2 * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the BSD-type
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
17 * Redistributions in binary form must reproduce the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer in the documentation and/or other materials provided
20 * with the distribution.
22 * Neither the name of the Network Appliance, Inc. nor the names of
23 * its contributors may be used to endorse or promote products
24 * derived from this software without specific prior written
27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 * Author: Tom Tucker <tom@opengridcomputing.com>
42 #include <linux/sunrpc/debug.h>
43 #include <linux/sunrpc/rpc_rdma.h>
44 #include <linux/spinlock.h>
45 #include <asm/unaligned.h>
46 #include <rdma/ib_verbs.h>
47 #include <rdma/rdma_cm.h>
48 #include <linux/sunrpc/svc_rdma.h>
50 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
53 * Replace the pages in the rq_argpages array with the pages from the SGE in
54 * the RDMA_RECV completion. The SGL should contain full pages up until the
57 static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
58 struct svc_rdma_op_ctxt *ctxt,
65 /* Swap the page in the SGE with the page in argpages */
66 page = ctxt->pages[0];
67 put_page(rqstp->rq_pages[0]);
68 rqstp->rq_pages[0] = page;
70 /* Set up the XDR head */
71 rqstp->rq_arg.head[0].iov_base = page_address(page);
72 rqstp->rq_arg.head[0].iov_len = min(byte_count, ctxt->sge[0].length);
73 rqstp->rq_arg.len = byte_count;
74 rqstp->rq_arg.buflen = byte_count;
76 /* Compute bytes past head in the SGL */
77 bc = byte_count - rqstp->rq_arg.head[0].iov_len;
79 /* If data remains, store it in the pagelist */
80 rqstp->rq_arg.page_len = bc;
81 rqstp->rq_arg.page_base = 0;
82 rqstp->rq_arg.pages = &rqstp->rq_pages[1];
84 while (bc && sge_no < ctxt->count) {
85 page = ctxt->pages[sge_no];
86 put_page(rqstp->rq_pages[sge_no]);
87 rqstp->rq_pages[sge_no] = page;
88 bc -= min(bc, ctxt->sge[sge_no].length);
89 rqstp->rq_arg.buflen += ctxt->sge[sge_no].length;
92 rqstp->rq_respages = &rqstp->rq_pages[sge_no];
94 /* We should never run out of SGE because the limit is defined to
95 * support the max allowed RPC data length
97 BUG_ON(bc && (sge_no == ctxt->count));
98 BUG_ON((rqstp->rq_arg.head[0].iov_len + rqstp->rq_arg.page_len)
100 BUG_ON(rqstp->rq_arg.len != byte_count);
102 /* If not all pages were used from the SGL, free the remaining ones */
104 while (sge_no < ctxt->count) {
105 page = ctxt->pages[sge_no++];
111 rqstp->rq_arg.tail[0].iov_base = NULL;
112 rqstp->rq_arg.tail[0].iov_len = 0;
116 int start; /* sge no for this chunk */
117 int count; /* sge count for this chunk */
120 /* Encode a read-chunk-list as an array of IB SGE
123 * - chunk[0]->position points to pages[0] at an offset of 0
124 * - pages[] is not physically or virtually contigous and consists of
125 * PAGE_SIZE elements.
128 * - sge array pointing into pages[] array.
129 * - chunk_sge array specifying sge index and count for each
130 * chunk in the read list
133 static int rdma_rcl_to_sge(struct svcxprt_rdma *xprt,
134 struct svc_rqst *rqstp,
135 struct svc_rdma_op_ctxt *head,
136 struct rpcrdma_msg *rmsgp,
138 struct chunk_sge *ch_sge_ary,
148 struct rpcrdma_read_chunk *ch;
153 ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
155 ch_bytes = ch->rc_target.rs_length;
156 head->arg.head[0] = rqstp->rq_arg.head[0];
157 head->arg.tail[0] = rqstp->rq_arg.tail[0];
158 head->arg.pages = &head->pages[head->count];
159 head->sge[0].length = head->count; /* save count of hdr pages */
160 head->arg.page_base = 0;
161 head->arg.page_len = ch_bytes;
162 head->arg.len = rqstp->rq_arg.len + ch_bytes;
163 head->arg.buflen = rqstp->rq_arg.buflen + ch_bytes;
165 ch_sge_ary[0].start = 0;
167 sge_bytes = min_t(int, PAGE_SIZE-page_off, ch_bytes);
169 ib_dma_map_page(xprt->sc_cm_id->device,
170 rqstp->rq_arg.pages[page_no],
173 sge[sge_no].length = sge_bytes;
174 sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
176 * Don't bump head->count here because the same page
177 * may be used by multiple SGE.
179 head->arg.pages[page_no] = rqstp->rq_arg.pages[page_no];
180 rqstp->rq_respages = &rqstp->rq_arg.pages[page_no+1];
182 byte_count -= sge_bytes;
183 ch_bytes -= sge_bytes;
186 * If all bytes for this chunk have been mapped to an
187 * SGE, move to the next SGE
190 ch_sge_ary[ch_no].count =
191 sge_no - ch_sge_ary[ch_no].start;
194 ch_sge_ary[ch_no].start = sge_no;
195 ch_bytes = ch->rc_target.rs_length;
196 /* If bytes remaining account for next chunk */
198 head->arg.page_len += ch_bytes;
199 head->arg.len += ch_bytes;
200 head->arg.buflen += ch_bytes;
204 * If this SGE consumed all of the page, move to the
207 if ((sge_bytes + page_off) == PAGE_SIZE) {
211 * If there are still bytes left to map, bump
217 page_off += sge_bytes;
219 BUG_ON(byte_count != 0);
223 static void rdma_set_ctxt_sge(struct svc_rdma_op_ctxt *ctxt,
231 for (i = 0; i < count; i++) {
232 ctxt->sge[i].addr = sge[i].addr;
233 ctxt->sge[i].length = sge[i].length;
234 *sgl_offset = *sgl_offset + sge[i].length;
238 static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count)
240 if ((RDMA_TRANSPORT_IWARP ==
241 rdma_node_get_transport(xprt->sc_cm_id->
246 return min_t(int, sge_count, xprt->sc_max_sge);
250 * Use RDMA_READ to read data from the advertised client buffer into the
251 * XDR stream starting at rq_arg.head[0].iov_base.
252 * Each chunk in the array
253 * contains the following fields:
254 * discrim - '1', This isn't used for data placement
255 * position - The xdr stream offset (the same for every chunk)
256 * handle - RMR for client memory region
257 * length - data transfer length
258 * offset - 64 bit tagged offset in remote memory region
260 * On our side, we need to read into a pagelist. The first page immediately
261 * follows the RPC header.
263 * This function returns 1 to indicate success. The data is not yet in
264 * the pagelist and therefore the RPC request must be deferred. The
265 * I/O completion will enqueue the transport again and
266 * svc_rdma_recvfrom will complete the request.
268 * NOTE: The ctxt must not be touched after the last WR has been posted
269 * because the I/O completion processing may occur on another
270 * processor and free / modify the context. Ne touche pas!
272 static int rdma_read_xdr(struct svcxprt_rdma *xprt,
273 struct rpcrdma_msg *rmsgp,
274 struct svc_rqst *rqstp,
275 struct svc_rdma_op_ctxt *hdr_ctxt)
277 struct ib_send_wr read_wr;
285 struct rpcrdma_read_chunk *ch;
286 struct svc_rdma_op_ctxt *ctxt = NULL;
287 struct svc_rdma_op_ctxt *head;
288 struct svc_rdma_op_ctxt *tmp_sge_ctxt;
289 struct svc_rdma_op_ctxt *tmp_ch_ctxt;
290 struct chunk_sge *ch_sge_ary;
292 /* If no read list is present, return 0 */
293 ch = svc_rdma_get_read_chunk(rmsgp);
297 /* Allocate temporary contexts to keep SGE */
298 BUG_ON(sizeof(struct ib_sge) < sizeof(struct chunk_sge));
299 tmp_sge_ctxt = svc_rdma_get_context(xprt);
300 sge = tmp_sge_ctxt->sge;
301 tmp_ch_ctxt = svc_rdma_get_context(xprt);
302 ch_sge_ary = (struct chunk_sge *)tmp_ch_ctxt->sge;
304 svc_rdma_rcl_chunk_counts(ch, &ch_count, &byte_count);
305 sge_count = rdma_rcl_to_sge(xprt, rqstp, hdr_ctxt, rmsgp,
307 ch_count, byte_count);
308 head = svc_rdma_get_context(xprt);
312 for (ch = (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0];
313 ch->rc_discrim != 0; ch++, ch_no++) {
318 ctxt->next = svc_rdma_get_context(xprt);
322 ctxt->direction = DMA_FROM_DEVICE;
323 clear_bit(RDMACTXT_F_READ_DONE, &ctxt->flags);
324 clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
326 /* Prepare READ WR */
327 memset(&read_wr, 0, sizeof read_wr);
328 ctxt->wr_op = IB_WR_RDMA_READ;
329 read_wr.wr_id = (unsigned long)ctxt;
330 read_wr.opcode = IB_WR_RDMA_READ;
331 read_wr.send_flags = IB_SEND_SIGNALED;
332 read_wr.wr.rdma.rkey = ch->rc_target.rs_handle;
333 read_wr.wr.rdma.remote_addr =
334 get_unaligned(&(ch->rc_target.rs_offset)) +
336 read_wr.sg_list = &sge[ch_sge_ary[ch_no].start];
338 rdma_read_max_sge(xprt, ch_sge_ary[ch_no].count);
339 rdma_set_ctxt_sge(ctxt, &sge[ch_sge_ary[ch_no].start],
342 if (((ch+1)->rc_discrim == 0) &&
343 (read_wr.num_sge == ch_sge_ary[ch_no].count)) {
345 * Mark the last RDMA_READ with a bit to
346 * indicate all RPC data has been fetched from
347 * the client and the RPC needs to be enqueued.
349 set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
350 ctxt->next = hdr_ctxt;
351 hdr_ctxt->next = head;
354 err = svc_rdma_send(xprt, &read_wr);
356 printk(KERN_ERR "svcrdma: Error posting send = %d\n",
359 * Break the circular list so free knows when
360 * to stop if the error happened to occur on
366 atomic_inc(&rdma_stat_read);
368 if (read_wr.num_sge < ch_sge_ary[ch_no].count) {
369 ch_sge_ary[ch_no].count -= read_wr.num_sge;
370 ch_sge_ary[ch_no].start += read_wr.num_sge;
378 svc_rdma_put_context(tmp_sge_ctxt, 0);
379 svc_rdma_put_context(tmp_ch_ctxt, 0);
381 /* Detach arg pages. svc_recv will replenish them */
382 for (ch_no = 0; &rqstp->rq_pages[ch_no] < rqstp->rq_respages; ch_no++)
383 rqstp->rq_pages[ch_no] = NULL;
386 * Detach res pages. svc_release must see a resused count of
387 * zero or it will attempt to put them.
389 while (rqstp->rq_resused)
390 rqstp->rq_respages[--rqstp->rq_resused] = NULL;
393 printk(KERN_ERR "svcrdma : RDMA_READ error = %d\n", err);
394 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
395 /* Free the linked list of read contexts */
396 while (head != NULL) {
398 svc_rdma_put_context(head, 1);
407 static int rdma_read_complete(struct svc_rqst *rqstp,
408 struct svc_rdma_op_ctxt *data)
410 struct svc_rdma_op_ctxt *head = data->next;
417 for (page_no = 0; page_no < head->count; page_no++) {
418 put_page(rqstp->rq_pages[page_no]);
419 rqstp->rq_pages[page_no] = head->pages[page_no];
421 /* Point rq_arg.pages past header */
422 rqstp->rq_arg.pages = &rqstp->rq_pages[head->sge[0].length];
423 rqstp->rq_arg.page_len = head->arg.page_len;
424 rqstp->rq_arg.page_base = head->arg.page_base;
426 /* rq_respages starts after the last arg page */
427 rqstp->rq_respages = &rqstp->rq_arg.pages[page_no];
428 rqstp->rq_resused = 0;
430 /* Rebuild rq_arg head and tail. */
431 rqstp->rq_arg.head[0] = head->arg.head[0];
432 rqstp->rq_arg.tail[0] = head->arg.tail[0];
433 rqstp->rq_arg.len = head->arg.len;
434 rqstp->rq_arg.buflen = head->arg.buflen;
436 /* XXX: What should this be? */
437 rqstp->rq_prot = IPPROTO_MAX;
440 * Free the contexts we used to build the RDMA_READ. We have
441 * to be careful here because the context list uses the same
442 * next pointer used to chain the contexts associated with the
445 data->next = NULL; /* terminate circular list */
448 svc_rdma_put_context(head, 0);
450 } while (head != NULL);
452 ret = rqstp->rq_arg.head[0].iov_len
453 + rqstp->rq_arg.page_len
454 + rqstp->rq_arg.tail[0].iov_len;
455 dprintk("svcrdma: deferred read ret=%d, rq_arg.len =%d, "
456 "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
457 ret, rqstp->rq_arg.len, rqstp->rq_arg.head[0].iov_base,
458 rqstp->rq_arg.head[0].iov_len);
460 /* Indicate that we've consumed an RQ credit */
461 rqstp->rq_xprt_ctxt = rqstp->rq_xprt;
462 svc_xprt_received(rqstp->rq_xprt);
467 * Set up the rqstp thread context to point to the RQ buffer. If
468 * necessary, pull additional data from the client with an RDMA_READ
471 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
473 struct svc_xprt *xprt = rqstp->rq_xprt;
474 struct svcxprt_rdma *rdma_xprt =
475 container_of(xprt, struct svcxprt_rdma, sc_xprt);
476 struct svc_rdma_op_ctxt *ctxt = NULL;
477 struct rpcrdma_msg *rmsgp;
481 dprintk("svcrdma: rqstp=%p\n", rqstp);
484 * The rq_xprt_ctxt indicates if we've consumed an RQ credit
485 * or not. It is used in the rdma xpo_release_rqst function to
486 * determine whether or not to return an RQ WQE to the RQ.
488 rqstp->rq_xprt_ctxt = NULL;
490 spin_lock_bh(&rdma_xprt->sc_read_complete_lock);
491 if (!list_empty(&rdma_xprt->sc_read_complete_q)) {
492 ctxt = list_entry(rdma_xprt->sc_read_complete_q.next,
493 struct svc_rdma_op_ctxt,
495 list_del_init(&ctxt->dto_q);
497 spin_unlock_bh(&rdma_xprt->sc_read_complete_lock);
499 return rdma_read_complete(rqstp, ctxt);
501 spin_lock_bh(&rdma_xprt->sc_rq_dto_lock);
502 if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
503 ctxt = list_entry(rdma_xprt->sc_rq_dto_q.next,
504 struct svc_rdma_op_ctxt,
506 list_del_init(&ctxt->dto_q);
508 atomic_inc(&rdma_stat_rq_starve);
509 clear_bit(XPT_DATA, &xprt->xpt_flags);
512 spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
514 /* This is the EAGAIN path. The svc_recv routine will
515 * return -EAGAIN, the nfsd thread will go to call into
516 * svc_recv again and we shouldn't be on the active
519 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
525 dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p, status=%d\n",
526 ctxt, rdma_xprt, rqstp, ctxt->wc_status);
527 BUG_ON(ctxt->wc_status != IB_WC_SUCCESS);
528 atomic_inc(&rdma_stat_recv);
530 /* Build up the XDR from the receive buffers. */
531 rdma_build_arg_xdr(rqstp, ctxt, ctxt->byte_len);
533 /* Decode the RDMA header. */
534 len = svc_rdma_xdr_decode_req(&rmsgp, rqstp);
535 rqstp->rq_xprt_hlen = len;
537 /* If the request is invalid, reply with an error */
540 (void)svc_rdma_send_error(rdma_xprt, rmsgp, ERR_VERS);
544 /* Read read-list data. If we would need to wait, defer
545 * it. Not that in this case, we don't return the RQ credit
546 * until after the read completes.
548 if (rdma_read_xdr(rdma_xprt, rmsgp, rqstp, ctxt)) {
549 svc_xprt_received(xprt);
553 /* Indicate we've consumed an RQ credit */
554 rqstp->rq_xprt_ctxt = rqstp->rq_xprt;
556 ret = rqstp->rq_arg.head[0].iov_len
557 + rqstp->rq_arg.page_len
558 + rqstp->rq_arg.tail[0].iov_len;
559 svc_rdma_put_context(ctxt, 0);
561 dprintk("svcrdma: ret = %d, rq_arg.len =%d, "
562 "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len = %zd\n",
563 ret, rqstp->rq_arg.len,
564 rqstp->rq_arg.head[0].iov_base,
565 rqstp->rq_arg.head[0].iov_len);
566 rqstp->rq_prot = IPPROTO_MAX;
567 svc_xprt_copy_addrs(rqstp, xprt);
568 svc_xprt_received(xprt);
573 svc_rdma_put_context(ctxt, 1);
574 /* Indicate we've consumed an RQ credit */
575 rqstp->rq_xprt_ctxt = rqstp->rq_xprt;
577 dprintk("svcrdma: transport %p is closing\n", xprt);
579 * Set the close bit and enqueue it. svc_recv will see the
580 * close bit and call svc_xprt_delete
582 set_bit(XPT_CLOSE, &xprt->xpt_flags);
583 svc_xprt_received(xprt);