1 /*
2  * Copyright (c) 2016, 2017 Oracle. All rights reserved.
3  * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
4  * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the BSD-type
10  * license below:
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  *
16  *      Redistributions of source code must retain the above copyright
17  *      notice, this list of conditions and the following disclaimer.
18  *
19  *      Redistributions in binary form must reproduce the above
20  *      copyright notice, this list of conditions and the following
21  *      disclaimer in the documentation and/or other materials provided
22  *      with the distribution.
23  *
24  *      Neither the name of the Network Appliance, Inc. nor the names of
25  *      its contributors may be used to endorse or promote products
26  *      derived from this software without specific prior written
27  *      permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40  *
41  * Author: Tom Tucker <tom@opengridcomputing.com>
42  */
43 
44 /* Operation
45  *
46  * The main entry point is svc_rdma_recvfrom. This is called from
47  * svc_recv when the transport indicates there is incoming data to
48  * be read. "Data Ready" is signaled when an RDMA Receive completes,
49  * or when a set of RDMA Reads complete.
50  *
51  * An svc_rqst is passed in. This structure contains an array of
52  * free pages (rq_pages) that will contain the incoming RPC message.
53  *
54  * Short messages are moved directly into svc_rqst::rq_arg, and
55  * the RPC Call is ready to be processed by the Upper Layer.
56  * svc_rdma_recvfrom returns the length of the RPC Call message,
57  * completing the reception of the RPC Call.
58  *
59  * However, when an incoming message has Read chunks,
60  * svc_rdma_recvfrom must post RDMA Reads to pull the RPC Call's
61  * data payload from the client. svc_rdma_recvfrom sets up the
62  * RDMA Reads using pages in svc_rqst::rq_pages, which are
63  * transferred to an svc_rdma_op_ctxt for the duration of the
64  * I/O. svc_rdma_recvfrom then returns zero, since the RPC message
65  * is still not yet ready.
66  *
67  * When the Read chunk payloads have become available on the
68  * server, "Data Ready" is raised again, and svc_recv calls
69  * svc_rdma_recvfrom again. This second call may use a different
70  * svc_rqst than the first one, thus any information that needs
71  * to be preserved across these two calls is kept in an
72  * svc_rdma_op_ctxt.
73  *
74  * The second call to svc_rdma_recvfrom performs final assembly
75  * of the RPC Call message, using the RDMA Read sink pages kept in
76  * the svc_rdma_op_ctxt. The xdr_buf is copied from the
77  * svc_rdma_op_ctxt to the second svc_rqst. The second call returns
78  * the length of the completed RPC Call message.
79  *
80  * Page Management
81  *
82  * Pages under I/O must be transferred from the first svc_rqst to an
83  * svc_rdma_op_ctxt before the first svc_rdma_recvfrom call returns.
84  *
85  * The first svc_rqst supplies pages for RDMA Reads. These are moved
86  * from rqstp::rq_pages into ctxt::pages. The consumed elements of
87  * the rq_pages array are set to NULL and refilled with the first
88  * svc_rdma_recvfrom call returns.
89  *
90  * During the second svc_rdma_recvfrom call, RDMA Read sink pages
91  * are transferred from the svc_rdma_op_ctxt to the second svc_rqst
92  * (see rdma_read_complete() below).
93  */
94 
95 #include <asm/unaligned.h>
96 #include <rdma/ib_verbs.h>
97 #include <rdma/rdma_cm.h>
98 
99 #include <linux/spinlock.h>
100 
101 #include <linux/sunrpc/xdr.h>
102 #include <linux/sunrpc/debug.h>
103 #include <linux/sunrpc/rpc_rdma.h>
104 #include <linux/sunrpc/svc_rdma.h>
105 
106 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
107 
108 /*
109  * Replace the pages in the rq_argpages array with the pages from the SGE in
110  * the RDMA_RECV completion. The SGL should contain full pages up until the
111  * last one.
112  */
113 static void rdma_build_arg_xdr(struct svc_rqst *rqstp,
114 			       struct svc_rdma_op_ctxt *ctxt,
115 			       u32 byte_count)
116 {
117 	struct page *page;
118 	u32 bc;
119 	int sge_no;
120 
121 	/* Swap the page in the SGE with the page in argpages */
122 	page = ctxt->pages[0];
123 	put_page(rqstp->rq_pages[0]);
124 	rqstp->rq_pages[0] = page;
125 
126 	/* Set up the XDR head */
127 	rqstp->rq_arg.head[0].iov_base = page_address(page);
128 	rqstp->rq_arg.head[0].iov_len =
129 		min_t(size_t, byte_count, ctxt->sge[0].length);
130 	rqstp->rq_arg.len = byte_count;
131 	rqstp->rq_arg.buflen = byte_count;
132 
133 	/* Compute bytes past head in the SGL */
134 	bc = byte_count - rqstp->rq_arg.head[0].iov_len;
135 
136 	/* If data remains, store it in the pagelist */
137 	rqstp->rq_arg.page_len = bc;
138 	rqstp->rq_arg.page_base = 0;
139 
140 	sge_no = 1;
141 	while (bc && sge_no < ctxt->count) {
142 		page = ctxt->pages[sge_no];
143 		put_page(rqstp->rq_pages[sge_no]);
144 		rqstp->rq_pages[sge_no] = page;
145 		bc -= min_t(u32, bc, ctxt->sge[sge_no].length);
146 		sge_no++;
147 	}
148 	rqstp->rq_respages = &rqstp->rq_pages[sge_no];
149 	rqstp->rq_next_page = rqstp->rq_respages + 1;
150 
151 	/* If not all pages were used from the SGL, free the remaining ones */
152 	bc = sge_no;
153 	while (sge_no < ctxt->count) {
154 		page = ctxt->pages[sge_no++];
155 		put_page(page);
156 	}
157 	ctxt->count = bc;
158 
159 	/* Set up tail */
160 	rqstp->rq_arg.tail[0].iov_base = NULL;
161 	rqstp->rq_arg.tail[0].iov_len = 0;
162 }
163 
164 /* This accommodates the largest possible Write chunk,
165  * in one segment.
166  */
167 #define MAX_BYTES_WRITE_SEG	((u32)(RPCSVC_MAXPAGES << PAGE_SHIFT))
168 
169 /* This accommodates the largest possible Position-Zero
170  * Read chunk or Reply chunk, in one segment.
171  */
172 #define MAX_BYTES_SPECIAL_SEG	((u32)((RPCSVC_MAXPAGES + 2) << PAGE_SHIFT))
173 
174 /* Sanity check the Read list.
175  *
176  * Implementation limits:
177  * - This implementation supports only one Read chunk.
178  *
179  * Sanity checks:
180  * - Read list does not overflow buffer.
181  * - Segment size limited by largest NFS data payload.
182  *
183  * The segment count is limited to how many segments can
184  * fit in the transport header without overflowing the
185  * buffer. That's about 40 Read segments for a 1KB inline
186  * threshold.
187  *
188  * Returns pointer to the following Write list.
189  */
190 static __be32 *xdr_check_read_list(__be32 *p, const __be32 *end)
191 {
192 	u32 position;
193 	bool first;
194 
195 	first = true;
196 	while (*p++ != xdr_zero) {
197 		if (first) {
198 			position = be32_to_cpup(p++);
199 			first = false;
200 		} else if (be32_to_cpup(p++) != position) {
201 			return NULL;
202 		}
203 		p++;	/* handle */
204 		if (be32_to_cpup(p++) > MAX_BYTES_SPECIAL_SEG)
205 			return NULL;
206 		p += 2;	/* offset */
207 
208 		if (p > end)
209 			return NULL;
210 	}
211 	return p;
212 }
213 
214 /* The segment count is limited to how many segments can
215  * fit in the transport header without overflowing the
216  * buffer. That's about 60 Write segments for a 1KB inline
217  * threshold.
218  */
219 static __be32 *xdr_check_write_chunk(__be32 *p, const __be32 *end,
220 				     u32 maxlen)
221 {
222 	u32 i, segcount;
223 
224 	segcount = be32_to_cpup(p++);
225 	for (i = 0; i < segcount; i++) {
226 		p++;	/* handle */
227 		if (be32_to_cpup(p++) > maxlen)
228 			return NULL;
229 		p += 2;	/* offset */
230 
231 		if (p > end)
232 			return NULL;
233 	}
234 
235 	return p;
236 }
237 
238 /* Sanity check the Write list.
239  *
240  * Implementation limits:
241  * - This implementation supports only one Write chunk.
242  *
243  * Sanity checks:
244  * - Write list does not overflow buffer.
245  * - Segment size limited by largest NFS data payload.
246  *
247  * Returns pointer to the following Reply chunk.
248  */
249 static __be32 *xdr_check_write_list(__be32 *p, const __be32 *end)
250 {
251 	u32 chcount;
252 
253 	chcount = 0;
254 	while (*p++ != xdr_zero) {
255 		p = xdr_check_write_chunk(p, end, MAX_BYTES_WRITE_SEG);
256 		if (!p)
257 			return NULL;
258 		if (chcount++ > 1)
259 			return NULL;
260 	}
261 	return p;
262 }
263 
264 /* Sanity check the Reply chunk.
265  *
266  * Sanity checks:
267  * - Reply chunk does not overflow buffer.
268  * - Segment size limited by largest NFS data payload.
269  *
270  * Returns pointer to the following RPC header.
271  */
272 static __be32 *xdr_check_reply_chunk(__be32 *p, const __be32 *end)
273 {
274 	if (*p++ != xdr_zero) {
275 		p = xdr_check_write_chunk(p, end, MAX_BYTES_SPECIAL_SEG);
276 		if (!p)
277 			return NULL;
278 	}
279 	return p;
280 }
281 
282 /* On entry, xdr->head[0].iov_base points to first byte in the
283  * RPC-over-RDMA header.
284  *
285  * On successful exit, head[0] points to first byte past the
286  * RPC-over-RDMA header. For RDMA_MSG, this is the RPC message.
287  * The length of the RPC-over-RDMA header is returned.
288  *
289  * Assumptions:
290  * - The transport header is entirely contained in the head iovec.
291  */
292 static int svc_rdma_xdr_decode_req(struct xdr_buf *rq_arg)
293 {
294 	__be32 *p, *end, *rdma_argp;
295 	unsigned int hdr_len;
296 	char *proc;
297 
298 	/* Verify that there's enough bytes for header + something */
299 	if (rq_arg->len <= RPCRDMA_HDRLEN_ERR)
300 		goto out_short;
301 
302 	rdma_argp = rq_arg->head[0].iov_base;
303 	if (*(rdma_argp + 1) != rpcrdma_version)
304 		goto out_version;
305 
306 	switch (*(rdma_argp + 3)) {
307 	case rdma_msg:
308 		proc = "RDMA_MSG";
309 		break;
310 	case rdma_nomsg:
311 		proc = "RDMA_NOMSG";
312 		break;
313 
314 	case rdma_done:
315 		goto out_drop;
316 
317 	case rdma_error:
318 		goto out_drop;
319 
320 	default:
321 		goto out_proc;
322 	}
323 
324 	end = (__be32 *)((unsigned long)rdma_argp + rq_arg->len);
325 	p = xdr_check_read_list(rdma_argp + 4, end);
326 	if (!p)
327 		goto out_inval;
328 	p = xdr_check_write_list(p, end);
329 	if (!p)
330 		goto out_inval;
331 	p = xdr_check_reply_chunk(p, end);
332 	if (!p)
333 		goto out_inval;
334 	if (p > end)
335 		goto out_inval;
336 
337 	rq_arg->head[0].iov_base = p;
338 	hdr_len = (unsigned long)p - (unsigned long)rdma_argp;
339 	rq_arg->head[0].iov_len -= hdr_len;
340 	rq_arg->len -= hdr_len;
341 	dprintk("svcrdma: received %s request for XID 0x%08x, hdr_len=%u\n",
342 		proc, be32_to_cpup(rdma_argp), hdr_len);
343 	return hdr_len;
344 
345 out_short:
346 	dprintk("svcrdma: header too short = %d\n", rq_arg->len);
347 	return -EINVAL;
348 
349 out_version:
350 	dprintk("svcrdma: bad xprt version: %u\n",
351 		be32_to_cpup(rdma_argp + 1));
352 	return -EPROTONOSUPPORT;
353 
354 out_drop:
355 	dprintk("svcrdma: dropping RDMA_DONE/ERROR message\n");
356 	return 0;
357 
358 out_proc:
359 	dprintk("svcrdma: bad rdma procedure (%u)\n",
360 		be32_to_cpup(rdma_argp + 3));
361 	return -EINVAL;
362 
363 out_inval:
364 	dprintk("svcrdma: failed to parse transport header\n");
365 	return -EINVAL;
366 }
367 
368 static void rdma_read_complete(struct svc_rqst *rqstp,
369 			       struct svc_rdma_op_ctxt *head)
370 {
371 	int page_no;
372 
373 	/* Copy RPC pages */
374 	for (page_no = 0; page_no < head->count; page_no++) {
375 		put_page(rqstp->rq_pages[page_no]);
376 		rqstp->rq_pages[page_no] = head->pages[page_no];
377 	}
378 
379 	/* Point rq_arg.pages past header */
380 	rqstp->rq_arg.pages = &rqstp->rq_pages[head->hdr_count];
381 	rqstp->rq_arg.page_len = head->arg.page_len;
382 
383 	/* rq_respages starts after the last arg page */
384 	rqstp->rq_respages = &rqstp->rq_pages[page_no];
385 	rqstp->rq_next_page = rqstp->rq_respages + 1;
386 
387 	/* Rebuild rq_arg head and tail. */
388 	rqstp->rq_arg.head[0] = head->arg.head[0];
389 	rqstp->rq_arg.tail[0] = head->arg.tail[0];
390 	rqstp->rq_arg.len = head->arg.len;
391 	rqstp->rq_arg.buflen = head->arg.buflen;
392 }
393 
394 static void svc_rdma_send_error(struct svcxprt_rdma *xprt,
395 				__be32 *rdma_argp, int status)
396 {
397 	struct svc_rdma_op_ctxt *ctxt;
398 	__be32 *p, *err_msgp;
399 	unsigned int length;
400 	struct page *page;
401 	int ret;
402 
403 	ret = svc_rdma_repost_recv(xprt, GFP_KERNEL);
404 	if (ret)
405 		return;
406 
407 	page = alloc_page(GFP_KERNEL);
408 	if (!page)
409 		return;
410 	err_msgp = page_address(page);
411 
412 	p = err_msgp;
413 	*p++ = *rdma_argp;
414 	*p++ = *(rdma_argp + 1);
415 	*p++ = xprt->sc_fc_credits;
416 	*p++ = rdma_error;
417 	if (status == -EPROTONOSUPPORT) {
418 		*p++ = err_vers;
419 		*p++ = rpcrdma_version;
420 		*p++ = rpcrdma_version;
421 	} else {
422 		*p++ = err_chunk;
423 	}
424 	length = (unsigned long)p - (unsigned long)err_msgp;
425 
426 	/* Map transport header; no RPC message payload */
427 	ctxt = svc_rdma_get_context(xprt);
428 	ret = svc_rdma_map_reply_hdr(xprt, ctxt, err_msgp, length);
429 	if (ret) {
430 		dprintk("svcrdma: Error %d mapping send for protocol error\n",
431 			ret);
432 		return;
433 	}
434 
435 	ret = svc_rdma_post_send_wr(xprt, ctxt, 1, 0);
436 	if (ret) {
437 		dprintk("svcrdma: Error %d posting send for protocol error\n",
438 			ret);
439 		svc_rdma_unmap_dma(ctxt);
440 		svc_rdma_put_context(ctxt, 1);
441 	}
442 }
443 
444 /* By convention, backchannel calls arrive via rdma_msg type
445  * messages, and never populate the chunk lists. This makes
446  * the RPC/RDMA header small and fixed in size, so it is
447  * straightforward to check the RPC header's direction field.
448  */
449 static bool svc_rdma_is_backchannel_reply(struct svc_xprt *xprt,
450 					  __be32 *rdma_resp)
451 {
452 	__be32 *p;
453 
454 	if (!xprt->xpt_bc_xprt)
455 		return false;
456 
457 	p = rdma_resp + 3;
458 	if (*p++ != rdma_msg)
459 		return false;
460 
461 	if (*p++ != xdr_zero)
462 		return false;
463 	if (*p++ != xdr_zero)
464 		return false;
465 	if (*p++ != xdr_zero)
466 		return false;
467 
468 	/* XID sanity */
469 	if (*p++ != *rdma_resp)
470 		return false;
471 	/* call direction */
472 	if (*p == cpu_to_be32(RPC_CALL))
473 		return false;
474 
475 	return true;
476 }
477 
478 /**
479  * svc_rdma_recvfrom - Receive an RPC call
480  * @rqstp: request structure into which to receive an RPC Call
481  *
482  * Returns:
483  *	The positive number of bytes in the RPC Call message,
484  *	%0 if there were no Calls ready to return,
485  *	%-EINVAL if the Read chunk data is too large,
486  *	%-ENOMEM if rdma_rw context pool was exhausted,
487  *	%-ENOTCONN if posting failed (connection is lost),
488  *	%-EIO if rdma_rw initialization failed (DMA mapping, etc).
489  *
490  * Called in a loop when XPT_DATA is set. XPT_DATA is cleared only
491  * when there are no remaining ctxt's to process.
492  *
493  * The next ctxt is removed from the "receive" lists.
494  *
495  * - If the ctxt completes a Read, then finish assembling the Call
496  *   message and return the number of bytes in the message.
497  *
498  * - If the ctxt completes a Receive, then construct the Call
499  *   message from the contents of the Receive buffer.
500  *
501  *   - If there are no Read chunks in this message, then finish
502  *     assembling the Call message and return the number of bytes
503  *     in the message.
504  *
505  *   - If there are Read chunks in this message, post Read WRs to
506  *     pull that payload and return 0.
507  */
508 int svc_rdma_recvfrom(struct svc_rqst *rqstp)
509 {
510 	struct svc_xprt *xprt = rqstp->rq_xprt;
511 	struct svcxprt_rdma *rdma_xprt =
512 		container_of(xprt, struct svcxprt_rdma, sc_xprt);
513 	struct svc_rdma_op_ctxt *ctxt;
514 	__be32 *p;
515 	int ret;
516 
517 	spin_lock(&rdma_xprt->sc_rq_dto_lock);
518 	if (!list_empty(&rdma_xprt->sc_read_complete_q)) {
519 		ctxt = list_first_entry(&rdma_xprt->sc_read_complete_q,
520 					struct svc_rdma_op_ctxt, list);
521 		list_del(&ctxt->list);
522 		spin_unlock(&rdma_xprt->sc_rq_dto_lock);
523 		rdma_read_complete(rqstp, ctxt);
524 		goto complete;
525 	} else if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
526 		ctxt = list_first_entry(&rdma_xprt->sc_rq_dto_q,
527 					struct svc_rdma_op_ctxt, list);
528 		list_del(&ctxt->list);
529 	} else {
530 		/* No new incoming requests, terminate the loop */
531 		clear_bit(XPT_DATA, &xprt->xpt_flags);
532 		spin_unlock(&rdma_xprt->sc_rq_dto_lock);
533 		return 0;
534 	}
535 	spin_unlock(&rdma_xprt->sc_rq_dto_lock);
536 
537 	dprintk("svcrdma: recvfrom: ctxt=%p on xprt=%p, rqstp=%p\n",
538 		ctxt, rdma_xprt, rqstp);
539 	atomic_inc(&rdma_stat_recv);
540 
541 	/* Build up the XDR from the receive buffers. */
542 	rdma_build_arg_xdr(rqstp, ctxt, ctxt->byte_len);
543 
544 	/* Decode the RDMA header. */
545 	p = (__be32 *)rqstp->rq_arg.head[0].iov_base;
546 	ret = svc_rdma_xdr_decode_req(&rqstp->rq_arg);
547 	if (ret < 0)
548 		goto out_err;
549 	if (ret == 0)
550 		goto out_drop;
551 	rqstp->rq_xprt_hlen = ret;
552 
553 	if (svc_rdma_is_backchannel_reply(xprt, p)) {
554 		ret = svc_rdma_handle_bc_reply(xprt->xpt_bc_xprt, p,
555 					       &rqstp->rq_arg);
556 		svc_rdma_put_context(ctxt, 0);
557 		if (ret)
558 			goto repost;
559 		return ret;
560 	}
561 
562 	p += rpcrdma_fixed_maxsz;
563 	if (*p != xdr_zero)
564 		goto out_readchunk;
565 
566 complete:
567 	svc_rdma_put_context(ctxt, 0);
568 	dprintk("svcrdma: recvfrom: xprt=%p, rqstp=%p, rq_arg.len=%u\n",
569 		rdma_xprt, rqstp, rqstp->rq_arg.len);
570 	rqstp->rq_prot = IPPROTO_MAX;
571 	svc_xprt_copy_addrs(rqstp, xprt);
572 	return rqstp->rq_arg.len;
573 
574 out_readchunk:
575 	ret = svc_rdma_recv_read_chunk(rdma_xprt, rqstp, ctxt, p);
576 	if (ret < 0)
577 		goto out_postfail;
578 	return 0;
579 
580 out_err:
581 	svc_rdma_send_error(rdma_xprt, p, ret);
582 	svc_rdma_put_context(ctxt, 0);
583 	return 0;
584 
585 out_postfail:
586 	if (ret == -EINVAL)
587 		svc_rdma_send_error(rdma_xprt, p, ret);
588 	svc_rdma_put_context(ctxt, 1);
589 	return ret;
590 
591 out_drop:
592 	svc_rdma_put_context(ctxt, 1);
593 repost:
594 	return svc_rdma_repost_recv(rdma_xprt, GFP_KERNEL);
595 }
596