1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2016-2018 Oracle. All rights reserved.
4 *
5 * Use the core R/W API to move RPC-over-RDMA Read and Write chunks.
6 */
7
8 #include <rdma/rw.h>
9
10 #include <linux/sunrpc/xdr.h>
11 #include <linux/sunrpc/rpc_rdma.h>
12 #include <linux/sunrpc/svc_rdma.h>
13
14 #include "xprt_rdma.h"
15 #include <trace/events/rpcrdma.h>
16
17 static void svc_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc);
18 static void svc_rdma_wc_read_done(struct ib_cq *cq, struct ib_wc *wc);
19
20 /* Each R/W context contains state for one chain of RDMA Read or
21 * Write Work Requests.
22 *
23 * Each WR chain handles a single contiguous server-side buffer,
24 * because scatterlist entries after the first have to start on
25 * page alignment. xdr_buf iovecs cannot guarantee alignment.
26 *
27 * Each WR chain handles only one R_key. Each RPC-over-RDMA segment
28 * from a client may contain a unique R_key, so each WR chain moves
29 * up to one segment at a time.
30 *
31 * The scatterlist makes this data structure over 4KB in size. To
32 * make it less likely to fail, and to handle the allocation for
33 * smaller I/O requests without disabling bottom-halves, these
34 * contexts are created on demand, but cached and reused until the
35 * controlling svcxprt_rdma is destroyed.
36 */
37 struct svc_rdma_rw_ctxt {
38 struct llist_node rw_node;
39 struct list_head rw_list;
40 struct rdma_rw_ctx rw_ctx;
41 unsigned int rw_nents;
42 struct sg_table rw_sg_table;
43 struct scatterlist rw_first_sgl[];
44 };
45
46 static inline struct svc_rdma_rw_ctxt *
svc_rdma_next_ctxt(struct list_head * list)47 svc_rdma_next_ctxt(struct list_head *list)
48 {
49 return list_first_entry_or_null(list, struct svc_rdma_rw_ctxt,
50 rw_list);
51 }
52
53 static struct svc_rdma_rw_ctxt *
svc_rdma_get_rw_ctxt(struct svcxprt_rdma * rdma,unsigned int sges)54 svc_rdma_get_rw_ctxt(struct svcxprt_rdma *rdma, unsigned int sges)
55 {
56 struct svc_rdma_rw_ctxt *ctxt;
57 struct llist_node *node;
58
59 spin_lock(&rdma->sc_rw_ctxt_lock);
60 node = llist_del_first(&rdma->sc_rw_ctxts);
61 spin_unlock(&rdma->sc_rw_ctxt_lock);
62 if (node) {
63 ctxt = llist_entry(node, struct svc_rdma_rw_ctxt, rw_node);
64 } else {
65 ctxt = kmalloc_node(struct_size(ctxt, rw_first_sgl, SG_CHUNK_SIZE),
66 GFP_KERNEL, ibdev_to_node(rdma->sc_cm_id->device));
67 if (!ctxt)
68 goto out_noctx;
69
70 INIT_LIST_HEAD(&ctxt->rw_list);
71 }
72
73 ctxt->rw_sg_table.sgl = ctxt->rw_first_sgl;
74 if (sg_alloc_table_chained(&ctxt->rw_sg_table, sges,
75 ctxt->rw_sg_table.sgl,
76 SG_CHUNK_SIZE))
77 goto out_free;
78 return ctxt;
79
80 out_free:
81 kfree(ctxt);
82 out_noctx:
83 trace_svcrdma_no_rwctx_err(rdma, sges);
84 return NULL;
85 }
86
__svc_rdma_put_rw_ctxt(struct svc_rdma_rw_ctxt * ctxt,struct llist_head * list)87 static void __svc_rdma_put_rw_ctxt(struct svc_rdma_rw_ctxt *ctxt,
88 struct llist_head *list)
89 {
90 sg_free_table_chained(&ctxt->rw_sg_table, SG_CHUNK_SIZE);
91 llist_add(&ctxt->rw_node, list);
92 }
93
svc_rdma_put_rw_ctxt(struct svcxprt_rdma * rdma,struct svc_rdma_rw_ctxt * ctxt)94 static void svc_rdma_put_rw_ctxt(struct svcxprt_rdma *rdma,
95 struct svc_rdma_rw_ctxt *ctxt)
96 {
97 __svc_rdma_put_rw_ctxt(ctxt, &rdma->sc_rw_ctxts);
98 }
99
100 /**
101 * svc_rdma_destroy_rw_ctxts - Free accumulated R/W contexts
102 * @rdma: transport about to be destroyed
103 *
104 */
svc_rdma_destroy_rw_ctxts(struct svcxprt_rdma * rdma)105 void svc_rdma_destroy_rw_ctxts(struct svcxprt_rdma *rdma)
106 {
107 struct svc_rdma_rw_ctxt *ctxt;
108 struct llist_node *node;
109
110 while ((node = llist_del_first(&rdma->sc_rw_ctxts)) != NULL) {
111 ctxt = llist_entry(node, struct svc_rdma_rw_ctxt, rw_node);
112 kfree(ctxt);
113 }
114 }
115
116 /**
117 * svc_rdma_rw_ctx_init - Prepare a R/W context for I/O
118 * @rdma: controlling transport instance
119 * @ctxt: R/W context to prepare
120 * @offset: RDMA offset
121 * @handle: RDMA tag/handle
122 * @direction: I/O direction
123 *
124 * Returns on success, the number of WQEs that will be needed
125 * on the workqueue, or a negative errno.
126 */
svc_rdma_rw_ctx_init(struct svcxprt_rdma * rdma,struct svc_rdma_rw_ctxt * ctxt,u64 offset,u32 handle,enum dma_data_direction direction)127 static int svc_rdma_rw_ctx_init(struct svcxprt_rdma *rdma,
128 struct svc_rdma_rw_ctxt *ctxt,
129 u64 offset, u32 handle,
130 enum dma_data_direction direction)
131 {
132 int ret;
133
134 ret = rdma_rw_ctx_init(&ctxt->rw_ctx, rdma->sc_qp, rdma->sc_port_num,
135 ctxt->rw_sg_table.sgl, ctxt->rw_nents,
136 0, offset, handle, direction);
137 if (unlikely(ret < 0)) {
138 svc_rdma_put_rw_ctxt(rdma, ctxt);
139 trace_svcrdma_dma_map_rw_err(rdma, ctxt->rw_nents, ret);
140 }
141 return ret;
142 }
143
144 /* A chunk context tracks all I/O for moving one Read or Write
145 * chunk. This is a set of rdma_rw's that handle data movement
146 * for all segments of one chunk.
147 *
148 * These are small, acquired with a single allocator call, and
149 * no more than one is needed per chunk. They are allocated on
150 * demand, and not cached.
151 */
152 struct svc_rdma_chunk_ctxt {
153 struct rpc_rdma_cid cc_cid;
154 struct ib_cqe cc_cqe;
155 struct svcxprt_rdma *cc_rdma;
156 struct list_head cc_rwctxts;
157 ktime_t cc_posttime;
158 int cc_sqecount;
159 enum ib_wc_status cc_status;
160 struct completion cc_done;
161 };
162
svc_rdma_cc_cid_init(struct svcxprt_rdma * rdma,struct rpc_rdma_cid * cid)163 static void svc_rdma_cc_cid_init(struct svcxprt_rdma *rdma,
164 struct rpc_rdma_cid *cid)
165 {
166 cid->ci_queue_id = rdma->sc_sq_cq->res.id;
167 cid->ci_completion_id = atomic_inc_return(&rdma->sc_completion_ids);
168 }
169
svc_rdma_cc_init(struct svcxprt_rdma * rdma,struct svc_rdma_chunk_ctxt * cc)170 static void svc_rdma_cc_init(struct svcxprt_rdma *rdma,
171 struct svc_rdma_chunk_ctxt *cc)
172 {
173 svc_rdma_cc_cid_init(rdma, &cc->cc_cid);
174 cc->cc_rdma = rdma;
175
176 INIT_LIST_HEAD(&cc->cc_rwctxts);
177 cc->cc_sqecount = 0;
178 }
179
180 /*
181 * The consumed rw_ctx's are cleaned and placed on a local llist so
182 * that only one atomic llist operation is needed to put them all
183 * back on the free list.
184 */
svc_rdma_cc_release(struct svc_rdma_chunk_ctxt * cc,enum dma_data_direction dir)185 static void svc_rdma_cc_release(struct svc_rdma_chunk_ctxt *cc,
186 enum dma_data_direction dir)
187 {
188 struct svcxprt_rdma *rdma = cc->cc_rdma;
189 struct llist_node *first, *last;
190 struct svc_rdma_rw_ctxt *ctxt;
191 LLIST_HEAD(free);
192
193 trace_svcrdma_cc_release(&cc->cc_cid, cc->cc_sqecount);
194
195 first = last = NULL;
196 while ((ctxt = svc_rdma_next_ctxt(&cc->cc_rwctxts)) != NULL) {
197 list_del(&ctxt->rw_list);
198
199 rdma_rw_ctx_destroy(&ctxt->rw_ctx, rdma->sc_qp,
200 rdma->sc_port_num, ctxt->rw_sg_table.sgl,
201 ctxt->rw_nents, dir);
202 __svc_rdma_put_rw_ctxt(ctxt, &free);
203
204 ctxt->rw_node.next = first;
205 first = &ctxt->rw_node;
206 if (!last)
207 last = first;
208 }
209 if (first)
210 llist_add_batch(first, last, &rdma->sc_rw_ctxts);
211 }
212
213 /* State for sending a Write or Reply chunk.
214 * - Tracks progress of writing one chunk over all its segments
215 * - Stores arguments for the SGL constructor functions
216 */
217 struct svc_rdma_write_info {
218 const struct svc_rdma_chunk *wi_chunk;
219
220 /* write state of this chunk */
221 unsigned int wi_seg_off;
222 unsigned int wi_seg_no;
223
224 /* SGL constructor arguments */
225 const struct xdr_buf *wi_xdr;
226 unsigned char *wi_base;
227 unsigned int wi_next_off;
228
229 struct svc_rdma_chunk_ctxt wi_cc;
230 };
231
232 static struct svc_rdma_write_info *
svc_rdma_write_info_alloc(struct svcxprt_rdma * rdma,const struct svc_rdma_chunk * chunk)233 svc_rdma_write_info_alloc(struct svcxprt_rdma *rdma,
234 const struct svc_rdma_chunk *chunk)
235 {
236 struct svc_rdma_write_info *info;
237
238 info = kmalloc_node(sizeof(*info), GFP_KERNEL,
239 ibdev_to_node(rdma->sc_cm_id->device));
240 if (!info)
241 return info;
242
243 info->wi_chunk = chunk;
244 info->wi_seg_off = 0;
245 info->wi_seg_no = 0;
246 svc_rdma_cc_init(rdma, &info->wi_cc);
247 info->wi_cc.cc_cqe.done = svc_rdma_write_done;
248 return info;
249 }
250
svc_rdma_write_info_free(struct svc_rdma_write_info * info)251 static void svc_rdma_write_info_free(struct svc_rdma_write_info *info)
252 {
253 svc_rdma_cc_release(&info->wi_cc, DMA_TO_DEVICE);
254 kfree(info);
255 }
256
257 /**
258 * svc_rdma_write_done - Write chunk completion
259 * @cq: controlling Completion Queue
260 * @wc: Work Completion
261 *
262 * Pages under I/O are freed by a subsequent Send completion.
263 */
svc_rdma_write_done(struct ib_cq * cq,struct ib_wc * wc)264 static void svc_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc)
265 {
266 struct ib_cqe *cqe = wc->wr_cqe;
267 struct svc_rdma_chunk_ctxt *cc =
268 container_of(cqe, struct svc_rdma_chunk_ctxt, cc_cqe);
269 struct svcxprt_rdma *rdma = cc->cc_rdma;
270 struct svc_rdma_write_info *info =
271 container_of(cc, struct svc_rdma_write_info, wi_cc);
272
273 switch (wc->status) {
274 case IB_WC_SUCCESS:
275 trace_svcrdma_wc_write(wc, &cc->cc_cid);
276 break;
277 case IB_WC_WR_FLUSH_ERR:
278 trace_svcrdma_wc_write_flush(wc, &cc->cc_cid);
279 break;
280 default:
281 trace_svcrdma_wc_write_err(wc, &cc->cc_cid);
282 }
283
284 svc_rdma_wake_send_waiters(rdma, cc->cc_sqecount);
285
286 if (unlikely(wc->status != IB_WC_SUCCESS))
287 svc_xprt_deferred_close(&rdma->sc_xprt);
288
289 svc_rdma_write_info_free(info);
290 }
291
292 /* State for pulling a Read chunk.
293 */
294 struct svc_rdma_read_info {
295 struct svc_rqst *ri_rqst;
296 struct svc_rdma_recv_ctxt *ri_readctxt;
297 unsigned int ri_pageno;
298 unsigned int ri_pageoff;
299 unsigned int ri_totalbytes;
300
301 struct svc_rdma_chunk_ctxt ri_cc;
302 };
303
304 static struct svc_rdma_read_info *
svc_rdma_read_info_alloc(struct svcxprt_rdma * rdma)305 svc_rdma_read_info_alloc(struct svcxprt_rdma *rdma)
306 {
307 struct svc_rdma_read_info *info;
308
309 info = kmalloc_node(sizeof(*info), GFP_KERNEL,
310 ibdev_to_node(rdma->sc_cm_id->device));
311 if (!info)
312 return info;
313
314 svc_rdma_cc_init(rdma, &info->ri_cc);
315 info->ri_cc.cc_cqe.done = svc_rdma_wc_read_done;
316 return info;
317 }
318
svc_rdma_read_info_free(struct svc_rdma_read_info * info)319 static void svc_rdma_read_info_free(struct svc_rdma_read_info *info)
320 {
321 svc_rdma_cc_release(&info->ri_cc, DMA_FROM_DEVICE);
322 kfree(info);
323 }
324
325 /**
326 * svc_rdma_wc_read_done - Handle completion of an RDMA Read ctx
327 * @cq: controlling Completion Queue
328 * @wc: Work Completion
329 *
330 */
svc_rdma_wc_read_done(struct ib_cq * cq,struct ib_wc * wc)331 static void svc_rdma_wc_read_done(struct ib_cq *cq, struct ib_wc *wc)
332 {
333 struct ib_cqe *cqe = wc->wr_cqe;
334 struct svc_rdma_chunk_ctxt *cc =
335 container_of(cqe, struct svc_rdma_chunk_ctxt, cc_cqe);
336 struct svc_rdma_read_info *info;
337
338 switch (wc->status) {
339 case IB_WC_SUCCESS:
340 info = container_of(cc, struct svc_rdma_read_info, ri_cc);
341 trace_svcrdma_wc_read(wc, &cc->cc_cid, info->ri_totalbytes,
342 cc->cc_posttime);
343 break;
344 case IB_WC_WR_FLUSH_ERR:
345 trace_svcrdma_wc_read_flush(wc, &cc->cc_cid);
346 break;
347 default:
348 trace_svcrdma_wc_read_err(wc, &cc->cc_cid);
349 }
350
351 svc_rdma_wake_send_waiters(cc->cc_rdma, cc->cc_sqecount);
352 cc->cc_status = wc->status;
353 complete(&cc->cc_done);
354 return;
355 }
356
357 /*
358 * Assumptions:
359 * - If ib_post_send() succeeds, only one completion is expected,
360 * even if one or more WRs are flushed. This is true when posting
361 * an rdma_rw_ctx or when posting a single signaled WR.
362 */
svc_rdma_post_chunk_ctxt(struct svc_rdma_chunk_ctxt * cc)363 static int svc_rdma_post_chunk_ctxt(struct svc_rdma_chunk_ctxt *cc)
364 {
365 struct svcxprt_rdma *rdma = cc->cc_rdma;
366 struct ib_send_wr *first_wr;
367 const struct ib_send_wr *bad_wr;
368 struct list_head *tmp;
369 struct ib_cqe *cqe;
370 int ret;
371
372 might_sleep();
373
374 if (cc->cc_sqecount > rdma->sc_sq_depth)
375 return -EINVAL;
376
377 first_wr = NULL;
378 cqe = &cc->cc_cqe;
379 list_for_each(tmp, &cc->cc_rwctxts) {
380 struct svc_rdma_rw_ctxt *ctxt;
381
382 ctxt = list_entry(tmp, struct svc_rdma_rw_ctxt, rw_list);
383 first_wr = rdma_rw_ctx_wrs(&ctxt->rw_ctx, rdma->sc_qp,
384 rdma->sc_port_num, cqe, first_wr);
385 cqe = NULL;
386 }
387
388 do {
389 if (atomic_sub_return(cc->cc_sqecount,
390 &rdma->sc_sq_avail) > 0) {
391 cc->cc_posttime = ktime_get();
392 ret = ib_post_send(rdma->sc_qp, first_wr, &bad_wr);
393 if (ret)
394 break;
395 return 0;
396 }
397
398 percpu_counter_inc(&svcrdma_stat_sq_starve);
399 trace_svcrdma_sq_full(rdma);
400 atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
401 wait_event(rdma->sc_send_wait,
402 atomic_read(&rdma->sc_sq_avail) > cc->cc_sqecount);
403 trace_svcrdma_sq_retry(rdma);
404 } while (1);
405
406 trace_svcrdma_sq_post_err(rdma, ret);
407 svc_xprt_deferred_close(&rdma->sc_xprt);
408
409 /* If even one was posted, there will be a completion. */
410 if (bad_wr != first_wr)
411 return 0;
412
413 atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
414 wake_up(&rdma->sc_send_wait);
415 return -ENOTCONN;
416 }
417
418 /* Build and DMA-map an SGL that covers one kvec in an xdr_buf
419 */
svc_rdma_vec_to_sg(struct svc_rdma_write_info * info,unsigned int len,struct svc_rdma_rw_ctxt * ctxt)420 static void svc_rdma_vec_to_sg(struct svc_rdma_write_info *info,
421 unsigned int len,
422 struct svc_rdma_rw_ctxt *ctxt)
423 {
424 struct scatterlist *sg = ctxt->rw_sg_table.sgl;
425
426 sg_set_buf(&sg[0], info->wi_base, len);
427 info->wi_base += len;
428
429 ctxt->rw_nents = 1;
430 }
431
432 /* Build and DMA-map an SGL that covers part of an xdr_buf's pagelist.
433 */
svc_rdma_pagelist_to_sg(struct svc_rdma_write_info * info,unsigned int remaining,struct svc_rdma_rw_ctxt * ctxt)434 static void svc_rdma_pagelist_to_sg(struct svc_rdma_write_info *info,
435 unsigned int remaining,
436 struct svc_rdma_rw_ctxt *ctxt)
437 {
438 unsigned int sge_no, sge_bytes, page_off, page_no;
439 const struct xdr_buf *xdr = info->wi_xdr;
440 struct scatterlist *sg;
441 struct page **page;
442
443 page_off = info->wi_next_off + xdr->page_base;
444 page_no = page_off >> PAGE_SHIFT;
445 page_off = offset_in_page(page_off);
446 page = xdr->pages + page_no;
447 info->wi_next_off += remaining;
448 sg = ctxt->rw_sg_table.sgl;
449 sge_no = 0;
450 do {
451 sge_bytes = min_t(unsigned int, remaining,
452 PAGE_SIZE - page_off);
453 sg_set_page(sg, *page, sge_bytes, page_off);
454
455 remaining -= sge_bytes;
456 sg = sg_next(sg);
457 page_off = 0;
458 sge_no++;
459 page++;
460 } while (remaining);
461
462 ctxt->rw_nents = sge_no;
463 }
464
465 /* Construct RDMA Write WRs to send a portion of an xdr_buf containing
466 * an RPC Reply.
467 */
468 static int
svc_rdma_build_writes(struct svc_rdma_write_info * info,void (* constructor)(struct svc_rdma_write_info * info,unsigned int len,struct svc_rdma_rw_ctxt * ctxt),unsigned int remaining)469 svc_rdma_build_writes(struct svc_rdma_write_info *info,
470 void (*constructor)(struct svc_rdma_write_info *info,
471 unsigned int len,
472 struct svc_rdma_rw_ctxt *ctxt),
473 unsigned int remaining)
474 {
475 struct svc_rdma_chunk_ctxt *cc = &info->wi_cc;
476 struct svcxprt_rdma *rdma = cc->cc_rdma;
477 const struct svc_rdma_segment *seg;
478 struct svc_rdma_rw_ctxt *ctxt;
479 int ret;
480
481 do {
482 unsigned int write_len;
483 u64 offset;
484
485 if (info->wi_seg_no >= info->wi_chunk->ch_segcount)
486 goto out_overflow;
487
488 seg = &info->wi_chunk->ch_segments[info->wi_seg_no];
489 write_len = min(remaining, seg->rs_length - info->wi_seg_off);
490 if (!write_len)
491 goto out_overflow;
492 ctxt = svc_rdma_get_rw_ctxt(rdma,
493 (write_len >> PAGE_SHIFT) + 2);
494 if (!ctxt)
495 return -ENOMEM;
496
497 constructor(info, write_len, ctxt);
498 offset = seg->rs_offset + info->wi_seg_off;
499 ret = svc_rdma_rw_ctx_init(rdma, ctxt, offset, seg->rs_handle,
500 DMA_TO_DEVICE);
501 if (ret < 0)
502 return -EIO;
503 percpu_counter_inc(&svcrdma_stat_write);
504
505 list_add(&ctxt->rw_list, &cc->cc_rwctxts);
506 cc->cc_sqecount += ret;
507 if (write_len == seg->rs_length - info->wi_seg_off) {
508 info->wi_seg_no++;
509 info->wi_seg_off = 0;
510 } else {
511 info->wi_seg_off += write_len;
512 }
513 remaining -= write_len;
514 } while (remaining);
515
516 return 0;
517
518 out_overflow:
519 trace_svcrdma_small_wrch_err(rdma, remaining, info->wi_seg_no,
520 info->wi_chunk->ch_segcount);
521 return -E2BIG;
522 }
523
524 /**
525 * svc_rdma_iov_write - Construct RDMA Writes from an iov
526 * @info: pointer to write arguments
527 * @iov: kvec to write
528 *
529 * Returns:
530 * On success, returns zero
531 * %-E2BIG if the client-provided Write chunk is too small
532 * %-ENOMEM if a resource has been exhausted
533 * %-EIO if an rdma-rw error occurred
534 */
svc_rdma_iov_write(struct svc_rdma_write_info * info,const struct kvec * iov)535 static int svc_rdma_iov_write(struct svc_rdma_write_info *info,
536 const struct kvec *iov)
537 {
538 info->wi_base = iov->iov_base;
539 return svc_rdma_build_writes(info, svc_rdma_vec_to_sg,
540 iov->iov_len);
541 }
542
543 /**
544 * svc_rdma_pages_write - Construct RDMA Writes from pages
545 * @info: pointer to write arguments
546 * @xdr: xdr_buf with pages to write
547 * @offset: offset into the content of @xdr
548 * @length: number of bytes to write
549 *
550 * Returns:
551 * On success, returns zero
552 * %-E2BIG if the client-provided Write chunk is too small
553 * %-ENOMEM if a resource has been exhausted
554 * %-EIO if an rdma-rw error occurred
555 */
svc_rdma_pages_write(struct svc_rdma_write_info * info,const struct xdr_buf * xdr,unsigned int offset,unsigned long length)556 static int svc_rdma_pages_write(struct svc_rdma_write_info *info,
557 const struct xdr_buf *xdr,
558 unsigned int offset,
559 unsigned long length)
560 {
561 info->wi_xdr = xdr;
562 info->wi_next_off = offset - xdr->head[0].iov_len;
563 return svc_rdma_build_writes(info, svc_rdma_pagelist_to_sg,
564 length);
565 }
566
567 /**
568 * svc_rdma_xb_write - Construct RDMA Writes to write an xdr_buf
569 * @xdr: xdr_buf to write
570 * @data: pointer to write arguments
571 *
572 * Returns:
573 * On success, returns zero
574 * %-E2BIG if the client-provided Write chunk is too small
575 * %-ENOMEM if a resource has been exhausted
576 * %-EIO if an rdma-rw error occurred
577 */
svc_rdma_xb_write(const struct xdr_buf * xdr,void * data)578 static int svc_rdma_xb_write(const struct xdr_buf *xdr, void *data)
579 {
580 struct svc_rdma_write_info *info = data;
581 int ret;
582
583 if (xdr->head[0].iov_len) {
584 ret = svc_rdma_iov_write(info, &xdr->head[0]);
585 if (ret < 0)
586 return ret;
587 }
588
589 if (xdr->page_len) {
590 ret = svc_rdma_pages_write(info, xdr, xdr->head[0].iov_len,
591 xdr->page_len);
592 if (ret < 0)
593 return ret;
594 }
595
596 if (xdr->tail[0].iov_len) {
597 ret = svc_rdma_iov_write(info, &xdr->tail[0]);
598 if (ret < 0)
599 return ret;
600 }
601
602 return xdr->len;
603 }
604
605 /**
606 * svc_rdma_send_write_chunk - Write all segments in a Write chunk
607 * @rdma: controlling RDMA transport
608 * @chunk: Write chunk provided by the client
609 * @xdr: xdr_buf containing the data payload
610 *
611 * Returns a non-negative number of bytes the chunk consumed, or
612 * %-E2BIG if the payload was larger than the Write chunk,
613 * %-EINVAL if client provided too many segments,
614 * %-ENOMEM if rdma_rw context pool was exhausted,
615 * %-ENOTCONN if posting failed (connection is lost),
616 * %-EIO if rdma_rw initialization failed (DMA mapping, etc).
617 */
svc_rdma_send_write_chunk(struct svcxprt_rdma * rdma,const struct svc_rdma_chunk * chunk,const struct xdr_buf * xdr)618 int svc_rdma_send_write_chunk(struct svcxprt_rdma *rdma,
619 const struct svc_rdma_chunk *chunk,
620 const struct xdr_buf *xdr)
621 {
622 struct svc_rdma_write_info *info;
623 struct svc_rdma_chunk_ctxt *cc;
624 int ret;
625
626 info = svc_rdma_write_info_alloc(rdma, chunk);
627 if (!info)
628 return -ENOMEM;
629 cc = &info->wi_cc;
630
631 ret = svc_rdma_xb_write(xdr, info);
632 if (ret != xdr->len)
633 goto out_err;
634
635 trace_svcrdma_post_write_chunk(&cc->cc_cid, cc->cc_sqecount);
636 ret = svc_rdma_post_chunk_ctxt(cc);
637 if (ret < 0)
638 goto out_err;
639 return xdr->len;
640
641 out_err:
642 svc_rdma_write_info_free(info);
643 return ret;
644 }
645
646 /**
647 * svc_rdma_send_reply_chunk - Write all segments in the Reply chunk
648 * @rdma: controlling RDMA transport
649 * @rctxt: Write and Reply chunks from client
650 * @xdr: xdr_buf containing an RPC Reply
651 *
652 * Returns a non-negative number of bytes the chunk consumed, or
653 * %-E2BIG if the payload was larger than the Reply chunk,
654 * %-EINVAL if client provided too many segments,
655 * %-ENOMEM if rdma_rw context pool was exhausted,
656 * %-ENOTCONN if posting failed (connection is lost),
657 * %-EIO if rdma_rw initialization failed (DMA mapping, etc).
658 */
svc_rdma_send_reply_chunk(struct svcxprt_rdma * rdma,const struct svc_rdma_recv_ctxt * rctxt,const struct xdr_buf * xdr)659 int svc_rdma_send_reply_chunk(struct svcxprt_rdma *rdma,
660 const struct svc_rdma_recv_ctxt *rctxt,
661 const struct xdr_buf *xdr)
662 {
663 struct svc_rdma_write_info *info;
664 struct svc_rdma_chunk_ctxt *cc;
665 struct svc_rdma_chunk *chunk;
666 int ret;
667
668 if (pcl_is_empty(&rctxt->rc_reply_pcl))
669 return 0;
670
671 chunk = pcl_first_chunk(&rctxt->rc_reply_pcl);
672 info = svc_rdma_write_info_alloc(rdma, chunk);
673 if (!info)
674 return -ENOMEM;
675 cc = &info->wi_cc;
676
677 ret = pcl_process_nonpayloads(&rctxt->rc_write_pcl, xdr,
678 svc_rdma_xb_write, info);
679 if (ret < 0)
680 goto out_err;
681
682 trace_svcrdma_post_reply_chunk(&cc->cc_cid, cc->cc_sqecount);
683 ret = svc_rdma_post_chunk_ctxt(cc);
684 if (ret < 0)
685 goto out_err;
686
687 return xdr->len;
688
689 out_err:
690 svc_rdma_write_info_free(info);
691 return ret;
692 }
693
694 /**
695 * svc_rdma_build_read_segment - Build RDMA Read WQEs to pull one RDMA segment
696 * @info: context for ongoing I/O
697 * @segment: co-ordinates of remote memory to be read
698 *
699 * Returns:
700 * %0: the Read WR chain was constructed successfully
701 * %-EINVAL: there were not enough rq_pages to finish
702 * %-ENOMEM: allocating a local resources failed
703 * %-EIO: a DMA mapping error occurred
704 */
svc_rdma_build_read_segment(struct svc_rdma_read_info * info,const struct svc_rdma_segment * segment)705 static int svc_rdma_build_read_segment(struct svc_rdma_read_info *info,
706 const struct svc_rdma_segment *segment)
707 {
708 struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
709 struct svc_rdma_chunk_ctxt *cc = &info->ri_cc;
710 struct svc_rqst *rqstp = info->ri_rqst;
711 unsigned int sge_no, seg_len, len;
712 struct svc_rdma_rw_ctxt *ctxt;
713 struct scatterlist *sg;
714 int ret;
715
716 len = segment->rs_length;
717 sge_no = PAGE_ALIGN(info->ri_pageoff + len) >> PAGE_SHIFT;
718 ctxt = svc_rdma_get_rw_ctxt(cc->cc_rdma, sge_no);
719 if (!ctxt)
720 return -ENOMEM;
721 ctxt->rw_nents = sge_no;
722
723 sg = ctxt->rw_sg_table.sgl;
724 for (sge_no = 0; sge_no < ctxt->rw_nents; sge_no++) {
725 seg_len = min_t(unsigned int, len,
726 PAGE_SIZE - info->ri_pageoff);
727
728 if (!info->ri_pageoff)
729 head->rc_page_count++;
730
731 sg_set_page(sg, rqstp->rq_pages[info->ri_pageno],
732 seg_len, info->ri_pageoff);
733 sg = sg_next(sg);
734
735 info->ri_pageoff += seg_len;
736 if (info->ri_pageoff == PAGE_SIZE) {
737 info->ri_pageno++;
738 info->ri_pageoff = 0;
739 }
740 len -= seg_len;
741
742 /* Safety check */
743 if (len &&
744 &rqstp->rq_pages[info->ri_pageno + 1] > rqstp->rq_page_end)
745 goto out_overrun;
746 }
747
748 ret = svc_rdma_rw_ctx_init(cc->cc_rdma, ctxt, segment->rs_offset,
749 segment->rs_handle, DMA_FROM_DEVICE);
750 if (ret < 0)
751 return -EIO;
752 percpu_counter_inc(&svcrdma_stat_read);
753
754 list_add(&ctxt->rw_list, &cc->cc_rwctxts);
755 cc->cc_sqecount += ret;
756 return 0;
757
758 out_overrun:
759 trace_svcrdma_page_overrun_err(cc->cc_rdma, rqstp, info->ri_pageno);
760 return -EINVAL;
761 }
762
763 /**
764 * svc_rdma_build_read_chunk - Build RDMA Read WQEs to pull one RDMA chunk
765 * @info: context for ongoing I/O
766 * @chunk: Read chunk to pull
767 *
768 * Return values:
769 * %0: the Read WR chain was constructed successfully
770 * %-EINVAL: there were not enough resources to finish
771 * %-ENOMEM: allocating a local resources failed
772 * %-EIO: a DMA mapping error occurred
773 */
svc_rdma_build_read_chunk(struct svc_rdma_read_info * info,const struct svc_rdma_chunk * chunk)774 static int svc_rdma_build_read_chunk(struct svc_rdma_read_info *info,
775 const struct svc_rdma_chunk *chunk)
776 {
777 const struct svc_rdma_segment *segment;
778 int ret;
779
780 ret = -EINVAL;
781 pcl_for_each_segment(segment, chunk) {
782 ret = svc_rdma_build_read_segment(info, segment);
783 if (ret < 0)
784 break;
785 info->ri_totalbytes += segment->rs_length;
786 }
787 return ret;
788 }
789
790 /**
791 * svc_rdma_copy_inline_range - Copy part of the inline content into pages
792 * @info: context for RDMA Reads
793 * @offset: offset into the Receive buffer of region to copy
794 * @remaining: length of region to copy
795 *
796 * Take a page at a time from rqstp->rq_pages and copy the inline
797 * content from the Receive buffer into that page. Update
798 * info->ri_pageno and info->ri_pageoff so that the next RDMA Read
799 * result will land contiguously with the copied content.
800 *
801 * Return values:
802 * %0: Inline content was successfully copied
803 * %-EINVAL: offset or length was incorrect
804 */
svc_rdma_copy_inline_range(struct svc_rdma_read_info * info,unsigned int offset,unsigned int remaining)805 static int svc_rdma_copy_inline_range(struct svc_rdma_read_info *info,
806 unsigned int offset,
807 unsigned int remaining)
808 {
809 struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
810 unsigned char *dst, *src = head->rc_recv_buf;
811 struct svc_rqst *rqstp = info->ri_rqst;
812 unsigned int page_no, numpages;
813
814 numpages = PAGE_ALIGN(info->ri_pageoff + remaining) >> PAGE_SHIFT;
815 for (page_no = 0; page_no < numpages; page_no++) {
816 unsigned int page_len;
817
818 page_len = min_t(unsigned int, remaining,
819 PAGE_SIZE - info->ri_pageoff);
820
821 if (!info->ri_pageoff)
822 head->rc_page_count++;
823
824 dst = page_address(rqstp->rq_pages[info->ri_pageno]);
825 memcpy(dst + info->ri_pageno, src + offset, page_len);
826
827 info->ri_totalbytes += page_len;
828 info->ri_pageoff += page_len;
829 if (info->ri_pageoff == PAGE_SIZE) {
830 info->ri_pageno++;
831 info->ri_pageoff = 0;
832 }
833 remaining -= page_len;
834 offset += page_len;
835 }
836
837 return -EINVAL;
838 }
839
840 /**
841 * svc_rdma_read_multiple_chunks - Construct RDMA Reads to pull data item Read chunks
842 * @info: context for RDMA Reads
843 *
844 * The chunk data lands in rqstp->rq_arg as a series of contiguous pages,
845 * like an incoming TCP call.
846 *
847 * Return values:
848 * %0: RDMA Read WQEs were successfully built
849 * %-EINVAL: client provided too many chunks or segments,
850 * %-ENOMEM: rdma_rw context pool was exhausted,
851 * %-ENOTCONN: posting failed (connection is lost),
852 * %-EIO: rdma_rw initialization failed (DMA mapping, etc).
853 */
svc_rdma_read_multiple_chunks(struct svc_rdma_read_info * info)854 static noinline int svc_rdma_read_multiple_chunks(struct svc_rdma_read_info *info)
855 {
856 struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
857 const struct svc_rdma_pcl *pcl = &head->rc_read_pcl;
858 struct xdr_buf *buf = &info->ri_rqst->rq_arg;
859 struct svc_rdma_chunk *chunk, *next;
860 unsigned int start, length;
861 int ret;
862
863 start = 0;
864 chunk = pcl_first_chunk(pcl);
865 length = chunk->ch_position;
866 ret = svc_rdma_copy_inline_range(info, start, length);
867 if (ret < 0)
868 return ret;
869
870 pcl_for_each_chunk(chunk, pcl) {
871 ret = svc_rdma_build_read_chunk(info, chunk);
872 if (ret < 0)
873 return ret;
874
875 next = pcl_next_chunk(pcl, chunk);
876 if (!next)
877 break;
878
879 start += length;
880 length = next->ch_position - info->ri_totalbytes;
881 ret = svc_rdma_copy_inline_range(info, start, length);
882 if (ret < 0)
883 return ret;
884 }
885
886 start += length;
887 length = head->rc_byte_len - start;
888 ret = svc_rdma_copy_inline_range(info, start, length);
889 if (ret < 0)
890 return ret;
891
892 buf->len += info->ri_totalbytes;
893 buf->buflen += info->ri_totalbytes;
894
895 buf->head[0].iov_base = page_address(info->ri_rqst->rq_pages[0]);
896 buf->head[0].iov_len = min_t(size_t, PAGE_SIZE, info->ri_totalbytes);
897 buf->pages = &info->ri_rqst->rq_pages[1];
898 buf->page_len = info->ri_totalbytes - buf->head[0].iov_len;
899 return 0;
900 }
901
902 /**
903 * svc_rdma_read_data_item - Construct RDMA Reads to pull data item Read chunks
904 * @info: context for RDMA Reads
905 *
906 * The chunk data lands in the page list of rqstp->rq_arg.pages.
907 *
908 * Currently NFSD does not look at the rqstp->rq_arg.tail[0] kvec.
909 * Therefore, XDR round-up of the Read chunk and trailing
910 * inline content must both be added at the end of the pagelist.
911 *
912 * Return values:
913 * %0: RDMA Read WQEs were successfully built
914 * %-EINVAL: client provided too many chunks or segments,
915 * %-ENOMEM: rdma_rw context pool was exhausted,
916 * %-ENOTCONN: posting failed (connection is lost),
917 * %-EIO: rdma_rw initialization failed (DMA mapping, etc).
918 */
svc_rdma_read_data_item(struct svc_rdma_read_info * info)919 static int svc_rdma_read_data_item(struct svc_rdma_read_info *info)
920 {
921 struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
922 struct xdr_buf *buf = &info->ri_rqst->rq_arg;
923 struct svc_rdma_chunk *chunk;
924 unsigned int length;
925 int ret;
926
927 chunk = pcl_first_chunk(&head->rc_read_pcl);
928 ret = svc_rdma_build_read_chunk(info, chunk);
929 if (ret < 0)
930 goto out;
931
932 /* Split the Receive buffer between the head and tail
933 * buffers at Read chunk's position. XDR roundup of the
934 * chunk is not included in either the pagelist or in
935 * the tail.
936 */
937 buf->tail[0].iov_base = buf->head[0].iov_base + chunk->ch_position;
938 buf->tail[0].iov_len = buf->head[0].iov_len - chunk->ch_position;
939 buf->head[0].iov_len = chunk->ch_position;
940
941 /* Read chunk may need XDR roundup (see RFC 8166, s. 3.4.5.2).
942 *
943 * If the client already rounded up the chunk length, the
944 * length does not change. Otherwise, the length of the page
945 * list is increased to include XDR round-up.
946 *
947 * Currently these chunks always start at page offset 0,
948 * thus the rounded-up length never crosses a page boundary.
949 */
950 buf->pages = &info->ri_rqst->rq_pages[0];
951 length = xdr_align_size(chunk->ch_length);
952 buf->page_len = length;
953 buf->len += length;
954 buf->buflen += length;
955
956 out:
957 return ret;
958 }
959
960 /**
961 * svc_rdma_read_chunk_range - Build RDMA Read WQEs for portion of a chunk
962 * @info: context for RDMA Reads
963 * @chunk: parsed Call chunk to pull
964 * @offset: offset of region to pull
965 * @length: length of region to pull
966 *
967 * Return values:
968 * %0: RDMA Read WQEs were successfully built
969 * %-EINVAL: there were not enough resources to finish
970 * %-ENOMEM: rdma_rw context pool was exhausted,
971 * %-ENOTCONN: posting failed (connection is lost),
972 * %-EIO: rdma_rw initialization failed (DMA mapping, etc).
973 */
svc_rdma_read_chunk_range(struct svc_rdma_read_info * info,const struct svc_rdma_chunk * chunk,unsigned int offset,unsigned int length)974 static int svc_rdma_read_chunk_range(struct svc_rdma_read_info *info,
975 const struct svc_rdma_chunk *chunk,
976 unsigned int offset, unsigned int length)
977 {
978 const struct svc_rdma_segment *segment;
979 int ret;
980
981 ret = -EINVAL;
982 pcl_for_each_segment(segment, chunk) {
983 struct svc_rdma_segment dummy;
984
985 if (offset > segment->rs_length) {
986 offset -= segment->rs_length;
987 continue;
988 }
989
990 dummy.rs_handle = segment->rs_handle;
991 dummy.rs_length = min_t(u32, length, segment->rs_length) - offset;
992 dummy.rs_offset = segment->rs_offset + offset;
993
994 ret = svc_rdma_build_read_segment(info, &dummy);
995 if (ret < 0)
996 break;
997
998 info->ri_totalbytes += dummy.rs_length;
999 length -= dummy.rs_length;
1000 offset = 0;
1001 }
1002 return ret;
1003 }
1004
1005 /**
1006 * svc_rdma_read_call_chunk - Build RDMA Read WQEs to pull a Long Message
1007 * @info: context for RDMA Reads
1008 *
1009 * Return values:
1010 * %0: RDMA Read WQEs were successfully built
1011 * %-EINVAL: there were not enough resources to finish
1012 * %-ENOMEM: rdma_rw context pool was exhausted,
1013 * %-ENOTCONN: posting failed (connection is lost),
1014 * %-EIO: rdma_rw initialization failed (DMA mapping, etc).
1015 */
svc_rdma_read_call_chunk(struct svc_rdma_read_info * info)1016 static int svc_rdma_read_call_chunk(struct svc_rdma_read_info *info)
1017 {
1018 struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
1019 const struct svc_rdma_chunk *call_chunk =
1020 pcl_first_chunk(&head->rc_call_pcl);
1021 const struct svc_rdma_pcl *pcl = &head->rc_read_pcl;
1022 struct svc_rdma_chunk *chunk, *next;
1023 unsigned int start, length;
1024 int ret;
1025
1026 if (pcl_is_empty(pcl))
1027 return svc_rdma_build_read_chunk(info, call_chunk);
1028
1029 start = 0;
1030 chunk = pcl_first_chunk(pcl);
1031 length = chunk->ch_position;
1032 ret = svc_rdma_read_chunk_range(info, call_chunk, start, length);
1033 if (ret < 0)
1034 return ret;
1035
1036 pcl_for_each_chunk(chunk, pcl) {
1037 ret = svc_rdma_build_read_chunk(info, chunk);
1038 if (ret < 0)
1039 return ret;
1040
1041 next = pcl_next_chunk(pcl, chunk);
1042 if (!next)
1043 break;
1044
1045 start += length;
1046 length = next->ch_position - info->ri_totalbytes;
1047 ret = svc_rdma_read_chunk_range(info, call_chunk,
1048 start, length);
1049 if (ret < 0)
1050 return ret;
1051 }
1052
1053 start += length;
1054 length = call_chunk->ch_length - start;
1055 return svc_rdma_read_chunk_range(info, call_chunk, start, length);
1056 }
1057
1058 /**
1059 * svc_rdma_read_special - Build RDMA Read WQEs to pull a Long Message
1060 * @info: context for RDMA Reads
1061 *
1062 * The start of the data lands in the first page just after the
1063 * Transport header, and the rest lands in rqstp->rq_arg.pages.
1064 *
1065 * Assumptions:
1066 * - A PZRC is never sent in an RDMA_MSG message, though it's
1067 * allowed by spec.
1068 *
1069 * Return values:
1070 * %0: RDMA Read WQEs were successfully built
1071 * %-EINVAL: client provided too many chunks or segments,
1072 * %-ENOMEM: rdma_rw context pool was exhausted,
1073 * %-ENOTCONN: posting failed (connection is lost),
1074 * %-EIO: rdma_rw initialization failed (DMA mapping, etc).
1075 */
svc_rdma_read_special(struct svc_rdma_read_info * info)1076 static noinline int svc_rdma_read_special(struct svc_rdma_read_info *info)
1077 {
1078 struct xdr_buf *buf = &info->ri_rqst->rq_arg;
1079 int ret;
1080
1081 ret = svc_rdma_read_call_chunk(info);
1082 if (ret < 0)
1083 goto out;
1084
1085 buf->len += info->ri_totalbytes;
1086 buf->buflen += info->ri_totalbytes;
1087
1088 buf->head[0].iov_base = page_address(info->ri_rqst->rq_pages[0]);
1089 buf->head[0].iov_len = min_t(size_t, PAGE_SIZE, info->ri_totalbytes);
1090 buf->pages = &info->ri_rqst->rq_pages[1];
1091 buf->page_len = info->ri_totalbytes - buf->head[0].iov_len;
1092
1093 out:
1094 return ret;
1095 }
1096
1097 /**
1098 * svc_rdma_process_read_list - Pull list of Read chunks from the client
1099 * @rdma: controlling RDMA transport
1100 * @rqstp: set of pages to use as Read sink buffers
1101 * @head: pages under I/O collect here
1102 *
1103 * The RPC/RDMA protocol assumes that the upper layer's XDR decoders
1104 * pull each Read chunk as they decode an incoming RPC message.
1105 *
1106 * On Linux, however, the server needs to have a fully-constructed RPC
1107 * message in rqstp->rq_arg when there is a positive return code from
1108 * ->xpo_recvfrom. So the Read list is safety-checked immediately when
1109 * it is received, then here the whole Read list is pulled all at once.
1110 * The ingress RPC message is fully reconstructed once all associated
1111 * RDMA Reads have completed.
1112 *
1113 * Return values:
1114 * %1: all needed RDMA Reads were posted successfully,
1115 * %-EINVAL: client provided too many chunks or segments,
1116 * %-ENOMEM: rdma_rw context pool was exhausted,
1117 * %-ENOTCONN: posting failed (connection is lost),
1118 * %-EIO: rdma_rw initialization failed (DMA mapping, etc).
1119 */
svc_rdma_process_read_list(struct svcxprt_rdma * rdma,struct svc_rqst * rqstp,struct svc_rdma_recv_ctxt * head)1120 int svc_rdma_process_read_list(struct svcxprt_rdma *rdma,
1121 struct svc_rqst *rqstp,
1122 struct svc_rdma_recv_ctxt *head)
1123 {
1124 struct svc_rdma_read_info *info;
1125 struct svc_rdma_chunk_ctxt *cc;
1126 int ret;
1127
1128 info = svc_rdma_read_info_alloc(rdma);
1129 if (!info)
1130 return -ENOMEM;
1131 cc = &info->ri_cc;
1132 info->ri_rqst = rqstp;
1133 info->ri_readctxt = head;
1134 info->ri_pageno = 0;
1135 info->ri_pageoff = 0;
1136 info->ri_totalbytes = 0;
1137
1138 if (pcl_is_empty(&head->rc_call_pcl)) {
1139 if (head->rc_read_pcl.cl_count == 1)
1140 ret = svc_rdma_read_data_item(info);
1141 else
1142 ret = svc_rdma_read_multiple_chunks(info);
1143 } else
1144 ret = svc_rdma_read_special(info);
1145 if (ret < 0)
1146 goto out_err;
1147
1148 trace_svcrdma_post_read_chunk(&cc->cc_cid, cc->cc_sqecount);
1149 init_completion(&cc->cc_done);
1150 ret = svc_rdma_post_chunk_ctxt(cc);
1151 if (ret < 0)
1152 goto out_err;
1153
1154 ret = 1;
1155 wait_for_completion(&cc->cc_done);
1156 if (cc->cc_status != IB_WC_SUCCESS)
1157 ret = -EIO;
1158
1159 /* rq_respages starts after the last arg page */
1160 rqstp->rq_respages = &rqstp->rq_pages[head->rc_page_count];
1161 rqstp->rq_next_page = rqstp->rq_respages + 1;
1162
1163 /* Ensure svc_rdma_recv_ctxt_put() does not try to release pages */
1164 head->rc_page_count = 0;
1165
1166 out_err:
1167 svc_rdma_read_info_free(info);
1168 return ret;
1169 }
1170