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 Registration Work
7 * Requests (FRWR). Also referred to sometimes as FRMR mode.
9 * FRWR features ordered asynchronous registration and deregistration
10 * of arbitrarily sized memory regions. This is the fastest and safest
11 * but most complex memory registration mode.
16 * A Memory Region is prepared for RDMA READ or WRITE using a FAST_REG
17 * Work Request (frmr_op_map). When the RDMA operation is finished, this
18 * Memory Region is invalidated using a LOCAL_INV Work Request
21 * Typically these Work Requests are not signaled, and neither are RDMA
22 * SEND Work Requests (with the exception of signaling occasionally to
23 * prevent provider work queue overflows). This greatly reduces HCA
26 * As an optimization, frwr_op_unmap marks MRs INVALID before the
27 * LOCAL_INV WR is posted. If posting succeeds, the MR is placed on
28 * rb_mws immediately so that no work (like managing a linked list
29 * under a spinlock) is needed in the completion upcall.
31 * But this means that frwr_op_map() can occasionally encounter an MR
32 * that is INVALID but the LOCAL_INV WR has not completed. Work Queue
33 * ordering prevents a subsequent FAST_REG WR from executing against
34 * that MR while it is still being invalidated.
39 * ->op_map and the transport connect worker cannot run at the same
40 * time, but ->op_unmap can fire while the transport connect worker
41 * is running. Thus MR recovery is handled in ->op_map, to guarantee
42 * that recovered MRs are owned by a sending RPC, and not one where
43 * ->op_unmap could fire at the same time transport reconnect is
46 * When the underlying transport disconnects, MRs are left in one of
49 * INVALID: The MR was not in use before the QP entered ERROR state.
50 * (Or, the LOCAL_INV WR has not completed or flushed yet).
52 * STALE: The MR was being registered or unregistered when the QP
53 * entered ERROR state, and the pending WR was flushed.
55 * VALID: The MR was registered before the QP entered ERROR state.
57 * When frwr_op_map encounters STALE and VALID MRs, they are recovered
58 * with ib_dereg_mr and then are re-initialized. Beause MR recovery
59 * allocates fresh resources, it is deferred to a workqueue, and the
60 * recovered MRs are placed back on the rb_mws list when recovery is
61 * complete. frwr_op_map allocates another MR for the current RPC while
62 * the broken MR is reset.
64 * To ensure that frwr_op_map doesn't encounter an MR that is marked
65 * INVALID but that is about to be flushed due to a previous transport
66 * disconnect, the transport connect worker attempts to drain all
67 * pending send queue WRs before the transport is reconnected.
70 #include <linux/sunrpc/rpc_rdma.h>
72 #include "xprt_rdma.h"
74 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
75 # define RPCDBG_FACILITY RPCDBG_TRANS
79 frwr_is_supported(struct rpcrdma_ia *ia)
81 struct ib_device_attr *attrs = &ia->ri_device->attrs;
83 if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
84 goto out_not_supported;
85 if (attrs->max_fast_reg_page_list_len == 0)
86 goto out_not_supported;
90 pr_info("rpcrdma: 'frwr' mode is not supported by device %s\n",
96 frwr_op_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mw *r)
98 unsigned int depth = ia->ri_max_frmr_depth;
99 struct rpcrdma_frmr *f = &r->frmr;
102 f->fr_mr = ib_alloc_mr(ia->ri_pd, IB_MR_TYPE_MEM_REG, depth);
103 if (IS_ERR(f->fr_mr))
106 r->mw_sg = kcalloc(depth, sizeof(*r->mw_sg), GFP_KERNEL);
110 sg_init_table(r->mw_sg, depth);
111 init_completion(&f->fr_linv_done);
115 rc = PTR_ERR(f->fr_mr);
116 dprintk("RPC: %s: ib_alloc_mr status %i\n",
122 dprintk("RPC: %s: sg allocation failure\n",
124 ib_dereg_mr(f->fr_mr);
129 frwr_op_release_mr(struct rpcrdma_mw *r)
133 /* Ensure MW is not on any rl_registered list */
134 if (!list_empty(&r->mw_list))
135 list_del(&r->mw_list);
137 rc = ib_dereg_mr(r->frmr.fr_mr);
139 pr_err("rpcrdma: final ib_dereg_mr for %p returned %i\n",
146 __frwr_reset_mr(struct rpcrdma_ia *ia, struct rpcrdma_mw *r)
148 struct rpcrdma_frmr *f = &r->frmr;
151 rc = ib_dereg_mr(f->fr_mr);
153 pr_warn("rpcrdma: ib_dereg_mr status %d, frwr %p orphaned\n",
158 f->fr_mr = ib_alloc_mr(ia->ri_pd, IB_MR_TYPE_MEM_REG,
159 ia->ri_max_frmr_depth);
160 if (IS_ERR(f->fr_mr)) {
161 pr_warn("rpcrdma: ib_alloc_mr status %ld, frwr %p orphaned\n",
162 PTR_ERR(f->fr_mr), r);
163 return PTR_ERR(f->fr_mr);
166 dprintk("RPC: %s: recovered FRMR %p\n", __func__, f);
167 f->fr_state = FRMR_IS_INVALID;
171 /* Reset of a single FRMR. Generate a fresh rkey by replacing the MR.
173 * There's no recovery if this fails. The FRMR is abandoned, but
174 * remains in rb_all. It will be cleaned up when the transport is
178 frwr_op_recover_mr(struct rpcrdma_mw *mw)
180 struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
181 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
184 rc = __frwr_reset_mr(ia, mw);
185 ib_dma_unmap_sg(ia->ri_device, mw->mw_sg, mw->mw_nents, mw->mw_dir);
189 rpcrdma_put_mw(r_xprt, mw);
190 r_xprt->rx_stats.mrs_recovered++;
194 pr_err("rpcrdma: FRMR reset failed %d, %p release\n", rc, mw);
195 r_xprt->rx_stats.mrs_orphaned++;
197 spin_lock(&r_xprt->rx_buf.rb_mwlock);
198 list_del(&mw->mw_all);
199 spin_unlock(&r_xprt->rx_buf.rb_mwlock);
201 frwr_op_release_mr(mw);
205 frwr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
206 struct rpcrdma_create_data_internal *cdata)
210 ia->ri_max_frmr_depth =
211 min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
212 ia->ri_device->attrs.max_fast_reg_page_list_len);
213 dprintk("RPC: %s: device's max FR page list len = %u\n",
214 __func__, ia->ri_max_frmr_depth);
216 /* Add room for frmr register and invalidate WRs.
217 * 1. FRMR reg WR for head
218 * 2. FRMR invalidate WR for head
219 * 3. N FRMR reg WRs for pagelist
220 * 4. N FRMR invalidate WRs for pagelist
221 * 5. FRMR reg WR for tail
222 * 6. FRMR invalidate WR for tail
223 * 7. The RDMA_SEND WR
227 /* Calculate N if the device max FRMR depth is smaller than
228 * RPCRDMA_MAX_DATA_SEGS.
230 if (ia->ri_max_frmr_depth < RPCRDMA_MAX_DATA_SEGS) {
231 delta = RPCRDMA_MAX_DATA_SEGS - ia->ri_max_frmr_depth;
233 depth += 2; /* FRMR reg + invalidate */
234 delta -= ia->ri_max_frmr_depth;
238 ep->rep_attr.cap.max_send_wr *= depth;
239 if (ep->rep_attr.cap.max_send_wr > ia->ri_device->attrs.max_qp_wr) {
240 cdata->max_requests = ia->ri_device->attrs.max_qp_wr / depth;
241 if (!cdata->max_requests)
243 ep->rep_attr.cap.max_send_wr = cdata->max_requests *
247 ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
248 ia->ri_max_frmr_depth);
252 /* FRWR mode conveys a list of pages per chunk segment. The
253 * maximum length of that list is the FRWR page list depth.
256 frwr_op_maxpages(struct rpcrdma_xprt *r_xprt)
258 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
260 return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
261 RPCRDMA_MAX_HDR_SEGS * ia->ri_max_frmr_depth);
265 __frwr_sendcompletion_flush(struct ib_wc *wc, struct rpcrdma_frmr *frmr,
268 frmr->fr_state = FRMR_IS_STALE;
269 if (wc->status != IB_WC_WR_FLUSH_ERR)
270 pr_err("rpcrdma: %s: %s (%u/0x%x)\n",
271 wr, ib_wc_status_msg(wc->status),
272 wc->status, wc->vendor_err);
276 * frwr_wc_fastreg - Invoked by RDMA provider for each polled FastReg WC
277 * @cq: completion queue (ignored)
282 frwr_wc_fastreg(struct ib_cq *cq, struct ib_wc *wc)
284 struct rpcrdma_frmr *frmr;
287 /* WARNING: Only wr_cqe and status are reliable at this point */
288 if (wc->status != IB_WC_SUCCESS) {
290 frmr = container_of(cqe, struct rpcrdma_frmr, fr_cqe);
291 __frwr_sendcompletion_flush(wc, frmr, "fastreg");
296 * frwr_wc_localinv - Invoked by RDMA provider for each polled LocalInv WC
297 * @cq: completion queue (ignored)
302 frwr_wc_localinv(struct ib_cq *cq, struct ib_wc *wc)
304 struct rpcrdma_frmr *frmr;
307 /* WARNING: Only wr_cqe and status are reliable at this point */
308 if (wc->status != IB_WC_SUCCESS) {
310 frmr = container_of(cqe, struct rpcrdma_frmr, fr_cqe);
311 __frwr_sendcompletion_flush(wc, frmr, "localinv");
316 * frwr_wc_localinv - Invoked by RDMA provider for each polled LocalInv WC
317 * @cq: completion queue (ignored)
320 * Awaken anyone waiting for an MR to finish being fenced.
323 frwr_wc_localinv_wake(struct ib_cq *cq, struct ib_wc *wc)
325 struct rpcrdma_frmr *frmr;
328 /* WARNING: Only wr_cqe and status are reliable at this point */
330 frmr = container_of(cqe, struct rpcrdma_frmr, fr_cqe);
331 if (wc->status != IB_WC_SUCCESS)
332 __frwr_sendcompletion_flush(wc, frmr, "localinv");
333 complete(&frmr->fr_linv_done);
336 /* Post a REG_MR Work Request to register a memory region
337 * for remote access via RDMA READ or RDMA WRITE.
340 frwr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
341 int nsegs, bool writing, struct rpcrdma_mw **out)
343 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
344 struct rpcrdma_mw *mw;
345 struct rpcrdma_frmr *frmr;
347 struct ib_reg_wr *reg_wr;
348 struct ib_send_wr *bad_wr;
349 int rc, i, n, dma_nents;
355 rpcrdma_defer_mr_recovery(mw);
356 mw = rpcrdma_get_mw(r_xprt);
359 } while (mw->frmr.fr_state != FRMR_IS_INVALID);
361 frmr->fr_state = FRMR_IS_VALID;
363 reg_wr = &frmr->fr_regwr;
365 if (nsegs > ia->ri_max_frmr_depth)
366 nsegs = ia->ri_max_frmr_depth;
367 for (i = 0; i < nsegs;) {
369 sg_set_page(&mw->mw_sg[i],
372 offset_in_page(seg->mr_offset));
374 sg_set_buf(&mw->mw_sg[i], seg->mr_offset,
380 /* Check for holes */
381 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
382 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
386 mw->mw_dir = rpcrdma_data_dir(writing);
390 dma_nents = ib_dma_map_sg(ia->ri_device,
391 mw->mw_sg, mw->mw_nents, mw->mw_dir);
395 n = ib_map_mr_sg(mr, mw->mw_sg, mw->mw_nents, NULL, PAGE_SIZE);
396 if (unlikely(n != mw->mw_nents))
399 dprintk("RPC: %s: Using frmr %p to map %u segments (%u bytes)\n",
400 __func__, frmr, mw->mw_nents, mr->length);
402 key = (u8)(mr->rkey & 0x000000FF);
403 ib_update_fast_reg_key(mr, ++key);
405 reg_wr->wr.next = NULL;
406 reg_wr->wr.opcode = IB_WR_REG_MR;
407 frmr->fr_cqe.done = frwr_wc_fastreg;
408 reg_wr->wr.wr_cqe = &frmr->fr_cqe;
409 reg_wr->wr.num_sge = 0;
410 reg_wr->wr.send_flags = 0;
412 reg_wr->key = mr->rkey;
413 reg_wr->access = writing ?
414 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
415 IB_ACCESS_REMOTE_READ;
417 DECR_CQCOUNT(&r_xprt->rx_ep);
418 rc = ib_post_send(ia->ri_id->qp, ®_wr->wr, &bad_wr);
422 mw->mw_handle = mr->rkey;
423 mw->mw_length = mr->length;
424 mw->mw_offset = mr->iova;
430 pr_err("rpcrdma: failed to dma map sg %p sg_nents %u\n",
431 mw->mw_sg, mw->mw_nents);
432 rpcrdma_defer_mr_recovery(mw);
436 pr_err("rpcrdma: failed to map mr %p (%u/%u)\n",
437 frmr->fr_mr, n, mw->mw_nents);
438 rpcrdma_defer_mr_recovery(mw);
442 pr_err("rpcrdma: FRMR registration ib_post_send returned %i\n", rc);
443 rpcrdma_defer_mr_recovery(mw);
447 static struct ib_send_wr *
448 __frwr_prepare_linv_wr(struct rpcrdma_mw *mw)
450 struct rpcrdma_frmr *f = &mw->frmr;
451 struct ib_send_wr *invalidate_wr;
453 dprintk("RPC: %s: invalidating frmr %p\n", __func__, f);
455 f->fr_state = FRMR_IS_INVALID;
456 invalidate_wr = &f->fr_invwr;
458 memset(invalidate_wr, 0, sizeof(*invalidate_wr));
459 f->fr_cqe.done = frwr_wc_localinv;
460 invalidate_wr->wr_cqe = &f->fr_cqe;
461 invalidate_wr->opcode = IB_WR_LOCAL_INV;
462 invalidate_wr->ex.invalidate_rkey = f->fr_mr->rkey;
464 return invalidate_wr;
467 /* Invalidate all memory regions that were registered for "req".
469 * Sleeps until it is safe for the host CPU to access the
470 * previously mapped memory regions.
472 * Caller ensures that req->rl_registered is not empty.
475 frwr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
477 struct ib_send_wr *invalidate_wrs, *pos, *prev, *bad_wr;
478 struct rpcrdma_rep *rep = req->rl_reply;
479 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
480 struct rpcrdma_mw *mw, *tmp;
481 struct rpcrdma_frmr *f;
484 dprintk("RPC: %s: req %p\n", __func__, req);
486 /* ORDER: Invalidate all of the req's MRs first
488 * Chain the LOCAL_INV Work Requests and post them with
489 * a single ib_post_send() call.
492 invalidate_wrs = pos = prev = NULL;
493 list_for_each_entry(mw, &req->rl_registered, mw_list) {
494 if ((rep->rr_wc_flags & IB_WC_WITH_INVALIDATE) &&
495 (mw->mw_handle == rep->rr_inv_rkey)) {
496 mw->frmr.fr_state = FRMR_IS_INVALID;
500 pos = __frwr_prepare_linv_wr(mw);
503 invalidate_wrs = pos;
512 /* Strong send queue ordering guarantees that when the
513 * last WR in the chain completes, all WRs in the chain
516 f->fr_invwr.send_flags = IB_SEND_SIGNALED;
517 f->fr_cqe.done = frwr_wc_localinv_wake;
518 reinit_completion(&f->fr_linv_done);
519 INIT_CQCOUNT(&r_xprt->rx_ep);
521 /* Transport disconnect drains the receive CQ before it
522 * replaces the QP. The RPC reply handler won't call us
523 * unless ri_id->qp is a valid pointer.
525 r_xprt->rx_stats.local_inv_needed++;
526 rc = ib_post_send(ia->ri_id->qp, invalidate_wrs, &bad_wr);
530 wait_for_completion(&f->fr_linv_done);
532 /* ORDER: Now DMA unmap all of the req's MRs, and return
533 * them to the free MW list.
536 list_for_each_entry_safe(mw, tmp, &req->rl_registered, mw_list) {
537 dprintk("RPC: %s: unmapping frmr %p\n",
538 __func__, &mw->frmr);
539 list_del_init(&mw->mw_list);
540 ib_dma_unmap_sg(ia->ri_device,
541 mw->mw_sg, mw->mw_nents, mw->mw_dir);
542 rpcrdma_put_mw(r_xprt, mw);
547 pr_err("rpcrdma: FRMR invalidate ib_post_send returned %i\n", rc);
548 rdma_disconnect(ia->ri_id);
550 /* Find and reset the MRs in the LOCAL_INV WRs that did not
551 * get posted. This is synchronous, and slow.
553 list_for_each_entry(mw, &req->rl_registered, mw_list) {
555 if (mw->frmr.fr_mr->rkey == bad_wr->ex.invalidate_rkey) {
556 __frwr_reset_mr(ia, mw);
557 bad_wr = bad_wr->next;
563 /* Use a slow, safe mechanism to invalidate all memory regions
564 * that were registered for "req".
567 frwr_op_unmap_safe(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
570 struct rpcrdma_mw *mw;
572 while (!list_empty(&req->rl_registered)) {
573 mw = list_first_entry(&req->rl_registered,
574 struct rpcrdma_mw, mw_list);
575 list_del_init(&mw->mw_list);
578 frwr_op_recover_mr(mw);
580 rpcrdma_defer_mr_recovery(mw);
584 const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops = {
585 .ro_map = frwr_op_map,
586 .ro_unmap_sync = frwr_op_unmap_sync,
587 .ro_unmap_safe = frwr_op_unmap_safe,
588 .ro_recover_mr = frwr_op_recover_mr,
589 .ro_open = frwr_op_open,
590 .ro_maxpages = frwr_op_maxpages,
591 .ro_init_mr = frwr_op_init_mr,
592 .ro_release_mr = frwr_op_release_mr,
593 .ro_displayname = "frwr",
594 .ro_send_w_inv_ok = RPCRDMA_CMP_F_SND_W_INV_OK,