xref: /linux/net/sunrpc/xprtrdma/svc_rdma_rw.c (revision e6a901a00822659181c93c86d8bbc2a17779fddc)
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 	unsigned int		rw_first_sgl_nents;
43 	struct sg_table		rw_sg_table;
44 	struct scatterlist	rw_first_sgl[];
45 };
46 
47 static inline struct svc_rdma_rw_ctxt *
48 svc_rdma_next_ctxt(struct list_head *list)
49 {
50 	return list_first_entry_or_null(list, struct svc_rdma_rw_ctxt,
51 					rw_list);
52 }
53 
54 static struct svc_rdma_rw_ctxt *
55 svc_rdma_get_rw_ctxt(struct svcxprt_rdma *rdma, unsigned int sges)
56 {
57 	struct ib_device *dev = rdma->sc_cm_id->device;
58 	unsigned int first_sgl_nents = dev->attrs.max_send_sge;
59 	struct svc_rdma_rw_ctxt *ctxt;
60 	struct llist_node *node;
61 
62 	spin_lock(&rdma->sc_rw_ctxt_lock);
63 	node = llist_del_first(&rdma->sc_rw_ctxts);
64 	spin_unlock(&rdma->sc_rw_ctxt_lock);
65 	if (node) {
66 		ctxt = llist_entry(node, struct svc_rdma_rw_ctxt, rw_node);
67 	} else {
68 		ctxt = kmalloc_node(struct_size(ctxt, rw_first_sgl, first_sgl_nents),
69 				    GFP_KERNEL, ibdev_to_node(dev));
70 		if (!ctxt)
71 			goto out_noctx;
72 
73 		INIT_LIST_HEAD(&ctxt->rw_list);
74 		ctxt->rw_first_sgl_nents = first_sgl_nents;
75 	}
76 
77 	ctxt->rw_sg_table.sgl = ctxt->rw_first_sgl;
78 	if (sg_alloc_table_chained(&ctxt->rw_sg_table, sges,
79 				   ctxt->rw_sg_table.sgl,
80 				   first_sgl_nents))
81 		goto out_free;
82 	return ctxt;
83 
84 out_free:
85 	kfree(ctxt);
86 out_noctx:
87 	trace_svcrdma_rwctx_empty(rdma, sges);
88 	return NULL;
89 }
90 
91 static void __svc_rdma_put_rw_ctxt(struct svc_rdma_rw_ctxt *ctxt,
92 				   struct llist_head *list)
93 {
94 	sg_free_table_chained(&ctxt->rw_sg_table, ctxt->rw_first_sgl_nents);
95 	llist_add(&ctxt->rw_node, list);
96 }
97 
98 static void svc_rdma_put_rw_ctxt(struct svcxprt_rdma *rdma,
99 				 struct svc_rdma_rw_ctxt *ctxt)
100 {
101 	__svc_rdma_put_rw_ctxt(ctxt, &rdma->sc_rw_ctxts);
102 }
103 
104 /**
105  * svc_rdma_destroy_rw_ctxts - Free accumulated R/W contexts
106  * @rdma: transport about to be destroyed
107  *
108  */
109 void svc_rdma_destroy_rw_ctxts(struct svcxprt_rdma *rdma)
110 {
111 	struct svc_rdma_rw_ctxt *ctxt;
112 	struct llist_node *node;
113 
114 	while ((node = llist_del_first(&rdma->sc_rw_ctxts)) != NULL) {
115 		ctxt = llist_entry(node, struct svc_rdma_rw_ctxt, rw_node);
116 		kfree(ctxt);
117 	}
118 }
119 
120 /**
121  * svc_rdma_rw_ctx_init - Prepare a R/W context for I/O
122  * @rdma: controlling transport instance
123  * @ctxt: R/W context to prepare
124  * @offset: RDMA offset
125  * @handle: RDMA tag/handle
126  * @direction: I/O direction
127  *
128  * Returns on success, the number of WQEs that will be needed
129  * on the workqueue, or a negative errno.
130  */
131 static int svc_rdma_rw_ctx_init(struct svcxprt_rdma *rdma,
132 				struct svc_rdma_rw_ctxt *ctxt,
133 				u64 offset, u32 handle,
134 				enum dma_data_direction direction)
135 {
136 	int ret;
137 
138 	ret = rdma_rw_ctx_init(&ctxt->rw_ctx, rdma->sc_qp, rdma->sc_port_num,
139 			       ctxt->rw_sg_table.sgl, ctxt->rw_nents,
140 			       0, offset, handle, direction);
141 	if (unlikely(ret < 0)) {
142 		trace_svcrdma_dma_map_rw_err(rdma, offset, handle,
143 					     ctxt->rw_nents, ret);
144 		svc_rdma_put_rw_ctxt(rdma, ctxt);
145 	}
146 	return ret;
147 }
148 
149 /**
150  * svc_rdma_cc_init - Initialize an svc_rdma_chunk_ctxt
151  * @rdma: controlling transport instance
152  * @cc: svc_rdma_chunk_ctxt to be initialized
153  */
154 void svc_rdma_cc_init(struct svcxprt_rdma *rdma,
155 		      struct svc_rdma_chunk_ctxt *cc)
156 {
157 	struct rpc_rdma_cid *cid = &cc->cc_cid;
158 
159 	if (unlikely(!cid->ci_completion_id))
160 		svc_rdma_send_cid_init(rdma, cid);
161 
162 	INIT_LIST_HEAD(&cc->cc_rwctxts);
163 	cc->cc_sqecount = 0;
164 }
165 
166 /**
167  * svc_rdma_cc_release - Release resources held by a svc_rdma_chunk_ctxt
168  * @rdma: controlling transport instance
169  * @cc: svc_rdma_chunk_ctxt to be released
170  * @dir: DMA direction
171  */
172 void svc_rdma_cc_release(struct svcxprt_rdma *rdma,
173 			 struct svc_rdma_chunk_ctxt *cc,
174 			 enum dma_data_direction dir)
175 {
176 	struct llist_node *first, *last;
177 	struct svc_rdma_rw_ctxt *ctxt;
178 	LLIST_HEAD(free);
179 
180 	trace_svcrdma_cc_release(&cc->cc_cid, cc->cc_sqecount);
181 
182 	first = last = NULL;
183 	while ((ctxt = svc_rdma_next_ctxt(&cc->cc_rwctxts)) != NULL) {
184 		list_del(&ctxt->rw_list);
185 
186 		rdma_rw_ctx_destroy(&ctxt->rw_ctx, rdma->sc_qp,
187 				    rdma->sc_port_num, ctxt->rw_sg_table.sgl,
188 				    ctxt->rw_nents, dir);
189 		__svc_rdma_put_rw_ctxt(ctxt, &free);
190 
191 		ctxt->rw_node.next = first;
192 		first = &ctxt->rw_node;
193 		if (!last)
194 			last = first;
195 	}
196 	if (first)
197 		llist_add_batch(first, last, &rdma->sc_rw_ctxts);
198 }
199 
200 static struct svc_rdma_write_info *
201 svc_rdma_write_info_alloc(struct svcxprt_rdma *rdma,
202 			  const struct svc_rdma_chunk *chunk)
203 {
204 	struct svc_rdma_write_info *info;
205 
206 	info = kzalloc_node(sizeof(*info), GFP_KERNEL,
207 			    ibdev_to_node(rdma->sc_cm_id->device));
208 	if (!info)
209 		return info;
210 
211 	info->wi_rdma = rdma;
212 	info->wi_chunk = chunk;
213 	svc_rdma_cc_init(rdma, &info->wi_cc);
214 	info->wi_cc.cc_cqe.done = svc_rdma_write_done;
215 	return info;
216 }
217 
218 static void svc_rdma_write_info_free_async(struct work_struct *work)
219 {
220 	struct svc_rdma_write_info *info;
221 
222 	info = container_of(work, struct svc_rdma_write_info, wi_work);
223 	svc_rdma_cc_release(info->wi_rdma, &info->wi_cc, DMA_TO_DEVICE);
224 	kfree(info);
225 }
226 
227 static void svc_rdma_write_info_free(struct svc_rdma_write_info *info)
228 {
229 	INIT_WORK(&info->wi_work, svc_rdma_write_info_free_async);
230 	queue_work(svcrdma_wq, &info->wi_work);
231 }
232 
233 /**
234  * svc_rdma_write_chunk_release - Release Write chunk I/O resources
235  * @rdma: controlling transport
236  * @ctxt: Send context that is being released
237  */
238 void svc_rdma_write_chunk_release(struct svcxprt_rdma *rdma,
239 				  struct svc_rdma_send_ctxt *ctxt)
240 {
241 	struct svc_rdma_write_info *info;
242 	struct svc_rdma_chunk_ctxt *cc;
243 
244 	while (!list_empty(&ctxt->sc_write_info_list)) {
245 		info = list_first_entry(&ctxt->sc_write_info_list,
246 					struct svc_rdma_write_info, wi_list);
247 		list_del(&info->wi_list);
248 
249 		cc = &info->wi_cc;
250 		svc_rdma_wake_send_waiters(rdma, cc->cc_sqecount);
251 		svc_rdma_write_info_free(info);
252 	}
253 }
254 
255 /**
256  * svc_rdma_reply_chunk_release - Release Reply chunk I/O resources
257  * @rdma: controlling transport
258  * @ctxt: Send context that is being released
259  */
260 void svc_rdma_reply_chunk_release(struct svcxprt_rdma *rdma,
261 				  struct svc_rdma_send_ctxt *ctxt)
262 {
263 	struct svc_rdma_chunk_ctxt *cc = &ctxt->sc_reply_info.wi_cc;
264 
265 	if (!cc->cc_sqecount)
266 		return;
267 	svc_rdma_cc_release(rdma, cc, DMA_TO_DEVICE);
268 }
269 
270 /**
271  * svc_rdma_reply_done - Reply chunk Write completion handler
272  * @cq: controlling Completion Queue
273  * @wc: Work Completion report
274  *
275  * Pages under I/O are released by a subsequent Send completion.
276  */
277 static void svc_rdma_reply_done(struct ib_cq *cq, struct ib_wc *wc)
278 {
279 	struct ib_cqe *cqe = wc->wr_cqe;
280 	struct svc_rdma_chunk_ctxt *cc =
281 			container_of(cqe, struct svc_rdma_chunk_ctxt, cc_cqe);
282 	struct svcxprt_rdma *rdma = cq->cq_context;
283 
284 	switch (wc->status) {
285 	case IB_WC_SUCCESS:
286 		trace_svcrdma_wc_reply(&cc->cc_cid);
287 		return;
288 	case IB_WC_WR_FLUSH_ERR:
289 		trace_svcrdma_wc_reply_flush(wc, &cc->cc_cid);
290 		break;
291 	default:
292 		trace_svcrdma_wc_reply_err(wc, &cc->cc_cid);
293 	}
294 
295 	svc_xprt_deferred_close(&rdma->sc_xprt);
296 }
297 
298 /**
299  * svc_rdma_write_done - Write chunk completion
300  * @cq: controlling Completion Queue
301  * @wc: Work Completion
302  *
303  * Pages under I/O are freed by a subsequent Send completion.
304  */
305 static void svc_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc)
306 {
307 	struct svcxprt_rdma *rdma = cq->cq_context;
308 	struct ib_cqe *cqe = wc->wr_cqe;
309 	struct svc_rdma_chunk_ctxt *cc =
310 			container_of(cqe, struct svc_rdma_chunk_ctxt, cc_cqe);
311 
312 	switch (wc->status) {
313 	case IB_WC_SUCCESS:
314 		trace_svcrdma_wc_write(&cc->cc_cid);
315 		return;
316 	case IB_WC_WR_FLUSH_ERR:
317 		trace_svcrdma_wc_write_flush(wc, &cc->cc_cid);
318 		break;
319 	default:
320 		trace_svcrdma_wc_write_err(wc, &cc->cc_cid);
321 	}
322 
323 	/* The RDMA Write has flushed, so the client won't get
324 	 * some of the outgoing RPC message. Signal the loss
325 	 * to the client by closing the connection.
326 	 */
327 	svc_xprt_deferred_close(&rdma->sc_xprt);
328 }
329 
330 /**
331  * svc_rdma_wc_read_done - Handle completion of an RDMA Read ctx
332  * @cq: controlling Completion Queue
333  * @wc: Work Completion
334  *
335  */
336 static void svc_rdma_wc_read_done(struct ib_cq *cq, struct ib_wc *wc)
337 {
338 	struct svcxprt_rdma *rdma = cq->cq_context;
339 	struct ib_cqe *cqe = wc->wr_cqe;
340 	struct svc_rdma_chunk_ctxt *cc =
341 			container_of(cqe, struct svc_rdma_chunk_ctxt, cc_cqe);
342 	struct svc_rdma_recv_ctxt *ctxt;
343 
344 	svc_rdma_wake_send_waiters(rdma, cc->cc_sqecount);
345 
346 	ctxt = container_of(cc, struct svc_rdma_recv_ctxt, rc_cc);
347 	switch (wc->status) {
348 	case IB_WC_SUCCESS:
349 		trace_svcrdma_wc_read(wc, &cc->cc_cid, ctxt->rc_readbytes,
350 				      cc->cc_posttime);
351 
352 		spin_lock(&rdma->sc_rq_dto_lock);
353 		list_add_tail(&ctxt->rc_list, &rdma->sc_read_complete_q);
354 		/* the unlock pairs with the smp_rmb in svc_xprt_ready */
355 		set_bit(XPT_DATA, &rdma->sc_xprt.xpt_flags);
356 		spin_unlock(&rdma->sc_rq_dto_lock);
357 		svc_xprt_enqueue(&rdma->sc_xprt);
358 		return;
359 	case IB_WC_WR_FLUSH_ERR:
360 		trace_svcrdma_wc_read_flush(wc, &cc->cc_cid);
361 		break;
362 	default:
363 		trace_svcrdma_wc_read_err(wc, &cc->cc_cid);
364 	}
365 
366 	/* The RDMA Read has flushed, so the incoming RPC message
367 	 * cannot be constructed and must be dropped. Signal the
368 	 * loss to the client by closing the connection.
369 	 */
370 	svc_rdma_cc_release(rdma, cc, DMA_FROM_DEVICE);
371 	svc_rdma_recv_ctxt_put(rdma, ctxt);
372 	svc_xprt_deferred_close(&rdma->sc_xprt);
373 }
374 
375 /*
376  * Assumptions:
377  * - If ib_post_send() succeeds, only one completion is expected,
378  *   even if one or more WRs are flushed. This is true when posting
379  *   an rdma_rw_ctx or when posting a single signaled WR.
380  */
381 static int svc_rdma_post_chunk_ctxt(struct svcxprt_rdma *rdma,
382 				    struct svc_rdma_chunk_ctxt *cc)
383 {
384 	struct ib_send_wr *first_wr;
385 	const struct ib_send_wr *bad_wr;
386 	struct list_head *tmp;
387 	struct ib_cqe *cqe;
388 	int ret;
389 
390 	might_sleep();
391 
392 	if (cc->cc_sqecount > rdma->sc_sq_depth)
393 		return -EINVAL;
394 
395 	first_wr = NULL;
396 	cqe = &cc->cc_cqe;
397 	list_for_each(tmp, &cc->cc_rwctxts) {
398 		struct svc_rdma_rw_ctxt *ctxt;
399 
400 		ctxt = list_entry(tmp, struct svc_rdma_rw_ctxt, rw_list);
401 		first_wr = rdma_rw_ctx_wrs(&ctxt->rw_ctx, rdma->sc_qp,
402 					   rdma->sc_port_num, cqe, first_wr);
403 		cqe = NULL;
404 	}
405 
406 	do {
407 		if (atomic_sub_return(cc->cc_sqecount,
408 				      &rdma->sc_sq_avail) > 0) {
409 			cc->cc_posttime = ktime_get();
410 			ret = ib_post_send(rdma->sc_qp, first_wr, &bad_wr);
411 			if (ret)
412 				break;
413 			return 0;
414 		}
415 
416 		percpu_counter_inc(&svcrdma_stat_sq_starve);
417 		trace_svcrdma_sq_full(rdma, &cc->cc_cid);
418 		atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
419 		wait_event(rdma->sc_send_wait,
420 			   atomic_read(&rdma->sc_sq_avail) > cc->cc_sqecount);
421 		trace_svcrdma_sq_retry(rdma, &cc->cc_cid);
422 	} while (1);
423 
424 	trace_svcrdma_sq_post_err(rdma, &cc->cc_cid, ret);
425 	svc_xprt_deferred_close(&rdma->sc_xprt);
426 
427 	/* If even one was posted, there will be a completion. */
428 	if (bad_wr != first_wr)
429 		return 0;
430 
431 	atomic_add(cc->cc_sqecount, &rdma->sc_sq_avail);
432 	wake_up(&rdma->sc_send_wait);
433 	return -ENOTCONN;
434 }
435 
436 /* Build and DMA-map an SGL that covers one kvec in an xdr_buf
437  */
438 static void svc_rdma_vec_to_sg(struct svc_rdma_write_info *info,
439 			       unsigned int len,
440 			       struct svc_rdma_rw_ctxt *ctxt)
441 {
442 	struct scatterlist *sg = ctxt->rw_sg_table.sgl;
443 
444 	sg_set_buf(&sg[0], info->wi_base, len);
445 	info->wi_base += len;
446 
447 	ctxt->rw_nents = 1;
448 }
449 
450 /* Build and DMA-map an SGL that covers part of an xdr_buf's pagelist.
451  */
452 static void svc_rdma_pagelist_to_sg(struct svc_rdma_write_info *info,
453 				    unsigned int remaining,
454 				    struct svc_rdma_rw_ctxt *ctxt)
455 {
456 	unsigned int sge_no, sge_bytes, page_off, page_no;
457 	const struct xdr_buf *xdr = info->wi_xdr;
458 	struct scatterlist *sg;
459 	struct page **page;
460 
461 	page_off = info->wi_next_off + xdr->page_base;
462 	page_no = page_off >> PAGE_SHIFT;
463 	page_off = offset_in_page(page_off);
464 	page = xdr->pages + page_no;
465 	info->wi_next_off += remaining;
466 	sg = ctxt->rw_sg_table.sgl;
467 	sge_no = 0;
468 	do {
469 		sge_bytes = min_t(unsigned int, remaining,
470 				  PAGE_SIZE - page_off);
471 		sg_set_page(sg, *page, sge_bytes, page_off);
472 
473 		remaining -= sge_bytes;
474 		sg = sg_next(sg);
475 		page_off = 0;
476 		sge_no++;
477 		page++;
478 	} while (remaining);
479 
480 	ctxt->rw_nents = sge_no;
481 }
482 
483 /* Construct RDMA Write WRs to send a portion of an xdr_buf containing
484  * an RPC Reply.
485  */
486 static int
487 svc_rdma_build_writes(struct svc_rdma_write_info *info,
488 		      void (*constructor)(struct svc_rdma_write_info *info,
489 					  unsigned int len,
490 					  struct svc_rdma_rw_ctxt *ctxt),
491 		      unsigned int remaining)
492 {
493 	struct svc_rdma_chunk_ctxt *cc = &info->wi_cc;
494 	struct svcxprt_rdma *rdma = info->wi_rdma;
495 	const struct svc_rdma_segment *seg;
496 	struct svc_rdma_rw_ctxt *ctxt;
497 	int ret;
498 
499 	do {
500 		unsigned int write_len;
501 		u64 offset;
502 
503 		if (info->wi_seg_no >= info->wi_chunk->ch_segcount)
504 			goto out_overflow;
505 
506 		seg = &info->wi_chunk->ch_segments[info->wi_seg_no];
507 		write_len = min(remaining, seg->rs_length - info->wi_seg_off);
508 		if (!write_len)
509 			goto out_overflow;
510 		ctxt = svc_rdma_get_rw_ctxt(rdma,
511 					    (write_len >> PAGE_SHIFT) + 2);
512 		if (!ctxt)
513 			return -ENOMEM;
514 
515 		constructor(info, write_len, ctxt);
516 		offset = seg->rs_offset + info->wi_seg_off;
517 		ret = svc_rdma_rw_ctx_init(rdma, ctxt, offset, seg->rs_handle,
518 					   DMA_TO_DEVICE);
519 		if (ret < 0)
520 			return -EIO;
521 		percpu_counter_inc(&svcrdma_stat_write);
522 
523 		list_add(&ctxt->rw_list, &cc->cc_rwctxts);
524 		cc->cc_sqecount += ret;
525 		if (write_len == seg->rs_length - info->wi_seg_off) {
526 			info->wi_seg_no++;
527 			info->wi_seg_off = 0;
528 		} else {
529 			info->wi_seg_off += write_len;
530 		}
531 		remaining -= write_len;
532 	} while (remaining);
533 
534 	return 0;
535 
536 out_overflow:
537 	trace_svcrdma_small_wrch_err(&cc->cc_cid, remaining, info->wi_seg_no,
538 				     info->wi_chunk->ch_segcount);
539 	return -E2BIG;
540 }
541 
542 /**
543  * svc_rdma_iov_write - Construct RDMA Writes from an iov
544  * @info: pointer to write arguments
545  * @iov: kvec to write
546  *
547  * Returns:
548  *   On success, returns zero
549  *   %-E2BIG if the client-provided Write chunk is too small
550  *   %-ENOMEM if a resource has been exhausted
551  *   %-EIO if an rdma-rw error occurred
552  */
553 static int svc_rdma_iov_write(struct svc_rdma_write_info *info,
554 			      const struct kvec *iov)
555 {
556 	info->wi_base = iov->iov_base;
557 	return svc_rdma_build_writes(info, svc_rdma_vec_to_sg,
558 				     iov->iov_len);
559 }
560 
561 /**
562  * svc_rdma_pages_write - Construct RDMA Writes from pages
563  * @info: pointer to write arguments
564  * @xdr: xdr_buf with pages to write
565  * @offset: offset into the content of @xdr
566  * @length: number of bytes to write
567  *
568  * Returns:
569  *   On success, returns zero
570  *   %-E2BIG if the client-provided Write chunk is too small
571  *   %-ENOMEM if a resource has been exhausted
572  *   %-EIO if an rdma-rw error occurred
573  */
574 static int svc_rdma_pages_write(struct svc_rdma_write_info *info,
575 				const struct xdr_buf *xdr,
576 				unsigned int offset,
577 				unsigned long length)
578 {
579 	info->wi_xdr = xdr;
580 	info->wi_next_off = offset - xdr->head[0].iov_len;
581 	return svc_rdma_build_writes(info, svc_rdma_pagelist_to_sg,
582 				     length);
583 }
584 
585 /**
586  * svc_rdma_xb_write - Construct RDMA Writes to write an xdr_buf
587  * @xdr: xdr_buf to write
588  * @data: pointer to write arguments
589  *
590  * Returns:
591  *   On success, returns zero
592  *   %-E2BIG if the client-provided Write chunk is too small
593  *   %-ENOMEM if a resource has been exhausted
594  *   %-EIO if an rdma-rw error occurred
595  */
596 static int svc_rdma_xb_write(const struct xdr_buf *xdr, void *data)
597 {
598 	struct svc_rdma_write_info *info = data;
599 	int ret;
600 
601 	if (xdr->head[0].iov_len) {
602 		ret = svc_rdma_iov_write(info, &xdr->head[0]);
603 		if (ret < 0)
604 			return ret;
605 	}
606 
607 	if (xdr->page_len) {
608 		ret = svc_rdma_pages_write(info, xdr, xdr->head[0].iov_len,
609 					   xdr->page_len);
610 		if (ret < 0)
611 			return ret;
612 	}
613 
614 	if (xdr->tail[0].iov_len) {
615 		ret = svc_rdma_iov_write(info, &xdr->tail[0]);
616 		if (ret < 0)
617 			return ret;
618 	}
619 
620 	return xdr->len;
621 }
622 
623 /* Link Write WRs for @chunk onto @sctxt's WR chain.
624  */
625 static int svc_rdma_prepare_write_chunk(struct svcxprt_rdma *rdma,
626 					struct svc_rdma_send_ctxt *sctxt,
627 					const struct svc_rdma_chunk *chunk,
628 					const struct xdr_buf *xdr)
629 {
630 	struct svc_rdma_write_info *info;
631 	struct svc_rdma_chunk_ctxt *cc;
632 	struct ib_send_wr *first_wr;
633 	struct xdr_buf payload;
634 	struct list_head *pos;
635 	struct ib_cqe *cqe;
636 	int ret;
637 
638 	if (xdr_buf_subsegment(xdr, &payload, chunk->ch_position,
639 			       chunk->ch_payload_length))
640 		return -EMSGSIZE;
641 
642 	info = svc_rdma_write_info_alloc(rdma, chunk);
643 	if (!info)
644 		return -ENOMEM;
645 	cc = &info->wi_cc;
646 
647 	ret = svc_rdma_xb_write(&payload, info);
648 	if (ret != payload.len)
649 		goto out_err;
650 
651 	ret = -EINVAL;
652 	if (unlikely(cc->cc_sqecount > rdma->sc_sq_depth))
653 		goto out_err;
654 
655 	first_wr = sctxt->sc_wr_chain;
656 	cqe = &cc->cc_cqe;
657 	list_for_each(pos, &cc->cc_rwctxts) {
658 		struct svc_rdma_rw_ctxt *rwc;
659 
660 		rwc = list_entry(pos, struct svc_rdma_rw_ctxt, rw_list);
661 		first_wr = rdma_rw_ctx_wrs(&rwc->rw_ctx, rdma->sc_qp,
662 					   rdma->sc_port_num, cqe, first_wr);
663 		cqe = NULL;
664 	}
665 	sctxt->sc_wr_chain = first_wr;
666 	sctxt->sc_sqecount += cc->cc_sqecount;
667 	list_add(&info->wi_list, &sctxt->sc_write_info_list);
668 
669 	trace_svcrdma_post_write_chunk(&cc->cc_cid, cc->cc_sqecount);
670 	return 0;
671 
672 out_err:
673 	svc_rdma_write_info_free(info);
674 	return ret;
675 }
676 
677 /**
678  * svc_rdma_prepare_write_list - Construct WR chain for sending Write list
679  * @rdma: controlling RDMA transport
680  * @write_pcl: Write list provisioned by the client
681  * @sctxt: Send WR resources
682  * @xdr: xdr_buf containing an RPC Reply message
683  *
684  * Returns zero on success, or a negative errno if one or more
685  * Write chunks could not be sent.
686  */
687 int svc_rdma_prepare_write_list(struct svcxprt_rdma *rdma,
688 				const struct svc_rdma_pcl *write_pcl,
689 				struct svc_rdma_send_ctxt *sctxt,
690 				const struct xdr_buf *xdr)
691 {
692 	struct svc_rdma_chunk *chunk;
693 	int ret;
694 
695 	pcl_for_each_chunk(chunk, write_pcl) {
696 		if (!chunk->ch_payload_length)
697 			break;
698 		ret = svc_rdma_prepare_write_chunk(rdma, sctxt, chunk, xdr);
699 		if (ret < 0)
700 			return ret;
701 	}
702 	return 0;
703 }
704 
705 /**
706  * svc_rdma_prepare_reply_chunk - Construct WR chain for writing the Reply chunk
707  * @rdma: controlling RDMA transport
708  * @write_pcl: Write chunk list provided by client
709  * @reply_pcl: Reply chunk provided by client
710  * @sctxt: Send WR resources
711  * @xdr: xdr_buf containing an RPC Reply
712  *
713  * Returns a non-negative number of bytes the chunk consumed, or
714  *	%-E2BIG if the payload was larger than the Reply chunk,
715  *	%-EINVAL if client provided too many segments,
716  *	%-ENOMEM if rdma_rw context pool was exhausted,
717  *	%-ENOTCONN if posting failed (connection is lost),
718  *	%-EIO if rdma_rw initialization failed (DMA mapping, etc).
719  */
720 int svc_rdma_prepare_reply_chunk(struct svcxprt_rdma *rdma,
721 				 const struct svc_rdma_pcl *write_pcl,
722 				 const struct svc_rdma_pcl *reply_pcl,
723 				 struct svc_rdma_send_ctxt *sctxt,
724 				 const struct xdr_buf *xdr)
725 {
726 	struct svc_rdma_write_info *info = &sctxt->sc_reply_info;
727 	struct svc_rdma_chunk_ctxt *cc = &info->wi_cc;
728 	struct ib_send_wr *first_wr;
729 	struct list_head *pos;
730 	struct ib_cqe *cqe;
731 	int ret;
732 
733 	info->wi_rdma = rdma;
734 	info->wi_chunk = pcl_first_chunk(reply_pcl);
735 	info->wi_seg_off = 0;
736 	info->wi_seg_no = 0;
737 	info->wi_cc.cc_cqe.done = svc_rdma_reply_done;
738 
739 	ret = pcl_process_nonpayloads(write_pcl, xdr,
740 				      svc_rdma_xb_write, info);
741 	if (ret < 0)
742 		return ret;
743 
744 	first_wr = sctxt->sc_wr_chain;
745 	cqe = &cc->cc_cqe;
746 	list_for_each(pos, &cc->cc_rwctxts) {
747 		struct svc_rdma_rw_ctxt *rwc;
748 
749 		rwc = list_entry(pos, struct svc_rdma_rw_ctxt, rw_list);
750 		first_wr = rdma_rw_ctx_wrs(&rwc->rw_ctx, rdma->sc_qp,
751 					   rdma->sc_port_num, cqe, first_wr);
752 		cqe = NULL;
753 	}
754 	sctxt->sc_wr_chain = first_wr;
755 	sctxt->sc_sqecount += cc->cc_sqecount;
756 
757 	trace_svcrdma_post_reply_chunk(&cc->cc_cid, cc->cc_sqecount);
758 	return xdr->len;
759 }
760 
761 /**
762  * svc_rdma_build_read_segment - Build RDMA Read WQEs to pull one RDMA segment
763  * @rqstp: RPC transaction context
764  * @head: context for ongoing I/O
765  * @segment: co-ordinates of remote memory to be read
766  *
767  * Returns:
768  *   %0: the Read WR chain was constructed successfully
769  *   %-EINVAL: there were not enough rq_pages to finish
770  *   %-ENOMEM: allocating a local resources failed
771  *   %-EIO: a DMA mapping error occurred
772  */
773 static int svc_rdma_build_read_segment(struct svc_rqst *rqstp,
774 				       struct svc_rdma_recv_ctxt *head,
775 				       const struct svc_rdma_segment *segment)
776 {
777 	struct svcxprt_rdma *rdma = svc_rdma_rqst_rdma(rqstp);
778 	struct svc_rdma_chunk_ctxt *cc = &head->rc_cc;
779 	unsigned int sge_no, seg_len, len;
780 	struct svc_rdma_rw_ctxt *ctxt;
781 	struct scatterlist *sg;
782 	int ret;
783 
784 	len = segment->rs_length;
785 	sge_no = PAGE_ALIGN(head->rc_pageoff + len) >> PAGE_SHIFT;
786 	ctxt = svc_rdma_get_rw_ctxt(rdma, sge_no);
787 	if (!ctxt)
788 		return -ENOMEM;
789 	ctxt->rw_nents = sge_no;
790 
791 	sg = ctxt->rw_sg_table.sgl;
792 	for (sge_no = 0; sge_no < ctxt->rw_nents; sge_no++) {
793 		seg_len = min_t(unsigned int, len,
794 				PAGE_SIZE - head->rc_pageoff);
795 
796 		if (!head->rc_pageoff)
797 			head->rc_page_count++;
798 
799 		sg_set_page(sg, rqstp->rq_pages[head->rc_curpage],
800 			    seg_len, head->rc_pageoff);
801 		sg = sg_next(sg);
802 
803 		head->rc_pageoff += seg_len;
804 		if (head->rc_pageoff == PAGE_SIZE) {
805 			head->rc_curpage++;
806 			head->rc_pageoff = 0;
807 		}
808 		len -= seg_len;
809 
810 		if (len && ((head->rc_curpage + 1) > ARRAY_SIZE(rqstp->rq_pages)))
811 			goto out_overrun;
812 	}
813 
814 	ret = svc_rdma_rw_ctx_init(rdma, ctxt, segment->rs_offset,
815 				   segment->rs_handle, DMA_FROM_DEVICE);
816 	if (ret < 0)
817 		return -EIO;
818 	percpu_counter_inc(&svcrdma_stat_read);
819 
820 	list_add(&ctxt->rw_list, &cc->cc_rwctxts);
821 	cc->cc_sqecount += ret;
822 	return 0;
823 
824 out_overrun:
825 	trace_svcrdma_page_overrun_err(&cc->cc_cid, head->rc_curpage);
826 	return -EINVAL;
827 }
828 
829 /**
830  * svc_rdma_build_read_chunk - Build RDMA Read WQEs to pull one RDMA chunk
831  * @rqstp: RPC transaction context
832  * @head: context for ongoing I/O
833  * @chunk: Read chunk to pull
834  *
835  * Return values:
836  *   %0: the Read WR chain was constructed successfully
837  *   %-EINVAL: there were not enough resources to finish
838  *   %-ENOMEM: allocating a local resources failed
839  *   %-EIO: a DMA mapping error occurred
840  */
841 static int svc_rdma_build_read_chunk(struct svc_rqst *rqstp,
842 				     struct svc_rdma_recv_ctxt *head,
843 				     const struct svc_rdma_chunk *chunk)
844 {
845 	const struct svc_rdma_segment *segment;
846 	int ret;
847 
848 	ret = -EINVAL;
849 	pcl_for_each_segment(segment, chunk) {
850 		ret = svc_rdma_build_read_segment(rqstp, head, segment);
851 		if (ret < 0)
852 			break;
853 		head->rc_readbytes += segment->rs_length;
854 	}
855 	return ret;
856 }
857 
858 /**
859  * svc_rdma_copy_inline_range - Copy part of the inline content into pages
860  * @rqstp: RPC transaction context
861  * @head: context for ongoing I/O
862  * @offset: offset into the Receive buffer of region to copy
863  * @remaining: length of region to copy
864  *
865  * Take a page at a time from rqstp->rq_pages and copy the inline
866  * content from the Receive buffer into that page. Update
867  * head->rc_curpage and head->rc_pageoff so that the next RDMA Read
868  * result will land contiguously with the copied content.
869  *
870  * Return values:
871  *   %0: Inline content was successfully copied
872  *   %-EINVAL: offset or length was incorrect
873  */
874 static int svc_rdma_copy_inline_range(struct svc_rqst *rqstp,
875 				      struct svc_rdma_recv_ctxt *head,
876 				      unsigned int offset,
877 				      unsigned int remaining)
878 {
879 	unsigned char *dst, *src = head->rc_recv_buf;
880 	unsigned int page_no, numpages;
881 
882 	numpages = PAGE_ALIGN(head->rc_pageoff + remaining) >> PAGE_SHIFT;
883 	for (page_no = 0; page_no < numpages; page_no++) {
884 		unsigned int page_len;
885 
886 		page_len = min_t(unsigned int, remaining,
887 				 PAGE_SIZE - head->rc_pageoff);
888 
889 		if (!head->rc_pageoff)
890 			head->rc_page_count++;
891 
892 		dst = page_address(rqstp->rq_pages[head->rc_curpage]);
893 		memcpy(dst + head->rc_curpage, src + offset, page_len);
894 
895 		head->rc_readbytes += page_len;
896 		head->rc_pageoff += page_len;
897 		if (head->rc_pageoff == PAGE_SIZE) {
898 			head->rc_curpage++;
899 			head->rc_pageoff = 0;
900 		}
901 		remaining -= page_len;
902 		offset += page_len;
903 	}
904 
905 	return -EINVAL;
906 }
907 
908 /**
909  * svc_rdma_read_multiple_chunks - Construct RDMA Reads to pull data item Read chunks
910  * @rqstp: RPC transaction context
911  * @head: context for ongoing I/O
912  *
913  * The chunk data lands in rqstp->rq_arg as a series of contiguous pages,
914  * like an incoming TCP call.
915  *
916  * Return values:
917  *   %0: RDMA Read WQEs were successfully built
918  *   %-EINVAL: client provided too many chunks or segments,
919  *   %-ENOMEM: rdma_rw context pool was exhausted,
920  *   %-ENOTCONN: posting failed (connection is lost),
921  *   %-EIO: rdma_rw initialization failed (DMA mapping, etc).
922  */
923 static noinline int
924 svc_rdma_read_multiple_chunks(struct svc_rqst *rqstp,
925 			      struct svc_rdma_recv_ctxt *head)
926 {
927 	const struct svc_rdma_pcl *pcl = &head->rc_read_pcl;
928 	struct svc_rdma_chunk *chunk, *next;
929 	unsigned int start, length;
930 	int ret;
931 
932 	start = 0;
933 	chunk = pcl_first_chunk(pcl);
934 	length = chunk->ch_position;
935 	ret = svc_rdma_copy_inline_range(rqstp, head, start, length);
936 	if (ret < 0)
937 		return ret;
938 
939 	pcl_for_each_chunk(chunk, pcl) {
940 		ret = svc_rdma_build_read_chunk(rqstp, head, chunk);
941 		if (ret < 0)
942 			return ret;
943 
944 		next = pcl_next_chunk(pcl, chunk);
945 		if (!next)
946 			break;
947 
948 		start += length;
949 		length = next->ch_position - head->rc_readbytes;
950 		ret = svc_rdma_copy_inline_range(rqstp, head, start, length);
951 		if (ret < 0)
952 			return ret;
953 	}
954 
955 	start += length;
956 	length = head->rc_byte_len - start;
957 	return svc_rdma_copy_inline_range(rqstp, head, start, length);
958 }
959 
960 /**
961  * svc_rdma_read_data_item - Construct RDMA Reads to pull data item Read chunks
962  * @rqstp: RPC transaction context
963  * @head: context for ongoing I/O
964  *
965  * The chunk data lands in the page list of rqstp->rq_arg.pages.
966  *
967  * Currently NFSD does not look at the rqstp->rq_arg.tail[0] kvec.
968  * Therefore, XDR round-up of the Read chunk and trailing
969  * inline content must both be added at the end of the pagelist.
970  *
971  * Return values:
972  *   %0: RDMA Read WQEs were successfully built
973  *   %-EINVAL: client provided too many chunks or segments,
974  *   %-ENOMEM: rdma_rw context pool was exhausted,
975  *   %-ENOTCONN: posting failed (connection is lost),
976  *   %-EIO: rdma_rw initialization failed (DMA mapping, etc).
977  */
978 static int svc_rdma_read_data_item(struct svc_rqst *rqstp,
979 				   struct svc_rdma_recv_ctxt *head)
980 {
981 	return svc_rdma_build_read_chunk(rqstp, head,
982 					 pcl_first_chunk(&head->rc_read_pcl));
983 }
984 
985 /**
986  * svc_rdma_read_chunk_range - Build RDMA Read WRs for portion of a chunk
987  * @rqstp: RPC transaction context
988  * @head: context for ongoing I/O
989  * @chunk: parsed Call chunk to pull
990  * @offset: offset of region to pull
991  * @length: length of region to pull
992  *
993  * Return values:
994  *   %0: RDMA Read WQEs were successfully built
995  *   %-EINVAL: there were not enough resources to finish
996  *   %-ENOMEM: rdma_rw context pool was exhausted,
997  *   %-ENOTCONN: posting failed (connection is lost),
998  *   %-EIO: rdma_rw initialization failed (DMA mapping, etc).
999  */
1000 static int svc_rdma_read_chunk_range(struct svc_rqst *rqstp,
1001 				     struct svc_rdma_recv_ctxt *head,
1002 				     const struct svc_rdma_chunk *chunk,
1003 				     unsigned int offset, unsigned int length)
1004 {
1005 	const struct svc_rdma_segment *segment;
1006 	int ret;
1007 
1008 	ret = -EINVAL;
1009 	pcl_for_each_segment(segment, chunk) {
1010 		struct svc_rdma_segment dummy;
1011 
1012 		if (offset > segment->rs_length) {
1013 			offset -= segment->rs_length;
1014 			continue;
1015 		}
1016 
1017 		dummy.rs_handle = segment->rs_handle;
1018 		dummy.rs_length = min_t(u32, length, segment->rs_length) - offset;
1019 		dummy.rs_offset = segment->rs_offset + offset;
1020 
1021 		ret = svc_rdma_build_read_segment(rqstp, head, &dummy);
1022 		if (ret < 0)
1023 			break;
1024 
1025 		head->rc_readbytes += dummy.rs_length;
1026 		length -= dummy.rs_length;
1027 		offset = 0;
1028 	}
1029 	return ret;
1030 }
1031 
1032 /**
1033  * svc_rdma_read_call_chunk - Build RDMA Read WQEs to pull a Long Message
1034  * @rqstp: RPC transaction context
1035  * @head: context for ongoing I/O
1036  *
1037  * Return values:
1038  *   %0: RDMA Read WQEs were successfully built
1039  *   %-EINVAL: there were not enough resources to finish
1040  *   %-ENOMEM: rdma_rw context pool was exhausted,
1041  *   %-ENOTCONN: posting failed (connection is lost),
1042  *   %-EIO: rdma_rw initialization failed (DMA mapping, etc).
1043  */
1044 static int svc_rdma_read_call_chunk(struct svc_rqst *rqstp,
1045 				    struct svc_rdma_recv_ctxt *head)
1046 {
1047 	const struct svc_rdma_chunk *call_chunk =
1048 			pcl_first_chunk(&head->rc_call_pcl);
1049 	const struct svc_rdma_pcl *pcl = &head->rc_read_pcl;
1050 	struct svc_rdma_chunk *chunk, *next;
1051 	unsigned int start, length;
1052 	int ret;
1053 
1054 	if (pcl_is_empty(pcl))
1055 		return svc_rdma_build_read_chunk(rqstp, head, call_chunk);
1056 
1057 	start = 0;
1058 	chunk = pcl_first_chunk(pcl);
1059 	length = chunk->ch_position;
1060 	ret = svc_rdma_read_chunk_range(rqstp, head, call_chunk,
1061 					start, length);
1062 	if (ret < 0)
1063 		return ret;
1064 
1065 	pcl_for_each_chunk(chunk, pcl) {
1066 		ret = svc_rdma_build_read_chunk(rqstp, head, chunk);
1067 		if (ret < 0)
1068 			return ret;
1069 
1070 		next = pcl_next_chunk(pcl, chunk);
1071 		if (!next)
1072 			break;
1073 
1074 		start += length;
1075 		length = next->ch_position - head->rc_readbytes;
1076 		ret = svc_rdma_read_chunk_range(rqstp, head, call_chunk,
1077 						start, length);
1078 		if (ret < 0)
1079 			return ret;
1080 	}
1081 
1082 	start += length;
1083 	length = call_chunk->ch_length - start;
1084 	return svc_rdma_read_chunk_range(rqstp, head, call_chunk,
1085 					 start, length);
1086 }
1087 
1088 /**
1089  * svc_rdma_read_special - Build RDMA Read WQEs to pull a Long Message
1090  * @rqstp: RPC transaction context
1091  * @head: context for ongoing I/O
1092  *
1093  * The start of the data lands in the first page just after the
1094  * Transport header, and the rest lands in rqstp->rq_arg.pages.
1095  *
1096  * Assumptions:
1097  *	- A PZRC is never sent in an RDMA_MSG message, though it's
1098  *	  allowed by spec.
1099  *
1100  * Return values:
1101  *   %0: RDMA Read WQEs were successfully built
1102  *   %-EINVAL: client provided too many chunks or segments,
1103  *   %-ENOMEM: rdma_rw context pool was exhausted,
1104  *   %-ENOTCONN: posting failed (connection is lost),
1105  *   %-EIO: rdma_rw initialization failed (DMA mapping, etc).
1106  */
1107 static noinline int svc_rdma_read_special(struct svc_rqst *rqstp,
1108 					  struct svc_rdma_recv_ctxt *head)
1109 {
1110 	return svc_rdma_read_call_chunk(rqstp, head);
1111 }
1112 
1113 /* Pages under I/O have been copied to head->rc_pages. Ensure that
1114  * svc_xprt_release() does not put them when svc_rdma_recvfrom()
1115  * returns. This has to be done after all Read WRs are constructed
1116  * to properly handle a page that happens to be part of I/O on behalf
1117  * of two different RDMA segments.
1118  *
1119  * Note: if the subsequent post_send fails, these pages have already
1120  * been moved to head->rc_pages and thus will be cleaned up by
1121  * svc_rdma_recv_ctxt_put().
1122  */
1123 static void svc_rdma_clear_rqst_pages(struct svc_rqst *rqstp,
1124 				      struct svc_rdma_recv_ctxt *head)
1125 {
1126 	unsigned int i;
1127 
1128 	for (i = 0; i < head->rc_page_count; i++) {
1129 		head->rc_pages[i] = rqstp->rq_pages[i];
1130 		rqstp->rq_pages[i] = NULL;
1131 	}
1132 }
1133 
1134 /**
1135  * svc_rdma_process_read_list - Pull list of Read chunks from the client
1136  * @rdma: controlling RDMA transport
1137  * @rqstp: set of pages to use as Read sink buffers
1138  * @head: pages under I/O collect here
1139  *
1140  * The RPC/RDMA protocol assumes that the upper layer's XDR decoders
1141  * pull each Read chunk as they decode an incoming RPC message.
1142  *
1143  * On Linux, however, the server needs to have a fully-constructed RPC
1144  * message in rqstp->rq_arg when there is a positive return code from
1145  * ->xpo_recvfrom. So the Read list is safety-checked immediately when
1146  * it is received, then here the whole Read list is pulled all at once.
1147  * The ingress RPC message is fully reconstructed once all associated
1148  * RDMA Reads have completed.
1149  *
1150  * Return values:
1151  *   %1: all needed RDMA Reads were posted successfully,
1152  *   %-EINVAL: client provided too many chunks or segments,
1153  *   %-ENOMEM: rdma_rw context pool was exhausted,
1154  *   %-ENOTCONN: posting failed (connection is lost),
1155  *   %-EIO: rdma_rw initialization failed (DMA mapping, etc).
1156  */
1157 int svc_rdma_process_read_list(struct svcxprt_rdma *rdma,
1158 			       struct svc_rqst *rqstp,
1159 			       struct svc_rdma_recv_ctxt *head)
1160 {
1161 	struct svc_rdma_chunk_ctxt *cc = &head->rc_cc;
1162 	int ret;
1163 
1164 	cc->cc_cqe.done = svc_rdma_wc_read_done;
1165 	cc->cc_sqecount = 0;
1166 	head->rc_pageoff = 0;
1167 	head->rc_curpage = 0;
1168 	head->rc_readbytes = 0;
1169 
1170 	if (pcl_is_empty(&head->rc_call_pcl)) {
1171 		if (head->rc_read_pcl.cl_count == 1)
1172 			ret = svc_rdma_read_data_item(rqstp, head);
1173 		else
1174 			ret = svc_rdma_read_multiple_chunks(rqstp, head);
1175 	} else
1176 		ret = svc_rdma_read_special(rqstp, head);
1177 	svc_rdma_clear_rqst_pages(rqstp, head);
1178 	if (ret < 0)
1179 		return ret;
1180 
1181 	trace_svcrdma_post_read_chunk(&cc->cc_cid, cc->cc_sqecount);
1182 	ret = svc_rdma_post_chunk_ctxt(rdma, cc);
1183 	return ret < 0 ? ret : 1;
1184 }
1185