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 flushed MRs are
40 * destroyed. 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 /* MRs are dynamically allocated, so simply clean up and release the MR.
92 * A replacement MR will subsequently be allocated on demand.
95 frwr_mr_recycle_worker(struct work_struct *work)
97 struct rpcrdma_mr *mr = container_of(work, struct rpcrdma_mr, mr_recycle);
98 struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
100 trace_xprtrdma_mr_recycle(mr);
102 if (mr->mr_dir != DMA_NONE) {
103 trace_xprtrdma_mr_unmap(mr);
104 ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device,
105 mr->mr_sg, mr->mr_nents, mr->mr_dir);
106 mr->mr_dir = DMA_NONE;
109 spin_lock(&r_xprt->rx_buf.rb_mrlock);
110 list_del(&mr->mr_all);
111 r_xprt->rx_stats.mrs_recycled++;
112 spin_unlock(&r_xprt->rx_buf.rb_mrlock);
117 /* frwr_reset - Place MRs back on the free list
118 * @req: request to reset
120 * Used after a failed marshal. For FRWR, this means the MRs
121 * don't have to be fully released and recreated.
123 * NB: This is safe only as long as none of @req's MRs are
124 * involved with an ongoing asynchronous FAST_REG or LOCAL_INV
127 void frwr_reset(struct rpcrdma_req *req)
129 struct rpcrdma_mr *mr;
131 while ((mr = rpcrdma_mr_pop(&req->rl_registered)))
136 * frwr_init_mr - Initialize one MR
137 * @ia: interface adapter
138 * @mr: generic MR to prepare for FRWR
140 * Returns zero if successful. Otherwise a negative errno
143 int frwr_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr)
145 unsigned int depth = ia->ri_max_frwr_depth;
146 struct scatterlist *sg;
150 /* NB: ib_alloc_mr and device drivers typically allocate
151 * memory with GFP_KERNEL.
153 frmr = ib_alloc_mr(ia->ri_pd, ia->ri_mrtype, depth);
157 sg = kcalloc(depth, sizeof(*sg), GFP_NOFS);
161 mr->frwr.fr_mr = frmr;
162 mr->mr_dir = DMA_NONE;
163 INIT_LIST_HEAD(&mr->mr_list);
164 INIT_WORK(&mr->mr_recycle, frwr_mr_recycle_worker);
165 init_completion(&mr->frwr.fr_linv_done);
167 sg_init_table(sg, depth);
173 trace_xprtrdma_frwr_alloc(mr, rc);
182 * frwr_open - Prepare an endpoint for use with FRWR
183 * @ia: interface adapter this endpoint will use
184 * @ep: endpoint to prepare
187 * ep->rep_attr.cap.max_send_wr
188 * ep->rep_attr.cap.max_recv_wr
189 * ep->rep_max_requests
192 * And these FRWR-related fields:
193 * ia->ri_max_frwr_depth
196 * On failure, a negative errno is returned.
198 int frwr_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep)
200 struct ib_device_attr *attrs = &ia->ri_id->device->attrs;
201 int max_qp_wr, depth, delta;
203 ia->ri_mrtype = IB_MR_TYPE_MEM_REG;
204 if (attrs->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
205 ia->ri_mrtype = IB_MR_TYPE_SG_GAPS;
207 /* Quirk: Some devices advertise a large max_fast_reg_page_list_len
208 * capability, but perform optimally when the MRs are not larger
211 if (attrs->max_sge_rd > 1)
212 ia->ri_max_frwr_depth = attrs->max_sge_rd;
214 ia->ri_max_frwr_depth = attrs->max_fast_reg_page_list_len;
215 if (ia->ri_max_frwr_depth > RPCRDMA_MAX_DATA_SEGS)
216 ia->ri_max_frwr_depth = RPCRDMA_MAX_DATA_SEGS;
217 dprintk("RPC: %s: max FR page list depth = %u\n",
218 __func__, ia->ri_max_frwr_depth);
220 /* Add room for frwr register and invalidate WRs.
221 * 1. FRWR reg WR for head
222 * 2. FRWR invalidate WR for head
223 * 3. N FRWR reg WRs for pagelist
224 * 4. N FRWR invalidate WRs for pagelist
225 * 5. FRWR reg WR for tail
226 * 6. FRWR invalidate WR for tail
227 * 7. The RDMA_SEND WR
231 /* Calculate N if the device max FRWR depth is smaller than
232 * RPCRDMA_MAX_DATA_SEGS.
234 if (ia->ri_max_frwr_depth < RPCRDMA_MAX_DATA_SEGS) {
235 delta = RPCRDMA_MAX_DATA_SEGS - ia->ri_max_frwr_depth;
237 depth += 2; /* FRWR reg + invalidate */
238 delta -= ia->ri_max_frwr_depth;
242 max_qp_wr = ia->ri_id->device->attrs.max_qp_wr;
243 max_qp_wr -= RPCRDMA_BACKWARD_WRS;
245 if (max_qp_wr < RPCRDMA_MIN_SLOT_TABLE)
247 if (ep->rep_max_requests > max_qp_wr)
248 ep->rep_max_requests = max_qp_wr;
249 ep->rep_attr.cap.max_send_wr = ep->rep_max_requests * depth;
250 if (ep->rep_attr.cap.max_send_wr > max_qp_wr) {
251 ep->rep_max_requests = max_qp_wr / depth;
252 if (!ep->rep_max_requests)
254 ep->rep_attr.cap.max_send_wr = ep->rep_max_requests * depth;
256 ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
257 ep->rep_attr.cap.max_send_wr += 1; /* for ib_drain_sq */
258 ep->rep_attr.cap.max_recv_wr = ep->rep_max_requests;
259 ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
260 ep->rep_attr.cap.max_recv_wr += 1; /* for ib_drain_rq */
263 DIV_ROUND_UP(RPCRDMA_MAX_DATA_SEGS, ia->ri_max_frwr_depth);
264 /* Reply chunks require segments for head and tail buffers */
265 ia->ri_max_segs += 2;
266 if (ia->ri_max_segs > RPCRDMA_MAX_HDR_SEGS)
267 ia->ri_max_segs = RPCRDMA_MAX_HDR_SEGS;
272 * frwr_maxpages - Compute size of largest payload
275 * Returns maximum size of an RPC message, in pages.
277 * FRWR mode conveys a list of pages per chunk segment. The
278 * maximum length of that list is the FRWR page list depth.
280 size_t frwr_maxpages(struct rpcrdma_xprt *r_xprt)
282 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
284 return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
285 (ia->ri_max_segs - 2) * ia->ri_max_frwr_depth);
289 * frwr_map - Register a memory region
290 * @r_xprt: controlling transport
291 * @seg: memory region co-ordinates
292 * @nsegs: number of segments remaining
293 * @writing: true when RDMA Write will be used
294 * @xid: XID of RPC using the registered memory
297 * Prepare a REG_MR Work Request to register a memory region
298 * for remote access via RDMA READ or RDMA WRITE.
300 * Returns the next segment or a negative errno pointer.
301 * On success, @mr is filled in.
303 struct rpcrdma_mr_seg *frwr_map(struct rpcrdma_xprt *r_xprt,
304 struct rpcrdma_mr_seg *seg,
305 int nsegs, bool writing, __be32 xid,
306 struct rpcrdma_mr *mr)
308 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
309 struct ib_reg_wr *reg_wr;
314 if (nsegs > ia->ri_max_frwr_depth)
315 nsegs = ia->ri_max_frwr_depth;
316 for (i = 0; i < nsegs;) {
318 sg_set_page(&mr->mr_sg[i],
321 offset_in_page(seg->mr_offset));
323 sg_set_buf(&mr->mr_sg[i], seg->mr_offset,
328 if (ia->ri_mrtype == IB_MR_TYPE_SG_GAPS)
330 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
331 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
334 mr->mr_dir = rpcrdma_data_dir(writing);
337 ib_dma_map_sg(ia->ri_id->device, mr->mr_sg, i, mr->mr_dir);
341 ibmr = mr->frwr.fr_mr;
342 n = ib_map_mr_sg(ibmr, mr->mr_sg, mr->mr_nents, NULL, PAGE_SIZE);
343 if (unlikely(n != mr->mr_nents))
346 ibmr->iova &= 0x00000000ffffffff;
347 ibmr->iova |= ((u64)be32_to_cpu(xid)) << 32;
348 key = (u8)(ibmr->rkey & 0x000000FF);
349 ib_update_fast_reg_key(ibmr, ++key);
351 reg_wr = &mr->frwr.fr_regwr;
353 reg_wr->key = ibmr->rkey;
354 reg_wr->access = writing ?
355 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
356 IB_ACCESS_REMOTE_READ;
358 mr->mr_handle = ibmr->rkey;
359 mr->mr_length = ibmr->length;
360 mr->mr_offset = ibmr->iova;
361 trace_xprtrdma_mr_map(mr);
366 mr->mr_dir = DMA_NONE;
367 trace_xprtrdma_frwr_sgerr(mr, i);
368 return ERR_PTR(-EIO);
371 trace_xprtrdma_frwr_maperr(mr, n);
372 return ERR_PTR(-EIO);
376 * frwr_wc_fastreg - Invoked by RDMA provider for a flushed FastReg WC
377 * @cq: completion queue (ignored)
381 static void frwr_wc_fastreg(struct ib_cq *cq, struct ib_wc *wc)
383 struct ib_cqe *cqe = wc->wr_cqe;
384 struct rpcrdma_frwr *frwr =
385 container_of(cqe, struct rpcrdma_frwr, fr_cqe);
387 /* WARNING: Only wr_cqe and status are reliable at this point */
388 trace_xprtrdma_wc_fastreg(wc, frwr);
389 /* The MR will get recycled when the associated req is retransmitted */
393 * frwr_send - post Send WR containing the RPC Call message
394 * @ia: interface adapter
395 * @req: Prepared RPC Call
397 * For FRWR, chain any FastReg WRs to the Send WR. Only a
398 * single ib_post_send call is needed to register memory
399 * and then post the Send WR.
401 * Returns the result of ib_post_send.
403 int frwr_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
405 struct ib_send_wr *post_wr;
406 struct rpcrdma_mr *mr;
408 post_wr = &req->rl_sendctx->sc_wr;
409 list_for_each_entry(mr, &req->rl_registered, mr_list) {
410 struct rpcrdma_frwr *frwr;
414 frwr->fr_cqe.done = frwr_wc_fastreg;
415 frwr->fr_regwr.wr.next = post_wr;
416 frwr->fr_regwr.wr.wr_cqe = &frwr->fr_cqe;
417 frwr->fr_regwr.wr.num_sge = 0;
418 frwr->fr_regwr.wr.opcode = IB_WR_REG_MR;
419 frwr->fr_regwr.wr.send_flags = 0;
421 post_wr = &frwr->fr_regwr.wr;
424 /* If ib_post_send fails, the next ->send_request for
425 * @req will queue these MRs for recovery.
427 return ib_post_send(ia->ri_id->qp, post_wr, NULL);
431 * frwr_reminv - handle a remotely invalidated mr on the @mrs list
432 * @rep: Received reply
433 * @mrs: list of MRs to check
436 void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs)
438 struct rpcrdma_mr *mr;
440 list_for_each_entry(mr, mrs, mr_list)
441 if (mr->mr_handle == rep->rr_inv_rkey) {
442 list_del_init(&mr->mr_list);
443 trace_xprtrdma_mr_remoteinv(mr);
445 break; /* only one invalidated MR per RPC */
449 static void __frwr_release_mr(struct ib_wc *wc, struct rpcrdma_mr *mr)
451 if (wc->status != IB_WC_SUCCESS)
452 rpcrdma_mr_recycle(mr);
458 * frwr_wc_localinv - Invoked by RDMA provider for a LOCAL_INV WC
459 * @cq: completion queue (ignored)
463 static void frwr_wc_localinv(struct ib_cq *cq, struct ib_wc *wc)
465 struct ib_cqe *cqe = wc->wr_cqe;
466 struct rpcrdma_frwr *frwr =
467 container_of(cqe, struct rpcrdma_frwr, fr_cqe);
468 struct rpcrdma_mr *mr = container_of(frwr, struct rpcrdma_mr, frwr);
470 /* WARNING: Only wr_cqe and status are reliable at this point */
471 trace_xprtrdma_wc_li(wc, frwr);
472 __frwr_release_mr(wc, mr);
476 * frwr_wc_localinv_wake - Invoked by RDMA provider for a LOCAL_INV WC
477 * @cq: completion queue (ignored)
480 * Awaken anyone waiting for an MR to finish being fenced.
482 static void frwr_wc_localinv_wake(struct ib_cq *cq, struct ib_wc *wc)
484 struct ib_cqe *cqe = wc->wr_cqe;
485 struct rpcrdma_frwr *frwr =
486 container_of(cqe, struct rpcrdma_frwr, fr_cqe);
487 struct rpcrdma_mr *mr = container_of(frwr, struct rpcrdma_mr, frwr);
489 /* WARNING: Only wr_cqe and status are reliable at this point */
490 trace_xprtrdma_wc_li_wake(wc, frwr);
491 __frwr_release_mr(wc, mr);
492 complete(&frwr->fr_linv_done);
496 * frwr_unmap_sync - invalidate memory regions that were registered for @req
497 * @r_xprt: controlling transport instance
498 * @req: rpcrdma_req with a non-empty list of MRs to process
500 * Sleeps until it is safe for the host CPU to access the previously mapped
501 * memory regions. This guarantees that registered MRs are properly fenced
502 * from the server before the RPC consumer accesses the data in them. It
503 * also ensures proper Send flow control: waking the next RPC waits until
504 * this RPC has relinquished all its Send Queue entries.
506 void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
508 struct ib_send_wr *first, **prev, *last;
509 const struct ib_send_wr *bad_wr;
510 struct rpcrdma_frwr *frwr;
511 struct rpcrdma_mr *mr;
514 /* ORDER: Invalidate all of the MRs first
516 * Chain the LOCAL_INV Work Requests and post them with
517 * a single ib_post_send() call.
521 while ((mr = rpcrdma_mr_pop(&req->rl_registered))) {
523 trace_xprtrdma_mr_localinv(mr);
524 r_xprt->rx_stats.local_inv_needed++;
527 frwr->fr_cqe.done = frwr_wc_localinv;
528 last = &frwr->fr_invwr;
530 last->wr_cqe = &frwr->fr_cqe;
531 last->sg_list = NULL;
533 last->opcode = IB_WR_LOCAL_INV;
534 last->send_flags = IB_SEND_SIGNALED;
535 last->ex.invalidate_rkey = mr->mr_handle;
541 /* Strong send queue ordering guarantees that when the
542 * last WR in the chain completes, all WRs in the chain
545 frwr->fr_cqe.done = frwr_wc_localinv_wake;
546 reinit_completion(&frwr->fr_linv_done);
548 /* Transport disconnect drains the receive CQ before it
549 * replaces the QP. The RPC reply handler won't call us
550 * unless ri_id->qp is a valid pointer.
553 rc = ib_post_send(r_xprt->rx_ia.ri_id->qp, first, &bad_wr);
554 trace_xprtrdma_post_send(req, rc);
556 /* The final LOCAL_INV WR in the chain is supposed to
557 * do the wake. If it was never posted, the wake will
558 * not happen, so don't wait in that case.
561 wait_for_completion(&frwr->fr_linv_done);
565 /* Recycle MRs in the LOCAL_INV chain that did not get posted.
568 frwr = container_of(bad_wr, struct rpcrdma_frwr,
570 mr = container_of(frwr, struct rpcrdma_mr, frwr);
571 bad_wr = bad_wr->next;
573 list_del_init(&mr->mr_list);
574 rpcrdma_mr_recycle(mr);
579 * frwr_wc_localinv_done - Invoked by RDMA provider for a signaled LOCAL_INV WC
580 * @cq: completion queue (ignored)
584 static void frwr_wc_localinv_done(struct ib_cq *cq, struct ib_wc *wc)
586 struct ib_cqe *cqe = wc->wr_cqe;
587 struct rpcrdma_frwr *frwr =
588 container_of(cqe, struct rpcrdma_frwr, fr_cqe);
589 struct rpcrdma_mr *mr = container_of(frwr, struct rpcrdma_mr, frwr);
590 struct rpcrdma_rep *rep = mr->mr_req->rl_reply;
592 /* WARNING: Only wr_cqe and status are reliable at this point */
593 trace_xprtrdma_wc_li_done(wc, frwr);
594 __frwr_release_mr(wc, mr);
596 /* Ensure @rep is generated before __frwr_release_mr */
598 rpcrdma_complete_rqst(rep);
602 * frwr_unmap_async - invalidate memory regions that were registered for @req
603 * @r_xprt: controlling transport instance
604 * @req: rpcrdma_req with a non-empty list of MRs to process
606 * This guarantees that registered MRs are properly fenced from the
607 * server before the RPC consumer accesses the data in them. It also
608 * ensures proper Send flow control: waking the next RPC waits until
609 * this RPC has relinquished all its Send Queue entries.
611 void frwr_unmap_async(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
613 struct ib_send_wr *first, *last, **prev;
614 const struct ib_send_wr *bad_wr;
615 struct rpcrdma_frwr *frwr;
616 struct rpcrdma_mr *mr;
619 /* Chain the LOCAL_INV Work Requests and post them with
620 * a single ib_post_send() call.
624 while ((mr = rpcrdma_mr_pop(&req->rl_registered))) {
626 trace_xprtrdma_mr_localinv(mr);
627 r_xprt->rx_stats.local_inv_needed++;
630 frwr->fr_cqe.done = frwr_wc_localinv;
631 last = &frwr->fr_invwr;
633 last->wr_cqe = &frwr->fr_cqe;
634 last->sg_list = NULL;
636 last->opcode = IB_WR_LOCAL_INV;
637 last->send_flags = IB_SEND_SIGNALED;
638 last->ex.invalidate_rkey = mr->mr_handle;
644 /* Strong send queue ordering guarantees that when the
645 * last WR in the chain completes, all WRs in the chain
646 * are complete. The last completion will wake up the
649 frwr->fr_cqe.done = frwr_wc_localinv_done;
651 /* Transport disconnect drains the receive CQ before it
652 * replaces the QP. The RPC reply handler won't call us
653 * unless ri_id->qp is a valid pointer.
656 rc = ib_post_send(r_xprt->rx_ia.ri_id->qp, first, &bad_wr);
657 trace_xprtrdma_post_send(req, rc);
661 /* Recycle MRs in the LOCAL_INV chain that did not get posted.
664 frwr = container_of(bad_wr, struct rpcrdma_frwr, fr_invwr);
665 mr = container_of(frwr, struct rpcrdma_mr, frwr);
666 bad_wr = bad_wr->next;
668 rpcrdma_mr_recycle(mr);
671 /* The final LOCAL_INV WR in the chain is supposed to
672 * do the wake. If it was never posted, the wake will
673 * not happen, so wake here in that case.
675 rpcrdma_complete_rqst(req->rl_reply);