1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3 * Copyright (c) 2016-2018 Oracle. All rights reserved.
4 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
5 * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
6 *
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the BSD-type
11 * license below:
12 *
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
15 * are met:
16 *
17 * Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 *
20 * Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials provided
23 * with the distribution.
24 *
25 * Neither the name of the Network Appliance, Inc. nor the names of
26 * its contributors may be used to endorse or promote products
27 * derived from this software without specific prior written
28 * permission.
29 *
30 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
31 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
32 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
33 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
34 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
35 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
36 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
37 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
38 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
39 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
40 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
41 *
42 * Author: Tom Tucker <tom@opengridcomputing.com>
43 */
44
45 /* Operation
46 *
47 * The main entry point is svc_rdma_sendto. This is called by the
48 * RPC server when an RPC Reply is ready to be transmitted to a client.
49 *
50 * The passed-in svc_rqst contains a struct xdr_buf which holds an
51 * XDR-encoded RPC Reply message. sendto must construct the RPC-over-RDMA
52 * transport header, post all Write WRs needed for this Reply, then post
53 * a Send WR conveying the transport header and the RPC message itself to
54 * the client.
55 *
56 * svc_rdma_sendto must fully transmit the Reply before returning, as
57 * the svc_rqst will be recycled as soon as sendto returns. Remaining
58 * resources referred to by the svc_rqst are also recycled at that time.
59 * Therefore any resources that must remain longer must be detached
60 * from the svc_rqst and released later.
61 *
62 * Page Management
63 *
64 * The I/O that performs Reply transmission is asynchronous, and may
65 * complete well after sendto returns. Thus pages under I/O must be
66 * removed from the svc_rqst before sendto returns.
67 *
68 * The logic here depends on Send Queue and completion ordering. Since
69 * the Send WR is always posted last, it will always complete last. Thus
70 * when it completes, it is guaranteed that all previous Write WRs have
71 * also completed.
72 *
73 * Write WRs are constructed and posted. Each Write segment gets its own
74 * svc_rdma_rw_ctxt, allowing the Write completion handler to find and
75 * DMA-unmap the pages under I/O for that Write segment. The Write
76 * completion handler does not release any pages.
77 *
78 * When the Send WR is constructed, it also gets its own svc_rdma_send_ctxt.
79 * The ownership of all of the Reply's pages are transferred into that
80 * ctxt, the Send WR is posted, and sendto returns.
81 *
82 * The svc_rdma_send_ctxt is presented when the Send WR completes. The
83 * Send completion handler finally releases the Reply's pages.
84 *
85 * This mechanism also assumes that completions on the transport's Send
86 * Completion Queue do not run in parallel. Otherwise a Write completion
87 * and Send completion running at the same time could release pages that
88 * are still DMA-mapped.
89 *
90 * Error Handling
91 *
92 * - If the Send WR is posted successfully, it will either complete
93 * successfully, or get flushed. Either way, the Send completion
94 * handler releases the Reply's pages.
95 * - If the Send WR cannot be not posted, the forward path releases
96 * the Reply's pages.
97 *
98 * This handles the case, without the use of page reference counting,
99 * where two different Write segments send portions of the same page.
100 */
101
102 #include <linux/spinlock.h>
103 #include <asm/unaligned.h>
104
105 #include <rdma/ib_verbs.h>
106 #include <rdma/rdma_cm.h>
107
108 #include <linux/sunrpc/debug.h>
109 #include <linux/sunrpc/rpc_rdma.h>
110 #include <linux/sunrpc/svc_rdma.h>
111
112 #include "xprt_rdma.h"
113 #include <trace/events/rpcrdma.h>
114
115 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
116
117 static void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc);
118
119 static inline struct svc_rdma_send_ctxt *
svc_rdma_next_send_ctxt(struct list_head * list)120 svc_rdma_next_send_ctxt(struct list_head *list)
121 {
122 return list_first_entry_or_null(list, struct svc_rdma_send_ctxt,
123 sc_list);
124 }
125
126 static struct svc_rdma_send_ctxt *
svc_rdma_send_ctxt_alloc(struct svcxprt_rdma * rdma)127 svc_rdma_send_ctxt_alloc(struct svcxprt_rdma *rdma)
128 {
129 struct svc_rdma_send_ctxt *ctxt;
130 dma_addr_t addr;
131 void *buffer;
132 size_t size;
133 int i;
134
135 size = sizeof(*ctxt);
136 size += rdma->sc_max_send_sges * sizeof(struct ib_sge);
137 ctxt = kmalloc(size, GFP_KERNEL);
138 if (!ctxt)
139 goto fail0;
140 buffer = kmalloc(rdma->sc_max_req_size, GFP_KERNEL);
141 if (!buffer)
142 goto fail1;
143 addr = ib_dma_map_single(rdma->sc_pd->device, buffer,
144 rdma->sc_max_req_size, DMA_TO_DEVICE);
145 if (ib_dma_mapping_error(rdma->sc_pd->device, addr))
146 goto fail2;
147
148 ctxt->sc_send_wr.next = NULL;
149 ctxt->sc_send_wr.wr_cqe = &ctxt->sc_cqe;
150 ctxt->sc_send_wr.sg_list = ctxt->sc_sges;
151 ctxt->sc_send_wr.send_flags = IB_SEND_SIGNALED;
152 ctxt->sc_cqe.done = svc_rdma_wc_send;
153 ctxt->sc_xprt_buf = buffer;
154 ctxt->sc_sges[0].addr = addr;
155
156 for (i = 0; i < rdma->sc_max_send_sges; i++)
157 ctxt->sc_sges[i].lkey = rdma->sc_pd->local_dma_lkey;
158 return ctxt;
159
160 fail2:
161 kfree(buffer);
162 fail1:
163 kfree(ctxt);
164 fail0:
165 return NULL;
166 }
167
168 /**
169 * svc_rdma_send_ctxts_destroy - Release all send_ctxt's for an xprt
170 * @rdma: svcxprt_rdma being torn down
171 *
172 */
svc_rdma_send_ctxts_destroy(struct svcxprt_rdma * rdma)173 void svc_rdma_send_ctxts_destroy(struct svcxprt_rdma *rdma)
174 {
175 struct svc_rdma_send_ctxt *ctxt;
176
177 while ((ctxt = svc_rdma_next_send_ctxt(&rdma->sc_send_ctxts))) {
178 list_del(&ctxt->sc_list);
179 ib_dma_unmap_single(rdma->sc_pd->device,
180 ctxt->sc_sges[0].addr,
181 rdma->sc_max_req_size,
182 DMA_TO_DEVICE);
183 kfree(ctxt->sc_xprt_buf);
184 kfree(ctxt);
185 }
186 }
187
188 /**
189 * svc_rdma_send_ctxt_get - Get a free send_ctxt
190 * @rdma: controlling svcxprt_rdma
191 *
192 * Returns a ready-to-use send_ctxt, or NULL if none are
193 * available and a fresh one cannot be allocated.
194 */
svc_rdma_send_ctxt_get(struct svcxprt_rdma * rdma)195 struct svc_rdma_send_ctxt *svc_rdma_send_ctxt_get(struct svcxprt_rdma *rdma)
196 {
197 struct svc_rdma_send_ctxt *ctxt;
198
199 spin_lock(&rdma->sc_send_lock);
200 ctxt = svc_rdma_next_send_ctxt(&rdma->sc_send_ctxts);
201 if (!ctxt)
202 goto out_empty;
203 list_del(&ctxt->sc_list);
204 spin_unlock(&rdma->sc_send_lock);
205
206 out:
207 ctxt->sc_send_wr.num_sge = 0;
208 ctxt->sc_cur_sge_no = 0;
209 ctxt->sc_page_count = 0;
210 return ctxt;
211
212 out_empty:
213 spin_unlock(&rdma->sc_send_lock);
214 ctxt = svc_rdma_send_ctxt_alloc(rdma);
215 if (!ctxt)
216 return NULL;
217 goto out;
218 }
219
220 /**
221 * svc_rdma_send_ctxt_put - Return send_ctxt to free list
222 * @rdma: controlling svcxprt_rdma
223 * @ctxt: object to return to the free list
224 *
225 * Pages left in sc_pages are DMA unmapped and released.
226 */
svc_rdma_send_ctxt_put(struct svcxprt_rdma * rdma,struct svc_rdma_send_ctxt * ctxt)227 void svc_rdma_send_ctxt_put(struct svcxprt_rdma *rdma,
228 struct svc_rdma_send_ctxt *ctxt)
229 {
230 struct ib_device *device = rdma->sc_cm_id->device;
231 unsigned int i;
232
233 /* The first SGE contains the transport header, which
234 * remains mapped until @ctxt is destroyed.
235 */
236 for (i = 1; i < ctxt->sc_send_wr.num_sge; i++)
237 ib_dma_unmap_page(device,
238 ctxt->sc_sges[i].addr,
239 ctxt->sc_sges[i].length,
240 DMA_TO_DEVICE);
241
242 for (i = 0; i < ctxt->sc_page_count; ++i)
243 put_page(ctxt->sc_pages[i]);
244
245 spin_lock(&rdma->sc_send_lock);
246 list_add(&ctxt->sc_list, &rdma->sc_send_ctxts);
247 spin_unlock(&rdma->sc_send_lock);
248 }
249
250 /**
251 * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC
252 * @cq: Completion Queue context
253 * @wc: Work Completion object
254 *
255 * NB: The svc_xprt/svcxprt_rdma is pinned whenever it's possible that
256 * the Send completion handler could be running.
257 */
svc_rdma_wc_send(struct ib_cq * cq,struct ib_wc * wc)258 static void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
259 {
260 struct svcxprt_rdma *rdma = cq->cq_context;
261 struct ib_cqe *cqe = wc->wr_cqe;
262 struct svc_rdma_send_ctxt *ctxt;
263
264 trace_svcrdma_wc_send(wc);
265
266 atomic_inc(&rdma->sc_sq_avail);
267 wake_up(&rdma->sc_send_wait);
268
269 ctxt = container_of(cqe, struct svc_rdma_send_ctxt, sc_cqe);
270 svc_rdma_send_ctxt_put(rdma, ctxt);
271
272 if (unlikely(wc->status != IB_WC_SUCCESS)) {
273 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
274 svc_xprt_enqueue(&rdma->sc_xprt);
275 }
276
277 svc_xprt_put(&rdma->sc_xprt);
278 }
279
280 /**
281 * svc_rdma_send - Post a single Send WR
282 * @rdma: transport on which to post the WR
283 * @wr: prepared Send WR to post
284 *
285 * Returns zero the Send WR was posted successfully. Otherwise, a
286 * negative errno is returned.
287 */
svc_rdma_send(struct svcxprt_rdma * rdma,struct ib_send_wr * wr)288 int svc_rdma_send(struct svcxprt_rdma *rdma, struct ib_send_wr *wr)
289 {
290 int ret;
291
292 might_sleep();
293
294 /* If the SQ is full, wait until an SQ entry is available */
295 while (1) {
296 if ((atomic_dec_return(&rdma->sc_sq_avail) < 0)) {
297 atomic_inc(&rdma_stat_sq_starve);
298 trace_svcrdma_sq_full(rdma);
299 atomic_inc(&rdma->sc_sq_avail);
300 wait_event(rdma->sc_send_wait,
301 atomic_read(&rdma->sc_sq_avail) > 1);
302 if (test_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags))
303 return -ENOTCONN;
304 trace_svcrdma_sq_retry(rdma);
305 continue;
306 }
307
308 svc_xprt_get(&rdma->sc_xprt);
309 trace_svcrdma_post_send(wr);
310 ret = ib_post_send(rdma->sc_qp, wr, NULL);
311 if (ret)
312 break;
313 return 0;
314 }
315
316 trace_svcrdma_sq_post_err(rdma, ret);
317 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
318 svc_xprt_put(&rdma->sc_xprt);
319 wake_up(&rdma->sc_send_wait);
320 return ret;
321 }
322
xdr_padsize(u32 len)323 static u32 xdr_padsize(u32 len)
324 {
325 return (len & 3) ? (4 - (len & 3)) : 0;
326 }
327
328 /* Returns length of transport header, in bytes.
329 */
svc_rdma_reply_hdr_len(__be32 * rdma_resp)330 static unsigned int svc_rdma_reply_hdr_len(__be32 *rdma_resp)
331 {
332 unsigned int nsegs;
333 __be32 *p;
334
335 p = rdma_resp;
336
337 /* RPC-over-RDMA V1 replies never have a Read list. */
338 p += rpcrdma_fixed_maxsz + 1;
339
340 /* Skip Write list. */
341 while (*p++ != xdr_zero) {
342 nsegs = be32_to_cpup(p++);
343 p += nsegs * rpcrdma_segment_maxsz;
344 }
345
346 /* Skip Reply chunk. */
347 if (*p++ != xdr_zero) {
348 nsegs = be32_to_cpup(p++);
349 p += nsegs * rpcrdma_segment_maxsz;
350 }
351
352 return (unsigned long)p - (unsigned long)rdma_resp;
353 }
354
355 /* One Write chunk is copied from Call transport header to Reply
356 * transport header. Each segment's length field is updated to
357 * reflect number of bytes consumed in the segment.
358 *
359 * Returns number of segments in this chunk.
360 */
xdr_encode_write_chunk(__be32 * dst,__be32 * src,unsigned int remaining)361 static unsigned int xdr_encode_write_chunk(__be32 *dst, __be32 *src,
362 unsigned int remaining)
363 {
364 unsigned int i, nsegs;
365 u32 seg_len;
366
367 /* Write list discriminator */
368 *dst++ = *src++;
369
370 /* number of segments in this chunk */
371 nsegs = be32_to_cpup(src);
372 *dst++ = *src++;
373
374 for (i = nsegs; i; i--) {
375 /* segment's RDMA handle */
376 *dst++ = *src++;
377
378 /* bytes returned in this segment */
379 seg_len = be32_to_cpu(*src);
380 if (remaining >= seg_len) {
381 /* entire segment was consumed */
382 *dst = *src;
383 remaining -= seg_len;
384 } else {
385 /* segment only partly filled */
386 *dst = cpu_to_be32(remaining);
387 remaining = 0;
388 }
389 dst++; src++;
390
391 /* segment's RDMA offset */
392 *dst++ = *src++;
393 *dst++ = *src++;
394 }
395
396 return nsegs;
397 }
398
399 /* The client provided a Write list in the Call message. Fill in
400 * the segments in the first Write chunk in the Reply's transport
401 * header with the number of bytes consumed in each segment.
402 * Remaining chunks are returned unused.
403 *
404 * Assumptions:
405 * - Client has provided only one Write chunk
406 */
svc_rdma_xdr_encode_write_list(__be32 * rdma_resp,__be32 * wr_ch,unsigned int consumed)407 static void svc_rdma_xdr_encode_write_list(__be32 *rdma_resp, __be32 *wr_ch,
408 unsigned int consumed)
409 {
410 unsigned int nsegs;
411 __be32 *p, *q;
412
413 /* RPC-over-RDMA V1 replies never have a Read list. */
414 p = rdma_resp + rpcrdma_fixed_maxsz + 1;
415
416 q = wr_ch;
417 while (*q != xdr_zero) {
418 nsegs = xdr_encode_write_chunk(p, q, consumed);
419 q += 2 + nsegs * rpcrdma_segment_maxsz;
420 p += 2 + nsegs * rpcrdma_segment_maxsz;
421 consumed = 0;
422 }
423
424 /* Terminate Write list */
425 *p++ = xdr_zero;
426
427 /* Reply chunk discriminator; may be replaced later */
428 *p = xdr_zero;
429 }
430
431 /* The client provided a Reply chunk in the Call message. Fill in
432 * the segments in the Reply chunk in the Reply message with the
433 * number of bytes consumed in each segment.
434 *
435 * Assumptions:
436 * - Reply can always fit in the provided Reply chunk
437 */
svc_rdma_xdr_encode_reply_chunk(__be32 * rdma_resp,__be32 * rp_ch,unsigned int consumed)438 static void svc_rdma_xdr_encode_reply_chunk(__be32 *rdma_resp, __be32 *rp_ch,
439 unsigned int consumed)
440 {
441 __be32 *p;
442
443 /* Find the Reply chunk in the Reply's xprt header.
444 * RPC-over-RDMA V1 replies never have a Read list.
445 */
446 p = rdma_resp + rpcrdma_fixed_maxsz + 1;
447
448 /* Skip past Write list */
449 while (*p++ != xdr_zero)
450 p += 1 + be32_to_cpup(p) * rpcrdma_segment_maxsz;
451
452 xdr_encode_write_chunk(p, rp_ch, consumed);
453 }
454
455 /* Parse the RPC Call's transport header.
456 */
svc_rdma_get_write_arrays(__be32 * rdma_argp,__be32 ** write,__be32 ** reply)457 static void svc_rdma_get_write_arrays(__be32 *rdma_argp,
458 __be32 **write, __be32 **reply)
459 {
460 __be32 *p;
461
462 p = rdma_argp + rpcrdma_fixed_maxsz;
463
464 /* Read list */
465 while (*p++ != xdr_zero)
466 p += 5;
467
468 /* Write list */
469 if (*p != xdr_zero) {
470 *write = p;
471 while (*p++ != xdr_zero)
472 p += 1 + be32_to_cpu(*p) * 4;
473 } else {
474 *write = NULL;
475 p++;
476 }
477
478 /* Reply chunk */
479 if (*p != xdr_zero)
480 *reply = p;
481 else
482 *reply = NULL;
483 }
484
svc_rdma_dma_map_page(struct svcxprt_rdma * rdma,struct svc_rdma_send_ctxt * ctxt,struct page * page,unsigned long offset,unsigned int len)485 static int svc_rdma_dma_map_page(struct svcxprt_rdma *rdma,
486 struct svc_rdma_send_ctxt *ctxt,
487 struct page *page,
488 unsigned long offset,
489 unsigned int len)
490 {
491 struct ib_device *dev = rdma->sc_cm_id->device;
492 dma_addr_t dma_addr;
493
494 dma_addr = ib_dma_map_page(dev, page, offset, len, DMA_TO_DEVICE);
495 if (ib_dma_mapping_error(dev, dma_addr))
496 goto out_maperr;
497
498 ctxt->sc_sges[ctxt->sc_cur_sge_no].addr = dma_addr;
499 ctxt->sc_sges[ctxt->sc_cur_sge_no].length = len;
500 ctxt->sc_send_wr.num_sge++;
501 return 0;
502
503 out_maperr:
504 trace_svcrdma_dma_map_page(rdma, page);
505 return -EIO;
506 }
507
508 /* ib_dma_map_page() is used here because svc_rdma_dma_unmap()
509 * handles DMA-unmap and it uses ib_dma_unmap_page() exclusively.
510 */
svc_rdma_dma_map_buf(struct svcxprt_rdma * rdma,struct svc_rdma_send_ctxt * ctxt,unsigned char * base,unsigned int len)511 static int svc_rdma_dma_map_buf(struct svcxprt_rdma *rdma,
512 struct svc_rdma_send_ctxt *ctxt,
513 unsigned char *base,
514 unsigned int len)
515 {
516 return svc_rdma_dma_map_page(rdma, ctxt, virt_to_page(base),
517 offset_in_page(base), len);
518 }
519
520 /**
521 * svc_rdma_sync_reply_hdr - DMA sync the transport header buffer
522 * @rdma: controlling transport
523 * @ctxt: send_ctxt for the Send WR
524 * @len: length of transport header
525 *
526 */
svc_rdma_sync_reply_hdr(struct svcxprt_rdma * rdma,struct svc_rdma_send_ctxt * ctxt,unsigned int len)527 void svc_rdma_sync_reply_hdr(struct svcxprt_rdma *rdma,
528 struct svc_rdma_send_ctxt *ctxt,
529 unsigned int len)
530 {
531 ctxt->sc_sges[0].length = len;
532 ctxt->sc_send_wr.num_sge++;
533 ib_dma_sync_single_for_device(rdma->sc_pd->device,
534 ctxt->sc_sges[0].addr, len,
535 DMA_TO_DEVICE);
536 }
537
538 /* If the xdr_buf has more elements than the device can
539 * transmit in a single RDMA Send, then the reply will
540 * have to be copied into a bounce buffer.
541 */
svc_rdma_pull_up_needed(struct svcxprt_rdma * rdma,struct xdr_buf * xdr,__be32 * wr_lst)542 static bool svc_rdma_pull_up_needed(struct svcxprt_rdma *rdma,
543 struct xdr_buf *xdr,
544 __be32 *wr_lst)
545 {
546 int elements;
547
548 /* xdr->head */
549 elements = 1;
550
551 /* xdr->pages */
552 if (!wr_lst) {
553 unsigned int remaining;
554 unsigned long pageoff;
555
556 pageoff = xdr->page_base & ~PAGE_MASK;
557 remaining = xdr->page_len;
558 while (remaining) {
559 ++elements;
560 remaining -= min_t(u32, PAGE_SIZE - pageoff,
561 remaining);
562 pageoff = 0;
563 }
564 }
565
566 /* xdr->tail */
567 if (xdr->tail[0].iov_len)
568 ++elements;
569
570 /* assume 1 SGE is needed for the transport header */
571 return elements >= rdma->sc_max_send_sges;
572 }
573
574 /* The device is not capable of sending the reply directly.
575 * Assemble the elements of @xdr into the transport header
576 * buffer.
577 */
svc_rdma_pull_up_reply_msg(struct svcxprt_rdma * rdma,struct svc_rdma_send_ctxt * ctxt,struct xdr_buf * xdr,__be32 * wr_lst)578 static int svc_rdma_pull_up_reply_msg(struct svcxprt_rdma *rdma,
579 struct svc_rdma_send_ctxt *ctxt,
580 struct xdr_buf *xdr, __be32 *wr_lst)
581 {
582 unsigned char *dst, *tailbase;
583 unsigned int taillen;
584
585 dst = ctxt->sc_xprt_buf;
586 dst += ctxt->sc_sges[0].length;
587
588 memcpy(dst, xdr->head[0].iov_base, xdr->head[0].iov_len);
589 dst += xdr->head[0].iov_len;
590
591 tailbase = xdr->tail[0].iov_base;
592 taillen = xdr->tail[0].iov_len;
593 if (wr_lst) {
594 u32 xdrpad;
595
596 xdrpad = xdr_padsize(xdr->page_len);
597 if (taillen && xdrpad) {
598 tailbase += xdrpad;
599 taillen -= xdrpad;
600 }
601 } else {
602 unsigned int len, remaining;
603 unsigned long pageoff;
604 struct page **ppages;
605
606 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
607 pageoff = xdr->page_base & ~PAGE_MASK;
608 remaining = xdr->page_len;
609 while (remaining) {
610 len = min_t(u32, PAGE_SIZE - pageoff, remaining);
611
612 memcpy(dst, page_address(*ppages) + pageoff, len);
613 remaining -= len;
614 dst += len;
615 pageoff = 0;
616 ppages++;
617 }
618 }
619
620 if (taillen)
621 memcpy(dst, tailbase, taillen);
622
623 ctxt->sc_sges[0].length += xdr->len;
624 ib_dma_sync_single_for_device(rdma->sc_pd->device,
625 ctxt->sc_sges[0].addr,
626 ctxt->sc_sges[0].length,
627 DMA_TO_DEVICE);
628
629 return 0;
630 }
631
632 /* svc_rdma_map_reply_msg - Map the buffer holding RPC message
633 * @rdma: controlling transport
634 * @ctxt: send_ctxt for the Send WR
635 * @xdr: prepared xdr_buf containing RPC message
636 * @wr_lst: pointer to Call header's Write list, or NULL
637 *
638 * Load the xdr_buf into the ctxt's sge array, and DMA map each
639 * element as it is added.
640 *
641 * Returns zero on success, or a negative errno on failure.
642 */
svc_rdma_map_reply_msg(struct svcxprt_rdma * rdma,struct svc_rdma_send_ctxt * ctxt,struct xdr_buf * xdr,__be32 * wr_lst)643 int svc_rdma_map_reply_msg(struct svcxprt_rdma *rdma,
644 struct svc_rdma_send_ctxt *ctxt,
645 struct xdr_buf *xdr, __be32 *wr_lst)
646 {
647 unsigned int len, remaining;
648 unsigned long page_off;
649 struct page **ppages;
650 unsigned char *base;
651 u32 xdr_pad;
652 int ret;
653
654 if (svc_rdma_pull_up_needed(rdma, xdr, wr_lst))
655 return svc_rdma_pull_up_reply_msg(rdma, ctxt, xdr, wr_lst);
656
657 ++ctxt->sc_cur_sge_no;
658 ret = svc_rdma_dma_map_buf(rdma, ctxt,
659 xdr->head[0].iov_base,
660 xdr->head[0].iov_len);
661 if (ret < 0)
662 return ret;
663
664 /* If a Write chunk is present, the xdr_buf's page list
665 * is not included inline. However the Upper Layer may
666 * have added XDR padding in the tail buffer, and that
667 * should not be included inline.
668 */
669 if (wr_lst) {
670 base = xdr->tail[0].iov_base;
671 len = xdr->tail[0].iov_len;
672 xdr_pad = xdr_padsize(xdr->page_len);
673
674 if (len && xdr_pad) {
675 base += xdr_pad;
676 len -= xdr_pad;
677 }
678
679 goto tail;
680 }
681
682 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
683 page_off = xdr->page_base & ~PAGE_MASK;
684 remaining = xdr->page_len;
685 while (remaining) {
686 len = min_t(u32, PAGE_SIZE - page_off, remaining);
687
688 ++ctxt->sc_cur_sge_no;
689 ret = svc_rdma_dma_map_page(rdma, ctxt, *ppages++,
690 page_off, len);
691 if (ret < 0)
692 return ret;
693
694 remaining -= len;
695 page_off = 0;
696 }
697
698 base = xdr->tail[0].iov_base;
699 len = xdr->tail[0].iov_len;
700 tail:
701 if (len) {
702 ++ctxt->sc_cur_sge_no;
703 ret = svc_rdma_dma_map_buf(rdma, ctxt, base, len);
704 if (ret < 0)
705 return ret;
706 }
707
708 return 0;
709 }
710
711 /* The svc_rqst and all resources it owns are released as soon as
712 * svc_rdma_sendto returns. Transfer pages under I/O to the ctxt
713 * so they are released by the Send completion handler.
714 */
svc_rdma_save_io_pages(struct svc_rqst * rqstp,struct svc_rdma_send_ctxt * ctxt)715 static void svc_rdma_save_io_pages(struct svc_rqst *rqstp,
716 struct svc_rdma_send_ctxt *ctxt)
717 {
718 int i, pages = rqstp->rq_next_page - rqstp->rq_respages;
719
720 ctxt->sc_page_count += pages;
721 for (i = 0; i < pages; i++) {
722 ctxt->sc_pages[i] = rqstp->rq_respages[i];
723 rqstp->rq_respages[i] = NULL;
724 }
725
726 /* Prevent svc_xprt_release from releasing pages in rq_pages */
727 rqstp->rq_next_page = rqstp->rq_respages;
728 }
729
730 /* Prepare the portion of the RPC Reply that will be transmitted
731 * via RDMA Send. The RPC-over-RDMA transport header is prepared
732 * in sc_sges[0], and the RPC xdr_buf is prepared in following sges.
733 *
734 * Depending on whether a Write list or Reply chunk is present,
735 * the server may send all, a portion of, or none of the xdr_buf.
736 * In the latter case, only the transport header (sc_sges[0]) is
737 * transmitted.
738 *
739 * RDMA Send is the last step of transmitting an RPC reply. Pages
740 * involved in the earlier RDMA Writes are here transferred out
741 * of the rqstp and into the sctxt's page array. These pages are
742 * DMA unmapped by each Write completion, but the subsequent Send
743 * completion finally releases these pages.
744 *
745 * Assumptions:
746 * - The Reply's transport header will never be larger than a page.
747 */
svc_rdma_send_reply_msg(struct svcxprt_rdma * rdma,struct svc_rdma_send_ctxt * sctxt,struct svc_rdma_recv_ctxt * rctxt,struct svc_rqst * rqstp,__be32 * wr_lst,__be32 * rp_ch)748 static int svc_rdma_send_reply_msg(struct svcxprt_rdma *rdma,
749 struct svc_rdma_send_ctxt *sctxt,
750 struct svc_rdma_recv_ctxt *rctxt,
751 struct svc_rqst *rqstp,
752 __be32 *wr_lst, __be32 *rp_ch)
753 {
754 int ret;
755
756 if (!rp_ch) {
757 ret = svc_rdma_map_reply_msg(rdma, sctxt,
758 &rqstp->rq_res, wr_lst);
759 if (ret < 0)
760 return ret;
761 }
762
763 svc_rdma_save_io_pages(rqstp, sctxt);
764
765 if (rctxt->rc_inv_rkey) {
766 sctxt->sc_send_wr.opcode = IB_WR_SEND_WITH_INV;
767 sctxt->sc_send_wr.ex.invalidate_rkey = rctxt->rc_inv_rkey;
768 } else {
769 sctxt->sc_send_wr.opcode = IB_WR_SEND;
770 }
771 dprintk("svcrdma: posting Send WR with %u sge(s)\n",
772 sctxt->sc_send_wr.num_sge);
773 return svc_rdma_send(rdma, &sctxt->sc_send_wr);
774 }
775
776 /* Given the client-provided Write and Reply chunks, the server was not
777 * able to form a complete reply. Return an RDMA_ERROR message so the
778 * client can retire this RPC transaction. As above, the Send completion
779 * routine releases payload pages that were part of a previous RDMA Write.
780 *
781 * Remote Invalidation is skipped for simplicity.
782 */
svc_rdma_send_error_msg(struct svcxprt_rdma * rdma,struct svc_rdma_send_ctxt * ctxt,struct svc_rqst * rqstp)783 static int svc_rdma_send_error_msg(struct svcxprt_rdma *rdma,
784 struct svc_rdma_send_ctxt *ctxt,
785 struct svc_rqst *rqstp)
786 {
787 __be32 *p;
788 int ret;
789
790 p = ctxt->sc_xprt_buf;
791 trace_svcrdma_err_chunk(*p);
792 p += 3;
793 *p++ = rdma_error;
794 *p = err_chunk;
795 svc_rdma_sync_reply_hdr(rdma, ctxt, RPCRDMA_HDRLEN_ERR);
796
797 svc_rdma_save_io_pages(rqstp, ctxt);
798
799 ctxt->sc_send_wr.opcode = IB_WR_SEND;
800 ret = svc_rdma_send(rdma, &ctxt->sc_send_wr);
801 if (ret) {
802 svc_rdma_send_ctxt_put(rdma, ctxt);
803 return ret;
804 }
805
806 return 0;
807 }
808
809 /**
810 * svc_rdma_sendto - Transmit an RPC reply
811 * @rqstp: processed RPC request, reply XDR already in ::rq_res
812 *
813 * Any resources still associated with @rqstp are released upon return.
814 * If no reply message was possible, the connection is closed.
815 *
816 * Returns:
817 * %0 if an RPC reply has been successfully posted,
818 * %-ENOMEM if a resource shortage occurred (connection is lost),
819 * %-ENOTCONN if posting failed (connection is lost).
820 */
svc_rdma_sendto(struct svc_rqst * rqstp)821 int svc_rdma_sendto(struct svc_rqst *rqstp)
822 {
823 struct svc_xprt *xprt = rqstp->rq_xprt;
824 struct svcxprt_rdma *rdma =
825 container_of(xprt, struct svcxprt_rdma, sc_xprt);
826 struct svc_rdma_recv_ctxt *rctxt = rqstp->rq_xprt_ctxt;
827 __be32 *p, *rdma_argp, *rdma_resp, *wr_lst, *rp_ch;
828 struct xdr_buf *xdr = &rqstp->rq_res;
829 struct svc_rdma_send_ctxt *sctxt;
830 int ret;
831
832 rdma_argp = rctxt->rc_recv_buf;
833 svc_rdma_get_write_arrays(rdma_argp, &wr_lst, &rp_ch);
834
835 /* Create the RDMA response header. xprt->xpt_mutex,
836 * acquired in svc_send(), serializes RPC replies. The
837 * code path below that inserts the credit grant value
838 * into each transport header runs only inside this
839 * critical section.
840 */
841 ret = -ENOMEM;
842 sctxt = svc_rdma_send_ctxt_get(rdma);
843 if (!sctxt)
844 goto err0;
845 rdma_resp = sctxt->sc_xprt_buf;
846
847 p = rdma_resp;
848 *p++ = *rdma_argp;
849 *p++ = *(rdma_argp + 1);
850 *p++ = rdma->sc_fc_credits;
851 *p++ = rp_ch ? rdma_nomsg : rdma_msg;
852
853 /* Start with empty chunks */
854 *p++ = xdr_zero;
855 *p++ = xdr_zero;
856 *p = xdr_zero;
857
858 if (wr_lst) {
859 /* XXX: Presume the client sent only one Write chunk */
860 unsigned long offset;
861 unsigned int length;
862
863 if (rctxt->rc_read_payload_length) {
864 offset = rctxt->rc_read_payload_offset;
865 length = rctxt->rc_read_payload_length;
866 } else {
867 offset = xdr->head[0].iov_len;
868 length = xdr->page_len;
869 }
870 ret = svc_rdma_send_write_chunk(rdma, wr_lst, xdr, offset,
871 length);
872 if (ret < 0)
873 goto err2;
874 svc_rdma_xdr_encode_write_list(rdma_resp, wr_lst, ret);
875 }
876 if (rp_ch) {
877 ret = svc_rdma_send_reply_chunk(rdma, rp_ch, wr_lst, xdr);
878 if (ret < 0)
879 goto err2;
880 svc_rdma_xdr_encode_reply_chunk(rdma_resp, rp_ch, ret);
881 }
882
883 svc_rdma_sync_reply_hdr(rdma, sctxt, svc_rdma_reply_hdr_len(rdma_resp));
884 ret = svc_rdma_send_reply_msg(rdma, sctxt, rctxt, rqstp,
885 wr_lst, rp_ch);
886 if (ret < 0)
887 goto err1;
888 return 0;
889
890 err2:
891 if (ret != -E2BIG && ret != -EINVAL)
892 goto err1;
893
894 ret = svc_rdma_send_error_msg(rdma, sctxt, rqstp);
895 if (ret < 0)
896 goto err1;
897 return 0;
898
899 err1:
900 svc_rdma_send_ctxt_put(rdma, sctxt);
901 err0:
902 trace_svcrdma_send_failed(rqstp, ret);
903 set_bit(XPT_CLOSE, &xprt->xpt_flags);
904 return -ENOTCONN;
905 }
906
907 /**
908 * svc_rdma_read_payload - special processing for a READ payload
909 * @rqstp: svc_rqst to operate on
910 * @offset: payload's byte offset in @xdr
911 * @length: size of payload, in bytes
912 *
913 * Returns zero on success.
914 *
915 * For the moment, just record the xdr_buf location of the READ
916 * payload. svc_rdma_sendto will use that location later when
917 * we actually send the payload.
918 */
svc_rdma_read_payload(struct svc_rqst * rqstp,unsigned int offset,unsigned int length)919 int svc_rdma_read_payload(struct svc_rqst *rqstp, unsigned int offset,
920 unsigned int length)
921 {
922 struct svc_rdma_recv_ctxt *rctxt = rqstp->rq_xprt_ctxt;
923
924 /* XXX: Just one READ payload slot for now, since our
925 * transport implementation currently supports only one
926 * Write chunk.
927 */
928 rctxt->rc_read_payload_offset = offset;
929 rctxt->rc_read_payload_length = length;
930
931 return 0;
932 }
933