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.
51 * VALID: The MR was registered before the QP entered ERROR state.
53 * FLUSHED_FR: The MR was being registered when the QP entered ERROR
54 * state, and the pending WR was flushed.
56 * FLUSHED_LI: The MR was being invalidated when the QP entered ERROR
57 * state, and the pending WR was flushed.
59 * When frwr_op_map encounters FLUSHED and VALID MRs, they are recovered
60 * with ib_dereg_mr and then are re-initialized. Because MR recovery
61 * allocates fresh resources, it is deferred to a workqueue, and the
62 * recovered MRs are placed back on the rb_mws list when recovery is
63 * complete. frwr_op_map allocates another MR for the current RPC while
64 * the broken MR is reset.
66 * To ensure that frwr_op_map doesn't encounter an MR that is marked
67 * INVALID but that is about to be flushed due to a previous transport
68 * disconnect, the transport connect worker attempts to drain all
69 * pending send queue WRs before the transport is reconnected.
72 #include <linux/sunrpc/rpc_rdma.h>
74 #include "xprt_rdma.h"
76 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
77 # define RPCDBG_FACILITY RPCDBG_TRANS
81 frwr_is_supported(struct rpcrdma_ia *ia)
83 struct ib_device_attr *attrs = &ia->ri_device->attrs;
85 if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
86 goto out_not_supported;
87 if (attrs->max_fast_reg_page_list_len == 0)
88 goto out_not_supported;
92 pr_info("rpcrdma: 'frwr' mode is not supported by device %s\n",
98 frwr_op_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mw *r)
100 unsigned int depth = ia->ri_max_frmr_depth;
101 struct rpcrdma_frmr *f = &r->frmr;
104 f->fr_mr = ib_alloc_mr(ia->ri_pd, ia->ri_mrtype, depth);
105 if (IS_ERR(f->fr_mr))
108 r->mw_sg = kcalloc(depth, sizeof(*r->mw_sg), GFP_KERNEL);
112 sg_init_table(r->mw_sg, depth);
113 init_completion(&f->fr_linv_done);
117 rc = PTR_ERR(f->fr_mr);
118 dprintk("RPC: %s: ib_alloc_mr status %i\n",
124 dprintk("RPC: %s: sg allocation failure\n",
126 ib_dereg_mr(f->fr_mr);
131 frwr_op_release_mr(struct rpcrdma_mw *r)
135 /* Ensure MW is not on any rl_registered list */
136 if (!list_empty(&r->mw_list))
137 list_del(&r->mw_list);
139 rc = ib_dereg_mr(r->frmr.fr_mr);
141 pr_err("rpcrdma: final ib_dereg_mr for %p returned %i\n",
148 __frwr_reset_mr(struct rpcrdma_ia *ia, struct rpcrdma_mw *r)
150 struct rpcrdma_frmr *f = &r->frmr;
153 rc = ib_dereg_mr(f->fr_mr);
155 pr_warn("rpcrdma: ib_dereg_mr status %d, frwr %p orphaned\n",
160 f->fr_mr = ib_alloc_mr(ia->ri_pd, ia->ri_mrtype,
161 ia->ri_max_frmr_depth);
162 if (IS_ERR(f->fr_mr)) {
163 pr_warn("rpcrdma: ib_alloc_mr status %ld, frwr %p orphaned\n",
164 PTR_ERR(f->fr_mr), r);
165 return PTR_ERR(f->fr_mr);
168 dprintk("RPC: %s: recovered FRMR %p\n", __func__, f);
169 f->fr_state = FRMR_IS_INVALID;
173 /* Reset of a single FRMR. Generate a fresh rkey by replacing the MR.
175 * There's no recovery if this fails. The FRMR is abandoned, but
176 * remains in rb_all. It will be cleaned up when the transport is
180 frwr_op_recover_mr(struct rpcrdma_mw *mw)
182 enum rpcrdma_frmr_state state = mw->frmr.fr_state;
183 struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
184 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
187 rc = __frwr_reset_mr(ia, mw);
188 if (state != FRMR_FLUSHED_LI)
189 ib_dma_unmap_sg(ia->ri_device,
190 mw->mw_sg, mw->mw_nents, mw->mw_dir);
194 rpcrdma_put_mw(r_xprt, mw);
195 r_xprt->rx_stats.mrs_recovered++;
199 pr_err("rpcrdma: FRMR reset failed %d, %p release\n", rc, mw);
200 r_xprt->rx_stats.mrs_orphaned++;
202 spin_lock(&r_xprt->rx_buf.rb_mwlock);
203 list_del(&mw->mw_all);
204 spin_unlock(&r_xprt->rx_buf.rb_mwlock);
206 frwr_op_release_mr(mw);
210 frwr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
211 struct rpcrdma_create_data_internal *cdata)
213 struct ib_device_attr *attrs = &ia->ri_device->attrs;
216 ia->ri_mrtype = IB_MR_TYPE_MEM_REG;
217 if (attrs->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
218 ia->ri_mrtype = IB_MR_TYPE_SG_GAPS;
220 ia->ri_max_frmr_depth =
221 min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
222 attrs->max_fast_reg_page_list_len);
223 dprintk("RPC: %s: device's max FR page list len = %u\n",
224 __func__, ia->ri_max_frmr_depth);
226 /* Add room for frmr register and invalidate WRs.
227 * 1. FRMR reg WR for head
228 * 2. FRMR invalidate WR for head
229 * 3. N FRMR reg WRs for pagelist
230 * 4. N FRMR invalidate WRs for pagelist
231 * 5. FRMR reg WR for tail
232 * 6. FRMR invalidate WR for tail
233 * 7. The RDMA_SEND WR
237 /* Calculate N if the device max FRMR depth is smaller than
238 * RPCRDMA_MAX_DATA_SEGS.
240 if (ia->ri_max_frmr_depth < RPCRDMA_MAX_DATA_SEGS) {
241 delta = RPCRDMA_MAX_DATA_SEGS - ia->ri_max_frmr_depth;
243 depth += 2; /* FRMR reg + invalidate */
244 delta -= ia->ri_max_frmr_depth;
248 ep->rep_attr.cap.max_send_wr *= depth;
249 if (ep->rep_attr.cap.max_send_wr > attrs->max_qp_wr) {
250 cdata->max_requests = attrs->max_qp_wr / depth;
251 if (!cdata->max_requests)
253 ep->rep_attr.cap.max_send_wr = cdata->max_requests *
257 ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
258 ia->ri_max_frmr_depth);
262 /* FRWR mode conveys a list of pages per chunk segment. The
263 * maximum length of that list is the FRWR page list depth.
266 frwr_op_maxpages(struct rpcrdma_xprt *r_xprt)
268 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
270 return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
271 RPCRDMA_MAX_HDR_SEGS * ia->ri_max_frmr_depth);
275 __frwr_sendcompletion_flush(struct ib_wc *wc, const char *wr)
277 if (wc->status != IB_WC_WR_FLUSH_ERR)
278 pr_err("rpcrdma: %s: %s (%u/0x%x)\n",
279 wr, ib_wc_status_msg(wc->status),
280 wc->status, wc->vendor_err);
284 * frwr_wc_fastreg - Invoked by RDMA provider for each polled FastReg WC
285 * @cq: completion queue (ignored)
290 frwr_wc_fastreg(struct ib_cq *cq, struct ib_wc *wc)
292 struct rpcrdma_frmr *frmr;
295 /* WARNING: Only wr_cqe and status are reliable at this point */
296 if (wc->status != IB_WC_SUCCESS) {
298 frmr = container_of(cqe, struct rpcrdma_frmr, fr_cqe);
299 frmr->fr_state = FRMR_FLUSHED_FR;
300 __frwr_sendcompletion_flush(wc, "fastreg");
305 * frwr_wc_localinv - Invoked by RDMA provider for each polled LocalInv WC
306 * @cq: completion queue (ignored)
311 frwr_wc_localinv(struct ib_cq *cq, struct ib_wc *wc)
313 struct rpcrdma_frmr *frmr;
316 /* WARNING: Only wr_cqe and status are reliable at this point */
317 if (wc->status != IB_WC_SUCCESS) {
319 frmr = container_of(cqe, struct rpcrdma_frmr, fr_cqe);
320 frmr->fr_state = FRMR_FLUSHED_LI;
321 __frwr_sendcompletion_flush(wc, "localinv");
326 * frwr_wc_localinv - Invoked by RDMA provider for each polled LocalInv WC
327 * @cq: completion queue (ignored)
330 * Awaken anyone waiting for an MR to finish being fenced.
333 frwr_wc_localinv_wake(struct ib_cq *cq, struct ib_wc *wc)
335 struct rpcrdma_frmr *frmr;
338 /* WARNING: Only wr_cqe and status are reliable at this point */
340 frmr = container_of(cqe, struct rpcrdma_frmr, fr_cqe);
341 if (wc->status != IB_WC_SUCCESS) {
342 frmr->fr_state = FRMR_FLUSHED_LI;
343 __frwr_sendcompletion_flush(wc, "localinv");
345 complete(&frmr->fr_linv_done);
348 /* Post a REG_MR Work Request to register a memory region
349 * for remote access via RDMA READ or RDMA WRITE.
352 frwr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
353 int nsegs, bool writing, struct rpcrdma_mw **out)
355 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
356 bool holes_ok = ia->ri_mrtype == IB_MR_TYPE_SG_GAPS;
357 struct rpcrdma_mw *mw;
358 struct rpcrdma_frmr *frmr;
360 struct ib_reg_wr *reg_wr;
361 struct ib_send_wr *bad_wr;
362 int rc, i, n, dma_nents;
368 rpcrdma_defer_mr_recovery(mw);
369 mw = rpcrdma_get_mw(r_xprt);
372 } while (mw->frmr.fr_state != FRMR_IS_INVALID);
374 frmr->fr_state = FRMR_IS_VALID;
376 reg_wr = &frmr->fr_regwr;
378 if (nsegs > ia->ri_max_frmr_depth)
379 nsegs = ia->ri_max_frmr_depth;
380 for (i = 0; i < nsegs;) {
382 sg_set_page(&mw->mw_sg[i],
385 offset_in_page(seg->mr_offset));
387 sg_set_buf(&mw->mw_sg[i], seg->mr_offset,
394 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
395 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
399 mw->mw_dir = rpcrdma_data_dir(writing);
403 dma_nents = ib_dma_map_sg(ia->ri_device,
404 mw->mw_sg, mw->mw_nents, mw->mw_dir);
408 n = ib_map_mr_sg(mr, mw->mw_sg, mw->mw_nents, NULL, PAGE_SIZE);
409 if (unlikely(n != mw->mw_nents))
412 dprintk("RPC: %s: Using frmr %p to map %u segments (%u bytes)\n",
413 __func__, frmr, mw->mw_nents, mr->length);
415 key = (u8)(mr->rkey & 0x000000FF);
416 ib_update_fast_reg_key(mr, ++key);
418 reg_wr->wr.next = NULL;
419 reg_wr->wr.opcode = IB_WR_REG_MR;
420 frmr->fr_cqe.done = frwr_wc_fastreg;
421 reg_wr->wr.wr_cqe = &frmr->fr_cqe;
422 reg_wr->wr.num_sge = 0;
423 reg_wr->wr.send_flags = 0;
425 reg_wr->key = mr->rkey;
426 reg_wr->access = writing ?
427 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
428 IB_ACCESS_REMOTE_READ;
430 rpcrdma_set_signaled(&r_xprt->rx_ep, ®_wr->wr);
431 rc = ib_post_send(ia->ri_id->qp, ®_wr->wr, &bad_wr);
435 mw->mw_handle = mr->rkey;
436 mw->mw_length = mr->length;
437 mw->mw_offset = mr->iova;
443 pr_err("rpcrdma: failed to dma map sg %p sg_nents %u\n",
444 mw->mw_sg, mw->mw_nents);
445 rpcrdma_defer_mr_recovery(mw);
449 pr_err("rpcrdma: failed to map mr %p (%u/%u)\n",
450 frmr->fr_mr, n, mw->mw_nents);
451 rpcrdma_defer_mr_recovery(mw);
455 pr_err("rpcrdma: FRMR registration ib_post_send returned %i\n", rc);
456 rpcrdma_defer_mr_recovery(mw);
460 /* Invalidate all memory regions that were registered for "req".
462 * Sleeps until it is safe for the host CPU to access the
463 * previously mapped memory regions.
465 * Caller ensures that req->rl_registered is not empty.
468 frwr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
470 struct ib_send_wr *first, **prev, *last, *bad_wr;
471 struct rpcrdma_rep *rep = req->rl_reply;
472 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
473 struct rpcrdma_mw *mw, *tmp;
474 struct rpcrdma_frmr *f;
477 dprintk("RPC: %s: req %p\n", __func__, req);
479 /* ORDER: Invalidate all of the req's MRs first
481 * Chain the LOCAL_INV Work Requests and post them with
482 * a single ib_post_send() call.
487 list_for_each_entry(mw, &req->rl_registered, mw_list) {
488 mw->frmr.fr_state = FRMR_IS_INVALID;
490 if ((rep->rr_wc_flags & IB_WC_WITH_INVALIDATE) &&
491 (mw->mw_handle == rep->rr_inv_rkey))
495 dprintk("RPC: %s: invalidating frmr %p\n",
498 f->fr_cqe.done = frwr_wc_localinv;
500 memset(last, 0, sizeof(*last));
501 last->wr_cqe = &f->fr_cqe;
502 last->opcode = IB_WR_LOCAL_INV;
503 last->ex.invalidate_rkey = mw->mw_handle;
512 /* Strong send queue ordering guarantees that when the
513 * last WR in the chain completes, all WRs in the chain
516 last->send_flags = IB_SEND_SIGNALED;
517 f->fr_cqe.done = frwr_wc_localinv_wake;
518 reinit_completion(&f->fr_linv_done);
520 /* Initialize CQ count, since there is always a signaled
521 * WR being posted here. The new cqcount depends on how
522 * many SQEs are about to be consumed.
524 rpcrdma_init_cqcount(&r_xprt->rx_ep, count);
526 /* Transport disconnect drains the receive CQ before it
527 * replaces the QP. The RPC reply handler won't call us
528 * unless ri_id->qp is a valid pointer.
530 r_xprt->rx_stats.local_inv_needed++;
531 rc = ib_post_send(ia->ri_id->qp, first, &bad_wr);
535 wait_for_completion(&f->fr_linv_done);
537 /* ORDER: Now DMA unmap all of the req's MRs, and return
538 * them to the free MW list.
541 list_for_each_entry_safe(mw, tmp, &req->rl_registered, mw_list) {
542 dprintk("RPC: %s: DMA unmapping frmr %p\n",
543 __func__, &mw->frmr);
544 list_del_init(&mw->mw_list);
545 ib_dma_unmap_sg(ia->ri_device,
546 mw->mw_sg, mw->mw_nents, mw->mw_dir);
547 rpcrdma_put_mw(r_xprt, mw);
552 pr_err("rpcrdma: FRMR invalidate ib_post_send returned %i\n", rc);
553 rdma_disconnect(ia->ri_id);
555 /* Find and reset the MRs in the LOCAL_INV WRs that did not
556 * get posted. This is synchronous, and slow.
558 list_for_each_entry(mw, &req->rl_registered, mw_list) {
560 if (mw->mw_handle == bad_wr->ex.invalidate_rkey) {
561 __frwr_reset_mr(ia, mw);
562 bad_wr = bad_wr->next;
568 /* Use a slow, safe mechanism to invalidate all memory regions
569 * that were registered for "req".
572 frwr_op_unmap_safe(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
575 struct rpcrdma_mw *mw;
577 while (!list_empty(&req->rl_registered)) {
578 mw = list_first_entry(&req->rl_registered,
579 struct rpcrdma_mw, mw_list);
580 list_del_init(&mw->mw_list);
583 frwr_op_recover_mr(mw);
585 rpcrdma_defer_mr_recovery(mw);
589 const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops = {
590 .ro_map = frwr_op_map,
591 .ro_unmap_sync = frwr_op_unmap_sync,
592 .ro_unmap_safe = frwr_op_unmap_safe,
593 .ro_recover_mr = frwr_op_recover_mr,
594 .ro_open = frwr_op_open,
595 .ro_maxpages = frwr_op_maxpages,
596 .ro_init_mr = frwr_op_init_mr,
597 .ro_release_mr = frwr_op_release_mr,
598 .ro_displayname = "frwr",
599 .ro_send_w_inv_ok = RPCRDMA_CMP_F_SND_W_INV_OK,