = bad_wr->next;
659 /* The final LOCAL_INV WR in the chain is supposed to
660 * do the wake. If it was never posted, the wake will
661 * not happen, so wake here in that case.
663 rpcrdma_complete_rqst(req->rl_reply);