1 /*
2 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
3 * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
4 *
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the BSD-type
9 * license below:
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 *
15 * Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 *
18 * Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials provided
21 * with the distribution.
22 *
23 * Neither the name of the Network Appliance, Inc. nor the names of
24 * its contributors may be used to endorse or promote products
25 * derived from this software without specific prior written
26 * permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 *
40 * Author: Tom Tucker <tom@opengridcomputing.com>
41 */
42
43 #include <linux/sunrpc/debug.h>
44 #include <linux/sunrpc/rpc_rdma.h>
45 #include <linux/spinlock.h>
46 #include <asm/unaligned.h>
47 #include <rdma/ib_verbs.h>
48 #include <rdma/rdma_cm.h>
49 #include <linux/sunrpc/svc_rdma.h>
50
51 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
52
53 /*
54 * Replace the pages in the rq_argpages array with the pages from the SGE in
55 * the RDMA_RECV completion. The SGL should contain full pages up until the
56 * last one.
57 */
rdma_build_arg_xdr(struct svc_rqst * rqstp,struct svc_rdma_op_ctxt * ctxt,u32 byte_count)58 static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
59 struct svc_rdma_op_ctxt *ctxt,
60 u32 byte_count)
61 {
62 struct rpcrdma_msg *rmsgp;
63 struct page *page;
64 u32 bc;
65 int sge_no;
66
67 /* Swap the page in the SGE with the page in argpages */
68 page = ctxt->pages[0];
69 put_page(rqstp->rq_pages[0]);
70 rqstp->rq_pages[0] = page;
71
72 /* Set up the XDR head */
73 rqstp->rq_arg.head[0].iov_base = page_address(page);
74 rqstp->rq_arg.head[0].iov_len =
75 min_t(size_t, byte_count, ctxt->sge[0].length);
76 rqstp->rq_arg.len = byte_count;
77 rqstp->rq_arg.buflen = byte_count;
78
79 /* Compute bytes past head in the SGL */
80 bc = byte_count - rqstp->rq_arg.head[0].iov_len;
81
82 /* If data remains, store it in the pagelist */
83 rqstp->rq_arg.page_len = bc;
84 rqstp->rq_arg.page_base = 0;
85
86 /* RDMA_NOMSG: RDMA READ data should land just after RDMA RECV data */
87 rmsgp = (struct rpcrdma_msg *)rqstp->rq_arg.head[0].iov_base;
88 if (rmsgp->rm_type == rdma_nomsg)
89 rqstp->rq_arg.pages = &rqstp->rq_pages[0];
90 else
91 rqstp->rq_arg.pages = &rqstp->rq_pages[1];
92
93 sge_no = 1;
94 while (bc && sge_no < ctxt->count) {
95 page = ctxt->pages[sge_no];
96 put_page(rqstp->rq_pages[sge_no]);
97 rqstp->rq_pages[sge_no] = page;
98 bc -= min_t(u32, bc, ctxt->sge[sge_no].length);
99 rqstp->rq_arg.buflen += ctxt->sge[sge_no].length;
100 sge_no++;
101 }
102 rqstp->rq_respages = &rqstp->rq_pages[sge_no];
103 rqstp->rq_next_page = rqstp->rq_respages + 1;
104
105 /* If not all pages were used from the SGL, free the remaining ones */
106 bc = sge_no;
107 while (sge_no < ctxt->count) {
108 page = ctxt->pages[sge_no++];
109 put_page(page);
110 }
111 ctxt->count = bc;
112
113 /* Set up tail */
114 rqstp->rq_arg.tail[0].iov_base = NULL;
115 rqstp->rq_arg.tail[0].iov_len = 0;
116 }
117
118 /* Issue an RDMA_READ using the local lkey to map the data sink */
rdma_read_chunk_lcl(struct svcxprt_rdma * xprt,struct svc_rqst * rqstp,struct svc_rdma_op_ctxt * head,int * page_no,u32 * page_offset,u32 rs_handle,u32 rs_length,u64 rs_offset,bool last)119 int rdma_read_chunk_lcl(struct svcxprt_rdma *xprt,
120 struct svc_rqst *rqstp,
121 struct svc_rdma_op_ctxt *head,
122 int *page_no,
123 u32 *page_offset,
124 u32 rs_handle,
125 u32 rs_length,
126 u64 rs_offset,
127 bool last)
128 {
129 struct ib_rdma_wr read_wr;
130 int pages_needed = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT;
131 struct svc_rdma_op_ctxt *ctxt = svc_rdma_get_context(xprt);
132 int ret, read, pno;
133 u32 pg_off = *page_offset;
134 u32 pg_no = *page_no;
135
136 ctxt->direction = DMA_FROM_DEVICE;
137 ctxt->read_hdr = head;
138 pages_needed = min_t(int, pages_needed, xprt->sc_max_sge_rd);
139 read = min_t(int, (pages_needed << PAGE_SHIFT) - *page_offset,
140 rs_length);
141
142 for (pno = 0; pno < pages_needed; pno++) {
143 int len = min_t(int, rs_length, PAGE_SIZE - pg_off);
144
145 head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no];
146 head->arg.page_len += len;
147
148 head->arg.len += len;
149 if (!pg_off)
150 head->count++;
151 rqstp->rq_respages = &rqstp->rq_arg.pages[pg_no+1];
152 rqstp->rq_next_page = rqstp->rq_respages + 1;
153 ctxt->sge[pno].addr =
154 ib_dma_map_page(xprt->sc_cm_id->device,
155 head->arg.pages[pg_no], pg_off,
156 PAGE_SIZE - pg_off,
157 DMA_FROM_DEVICE);
158 ret = ib_dma_mapping_error(xprt->sc_cm_id->device,
159 ctxt->sge[pno].addr);
160 if (ret)
161 goto err;
162 svc_rdma_count_mappings(xprt, ctxt);
163
164 ctxt->sge[pno].lkey = xprt->sc_pd->local_dma_lkey;
165 ctxt->sge[pno].length = len;
166 ctxt->count++;
167
168 /* adjust offset and wrap to next page if needed */
169 pg_off += len;
170 if (pg_off == PAGE_SIZE) {
171 pg_off = 0;
172 pg_no++;
173 }
174 rs_length -= len;
175 }
176
177 if (last && rs_length == 0)
178 set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
179 else
180 clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
181
182 memset(&read_wr, 0, sizeof(read_wr));
183 ctxt->cqe.done = svc_rdma_wc_read;
184 read_wr.wr.wr_cqe = &ctxt->cqe;
185 read_wr.wr.opcode = IB_WR_RDMA_READ;
186 read_wr.wr.send_flags = IB_SEND_SIGNALED;
187 read_wr.rkey = rs_handle;
188 read_wr.remote_addr = rs_offset;
189 read_wr.wr.sg_list = ctxt->sge;
190 read_wr.wr.num_sge = pages_needed;
191
192 ret = svc_rdma_send(xprt, &read_wr.wr);
193 if (ret) {
194 pr_err("svcrdma: Error %d posting RDMA_READ\n", ret);
195 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
196 goto err;
197 }
198
199 /* return current location in page array */
200 *page_no = pg_no;
201 *page_offset = pg_off;
202 ret = read;
203 atomic_inc(&rdma_stat_read);
204 return ret;
205 err:
206 svc_rdma_unmap_dma(ctxt);
207 svc_rdma_put_context(ctxt, 0);
208 return ret;
209 }
210
211 /* Issue an RDMA_READ using an FRMR to map the data sink */
rdma_read_chunk_frmr(struct svcxprt_rdma * xprt,struct svc_rqst * rqstp,struct svc_rdma_op_ctxt * head,int * page_no,u32 * page_offset,u32 rs_handle,u32 rs_length,u64 rs_offset,bool last)212 int rdma_read_chunk_frmr(struct svcxprt_rdma *xprt,
213 struct svc_rqst *rqstp,
214 struct svc_rdma_op_ctxt *head,
215 int *page_no,
216 u32 *page_offset,
217 u32 rs_handle,
218 u32 rs_length,
219 u64 rs_offset,
220 bool last)
221 {
222 struct ib_rdma_wr read_wr;
223 struct ib_send_wr inv_wr;
224 struct ib_reg_wr reg_wr;
225 u8 key;
226 int nents = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT;
227 struct svc_rdma_op_ctxt *ctxt = svc_rdma_get_context(xprt);
228 struct svc_rdma_fastreg_mr *frmr = svc_rdma_get_frmr(xprt);
229 int ret, read, pno, dma_nents, n;
230 u32 pg_off = *page_offset;
231 u32 pg_no = *page_no;
232
233 if (IS_ERR(frmr))
234 return -ENOMEM;
235
236 ctxt->direction = DMA_FROM_DEVICE;
237 ctxt->frmr = frmr;
238 nents = min_t(unsigned int, nents, xprt->sc_frmr_pg_list_len);
239 read = min_t(int, (nents << PAGE_SHIFT) - *page_offset, rs_length);
240
241 frmr->direction = DMA_FROM_DEVICE;
242 frmr->access_flags = (IB_ACCESS_LOCAL_WRITE|IB_ACCESS_REMOTE_WRITE);
243 frmr->sg_nents = nents;
244
245 for (pno = 0; pno < nents; pno++) {
246 int len = min_t(int, rs_length, PAGE_SIZE - pg_off);
247
248 head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no];
249 head->arg.page_len += len;
250 head->arg.len += len;
251 if (!pg_off)
252 head->count++;
253
254 sg_set_page(&frmr->sg[pno], rqstp->rq_arg.pages[pg_no],
255 len, pg_off);
256
257 rqstp->rq_respages = &rqstp->rq_arg.pages[pg_no+1];
258 rqstp->rq_next_page = rqstp->rq_respages + 1;
259
260 /* adjust offset and wrap to next page if needed */
261 pg_off += len;
262 if (pg_off == PAGE_SIZE) {
263 pg_off = 0;
264 pg_no++;
265 }
266 rs_length -= len;
267 }
268
269 if (last && rs_length == 0)
270 set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
271 else
272 clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
273
274 dma_nents = ib_dma_map_sg(xprt->sc_cm_id->device,
275 frmr->sg, frmr->sg_nents,
276 frmr->direction);
277 if (!dma_nents) {
278 pr_err("svcrdma: failed to dma map sg %p\n",
279 frmr->sg);
280 return -ENOMEM;
281 }
282 atomic_inc(&xprt->sc_dma_used);
283
284 n = ib_map_mr_sg(frmr->mr, frmr->sg, frmr->sg_nents, NULL, PAGE_SIZE);
285 if (unlikely(n != frmr->sg_nents)) {
286 pr_err("svcrdma: failed to map mr %p (%d/%d elements)\n",
287 frmr->mr, n, frmr->sg_nents);
288 return n < 0 ? n : -EINVAL;
289 }
290
291 /* Bump the key */
292 key = (u8)(frmr->mr->lkey & 0x000000FF);
293 ib_update_fast_reg_key(frmr->mr, ++key);
294
295 ctxt->sge[0].addr = frmr->mr->iova;
296 ctxt->sge[0].lkey = frmr->mr->lkey;
297 ctxt->sge[0].length = frmr->mr->length;
298 ctxt->count = 1;
299 ctxt->read_hdr = head;
300
301 /* Prepare REG WR */
302 ctxt->reg_cqe.done = svc_rdma_wc_reg;
303 reg_wr.wr.wr_cqe = &ctxt->reg_cqe;
304 reg_wr.wr.opcode = IB_WR_REG_MR;
305 reg_wr.wr.send_flags = IB_SEND_SIGNALED;
306 reg_wr.wr.num_sge = 0;
307 reg_wr.mr = frmr->mr;
308 reg_wr.key = frmr->mr->lkey;
309 reg_wr.access = frmr->access_flags;
310 reg_wr.wr.next = &read_wr.wr;
311
312 /* Prepare RDMA_READ */
313 memset(&read_wr, 0, sizeof(read_wr));
314 ctxt->cqe.done = svc_rdma_wc_read;
315 read_wr.wr.wr_cqe = &ctxt->cqe;
316 read_wr.wr.send_flags = IB_SEND_SIGNALED;
317 read_wr.rkey = rs_handle;
318 read_wr.remote_addr = rs_offset;
319 read_wr.wr.sg_list = ctxt->sge;
320 read_wr.wr.num_sge = 1;
321 if (xprt->sc_dev_caps & SVCRDMA_DEVCAP_READ_W_INV) {
322 read_wr.wr.opcode = IB_WR_RDMA_READ_WITH_INV;
323 read_wr.wr.ex.invalidate_rkey = ctxt->frmr->mr->lkey;
324 } else {
325 read_wr.wr.opcode = IB_WR_RDMA_READ;
326 read_wr.wr.next = &inv_wr;
327 /* Prepare invalidate */
328 memset(&inv_wr, 0, sizeof(inv_wr));
329 ctxt->inv_cqe.done = svc_rdma_wc_inv;
330 inv_wr.wr_cqe = &ctxt->inv_cqe;
331 inv_wr.opcode = IB_WR_LOCAL_INV;
332 inv_wr.send_flags = IB_SEND_SIGNALED | IB_SEND_FENCE;
333 inv_wr.ex.invalidate_rkey = frmr->mr->lkey;
334 }
335
336 /* Post the chain */
337 ret = svc_rdma_send(xprt, ®_wr.wr);
338 if (ret) {
339 pr_err("svcrdma: Error %d posting RDMA_READ\n", ret);
340 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
341 goto err;
342 }
343
344 /* return current location in page array */
345 *page_no = pg_no;
346 *page_offset = pg_off;
347 ret = read;
348 atomic_inc(&rdma_stat_read);
349 return ret;
350 err:
351 svc_rdma_put_context(ctxt, 0);
352 svc_rdma_put_frmr(xprt, frmr);
353 return ret;
354 }
355
356 static unsigned int
rdma_rcl_chunk_count(struct rpcrdma_read_chunk * ch)357 rdma_rcl_chunk_count(struct rpcrdma_read_chunk *ch)
358 {
359 unsigned int count;
360
361 for (count = 0; ch->rc_discrim != xdr_zero; ch++)
362 count++;
363 return count;
364 }
365
366 /* If there was additional inline content, append it to the end of arg.pages.
367 * Tail copy has to be done after the reader function has determined how many
368 * pages are needed for RDMA READ.
369 */
370 static int
rdma_copy_tail(struct svc_rqst * rqstp,struct svc_rdma_op_ctxt * head,u32 position,u32 byte_count,u32 page_offset,int page_no)371 rdma_copy_tail(struct svc_rqst *rqstp, struct svc_rdma_op_ctxt *head,
372 u32 position, u32 byte_count, u32 page_offset, int page_no)
373 {
374 char *srcp, *destp;
375 int ret;
376
377 ret = 0;
378 srcp = head->arg.head[0].iov_base + position;
379 byte_count = head->arg.head[0].iov_len - position;
380 if (byte_count > PAGE_SIZE) {
381 dprintk("svcrdma: large tail unsupported\n");
382 return 0;
383 }
384
385 /* Fit as much of the tail on the current page as possible */
386 if (page_offset != PAGE_SIZE) {
387 destp = page_address(rqstp->rq_arg.pages[page_no]);
388 destp += page_offset;
389 while (byte_count--) {
390 *destp++ = *srcp++;
391 page_offset++;
392 if (page_offset == PAGE_SIZE && byte_count)
393 goto more;
394 }
395 goto done;
396 }
397
398 more:
399 /* Fit the rest on the next page */
400 page_no++;
401 destp = page_address(rqstp->rq_arg.pages[page_no]);
402 while (byte_count--)
403 *destp++ = *srcp++;
404
405 rqstp->rq_respages = &rqstp->rq_arg.pages[page_no+1];
406 rqstp->rq_next_page = rqstp->rq_respages + 1;
407
408 done:
409 byte_count = head->arg.head[0].iov_len - position;
410 head->arg.page_len += byte_count;
411 head->arg.len += byte_count;
412 head->arg.buflen += byte_count;
413 return 1;
414 }
415
rdma_read_chunks(struct svcxprt_rdma * xprt,struct rpcrdma_msg * rmsgp,struct svc_rqst * rqstp,struct svc_rdma_op_ctxt * head)416 static int rdma_read_chunks(struct svcxprt_rdma *xprt,
417 struct rpcrdma_msg *rmsgp,
418 struct svc_rqst *rqstp,
419 struct svc_rdma_op_ctxt *head)
420 {
421 int page_no, ret;
422 struct rpcrdma_read_chunk *ch;
423 u32 handle, page_offset, byte_count;
424 u32 position;
425 u64 rs_offset;
426 bool last;
427
428 /* If no read list is present, return 0 */
429 ch = svc_rdma_get_read_chunk(rmsgp);
430 if (!ch)
431 return 0;
432
433 if (rdma_rcl_chunk_count(ch) > RPCSVC_MAXPAGES)
434 return -EINVAL;
435
436 /* The request is completed when the RDMA_READs complete. The
437 * head context keeps all the pages that comprise the
438 * request.
439 */
440 head->arg.head[0] = rqstp->rq_arg.head[0];
441 head->arg.tail[0] = rqstp->rq_arg.tail[0];
442 head->hdr_count = head->count;
443 head->arg.page_base = 0;
444 head->arg.page_len = 0;
445 head->arg.len = rqstp->rq_arg.len;
446 head->arg.buflen = rqstp->rq_arg.buflen;
447
448 /* RDMA_NOMSG: RDMA READ data should land just after RDMA RECV data */
449 position = be32_to_cpu(ch->rc_position);
450 if (position == 0) {
451 head->arg.pages = &head->pages[0];
452 page_offset = head->byte_len;
453 } else {
454 head->arg.pages = &head->pages[head->count];
455 page_offset = 0;
456 }
457
458 ret = 0;
459 page_no = 0;
460 for (; ch->rc_discrim != xdr_zero; ch++) {
461 if (be32_to_cpu(ch->rc_position) != position)
462 goto err;
463
464 handle = be32_to_cpu(ch->rc_target.rs_handle),
465 byte_count = be32_to_cpu(ch->rc_target.rs_length);
466 xdr_decode_hyper((__be32 *)&ch->rc_target.rs_offset,
467 &rs_offset);
468
469 while (byte_count > 0) {
470 last = (ch + 1)->rc_discrim == xdr_zero;
471 ret = xprt->sc_reader(xprt, rqstp, head,
472 &page_no, &page_offset,
473 handle, byte_count,
474 rs_offset, last);
475 if (ret < 0)
476 goto err;
477 byte_count -= ret;
478 rs_offset += ret;
479 head->arg.buflen += ret;
480 }
481 }
482
483 /* Read list may need XDR round-up (see RFC 5666, s. 3.7) */
484 if (page_offset & 3) {
485 u32 pad = 4 - (page_offset & 3);
486
487 head->arg.tail[0].iov_len += pad;
488 head->arg.len += pad;
489 head->arg.buflen += pad;
490 page_offset += pad;
491 }
492
493 ret = 1;
494 if (position && position < head->arg.head[0].iov_len)
495 ret = rdma_copy_tail(rqstp, head, position,
496 byte_count, page_offset, page_no);
497 head->arg.head[0].iov_len = position;
498 head->position = position;
499
500 err:
501 /* Detach arg pages. svc_recv will replenish them */
502 for (page_no = 0;
503 &rqstp->rq_pages[page_no] < rqstp->rq_respages; page_no++)
504 rqstp->rq_pages[page_no] = NULL;
505
506 return ret;
507 }
508
rdma_read_complete(struct svc_rqst * rqstp,struct svc_rdma_op_ctxt * head)509 static void rdma_read_complete(struct svc_rqst *rqstp,
510 struct svc_rdma_op_ctxt *head)
511 {
512 int page_no;
513
514 /* Copy RPC pages */
515 for (page_no = 0; page_no < head->count; page_no++) {
516 put_page(rqstp->rq_pages[page_no]);
517 rqstp->rq_pages[page_no] = head->pages[page_no];
518 }
519
520 /* Adjustments made for RDMA_NOMSG type requests */
521 if (head->position == 0) {
522 if (head->arg.len <= head->sge[0].length) {
523 head->arg.head[0].iov_len = head->arg.len -
524 head->byte_len;
525 head->arg.page_len = 0;
526 } else {
527 head->arg.head[0].iov_len = head->sge[0].length -
528 head->byte_len;
529 head->arg.page_len = head->arg.len -
530 head->sge[0].length;
531 }
532 }
533
534 /* Point rq_arg.pages past header */
535 rqstp->rq_arg.pages = &rqstp->rq_pages[head->hdr_count];
536 rqstp->rq_arg.page_len = head->arg.page_len;
537 rqstp->rq_arg.page_base = head->arg.page_base;
538
539 /* rq_respages starts after the last arg page */
540 rqstp->rq_respages = &rqstp->rq_pages[page_no];
541 rqstp->rq_next_page = rqstp->rq_respages + 1;
542
543 /* Rebuild rq_arg head and tail. */
544 rqstp->rq_arg.head[0] = head->arg.head[0];
545 rqstp->rq_arg.tail[0] = head->arg.tail[0];
546 rqstp->rq_arg.len = head->arg.len;
547 rqstp->rq_arg.buflen = head->arg.buflen;
548 }
549
550 /* By convention, backchannel calls arrive via rdma_msg type
551 * messages, and never populate the chunk lists. This makes
552 * the RPC/RDMA header small and fixed in size, so it is
553 * straightforward to check the RPC header's direction field.
554 */
555 static bool
svc_rdma_is_backchannel_reply(struct svc_xprt * xprt,struct rpcrdma_msg * rmsgp)556 svc_rdma_is_backchannel_reply(struct svc_xprt *xprt, struct rpcrdma_msg *rmsgp)
557 {
558 __be32 *p = (__be32 *)rmsgp;
559
560 if (!xprt->xpt_bc_xprt)
561 return false;
562
563 if (rmsgp->rm_type != rdma_msg)
564 return false;
565 if (rmsgp->rm_body.rm_chunks[0] != xdr_zero)
566 return false;
567 if (rmsgp->rm_body.rm_chunks[1] != xdr_zero)
568 return false;
569 if (rmsgp->rm_body.rm_chunks[2] != xdr_zero)
570 return false;
571
572 /* sanity */
573 if (p[7] != rmsgp->rm_xid)
574 return false;
575 /* call direction */
576 if (p[8] == cpu_to_be32(RPC_CALL))
577 return false;
578
579 return true;
580 }
581
582 /*
583 * Set up the rqstp thread context to point to the RQ buffer. If
584 * necessary, pull additional data from the client with an RDMA_READ
585 * request.
586 */
svc_rdma_recvfrom(struct svc_rqst * rqstp)587 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
588 {
589 struct svc_xprt *xprt = rqstp->rq_xprt;
590 struct svcxprt_rdma *rdma_xprt =
591 container_of(xprt, struct svcxprt_rdma, sc_xprt);
592 struct svc_rdma_op_ctxt *ctxt = NULL;
593 struct rpcrdma_msg *rmsgp;
594 int ret = 0;
595
596 dprintk("svcrdma: rqstp=%p\n", rqstp);
597
598 spin_lock_bh(&rdma_xprt->sc_rq_dto_lock);
599 if (!list_empty(&rdma_xprt->sc_read_complete_q)) {
600 ctxt = list_entry(rdma_xprt->sc_read_complete_q.next,
601 struct svc_rdma_op_ctxt,
602 dto_q);
603 list_del_init(&ctxt->dto_q);
604 spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
605 rdma_read_complete(rqstp, ctxt);
606 goto complete;
607 } else if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
608 ctxt = list_entry(rdma_xprt->sc_rq_dto_q.next,
609 struct svc_rdma_op_ctxt,
610 dto_q);
611 list_del_init(&ctxt->dto_q);
612 } else {
613 atomic_inc(&rdma_stat_rq_starve);
614 clear_bit(XPT_DATA, &xprt->xpt_flags);
615 ctxt = NULL;
616 }
617 spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
618 if (!ctxt) {
619 /* This is the EAGAIN path. The svc_recv routine will
620 * return -EAGAIN, the nfsd thread will go to call into
621 * svc_recv again and we shouldn't be on the active
622 * transport list
623 */
624 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
625 goto defer;
626 goto out;
627 }
628 dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p, status=%d\n",
629 ctxt, rdma_xprt, rqstp, ctxt->wc_status);
630 atomic_inc(&rdma_stat_recv);
631
632 /* Build up the XDR from the receive buffers. */
633 rdma_build_arg_xdr(rqstp, ctxt, ctxt->byte_len);
634
635 /* Decode the RDMA header. */
636 rmsgp = (struct rpcrdma_msg *)rqstp->rq_arg.head[0].iov_base;
637 ret = svc_rdma_xdr_decode_req(&rqstp->rq_arg);
638 if (ret < 0)
639 goto out_err;
640 if (ret == 0)
641 goto out_drop;
642 rqstp->rq_xprt_hlen = ret;
643
644 if (svc_rdma_is_backchannel_reply(xprt, rmsgp)) {
645 ret = svc_rdma_handle_bc_reply(xprt->xpt_bc_xprt, rmsgp,
646 &rqstp->rq_arg);
647 svc_rdma_put_context(ctxt, 0);
648 if (ret)
649 goto repost;
650 return ret;
651 }
652
653 /* Read read-list data. */
654 ret = rdma_read_chunks(rdma_xprt, rmsgp, rqstp, ctxt);
655 if (ret > 0) {
656 /* read-list posted, defer until data received from client. */
657 goto defer;
658 } else if (ret < 0) {
659 /* Post of read-list failed, free context. */
660 svc_rdma_put_context(ctxt, 1);
661 return 0;
662 }
663
664 complete:
665 ret = rqstp->rq_arg.head[0].iov_len
666 + rqstp->rq_arg.page_len
667 + rqstp->rq_arg.tail[0].iov_len;
668 svc_rdma_put_context(ctxt, 0);
669 out:
670 dprintk("svcrdma: ret=%d, rq_arg.len=%u, "
671 "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len=%zd\n",
672 ret, rqstp->rq_arg.len,
673 rqstp->rq_arg.head[0].iov_base,
674 rqstp->rq_arg.head[0].iov_len);
675 rqstp->rq_prot = IPPROTO_MAX;
676 svc_xprt_copy_addrs(rqstp, xprt);
677 return ret;
678
679 out_err:
680 svc_rdma_send_error(rdma_xprt, rmsgp, ret);
681 svc_rdma_put_context(ctxt, 0);
682 return 0;
683
684 defer:
685 return 0;
686
687 out_drop:
688 svc_rdma_put_context(ctxt, 1);
689 repost:
690 return svc_rdma_repost_recv(rdma_xprt, GFP_KERNEL);
691 }
692