Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (C) 2017, Microsoft Corporation.
4 *
5 * Author(s): Long Li <longli@microsoft.com>
6 */
7#include <linux/module.h>
8#include <linux/highmem.h>
9#include "smbdirect.h"
10#include "cifs_debug.h"
11#include "cifsproto.h"
12#include "smb2proto.h"
13
14static struct smbd_response *get_empty_queue_buffer(
15 struct smbd_connection *info);
16static struct smbd_response *get_receive_buffer(
17 struct smbd_connection *info);
18static void put_receive_buffer(
19 struct smbd_connection *info,
20 struct smbd_response *response);
21static int allocate_receive_buffers(struct smbd_connection *info, int num_buf);
22static void destroy_receive_buffers(struct smbd_connection *info);
23
24static void put_empty_packet(
25 struct smbd_connection *info, struct smbd_response *response);
26static void enqueue_reassembly(
27 struct smbd_connection *info,
28 struct smbd_response *response, int data_length);
29static struct smbd_response *_get_first_reassembly(
30 struct smbd_connection *info);
31
32static int smbd_post_recv(
33 struct smbd_connection *info,
34 struct smbd_response *response);
35
36static int smbd_post_send_empty(struct smbd_connection *info);
37static int smbd_post_send_data(
38 struct smbd_connection *info,
39 struct kvec *iov, int n_vec, int remaining_data_length);
40static int smbd_post_send_page(struct smbd_connection *info,
41 struct page *page, unsigned long offset,
42 size_t size, int remaining_data_length);
43
44static void destroy_mr_list(struct smbd_connection *info);
45static int allocate_mr_list(struct smbd_connection *info);
46
47/* SMBD version number */
48#define SMBD_V1 0x0100
49
50/* Port numbers for SMBD transport */
51#define SMB_PORT 445
52#define SMBD_PORT 5445
53
54/* Address lookup and resolve timeout in ms */
55#define RDMA_RESOLVE_TIMEOUT 5000
56
57/* SMBD negotiation timeout in seconds */
58#define SMBD_NEGOTIATE_TIMEOUT 120
59
60/* SMBD minimum receive size and fragmented sized defined in [MS-SMBD] */
61#define SMBD_MIN_RECEIVE_SIZE 128
62#define SMBD_MIN_FRAGMENTED_SIZE 131072
63
64/*
65 * Default maximum number of RDMA read/write outstanding on this connection
66 * This value is possibly decreased during QP creation on hardware limit
67 */
68#define SMBD_CM_RESPONDER_RESOURCES 32
69
70/* Maximum number of retries on data transfer operations */
71#define SMBD_CM_RETRY 6
72/* No need to retry on Receiver Not Ready since SMBD manages credits */
73#define SMBD_CM_RNR_RETRY 0
74
75/*
76 * User configurable initial values per SMBD transport connection
77 * as defined in [MS-SMBD] 3.1.1.1
78 * Those may change after a SMBD negotiation
79 */
80/* The local peer's maximum number of credits to grant to the peer */
81int smbd_receive_credit_max = 255;
82
83/* The remote peer's credit request of local peer */
84int smbd_send_credit_target = 255;
85
86/* The maximum single message size can be sent to remote peer */
87int smbd_max_send_size = 1364;
88
89/* The maximum fragmented upper-layer payload receive size supported */
90int smbd_max_fragmented_recv_size = 1024 * 1024;
91
92/* The maximum single-message size which can be received */
93int smbd_max_receive_size = 1364;
94
95/* The timeout to initiate send of a keepalive message on idle */
96int smbd_keep_alive_interval = 120;
97
98/*
99 * User configurable initial values for RDMA transport
100 * The actual values used may be lower and are limited to hardware capabilities
101 */
102/* Default maximum number of pages in a single RDMA write/read */
103int smbd_max_frmr_depth = 2048;
104
105/* If payload is less than this byte, use RDMA send/recv not read/write */
106int rdma_readwrite_threshold = 4096;
107
108/* Transport logging functions
109 * Logging are defined as classes. They can be OR'ed to define the actual
110 * logging level via module parameter smbd_logging_class
111 * e.g. cifs.smbd_logging_class=0xa0 will log all log_rdma_recv() and
112 * log_rdma_event()
113 */
114#define LOG_OUTGOING 0x1
115#define LOG_INCOMING 0x2
116#define LOG_READ 0x4
117#define LOG_WRITE 0x8
118#define LOG_RDMA_SEND 0x10
119#define LOG_RDMA_RECV 0x20
120#define LOG_KEEP_ALIVE 0x40
121#define LOG_RDMA_EVENT 0x80
122#define LOG_RDMA_MR 0x100
123static unsigned int smbd_logging_class;
124module_param(smbd_logging_class, uint, 0644);
125MODULE_PARM_DESC(smbd_logging_class,
126 "Logging class for SMBD transport 0x0 to 0x100");
127
128#define ERR 0x0
129#define INFO 0x1
130static unsigned int smbd_logging_level = ERR;
131module_param(smbd_logging_level, uint, 0644);
132MODULE_PARM_DESC(smbd_logging_level,
133 "Logging level for SMBD transport, 0 (default): error, 1: info");
134
135#define log_rdma(level, class, fmt, args...) \
136do { \
137 if (level <= smbd_logging_level || class & smbd_logging_class) \
138 cifs_dbg(VFS, "%s:%d " fmt, __func__, __LINE__, ##args);\
139} while (0)
140
141#define log_outgoing(level, fmt, args...) \
142 log_rdma(level, LOG_OUTGOING, fmt, ##args)
143#define log_incoming(level, fmt, args...) \
144 log_rdma(level, LOG_INCOMING, fmt, ##args)
145#define log_read(level, fmt, args...) log_rdma(level, LOG_READ, fmt, ##args)
146#define log_write(level, fmt, args...) log_rdma(level, LOG_WRITE, fmt, ##args)
147#define log_rdma_send(level, fmt, args...) \
148 log_rdma(level, LOG_RDMA_SEND, fmt, ##args)
149#define log_rdma_recv(level, fmt, args...) \
150 log_rdma(level, LOG_RDMA_RECV, fmt, ##args)
151#define log_keep_alive(level, fmt, args...) \
152 log_rdma(level, LOG_KEEP_ALIVE, fmt, ##args)
153#define log_rdma_event(level, fmt, args...) \
154 log_rdma(level, LOG_RDMA_EVENT, fmt, ##args)
155#define log_rdma_mr(level, fmt, args...) \
156 log_rdma(level, LOG_RDMA_MR, fmt, ##args)
157
158static void smbd_disconnect_rdma_work(struct work_struct *work)
159{
160 struct smbd_connection *info =
161 container_of(work, struct smbd_connection, disconnect_work);
162
163 if (info->transport_status == SMBD_CONNECTED) {
164 info->transport_status = SMBD_DISCONNECTING;
165 rdma_disconnect(info->id);
166 }
167}
168
169static void smbd_disconnect_rdma_connection(struct smbd_connection *info)
170{
171 queue_work(info->workqueue, &info->disconnect_work);
172}
173
174/* Upcall from RDMA CM */
175static int smbd_conn_upcall(
176 struct rdma_cm_id *id, struct rdma_cm_event *event)
177{
178 struct smbd_connection *info = id->context;
179
180 log_rdma_event(INFO, "event=%d status=%d\n",
181 event->event, event->status);
182
183 switch (event->event) {
184 case RDMA_CM_EVENT_ADDR_RESOLVED:
185 case RDMA_CM_EVENT_ROUTE_RESOLVED:
186 info->ri_rc = 0;
187 complete(&info->ri_done);
188 break;
189
190 case RDMA_CM_EVENT_ADDR_ERROR:
191 info->ri_rc = -EHOSTUNREACH;
192 complete(&info->ri_done);
193 break;
194
195 case RDMA_CM_EVENT_ROUTE_ERROR:
196 info->ri_rc = -ENETUNREACH;
197 complete(&info->ri_done);
198 break;
199
200 case RDMA_CM_EVENT_ESTABLISHED:
201 log_rdma_event(INFO, "connected event=%d\n", event->event);
202 info->transport_status = SMBD_CONNECTED;
203 wake_up_interruptible(&info->conn_wait);
204 break;
205
206 case RDMA_CM_EVENT_CONNECT_ERROR:
207 case RDMA_CM_EVENT_UNREACHABLE:
208 case RDMA_CM_EVENT_REJECTED:
209 log_rdma_event(INFO, "connecting failed event=%d\n", event->event);
210 info->transport_status = SMBD_DISCONNECTED;
211 wake_up_interruptible(&info->conn_wait);
212 break;
213
214 case RDMA_CM_EVENT_DEVICE_REMOVAL:
215 case RDMA_CM_EVENT_DISCONNECTED:
216 /* This happenes when we fail the negotiation */
217 if (info->transport_status == SMBD_NEGOTIATE_FAILED) {
218 info->transport_status = SMBD_DISCONNECTED;
219 wake_up(&info->conn_wait);
220 break;
221 }
222
223 info->transport_status = SMBD_DISCONNECTED;
224 wake_up_interruptible(&info->disconn_wait);
225 wake_up_interruptible(&info->wait_reassembly_queue);
226 wake_up_interruptible_all(&info->wait_send_queue);
227 break;
228
229 default:
230 break;
231 }
232
233 return 0;
234}
235
236/* Upcall from RDMA QP */
237static void
238smbd_qp_async_error_upcall(struct ib_event *event, void *context)
239{
240 struct smbd_connection *info = context;
241
242 log_rdma_event(ERR, "%s on device %s info %p\n",
243 ib_event_msg(event->event), event->device->name, info);
244
245 switch (event->event) {
246 case IB_EVENT_CQ_ERR:
247 case IB_EVENT_QP_FATAL:
248 smbd_disconnect_rdma_connection(info);
249 break;
250
251 default:
252 break;
253 }
254}
255
256static inline void *smbd_request_payload(struct smbd_request *request)
257{
258 return (void *)request->packet;
259}
260
261static inline void *smbd_response_payload(struct smbd_response *response)
262{
263 return (void *)response->packet;
264}
265
266/* Called when a RDMA send is done */
267static void send_done(struct ib_cq *cq, struct ib_wc *wc)
268{
269 int i;
270 struct smbd_request *request =
271 container_of(wc->wr_cqe, struct smbd_request, cqe);
272
273 log_rdma_send(INFO, "smbd_request 0x%p completed wc->status=%d\n",
274 request, wc->status);
275
276 if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_SEND) {
277 log_rdma_send(ERR, "wc->status=%d wc->opcode=%d\n",
278 wc->status, wc->opcode);
279 smbd_disconnect_rdma_connection(request->info);
280 }
281
282 for (i = 0; i < request->num_sge; i++)
283 ib_dma_unmap_single(request->info->id->device,
284 request->sge[i].addr,
285 request->sge[i].length,
286 DMA_TO_DEVICE);
287
288 if (atomic_dec_and_test(&request->info->send_pending))
289 wake_up(&request->info->wait_send_pending);
290
291 wake_up(&request->info->wait_post_send);
292
293 mempool_free(request, request->info->request_mempool);
294}
295
296static void dump_smbd_negotiate_resp(struct smbd_negotiate_resp *resp)
297{
298 log_rdma_event(INFO, "resp message min_version %u max_version %u negotiated_version %u credits_requested %u credits_granted %u status %u max_readwrite_size %u preferred_send_size %u max_receive_size %u max_fragmented_size %u\n",
299 resp->min_version, resp->max_version,
300 resp->negotiated_version, resp->credits_requested,
301 resp->credits_granted, resp->status,
302 resp->max_readwrite_size, resp->preferred_send_size,
303 resp->max_receive_size, resp->max_fragmented_size);
304}
305
306/*
307 * Process a negotiation response message, according to [MS-SMBD]3.1.5.7
308 * response, packet_length: the negotiation response message
309 * return value: true if negotiation is a success, false if failed
310 */
311static bool process_negotiation_response(
312 struct smbd_response *response, int packet_length)
313{
314 struct smbd_connection *info = response->info;
315 struct smbd_negotiate_resp *packet = smbd_response_payload(response);
316
317 if (packet_length < sizeof(struct smbd_negotiate_resp)) {
318 log_rdma_event(ERR,
319 "error: packet_length=%d\n", packet_length);
320 return false;
321 }
322
323 if (le16_to_cpu(packet->negotiated_version) != SMBD_V1) {
324 log_rdma_event(ERR, "error: negotiated_version=%x\n",
325 le16_to_cpu(packet->negotiated_version));
326 return false;
327 }
328 info->protocol = le16_to_cpu(packet->negotiated_version);
329
330 if (packet->credits_requested == 0) {
331 log_rdma_event(ERR, "error: credits_requested==0\n");
332 return false;
333 }
334 info->receive_credit_target = le16_to_cpu(packet->credits_requested);
335
336 if (packet->credits_granted == 0) {
337 log_rdma_event(ERR, "error: credits_granted==0\n");
338 return false;
339 }
340 atomic_set(&info->send_credits, le16_to_cpu(packet->credits_granted));
341
342 atomic_set(&info->receive_credits, 0);
343
344 if (le32_to_cpu(packet->preferred_send_size) > info->max_receive_size) {
345 log_rdma_event(ERR, "error: preferred_send_size=%d\n",
346 le32_to_cpu(packet->preferred_send_size));
347 return false;
348 }
349 info->max_receive_size = le32_to_cpu(packet->preferred_send_size);
350
351 if (le32_to_cpu(packet->max_receive_size) < SMBD_MIN_RECEIVE_SIZE) {
352 log_rdma_event(ERR, "error: max_receive_size=%d\n",
353 le32_to_cpu(packet->max_receive_size));
354 return false;
355 }
356 info->max_send_size = min_t(int, info->max_send_size,
357 le32_to_cpu(packet->max_receive_size));
358
359 if (le32_to_cpu(packet->max_fragmented_size) <
360 SMBD_MIN_FRAGMENTED_SIZE) {
361 log_rdma_event(ERR, "error: max_fragmented_size=%d\n",
362 le32_to_cpu(packet->max_fragmented_size));
363 return false;
364 }
365 info->max_fragmented_send_size =
366 le32_to_cpu(packet->max_fragmented_size);
367 info->rdma_readwrite_threshold =
368 rdma_readwrite_threshold > info->max_fragmented_send_size ?
369 info->max_fragmented_send_size :
370 rdma_readwrite_threshold;
371
372
373 info->max_readwrite_size = min_t(u32,
374 le32_to_cpu(packet->max_readwrite_size),
375 info->max_frmr_depth * PAGE_SIZE);
376 info->max_frmr_depth = info->max_readwrite_size / PAGE_SIZE;
377
378 return true;
379}
380
381static void smbd_post_send_credits(struct work_struct *work)
382{
383 int ret = 0;
384 int use_receive_queue = 1;
385 int rc;
386 struct smbd_response *response;
387 struct smbd_connection *info =
388 container_of(work, struct smbd_connection,
389 post_send_credits_work);
390
391 if (info->transport_status != SMBD_CONNECTED) {
392 wake_up(&info->wait_receive_queues);
393 return;
394 }
395
396 if (info->receive_credit_target >
397 atomic_read(&info->receive_credits)) {
398 while (true) {
399 if (use_receive_queue)
400 response = get_receive_buffer(info);
401 else
402 response = get_empty_queue_buffer(info);
403 if (!response) {
404 /* now switch to emtpy packet queue */
405 if (use_receive_queue) {
406 use_receive_queue = 0;
407 continue;
408 } else
409 break;
410 }
411
412 response->type = SMBD_TRANSFER_DATA;
413 response->first_segment = false;
414 rc = smbd_post_recv(info, response);
415 if (rc) {
416 log_rdma_recv(ERR,
417 "post_recv failed rc=%d\n", rc);
418 put_receive_buffer(info, response);
419 break;
420 }
421
422 ret++;
423 }
424 }
425
426 spin_lock(&info->lock_new_credits_offered);
427 info->new_credits_offered += ret;
428 spin_unlock(&info->lock_new_credits_offered);
429
430 /* Promptly send an immediate packet as defined in [MS-SMBD] 3.1.1.1 */
431 info->send_immediate = true;
432 if (atomic_read(&info->receive_credits) <
433 info->receive_credit_target - 1) {
434 if (info->keep_alive_requested == KEEP_ALIVE_PENDING ||
435 info->send_immediate) {
436 log_keep_alive(INFO, "send an empty message\n");
437 smbd_post_send_empty(info);
438 }
439 }
440}
441
442/* Called from softirq, when recv is done */
443static void recv_done(struct ib_cq *cq, struct ib_wc *wc)
444{
445 struct smbd_data_transfer *data_transfer;
446 struct smbd_response *response =
447 container_of(wc->wr_cqe, struct smbd_response, cqe);
448 struct smbd_connection *info = response->info;
449 int data_length = 0;
450
451 log_rdma_recv(INFO, "response=0x%p type=%d wc status=%d wc opcode %d byte_len=%d pkey_index=%u\n",
452 response, response->type, wc->status, wc->opcode,
453 wc->byte_len, wc->pkey_index);
454
455 if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_RECV) {
456 log_rdma_recv(INFO, "wc->status=%d opcode=%d\n",
457 wc->status, wc->opcode);
458 smbd_disconnect_rdma_connection(info);
459 goto error;
460 }
461
462 ib_dma_sync_single_for_cpu(
463 wc->qp->device,
464 response->sge.addr,
465 response->sge.length,
466 DMA_FROM_DEVICE);
467
468 switch (response->type) {
469 /* SMBD negotiation response */
470 case SMBD_NEGOTIATE_RESP:
471 dump_smbd_negotiate_resp(smbd_response_payload(response));
472 info->full_packet_received = true;
473 info->negotiate_done =
474 process_negotiation_response(response, wc->byte_len);
475 complete(&info->negotiate_completion);
476 break;
477
478 /* SMBD data transfer packet */
479 case SMBD_TRANSFER_DATA:
480 data_transfer = smbd_response_payload(response);
481 data_length = le32_to_cpu(data_transfer->data_length);
482
483 /*
484 * If this is a packet with data playload place the data in
485 * reassembly queue and wake up the reading thread
486 */
487 if (data_length) {
488 if (info->full_packet_received)
489 response->first_segment = true;
490
491 if (le32_to_cpu(data_transfer->remaining_data_length))
492 info->full_packet_received = false;
493 else
494 info->full_packet_received = true;
495
496 enqueue_reassembly(
497 info,
498 response,
499 data_length);
500 } else
501 put_empty_packet(info, response);
502
503 if (data_length)
504 wake_up_interruptible(&info->wait_reassembly_queue);
505
506 atomic_dec(&info->receive_credits);
507 info->receive_credit_target =
508 le16_to_cpu(data_transfer->credits_requested);
509 if (le16_to_cpu(data_transfer->credits_granted)) {
510 atomic_add(le16_to_cpu(data_transfer->credits_granted),
511 &info->send_credits);
512 /*
513 * We have new send credits granted from remote peer
514 * If any sender is waiting for credits, unblock it
515 */
516 wake_up_interruptible(&info->wait_send_queue);
517 }
518
519 log_incoming(INFO, "data flags %d data_offset %d data_length %d remaining_data_length %d\n",
520 le16_to_cpu(data_transfer->flags),
521 le32_to_cpu(data_transfer->data_offset),
522 le32_to_cpu(data_transfer->data_length),
523 le32_to_cpu(data_transfer->remaining_data_length));
524
525 /* Send a KEEP_ALIVE response right away if requested */
526 info->keep_alive_requested = KEEP_ALIVE_NONE;
527 if (le16_to_cpu(data_transfer->flags) &
528 SMB_DIRECT_RESPONSE_REQUESTED) {
529 info->keep_alive_requested = KEEP_ALIVE_PENDING;
530 }
531
532 return;
533
534 default:
535 log_rdma_recv(ERR,
536 "unexpected response type=%d\n", response->type);
537 }
538
539error:
540 put_receive_buffer(info, response);
541}
542
543static struct rdma_cm_id *smbd_create_id(
544 struct smbd_connection *info,
545 struct sockaddr *dstaddr, int port)
546{
547 struct rdma_cm_id *id;
548 int rc;
549 __be16 *sport;
550
551 id = rdma_create_id(&init_net, smbd_conn_upcall, info,
552 RDMA_PS_TCP, IB_QPT_RC);
553 if (IS_ERR(id)) {
554 rc = PTR_ERR(id);
555 log_rdma_event(ERR, "rdma_create_id() failed %i\n", rc);
556 return id;
557 }
558
559 if (dstaddr->sa_family == AF_INET6)
560 sport = &((struct sockaddr_in6 *)dstaddr)->sin6_port;
561 else
562 sport = &((struct sockaddr_in *)dstaddr)->sin_port;
563
564 *sport = htons(port);
565
566 init_completion(&info->ri_done);
567 info->ri_rc = -ETIMEDOUT;
568
569 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)dstaddr,
570 RDMA_RESOLVE_TIMEOUT);
571 if (rc) {
572 log_rdma_event(ERR, "rdma_resolve_addr() failed %i\n", rc);
573 goto out;
574 }
575 rc = wait_for_completion_interruptible_timeout(
576 &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT));
577 /* e.g. if interrupted returns -ERESTARTSYS */
578 if (rc < 0) {
579 log_rdma_event(ERR, "rdma_resolve_addr timeout rc: %i\n", rc);
580 goto out;
581 }
582 rc = info->ri_rc;
583 if (rc) {
584 log_rdma_event(ERR, "rdma_resolve_addr() completed %i\n", rc);
585 goto out;
586 }
587
588 info->ri_rc = -ETIMEDOUT;
589 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
590 if (rc) {
591 log_rdma_event(ERR, "rdma_resolve_route() failed %i\n", rc);
592 goto out;
593 }
594 rc = wait_for_completion_interruptible_timeout(
595 &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT));
596 /* e.g. if interrupted returns -ERESTARTSYS */
597 if (rc < 0) {
598 log_rdma_event(ERR, "rdma_resolve_addr timeout rc: %i\n", rc);
599 goto out;
600 }
601 rc = info->ri_rc;
602 if (rc) {
603 log_rdma_event(ERR, "rdma_resolve_route() completed %i\n", rc);
604 goto out;
605 }
606
607 return id;
608
609out:
610 rdma_destroy_id(id);
611 return ERR_PTR(rc);
612}
613
614/*
615 * Test if FRWR (Fast Registration Work Requests) is supported on the device
616 * This implementation requries FRWR on RDMA read/write
617 * return value: true if it is supported
618 */
619static bool frwr_is_supported(struct ib_device_attr *attrs)
620{
621 if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
622 return false;
623 if (attrs->max_fast_reg_page_list_len == 0)
624 return false;
625 return true;
626}
627
628static int smbd_ia_open(
629 struct smbd_connection *info,
630 struct sockaddr *dstaddr, int port)
631{
632 int rc;
633
634 info->id = smbd_create_id(info, dstaddr, port);
635 if (IS_ERR(info->id)) {
636 rc = PTR_ERR(info->id);
637 goto out1;
638 }
639
640 if (!frwr_is_supported(&info->id->device->attrs)) {
641 log_rdma_event(ERR, "Fast Registration Work Requests (FRWR) is not supported\n");
642 log_rdma_event(ERR, "Device capability flags = %llx max_fast_reg_page_list_len = %u\n",
643 info->id->device->attrs.device_cap_flags,
644 info->id->device->attrs.max_fast_reg_page_list_len);
645 rc = -EPROTONOSUPPORT;
646 goto out2;
647 }
648 info->max_frmr_depth = min_t(int,
649 smbd_max_frmr_depth,
650 info->id->device->attrs.max_fast_reg_page_list_len);
651 info->mr_type = IB_MR_TYPE_MEM_REG;
652 if (info->id->device->attrs.kernel_cap_flags & IBK_SG_GAPS_REG)
653 info->mr_type = IB_MR_TYPE_SG_GAPS;
654
655 info->pd = ib_alloc_pd(info->id->device, 0);
656 if (IS_ERR(info->pd)) {
657 rc = PTR_ERR(info->pd);
658 log_rdma_event(ERR, "ib_alloc_pd() returned %d\n", rc);
659 goto out2;
660 }
661
662 return 0;
663
664out2:
665 rdma_destroy_id(info->id);
666 info->id = NULL;
667
668out1:
669 return rc;
670}
671
672/*
673 * Send a negotiation request message to the peer
674 * The negotiation procedure is in [MS-SMBD] 3.1.5.2 and 3.1.5.3
675 * After negotiation, the transport is connected and ready for
676 * carrying upper layer SMB payload
677 */
678static int smbd_post_send_negotiate_req(struct smbd_connection *info)
679{
680 struct ib_send_wr send_wr;
681 int rc = -ENOMEM;
682 struct smbd_request *request;
683 struct smbd_negotiate_req *packet;
684
685 request = mempool_alloc(info->request_mempool, GFP_KERNEL);
686 if (!request)
687 return rc;
688
689 request->info = info;
690
691 packet = smbd_request_payload(request);
692 packet->min_version = cpu_to_le16(SMBD_V1);
693 packet->max_version = cpu_to_le16(SMBD_V1);
694 packet->reserved = 0;
695 packet->credits_requested = cpu_to_le16(info->send_credit_target);
696 packet->preferred_send_size = cpu_to_le32(info->max_send_size);
697 packet->max_receive_size = cpu_to_le32(info->max_receive_size);
698 packet->max_fragmented_size =
699 cpu_to_le32(info->max_fragmented_recv_size);
700
701 request->num_sge = 1;
702 request->sge[0].addr = ib_dma_map_single(
703 info->id->device, (void *)packet,
704 sizeof(*packet), DMA_TO_DEVICE);
705 if (ib_dma_mapping_error(info->id->device, request->sge[0].addr)) {
706 rc = -EIO;
707 goto dma_mapping_failed;
708 }
709
710 request->sge[0].length = sizeof(*packet);
711 request->sge[0].lkey = info->pd->local_dma_lkey;
712
713 ib_dma_sync_single_for_device(
714 info->id->device, request->sge[0].addr,
715 request->sge[0].length, DMA_TO_DEVICE);
716
717 request->cqe.done = send_done;
718
719 send_wr.next = NULL;
720 send_wr.wr_cqe = &request->cqe;
721 send_wr.sg_list = request->sge;
722 send_wr.num_sge = request->num_sge;
723 send_wr.opcode = IB_WR_SEND;
724 send_wr.send_flags = IB_SEND_SIGNALED;
725
726 log_rdma_send(INFO, "sge addr=0x%llx length=%u lkey=0x%x\n",
727 request->sge[0].addr,
728 request->sge[0].length, request->sge[0].lkey);
729
730 atomic_inc(&info->send_pending);
731 rc = ib_post_send(info->id->qp, &send_wr, NULL);
732 if (!rc)
733 return 0;
734
735 /* if we reach here, post send failed */
736 log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc);
737 atomic_dec(&info->send_pending);
738 ib_dma_unmap_single(info->id->device, request->sge[0].addr,
739 request->sge[0].length, DMA_TO_DEVICE);
740
741 smbd_disconnect_rdma_connection(info);
742
743dma_mapping_failed:
744 mempool_free(request, info->request_mempool);
745 return rc;
746}
747
748/*
749 * Extend the credits to remote peer
750 * This implements [MS-SMBD] 3.1.5.9
751 * The idea is that we should extend credits to remote peer as quickly as
752 * it's allowed, to maintain data flow. We allocate as much receive
753 * buffer as possible, and extend the receive credits to remote peer
754 * return value: the new credtis being granted.
755 */
756static int manage_credits_prior_sending(struct smbd_connection *info)
757{
758 int new_credits;
759
760 spin_lock(&info->lock_new_credits_offered);
761 new_credits = info->new_credits_offered;
762 info->new_credits_offered = 0;
763 spin_unlock(&info->lock_new_credits_offered);
764
765 return new_credits;
766}
767
768/*
769 * Check if we need to send a KEEP_ALIVE message
770 * The idle connection timer triggers a KEEP_ALIVE message when expires
771 * SMB_DIRECT_RESPONSE_REQUESTED is set in the message flag to have peer send
772 * back a response.
773 * return value:
774 * 1 if SMB_DIRECT_RESPONSE_REQUESTED needs to be set
775 * 0: otherwise
776 */
777static int manage_keep_alive_before_sending(struct smbd_connection *info)
778{
779 if (info->keep_alive_requested == KEEP_ALIVE_PENDING) {
780 info->keep_alive_requested = KEEP_ALIVE_SENT;
781 return 1;
782 }
783 return 0;
784}
785
786/* Post the send request */
787static int smbd_post_send(struct smbd_connection *info,
788 struct smbd_request *request)
789{
790 struct ib_send_wr send_wr;
791 int rc, i;
792
793 for (i = 0; i < request->num_sge; i++) {
794 log_rdma_send(INFO,
795 "rdma_request sge[%d] addr=0x%llx length=%u\n",
796 i, request->sge[i].addr, request->sge[i].length);
797 ib_dma_sync_single_for_device(
798 info->id->device,
799 request->sge[i].addr,
800 request->sge[i].length,
801 DMA_TO_DEVICE);
802 }
803
804 request->cqe.done = send_done;
805
806 send_wr.next = NULL;
807 send_wr.wr_cqe = &request->cqe;
808 send_wr.sg_list = request->sge;
809 send_wr.num_sge = request->num_sge;
810 send_wr.opcode = IB_WR_SEND;
811 send_wr.send_flags = IB_SEND_SIGNALED;
812
813 rc = ib_post_send(info->id->qp, &send_wr, NULL);
814 if (rc) {
815 log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc);
816 smbd_disconnect_rdma_connection(info);
817 rc = -EAGAIN;
818 } else
819 /* Reset timer for idle connection after packet is sent */
820 mod_delayed_work(info->workqueue, &info->idle_timer_work,
821 info->keep_alive_interval*HZ);
822
823 return rc;
824}
825
826static int smbd_post_send_sgl(struct smbd_connection *info,
827 struct scatterlist *sgl, int data_length, int remaining_data_length)
828{
829 int num_sgs;
830 int i, rc;
831 int header_length;
832 struct smbd_request *request;
833 struct smbd_data_transfer *packet;
834 int new_credits;
835 struct scatterlist *sg;
836
837wait_credit:
838 /* Wait for send credits. A SMBD packet needs one credit */
839 rc = wait_event_interruptible(info->wait_send_queue,
840 atomic_read(&info->send_credits) > 0 ||
841 info->transport_status != SMBD_CONNECTED);
842 if (rc)
843 goto err_wait_credit;
844
845 if (info->transport_status != SMBD_CONNECTED) {
846 log_outgoing(ERR, "disconnected not sending on wait_credit\n");
847 rc = -EAGAIN;
848 goto err_wait_credit;
849 }
850 if (unlikely(atomic_dec_return(&info->send_credits) < 0)) {
851 atomic_inc(&info->send_credits);
852 goto wait_credit;
853 }
854
855wait_send_queue:
856 wait_event(info->wait_post_send,
857 atomic_read(&info->send_pending) < info->send_credit_target ||
858 info->transport_status != SMBD_CONNECTED);
859
860 if (info->transport_status != SMBD_CONNECTED) {
861 log_outgoing(ERR, "disconnected not sending on wait_send_queue\n");
862 rc = -EAGAIN;
863 goto err_wait_send_queue;
864 }
865
866 if (unlikely(atomic_inc_return(&info->send_pending) >
867 info->send_credit_target)) {
868 atomic_dec(&info->send_pending);
869 goto wait_send_queue;
870 }
871
872 request = mempool_alloc(info->request_mempool, GFP_KERNEL);
873 if (!request) {
874 rc = -ENOMEM;
875 goto err_alloc;
876 }
877
878 request->info = info;
879
880 /* Fill in the packet header */
881 packet = smbd_request_payload(request);
882 packet->credits_requested = cpu_to_le16(info->send_credit_target);
883
884 new_credits = manage_credits_prior_sending(info);
885 atomic_add(new_credits, &info->receive_credits);
886 packet->credits_granted = cpu_to_le16(new_credits);
887
888 info->send_immediate = false;
889
890 packet->flags = 0;
891 if (manage_keep_alive_before_sending(info))
892 packet->flags |= cpu_to_le16(SMB_DIRECT_RESPONSE_REQUESTED);
893
894 packet->reserved = 0;
895 if (!data_length)
896 packet->data_offset = 0;
897 else
898 packet->data_offset = cpu_to_le32(24);
899 packet->data_length = cpu_to_le32(data_length);
900 packet->remaining_data_length = cpu_to_le32(remaining_data_length);
901 packet->padding = 0;
902
903 log_outgoing(INFO, "credits_requested=%d credits_granted=%d data_offset=%d data_length=%d remaining_data_length=%d\n",
904 le16_to_cpu(packet->credits_requested),
905 le16_to_cpu(packet->credits_granted),
906 le32_to_cpu(packet->data_offset),
907 le32_to_cpu(packet->data_length),
908 le32_to_cpu(packet->remaining_data_length));
909
910 /* Map the packet to DMA */
911 header_length = sizeof(struct smbd_data_transfer);
912 /* If this is a packet without payload, don't send padding */
913 if (!data_length)
914 header_length = offsetof(struct smbd_data_transfer, padding);
915
916 request->num_sge = 1;
917 request->sge[0].addr = ib_dma_map_single(info->id->device,
918 (void *)packet,
919 header_length,
920 DMA_TO_DEVICE);
921 if (ib_dma_mapping_error(info->id->device, request->sge[0].addr)) {
922 rc = -EIO;
923 request->sge[0].addr = 0;
924 goto err_dma;
925 }
926
927 request->sge[0].length = header_length;
928 request->sge[0].lkey = info->pd->local_dma_lkey;
929
930 /* Fill in the packet data payload */
931 num_sgs = sgl ? sg_nents(sgl) : 0;
932 for_each_sg(sgl, sg, num_sgs, i) {
933 request->sge[i+1].addr =
934 ib_dma_map_page(info->id->device, sg_page(sg),
935 sg->offset, sg->length, DMA_TO_DEVICE);
936 if (ib_dma_mapping_error(
937 info->id->device, request->sge[i+1].addr)) {
938 rc = -EIO;
939 request->sge[i+1].addr = 0;
940 goto err_dma;
941 }
942 request->sge[i+1].length = sg->length;
943 request->sge[i+1].lkey = info->pd->local_dma_lkey;
944 request->num_sge++;
945 }
946
947 rc = smbd_post_send(info, request);
948 if (!rc)
949 return 0;
950
951err_dma:
952 for (i = 0; i < request->num_sge; i++)
953 if (request->sge[i].addr)
954 ib_dma_unmap_single(info->id->device,
955 request->sge[i].addr,
956 request->sge[i].length,
957 DMA_TO_DEVICE);
958 mempool_free(request, info->request_mempool);
959
960 /* roll back receive credits and credits to be offered */
961 spin_lock(&info->lock_new_credits_offered);
962 info->new_credits_offered += new_credits;
963 spin_unlock(&info->lock_new_credits_offered);
964 atomic_sub(new_credits, &info->receive_credits);
965
966err_alloc:
967 if (atomic_dec_and_test(&info->send_pending))
968 wake_up(&info->wait_send_pending);
969
970err_wait_send_queue:
971 /* roll back send credits and pending */
972 atomic_inc(&info->send_credits);
973
974err_wait_credit:
975 return rc;
976}
977
978/*
979 * Send a page
980 * page: the page to send
981 * offset: offset in the page to send
982 * size: length in the page to send
983 * remaining_data_length: remaining data to send in this payload
984 */
985static int smbd_post_send_page(struct smbd_connection *info, struct page *page,
986 unsigned long offset, size_t size, int remaining_data_length)
987{
988 struct scatterlist sgl;
989
990 sg_init_table(&sgl, 1);
991 sg_set_page(&sgl, page, size, offset);
992
993 return smbd_post_send_sgl(info, &sgl, size, remaining_data_length);
994}
995
996/*
997 * Send an empty message
998 * Empty message is used to extend credits to peer to for keep live
999 * while there is no upper layer payload to send at the time
1000 */
1001static int smbd_post_send_empty(struct smbd_connection *info)
1002{
1003 info->count_send_empty++;
1004 return smbd_post_send_sgl(info, NULL, 0, 0);
1005}
1006
1007/*
1008 * Send a data buffer
1009 * iov: the iov array describing the data buffers
1010 * n_vec: number of iov array
1011 * remaining_data_length: remaining data to send following this packet
1012 * in segmented SMBD packet
1013 */
1014static int smbd_post_send_data(
1015 struct smbd_connection *info, struct kvec *iov, int n_vec,
1016 int remaining_data_length)
1017{
1018 int i;
1019 u32 data_length = 0;
1020 struct scatterlist sgl[SMBDIRECT_MAX_SEND_SGE - 1];
1021
1022 if (n_vec > SMBDIRECT_MAX_SEND_SGE - 1) {
1023 cifs_dbg(VFS, "Can't fit data to SGL, n_vec=%d\n", n_vec);
1024 return -EINVAL;
1025 }
1026
1027 sg_init_table(sgl, n_vec);
1028 for (i = 0; i < n_vec; i++) {
1029 data_length += iov[i].iov_len;
1030 sg_set_buf(&sgl[i], iov[i].iov_base, iov[i].iov_len);
1031 }
1032
1033 return smbd_post_send_sgl(info, sgl, data_length, remaining_data_length);
1034}
1035
1036/*
1037 * Post a receive request to the transport
1038 * The remote peer can only send data when a receive request is posted
1039 * The interaction is controlled by send/receive credit system
1040 */
1041static int smbd_post_recv(
1042 struct smbd_connection *info, struct smbd_response *response)
1043{
1044 struct ib_recv_wr recv_wr;
1045 int rc = -EIO;
1046
1047 response->sge.addr = ib_dma_map_single(
1048 info->id->device, response->packet,
1049 info->max_receive_size, DMA_FROM_DEVICE);
1050 if (ib_dma_mapping_error(info->id->device, response->sge.addr))
1051 return rc;
1052
1053 response->sge.length = info->max_receive_size;
1054 response->sge.lkey = info->pd->local_dma_lkey;
1055
1056 response->cqe.done = recv_done;
1057
1058 recv_wr.wr_cqe = &response->cqe;
1059 recv_wr.next = NULL;
1060 recv_wr.sg_list = &response->sge;
1061 recv_wr.num_sge = 1;
1062
1063 rc = ib_post_recv(info->id->qp, &recv_wr, NULL);
1064 if (rc) {
1065 ib_dma_unmap_single(info->id->device, response->sge.addr,
1066 response->sge.length, DMA_FROM_DEVICE);
1067 smbd_disconnect_rdma_connection(info);
1068 log_rdma_recv(ERR, "ib_post_recv failed rc=%d\n", rc);
1069 }
1070
1071 return rc;
1072}
1073
1074/* Perform SMBD negotiate according to [MS-SMBD] 3.1.5.2 */
1075static int smbd_negotiate(struct smbd_connection *info)
1076{
1077 int rc;
1078 struct smbd_response *response = get_receive_buffer(info);
1079
1080 response->type = SMBD_NEGOTIATE_RESP;
1081 rc = smbd_post_recv(info, response);
1082 log_rdma_event(INFO, "smbd_post_recv rc=%d iov.addr=0x%llx iov.length=%u iov.lkey=0x%x\n",
1083 rc, response->sge.addr,
1084 response->sge.length, response->sge.lkey);
1085 if (rc)
1086 return rc;
1087
1088 init_completion(&info->negotiate_completion);
1089 info->negotiate_done = false;
1090 rc = smbd_post_send_negotiate_req(info);
1091 if (rc)
1092 return rc;
1093
1094 rc = wait_for_completion_interruptible_timeout(
1095 &info->negotiate_completion, SMBD_NEGOTIATE_TIMEOUT * HZ);
1096 log_rdma_event(INFO, "wait_for_completion_timeout rc=%d\n", rc);
1097
1098 if (info->negotiate_done)
1099 return 0;
1100
1101 if (rc == 0)
1102 rc = -ETIMEDOUT;
1103 else if (rc == -ERESTARTSYS)
1104 rc = -EINTR;
1105 else
1106 rc = -ENOTCONN;
1107
1108 return rc;
1109}
1110
1111static void put_empty_packet(
1112 struct smbd_connection *info, struct smbd_response *response)
1113{
1114 spin_lock(&info->empty_packet_queue_lock);
1115 list_add_tail(&response->list, &info->empty_packet_queue);
1116 info->count_empty_packet_queue++;
1117 spin_unlock(&info->empty_packet_queue_lock);
1118
1119 queue_work(info->workqueue, &info->post_send_credits_work);
1120}
1121
1122/*
1123 * Implement Connection.FragmentReassemblyBuffer defined in [MS-SMBD] 3.1.1.1
1124 * This is a queue for reassembling upper layer payload and present to upper
1125 * layer. All the inncoming payload go to the reassembly queue, regardless of
1126 * if reassembly is required. The uuper layer code reads from the queue for all
1127 * incoming payloads.
1128 * Put a received packet to the reassembly queue
1129 * response: the packet received
1130 * data_length: the size of payload in this packet
1131 */
1132static void enqueue_reassembly(
1133 struct smbd_connection *info,
1134 struct smbd_response *response,
1135 int data_length)
1136{
1137 spin_lock(&info->reassembly_queue_lock);
1138 list_add_tail(&response->list, &info->reassembly_queue);
1139 info->reassembly_queue_length++;
1140 /*
1141 * Make sure reassembly_data_length is updated after list and
1142 * reassembly_queue_length are updated. On the dequeue side
1143 * reassembly_data_length is checked without a lock to determine
1144 * if reassembly_queue_length and list is up to date
1145 */
1146 virt_wmb();
1147 info->reassembly_data_length += data_length;
1148 spin_unlock(&info->reassembly_queue_lock);
1149 info->count_reassembly_queue++;
1150 info->count_enqueue_reassembly_queue++;
1151}
1152
1153/*
1154 * Get the first entry at the front of reassembly queue
1155 * Caller is responsible for locking
1156 * return value: the first entry if any, NULL if queue is empty
1157 */
1158static struct smbd_response *_get_first_reassembly(struct smbd_connection *info)
1159{
1160 struct smbd_response *ret = NULL;
1161
1162 if (!list_empty(&info->reassembly_queue)) {
1163 ret = list_first_entry(
1164 &info->reassembly_queue,
1165 struct smbd_response, list);
1166 }
1167 return ret;
1168}
1169
1170static struct smbd_response *get_empty_queue_buffer(
1171 struct smbd_connection *info)
1172{
1173 struct smbd_response *ret = NULL;
1174 unsigned long flags;
1175
1176 spin_lock_irqsave(&info->empty_packet_queue_lock, flags);
1177 if (!list_empty(&info->empty_packet_queue)) {
1178 ret = list_first_entry(
1179 &info->empty_packet_queue,
1180 struct smbd_response, list);
1181 list_del(&ret->list);
1182 info->count_empty_packet_queue--;
1183 }
1184 spin_unlock_irqrestore(&info->empty_packet_queue_lock, flags);
1185
1186 return ret;
1187}
1188
1189/*
1190 * Get a receive buffer
1191 * For each remote send, we need to post a receive. The receive buffers are
1192 * pre-allocated in advance.
1193 * return value: the receive buffer, NULL if none is available
1194 */
1195static struct smbd_response *get_receive_buffer(struct smbd_connection *info)
1196{
1197 struct smbd_response *ret = NULL;
1198 unsigned long flags;
1199
1200 spin_lock_irqsave(&info->receive_queue_lock, flags);
1201 if (!list_empty(&info->receive_queue)) {
1202 ret = list_first_entry(
1203 &info->receive_queue,
1204 struct smbd_response, list);
1205 list_del(&ret->list);
1206 info->count_receive_queue--;
1207 info->count_get_receive_buffer++;
1208 }
1209 spin_unlock_irqrestore(&info->receive_queue_lock, flags);
1210
1211 return ret;
1212}
1213
1214/*
1215 * Return a receive buffer
1216 * Upon returning of a receive buffer, we can post new receive and extend
1217 * more receive credits to remote peer. This is done immediately after a
1218 * receive buffer is returned.
1219 */
1220static void put_receive_buffer(
1221 struct smbd_connection *info, struct smbd_response *response)
1222{
1223 unsigned long flags;
1224
1225 ib_dma_unmap_single(info->id->device, response->sge.addr,
1226 response->sge.length, DMA_FROM_DEVICE);
1227
1228 spin_lock_irqsave(&info->receive_queue_lock, flags);
1229 list_add_tail(&response->list, &info->receive_queue);
1230 info->count_receive_queue++;
1231 info->count_put_receive_buffer++;
1232 spin_unlock_irqrestore(&info->receive_queue_lock, flags);
1233
1234 queue_work(info->workqueue, &info->post_send_credits_work);
1235}
1236
1237/* Preallocate all receive buffer on transport establishment */
1238static int allocate_receive_buffers(struct smbd_connection *info, int num_buf)
1239{
1240 int i;
1241 struct smbd_response *response;
1242
1243 INIT_LIST_HEAD(&info->reassembly_queue);
1244 spin_lock_init(&info->reassembly_queue_lock);
1245 info->reassembly_data_length = 0;
1246 info->reassembly_queue_length = 0;
1247
1248 INIT_LIST_HEAD(&info->receive_queue);
1249 spin_lock_init(&info->receive_queue_lock);
1250 info->count_receive_queue = 0;
1251
1252 INIT_LIST_HEAD(&info->empty_packet_queue);
1253 spin_lock_init(&info->empty_packet_queue_lock);
1254 info->count_empty_packet_queue = 0;
1255
1256 init_waitqueue_head(&info->wait_receive_queues);
1257
1258 for (i = 0; i < num_buf; i++) {
1259 response = mempool_alloc(info->response_mempool, GFP_KERNEL);
1260 if (!response)
1261 goto allocate_failed;
1262
1263 response->info = info;
1264 list_add_tail(&response->list, &info->receive_queue);
1265 info->count_receive_queue++;
1266 }
1267
1268 return 0;
1269
1270allocate_failed:
1271 while (!list_empty(&info->receive_queue)) {
1272 response = list_first_entry(
1273 &info->receive_queue,
1274 struct smbd_response, list);
1275 list_del(&response->list);
1276 info->count_receive_queue--;
1277
1278 mempool_free(response, info->response_mempool);
1279 }
1280 return -ENOMEM;
1281}
1282
1283static void destroy_receive_buffers(struct smbd_connection *info)
1284{
1285 struct smbd_response *response;
1286
1287 while ((response = get_receive_buffer(info)))
1288 mempool_free(response, info->response_mempool);
1289
1290 while ((response = get_empty_queue_buffer(info)))
1291 mempool_free(response, info->response_mempool);
1292}
1293
1294/* Implement idle connection timer [MS-SMBD] 3.1.6.2 */
1295static void idle_connection_timer(struct work_struct *work)
1296{
1297 struct smbd_connection *info = container_of(
1298 work, struct smbd_connection,
1299 idle_timer_work.work);
1300
1301 if (info->keep_alive_requested != KEEP_ALIVE_NONE) {
1302 log_keep_alive(ERR,
1303 "error status info->keep_alive_requested=%d\n",
1304 info->keep_alive_requested);
1305 smbd_disconnect_rdma_connection(info);
1306 return;
1307 }
1308
1309 log_keep_alive(INFO, "about to send an empty idle message\n");
1310 smbd_post_send_empty(info);
1311
1312 /* Setup the next idle timeout work */
1313 queue_delayed_work(info->workqueue, &info->idle_timer_work,
1314 info->keep_alive_interval*HZ);
1315}
1316
1317/*
1318 * Destroy the transport and related RDMA and memory resources
1319 * Need to go through all the pending counters and make sure on one is using
1320 * the transport while it is destroyed
1321 */
1322void smbd_destroy(struct TCP_Server_Info *server)
1323{
1324 struct smbd_connection *info = server->smbd_conn;
1325 struct smbd_response *response;
1326 unsigned long flags;
1327
1328 if (!info) {
1329 log_rdma_event(INFO, "rdma session already destroyed\n");
1330 return;
1331 }
1332
1333 log_rdma_event(INFO, "destroying rdma session\n");
1334 if (info->transport_status != SMBD_DISCONNECTED) {
1335 rdma_disconnect(server->smbd_conn->id);
1336 log_rdma_event(INFO, "wait for transport being disconnected\n");
1337 wait_event_interruptible(
1338 info->disconn_wait,
1339 info->transport_status == SMBD_DISCONNECTED);
1340 }
1341
1342 log_rdma_event(INFO, "destroying qp\n");
1343 ib_drain_qp(info->id->qp);
1344 rdma_destroy_qp(info->id);
1345
1346 log_rdma_event(INFO, "cancelling idle timer\n");
1347 cancel_delayed_work_sync(&info->idle_timer_work);
1348
1349 log_rdma_event(INFO, "wait for all send posted to IB to finish\n");
1350 wait_event(info->wait_send_pending,
1351 atomic_read(&info->send_pending) == 0);
1352
1353 /* It's not possible for upper layer to get to reassembly */
1354 log_rdma_event(INFO, "drain the reassembly queue\n");
1355 do {
1356 spin_lock_irqsave(&info->reassembly_queue_lock, flags);
1357 response = _get_first_reassembly(info);
1358 if (response) {
1359 list_del(&response->list);
1360 spin_unlock_irqrestore(
1361 &info->reassembly_queue_lock, flags);
1362 put_receive_buffer(info, response);
1363 } else
1364 spin_unlock_irqrestore(
1365 &info->reassembly_queue_lock, flags);
1366 } while (response);
1367 info->reassembly_data_length = 0;
1368
1369 log_rdma_event(INFO, "free receive buffers\n");
1370 wait_event(info->wait_receive_queues,
1371 info->count_receive_queue + info->count_empty_packet_queue
1372 == info->receive_credit_max);
1373 destroy_receive_buffers(info);
1374
1375 /*
1376 * For performance reasons, memory registration and deregistration
1377 * are not locked by srv_mutex. It is possible some processes are
1378 * blocked on transport srv_mutex while holding memory registration.
1379 * Release the transport srv_mutex to allow them to hit the failure
1380 * path when sending data, and then release memory registartions.
1381 */
1382 log_rdma_event(INFO, "freeing mr list\n");
1383 wake_up_interruptible_all(&info->wait_mr);
1384 while (atomic_read(&info->mr_used_count)) {
1385 cifs_server_unlock(server);
1386 msleep(1000);
1387 cifs_server_lock(server);
1388 }
1389 destroy_mr_list(info);
1390
1391 ib_free_cq(info->send_cq);
1392 ib_free_cq(info->recv_cq);
1393 ib_dealloc_pd(info->pd);
1394 rdma_destroy_id(info->id);
1395
1396 /* free mempools */
1397 mempool_destroy(info->request_mempool);
1398 kmem_cache_destroy(info->request_cache);
1399
1400 mempool_destroy(info->response_mempool);
1401 kmem_cache_destroy(info->response_cache);
1402
1403 info->transport_status = SMBD_DESTROYED;
1404
1405 destroy_workqueue(info->workqueue);
1406 log_rdma_event(INFO, "rdma session destroyed\n");
1407 kfree(info);
1408 server->smbd_conn = NULL;
1409}
1410
1411/*
1412 * Reconnect this SMBD connection, called from upper layer
1413 * return value: 0 on success, or actual error code
1414 */
1415int smbd_reconnect(struct TCP_Server_Info *server)
1416{
1417 log_rdma_event(INFO, "reconnecting rdma session\n");
1418
1419 if (!server->smbd_conn) {
1420 log_rdma_event(INFO, "rdma session already destroyed\n");
1421 goto create_conn;
1422 }
1423
1424 /*
1425 * This is possible if transport is disconnected and we haven't received
1426 * notification from RDMA, but upper layer has detected timeout
1427 */
1428 if (server->smbd_conn->transport_status == SMBD_CONNECTED) {
1429 log_rdma_event(INFO, "disconnecting transport\n");
1430 smbd_destroy(server);
1431 }
1432
1433create_conn:
1434 log_rdma_event(INFO, "creating rdma session\n");
1435 server->smbd_conn = smbd_get_connection(
1436 server, (struct sockaddr *) &server->dstaddr);
1437
1438 if (server->smbd_conn)
1439 cifs_dbg(VFS, "RDMA transport re-established\n");
1440
1441 return server->smbd_conn ? 0 : -ENOENT;
1442}
1443
1444static void destroy_caches_and_workqueue(struct smbd_connection *info)
1445{
1446 destroy_receive_buffers(info);
1447 destroy_workqueue(info->workqueue);
1448 mempool_destroy(info->response_mempool);
1449 kmem_cache_destroy(info->response_cache);
1450 mempool_destroy(info->request_mempool);
1451 kmem_cache_destroy(info->request_cache);
1452}
1453
1454#define MAX_NAME_LEN 80
1455static int allocate_caches_and_workqueue(struct smbd_connection *info)
1456{
1457 char name[MAX_NAME_LEN];
1458 int rc;
1459
1460 scnprintf(name, MAX_NAME_LEN, "smbd_request_%p", info);
1461 info->request_cache =
1462 kmem_cache_create(
1463 name,
1464 sizeof(struct smbd_request) +
1465 sizeof(struct smbd_data_transfer),
1466 0, SLAB_HWCACHE_ALIGN, NULL);
1467 if (!info->request_cache)
1468 return -ENOMEM;
1469
1470 info->request_mempool =
1471 mempool_create(info->send_credit_target, mempool_alloc_slab,
1472 mempool_free_slab, info->request_cache);
1473 if (!info->request_mempool)
1474 goto out1;
1475
1476 scnprintf(name, MAX_NAME_LEN, "smbd_response_%p", info);
1477 info->response_cache =
1478 kmem_cache_create(
1479 name,
1480 sizeof(struct smbd_response) +
1481 info->max_receive_size,
1482 0, SLAB_HWCACHE_ALIGN, NULL);
1483 if (!info->response_cache)
1484 goto out2;
1485
1486 info->response_mempool =
1487 mempool_create(info->receive_credit_max, mempool_alloc_slab,
1488 mempool_free_slab, info->response_cache);
1489 if (!info->response_mempool)
1490 goto out3;
1491
1492 scnprintf(name, MAX_NAME_LEN, "smbd_%p", info);
1493 info->workqueue = create_workqueue(name);
1494 if (!info->workqueue)
1495 goto out4;
1496
1497 rc = allocate_receive_buffers(info, info->receive_credit_max);
1498 if (rc) {
1499 log_rdma_event(ERR, "failed to allocate receive buffers\n");
1500 goto out5;
1501 }
1502
1503 return 0;
1504
1505out5:
1506 destroy_workqueue(info->workqueue);
1507out4:
1508 mempool_destroy(info->response_mempool);
1509out3:
1510 kmem_cache_destroy(info->response_cache);
1511out2:
1512 mempool_destroy(info->request_mempool);
1513out1:
1514 kmem_cache_destroy(info->request_cache);
1515 return -ENOMEM;
1516}
1517
1518/* Create a SMBD connection, called by upper layer */
1519static struct smbd_connection *_smbd_get_connection(
1520 struct TCP_Server_Info *server, struct sockaddr *dstaddr, int port)
1521{
1522 int rc;
1523 struct smbd_connection *info;
1524 struct rdma_conn_param conn_param;
1525 struct ib_qp_init_attr qp_attr;
1526 struct sockaddr_in *addr_in = (struct sockaddr_in *) dstaddr;
1527 struct ib_port_immutable port_immutable;
1528 u32 ird_ord_hdr[2];
1529
1530 info = kzalloc(sizeof(struct smbd_connection), GFP_KERNEL);
1531 if (!info)
1532 return NULL;
1533
1534 info->transport_status = SMBD_CONNECTING;
1535 rc = smbd_ia_open(info, dstaddr, port);
1536 if (rc) {
1537 log_rdma_event(INFO, "smbd_ia_open rc=%d\n", rc);
1538 goto create_id_failed;
1539 }
1540
1541 if (smbd_send_credit_target > info->id->device->attrs.max_cqe ||
1542 smbd_send_credit_target > info->id->device->attrs.max_qp_wr) {
1543 log_rdma_event(ERR, "consider lowering send_credit_target = %d. Possible CQE overrun, device reporting max_cqe %d max_qp_wr %d\n",
1544 smbd_send_credit_target,
1545 info->id->device->attrs.max_cqe,
1546 info->id->device->attrs.max_qp_wr);
1547 goto config_failed;
1548 }
1549
1550 if (smbd_receive_credit_max > info->id->device->attrs.max_cqe ||
1551 smbd_receive_credit_max > info->id->device->attrs.max_qp_wr) {
1552 log_rdma_event(ERR, "consider lowering receive_credit_max = %d. Possible CQE overrun, device reporting max_cqe %d max_qp_wr %d\n",
1553 smbd_receive_credit_max,
1554 info->id->device->attrs.max_cqe,
1555 info->id->device->attrs.max_qp_wr);
1556 goto config_failed;
1557 }
1558
1559 info->receive_credit_max = smbd_receive_credit_max;
1560 info->send_credit_target = smbd_send_credit_target;
1561 info->max_send_size = smbd_max_send_size;
1562 info->max_fragmented_recv_size = smbd_max_fragmented_recv_size;
1563 info->max_receive_size = smbd_max_receive_size;
1564 info->keep_alive_interval = smbd_keep_alive_interval;
1565
1566 if (info->id->device->attrs.max_send_sge < SMBDIRECT_MAX_SEND_SGE ||
1567 info->id->device->attrs.max_recv_sge < SMBDIRECT_MAX_RECV_SGE) {
1568 log_rdma_event(ERR,
1569 "device %.*s max_send_sge/max_recv_sge = %d/%d too small\n",
1570 IB_DEVICE_NAME_MAX,
1571 info->id->device->name,
1572 info->id->device->attrs.max_send_sge,
1573 info->id->device->attrs.max_recv_sge);
1574 goto config_failed;
1575 }
1576
1577 info->send_cq = NULL;
1578 info->recv_cq = NULL;
1579 info->send_cq =
1580 ib_alloc_cq_any(info->id->device, info,
1581 info->send_credit_target, IB_POLL_SOFTIRQ);
1582 if (IS_ERR(info->send_cq)) {
1583 info->send_cq = NULL;
1584 goto alloc_cq_failed;
1585 }
1586
1587 info->recv_cq =
1588 ib_alloc_cq_any(info->id->device, info,
1589 info->receive_credit_max, IB_POLL_SOFTIRQ);
1590 if (IS_ERR(info->recv_cq)) {
1591 info->recv_cq = NULL;
1592 goto alloc_cq_failed;
1593 }
1594
1595 memset(&qp_attr, 0, sizeof(qp_attr));
1596 qp_attr.event_handler = smbd_qp_async_error_upcall;
1597 qp_attr.qp_context = info;
1598 qp_attr.cap.max_send_wr = info->send_credit_target;
1599 qp_attr.cap.max_recv_wr = info->receive_credit_max;
1600 qp_attr.cap.max_send_sge = SMBDIRECT_MAX_SEND_SGE;
1601 qp_attr.cap.max_recv_sge = SMBDIRECT_MAX_RECV_SGE;
1602 qp_attr.cap.max_inline_data = 0;
1603 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1604 qp_attr.qp_type = IB_QPT_RC;
1605 qp_attr.send_cq = info->send_cq;
1606 qp_attr.recv_cq = info->recv_cq;
1607 qp_attr.port_num = ~0;
1608
1609 rc = rdma_create_qp(info->id, info->pd, &qp_attr);
1610 if (rc) {
1611 log_rdma_event(ERR, "rdma_create_qp failed %i\n", rc);
1612 goto create_qp_failed;
1613 }
1614
1615 memset(&conn_param, 0, sizeof(conn_param));
1616 conn_param.initiator_depth = 0;
1617
1618 conn_param.responder_resources =
1619 info->id->device->attrs.max_qp_rd_atom
1620 < SMBD_CM_RESPONDER_RESOURCES ?
1621 info->id->device->attrs.max_qp_rd_atom :
1622 SMBD_CM_RESPONDER_RESOURCES;
1623 info->responder_resources = conn_param.responder_resources;
1624 log_rdma_mr(INFO, "responder_resources=%d\n",
1625 info->responder_resources);
1626
1627 /* Need to send IRD/ORD in private data for iWARP */
1628 info->id->device->ops.get_port_immutable(
1629 info->id->device, info->id->port_num, &port_immutable);
1630 if (port_immutable.core_cap_flags & RDMA_CORE_PORT_IWARP) {
1631 ird_ord_hdr[0] = info->responder_resources;
1632 ird_ord_hdr[1] = 1;
1633 conn_param.private_data = ird_ord_hdr;
1634 conn_param.private_data_len = sizeof(ird_ord_hdr);
1635 } else {
1636 conn_param.private_data = NULL;
1637 conn_param.private_data_len = 0;
1638 }
1639
1640 conn_param.retry_count = SMBD_CM_RETRY;
1641 conn_param.rnr_retry_count = SMBD_CM_RNR_RETRY;
1642 conn_param.flow_control = 0;
1643
1644 log_rdma_event(INFO, "connecting to IP %pI4 port %d\n",
1645 &addr_in->sin_addr, port);
1646
1647 init_waitqueue_head(&info->conn_wait);
1648 init_waitqueue_head(&info->disconn_wait);
1649 init_waitqueue_head(&info->wait_reassembly_queue);
1650 rc = rdma_connect(info->id, &conn_param);
1651 if (rc) {
1652 log_rdma_event(ERR, "rdma_connect() failed with %i\n", rc);
1653 goto rdma_connect_failed;
1654 }
1655
1656 wait_event_interruptible(
1657 info->conn_wait, info->transport_status != SMBD_CONNECTING);
1658
1659 if (info->transport_status != SMBD_CONNECTED) {
1660 log_rdma_event(ERR, "rdma_connect failed port=%d\n", port);
1661 goto rdma_connect_failed;
1662 }
1663
1664 log_rdma_event(INFO, "rdma_connect connected\n");
1665
1666 rc = allocate_caches_and_workqueue(info);
1667 if (rc) {
1668 log_rdma_event(ERR, "cache allocation failed\n");
1669 goto allocate_cache_failed;
1670 }
1671
1672 init_waitqueue_head(&info->wait_send_queue);
1673 INIT_DELAYED_WORK(&info->idle_timer_work, idle_connection_timer);
1674 queue_delayed_work(info->workqueue, &info->idle_timer_work,
1675 info->keep_alive_interval*HZ);
1676
1677 init_waitqueue_head(&info->wait_send_pending);
1678 atomic_set(&info->send_pending, 0);
1679
1680 init_waitqueue_head(&info->wait_post_send);
1681
1682 INIT_WORK(&info->disconnect_work, smbd_disconnect_rdma_work);
1683 INIT_WORK(&info->post_send_credits_work, smbd_post_send_credits);
1684 info->new_credits_offered = 0;
1685 spin_lock_init(&info->lock_new_credits_offered);
1686
1687 rc = smbd_negotiate(info);
1688 if (rc) {
1689 log_rdma_event(ERR, "smbd_negotiate rc=%d\n", rc);
1690 goto negotiation_failed;
1691 }
1692
1693 rc = allocate_mr_list(info);
1694 if (rc) {
1695 log_rdma_mr(ERR, "memory registration allocation failed\n");
1696 goto allocate_mr_failed;
1697 }
1698
1699 return info;
1700
1701allocate_mr_failed:
1702 /* At this point, need to a full transport shutdown */
1703 smbd_destroy(server);
1704 return NULL;
1705
1706negotiation_failed:
1707 cancel_delayed_work_sync(&info->idle_timer_work);
1708 destroy_caches_and_workqueue(info);
1709 info->transport_status = SMBD_NEGOTIATE_FAILED;
1710 init_waitqueue_head(&info->conn_wait);
1711 rdma_disconnect(info->id);
1712 wait_event(info->conn_wait,
1713 info->transport_status == SMBD_DISCONNECTED);
1714
1715allocate_cache_failed:
1716rdma_connect_failed:
1717 rdma_destroy_qp(info->id);
1718
1719create_qp_failed:
1720alloc_cq_failed:
1721 if (info->send_cq)
1722 ib_free_cq(info->send_cq);
1723 if (info->recv_cq)
1724 ib_free_cq(info->recv_cq);
1725
1726config_failed:
1727 ib_dealloc_pd(info->pd);
1728 rdma_destroy_id(info->id);
1729
1730create_id_failed:
1731 kfree(info);
1732 return NULL;
1733}
1734
1735struct smbd_connection *smbd_get_connection(
1736 struct TCP_Server_Info *server, struct sockaddr *dstaddr)
1737{
1738 struct smbd_connection *ret;
1739 int port = SMBD_PORT;
1740
1741try_again:
1742 ret = _smbd_get_connection(server, dstaddr, port);
1743
1744 /* Try SMB_PORT if SMBD_PORT doesn't work */
1745 if (!ret && port == SMBD_PORT) {
1746 port = SMB_PORT;
1747 goto try_again;
1748 }
1749 return ret;
1750}
1751
1752/*
1753 * Receive data from receive reassembly queue
1754 * All the incoming data packets are placed in reassembly queue
1755 * buf: the buffer to read data into
1756 * size: the length of data to read
1757 * return value: actual data read
1758 * Note: this implementation copies the data from reassebmly queue to receive
1759 * buffers used by upper layer. This is not the optimal code path. A better way
1760 * to do it is to not have upper layer allocate its receive buffers but rather
1761 * borrow the buffer from reassembly queue, and return it after data is
1762 * consumed. But this will require more changes to upper layer code, and also
1763 * need to consider packet boundaries while they still being reassembled.
1764 */
1765static int smbd_recv_buf(struct smbd_connection *info, char *buf,
1766 unsigned int size)
1767{
1768 struct smbd_response *response;
1769 struct smbd_data_transfer *data_transfer;
1770 int to_copy, to_read, data_read, offset;
1771 u32 data_length, remaining_data_length, data_offset;
1772 int rc;
1773
1774again:
1775 /*
1776 * No need to hold the reassembly queue lock all the time as we are
1777 * the only one reading from the front of the queue. The transport
1778 * may add more entries to the back of the queue at the same time
1779 */
1780 log_read(INFO, "size=%d info->reassembly_data_length=%d\n", size,
1781 info->reassembly_data_length);
1782 if (info->reassembly_data_length >= size) {
1783 int queue_length;
1784 int queue_removed = 0;
1785
1786 /*
1787 * Need to make sure reassembly_data_length is read before
1788 * reading reassembly_queue_length and calling
1789 * _get_first_reassembly. This call is lock free
1790 * as we never read at the end of the queue which are being
1791 * updated in SOFTIRQ as more data is received
1792 */
1793 virt_rmb();
1794 queue_length = info->reassembly_queue_length;
1795 data_read = 0;
1796 to_read = size;
1797 offset = info->first_entry_offset;
1798 while (data_read < size) {
1799 response = _get_first_reassembly(info);
1800 data_transfer = smbd_response_payload(response);
1801 data_length = le32_to_cpu(data_transfer->data_length);
1802 remaining_data_length =
1803 le32_to_cpu(
1804 data_transfer->remaining_data_length);
1805 data_offset = le32_to_cpu(data_transfer->data_offset);
1806
1807 /*
1808 * The upper layer expects RFC1002 length at the
1809 * beginning of the payload. Return it to indicate
1810 * the total length of the packet. This minimize the
1811 * change to upper layer packet processing logic. This
1812 * will be eventually remove when an intermediate
1813 * transport layer is added
1814 */
1815 if (response->first_segment && size == 4) {
1816 unsigned int rfc1002_len =
1817 data_length + remaining_data_length;
1818 *((__be32 *)buf) = cpu_to_be32(rfc1002_len);
1819 data_read = 4;
1820 response->first_segment = false;
1821 log_read(INFO, "returning rfc1002 length %d\n",
1822 rfc1002_len);
1823 goto read_rfc1002_done;
1824 }
1825
1826 to_copy = min_t(int, data_length - offset, to_read);
1827 memcpy(
1828 buf + data_read,
1829 (char *)data_transfer + data_offset + offset,
1830 to_copy);
1831
1832 /* move on to the next buffer? */
1833 if (to_copy == data_length - offset) {
1834 queue_length--;
1835 /*
1836 * No need to lock if we are not at the
1837 * end of the queue
1838 */
1839 if (queue_length)
1840 list_del(&response->list);
1841 else {
1842 spin_lock_irq(
1843 &info->reassembly_queue_lock);
1844 list_del(&response->list);
1845 spin_unlock_irq(
1846 &info->reassembly_queue_lock);
1847 }
1848 queue_removed++;
1849 info->count_reassembly_queue--;
1850 info->count_dequeue_reassembly_queue++;
1851 put_receive_buffer(info, response);
1852 offset = 0;
1853 log_read(INFO, "put_receive_buffer offset=0\n");
1854 } else
1855 offset += to_copy;
1856
1857 to_read -= to_copy;
1858 data_read += to_copy;
1859
1860 log_read(INFO, "_get_first_reassembly memcpy %d bytes data_transfer_length-offset=%d after that to_read=%d data_read=%d offset=%d\n",
1861 to_copy, data_length - offset,
1862 to_read, data_read, offset);
1863 }
1864
1865 spin_lock_irq(&info->reassembly_queue_lock);
1866 info->reassembly_data_length -= data_read;
1867 info->reassembly_queue_length -= queue_removed;
1868 spin_unlock_irq(&info->reassembly_queue_lock);
1869
1870 info->first_entry_offset = offset;
1871 log_read(INFO, "returning to thread data_read=%d reassembly_data_length=%d first_entry_offset=%d\n",
1872 data_read, info->reassembly_data_length,
1873 info->first_entry_offset);
1874read_rfc1002_done:
1875 return data_read;
1876 }
1877
1878 log_read(INFO, "wait_event on more data\n");
1879 rc = wait_event_interruptible(
1880 info->wait_reassembly_queue,
1881 info->reassembly_data_length >= size ||
1882 info->transport_status != SMBD_CONNECTED);
1883 /* Don't return any data if interrupted */
1884 if (rc)
1885 return rc;
1886
1887 if (info->transport_status != SMBD_CONNECTED) {
1888 log_read(ERR, "disconnected\n");
1889 return -ECONNABORTED;
1890 }
1891
1892 goto again;
1893}
1894
1895/*
1896 * Receive a page from receive reassembly queue
1897 * page: the page to read data into
1898 * to_read: the length of data to read
1899 * return value: actual data read
1900 */
1901static int smbd_recv_page(struct smbd_connection *info,
1902 struct page *page, unsigned int page_offset,
1903 unsigned int to_read)
1904{
1905 int ret;
1906 char *to_address;
1907 void *page_address;
1908
1909 /* make sure we have the page ready for read */
1910 ret = wait_event_interruptible(
1911 info->wait_reassembly_queue,
1912 info->reassembly_data_length >= to_read ||
1913 info->transport_status != SMBD_CONNECTED);
1914 if (ret)
1915 return ret;
1916
1917 /* now we can read from reassembly queue and not sleep */
1918 page_address = kmap_atomic(page);
1919 to_address = (char *) page_address + page_offset;
1920
1921 log_read(INFO, "reading from page=%p address=%p to_read=%d\n",
1922 page, to_address, to_read);
1923
1924 ret = smbd_recv_buf(info, to_address, to_read);
1925 kunmap_atomic(page_address);
1926
1927 return ret;
1928}
1929
1930/*
1931 * Receive data from transport
1932 * msg: a msghdr point to the buffer, can be ITER_KVEC or ITER_BVEC
1933 * return: total bytes read, or 0. SMB Direct will not do partial read.
1934 */
1935int smbd_recv(struct smbd_connection *info, struct msghdr *msg)
1936{
1937 char *buf;
1938 struct page *page;
1939 unsigned int to_read, page_offset;
1940 int rc;
1941
1942 if (iov_iter_rw(&msg->msg_iter) == WRITE) {
1943 /* It's a bug in upper layer to get there */
1944 cifs_dbg(VFS, "Invalid msg iter dir %u\n",
1945 iov_iter_rw(&msg->msg_iter));
1946 rc = -EINVAL;
1947 goto out;
1948 }
1949
1950 switch (iov_iter_type(&msg->msg_iter)) {
1951 case ITER_KVEC:
1952 buf = msg->msg_iter.kvec->iov_base;
1953 to_read = msg->msg_iter.kvec->iov_len;
1954 rc = smbd_recv_buf(info, buf, to_read);
1955 break;
1956
1957 case ITER_BVEC:
1958 page = msg->msg_iter.bvec->bv_page;
1959 page_offset = msg->msg_iter.bvec->bv_offset;
1960 to_read = msg->msg_iter.bvec->bv_len;
1961 rc = smbd_recv_page(info, page, page_offset, to_read);
1962 break;
1963
1964 default:
1965 /* It's a bug in upper layer to get there */
1966 cifs_dbg(VFS, "Invalid msg type %d\n",
1967 iov_iter_type(&msg->msg_iter));
1968 rc = -EINVAL;
1969 }
1970
1971out:
1972 /* SMBDirect will read it all or nothing */
1973 if (rc > 0)
1974 msg->msg_iter.count = 0;
1975 return rc;
1976}
1977
1978/*
1979 * Send data to transport
1980 * Each rqst is transported as a SMBDirect payload
1981 * rqst: the data to write
1982 * return value: 0 if successfully write, otherwise error code
1983 */
1984int smbd_send(struct TCP_Server_Info *server,
1985 int num_rqst, struct smb_rqst *rqst_array)
1986{
1987 struct smbd_connection *info = server->smbd_conn;
1988 struct kvec vecs[SMBDIRECT_MAX_SEND_SGE - 1];
1989 int nvecs;
1990 int size;
1991 unsigned int buflen, remaining_data_length;
1992 unsigned int offset, remaining_vec_data_length;
1993 int start, i, j;
1994 int max_iov_size =
1995 info->max_send_size - sizeof(struct smbd_data_transfer);
1996 struct kvec *iov;
1997 int rc;
1998 struct smb_rqst *rqst;
1999 int rqst_idx;
2000
2001 if (info->transport_status != SMBD_CONNECTED)
2002 return -EAGAIN;
2003
2004 /*
2005 * Add in the page array if there is one. The caller needs to set
2006 * rq_tailsz to PAGE_SIZE when the buffer has multiple pages and
2007 * ends at page boundary
2008 */
2009 remaining_data_length = 0;
2010 for (i = 0; i < num_rqst; i++)
2011 remaining_data_length += smb_rqst_len(server, &rqst_array[i]);
2012
2013 if (unlikely(remaining_data_length > info->max_fragmented_send_size)) {
2014 /* assertion: payload never exceeds negotiated maximum */
2015 log_write(ERR, "payload size %d > max size %d\n",
2016 remaining_data_length, info->max_fragmented_send_size);
2017 return -EINVAL;
2018 }
2019
2020 log_write(INFO, "num_rqst=%d total length=%u\n",
2021 num_rqst, remaining_data_length);
2022
2023 rqst_idx = 0;
2024 do {
2025 rqst = &rqst_array[rqst_idx];
2026 iov = rqst->rq_iov;
2027
2028 cifs_dbg(FYI, "Sending smb (RDMA): idx=%d smb_len=%lu\n",
2029 rqst_idx, smb_rqst_len(server, rqst));
2030 remaining_vec_data_length = 0;
2031 for (i = 0; i < rqst->rq_nvec; i++) {
2032 remaining_vec_data_length += iov[i].iov_len;
2033 dump_smb(iov[i].iov_base, iov[i].iov_len);
2034 }
2035
2036 log_write(INFO, "rqst_idx=%d nvec=%d rqst->rq_npages=%d rq_pagesz=%d rq_tailsz=%d buflen=%lu\n",
2037 rqst_idx, rqst->rq_nvec,
2038 rqst->rq_npages, rqst->rq_pagesz,
2039 rqst->rq_tailsz, smb_rqst_len(server, rqst));
2040
2041 start = 0;
2042 offset = 0;
2043 do {
2044 buflen = 0;
2045 i = start;
2046 j = 0;
2047 while (i < rqst->rq_nvec &&
2048 j < SMBDIRECT_MAX_SEND_SGE - 1 &&
2049 buflen < max_iov_size) {
2050
2051 vecs[j].iov_base = iov[i].iov_base + offset;
2052 if (buflen + iov[i].iov_len > max_iov_size) {
2053 vecs[j].iov_len =
2054 max_iov_size - iov[i].iov_len;
2055 buflen = max_iov_size;
2056 offset = vecs[j].iov_len;
2057 } else {
2058 vecs[j].iov_len =
2059 iov[i].iov_len - offset;
2060 buflen += vecs[j].iov_len;
2061 offset = 0;
2062 ++i;
2063 }
2064 ++j;
2065 }
2066
2067 remaining_vec_data_length -= buflen;
2068 remaining_data_length -= buflen;
2069 log_write(INFO, "sending %s iov[%d] from start=%d nvecs=%d remaining_data_length=%d\n",
2070 remaining_vec_data_length > 0 ?
2071 "partial" : "complete",
2072 rqst->rq_nvec, start, j,
2073 remaining_data_length);
2074
2075 start = i;
2076 rc = smbd_post_send_data(info, vecs, j, remaining_data_length);
2077 if (rc)
2078 goto done;
2079 } while (remaining_vec_data_length > 0);
2080
2081 /* now sending pages if there are any */
2082 for (i = 0; i < rqst->rq_npages; i++) {
2083 rqst_page_get_length(rqst, i, &buflen, &offset);
2084 nvecs = (buflen + max_iov_size - 1) / max_iov_size;
2085 log_write(INFO, "sending pages buflen=%d nvecs=%d\n",
2086 buflen, nvecs);
2087 for (j = 0; j < nvecs; j++) {
2088 size = min_t(unsigned int, max_iov_size, remaining_data_length);
2089 remaining_data_length -= size;
2090 log_write(INFO, "sending pages i=%d offset=%d size=%d remaining_data_length=%d\n",
2091 i, j * max_iov_size + offset, size,
2092 remaining_data_length);
2093 rc = smbd_post_send_page(
2094 info, rqst->rq_pages[i],
2095 j*max_iov_size + offset,
2096 size, remaining_data_length);
2097 if (rc)
2098 goto done;
2099 }
2100 }
2101 } while (++rqst_idx < num_rqst);
2102
2103done:
2104 /*
2105 * As an optimization, we don't wait for individual I/O to finish
2106 * before sending the next one.
2107 * Send them all and wait for pending send count to get to 0
2108 * that means all the I/Os have been out and we are good to return
2109 */
2110
2111 wait_event(info->wait_send_pending,
2112 atomic_read(&info->send_pending) == 0);
2113
2114 return rc;
2115}
2116
2117static void register_mr_done(struct ib_cq *cq, struct ib_wc *wc)
2118{
2119 struct smbd_mr *mr;
2120 struct ib_cqe *cqe;
2121
2122 if (wc->status) {
2123 log_rdma_mr(ERR, "status=%d\n", wc->status);
2124 cqe = wc->wr_cqe;
2125 mr = container_of(cqe, struct smbd_mr, cqe);
2126 smbd_disconnect_rdma_connection(mr->conn);
2127 }
2128}
2129
2130/*
2131 * The work queue function that recovers MRs
2132 * We need to call ib_dereg_mr() and ib_alloc_mr() before this MR can be used
2133 * again. Both calls are slow, so finish them in a workqueue. This will not
2134 * block I/O path.
2135 * There is one workqueue that recovers MRs, there is no need to lock as the
2136 * I/O requests calling smbd_register_mr will never update the links in the
2137 * mr_list.
2138 */
2139static void smbd_mr_recovery_work(struct work_struct *work)
2140{
2141 struct smbd_connection *info =
2142 container_of(work, struct smbd_connection, mr_recovery_work);
2143 struct smbd_mr *smbdirect_mr;
2144 int rc;
2145
2146 list_for_each_entry(smbdirect_mr, &info->mr_list, list) {
2147 if (smbdirect_mr->state == MR_ERROR) {
2148
2149 /* recover this MR entry */
2150 rc = ib_dereg_mr(smbdirect_mr->mr);
2151 if (rc) {
2152 log_rdma_mr(ERR,
2153 "ib_dereg_mr failed rc=%x\n",
2154 rc);
2155 smbd_disconnect_rdma_connection(info);
2156 continue;
2157 }
2158
2159 smbdirect_mr->mr = ib_alloc_mr(
2160 info->pd, info->mr_type,
2161 info->max_frmr_depth);
2162 if (IS_ERR(smbdirect_mr->mr)) {
2163 log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x max_frmr_depth=%x\n",
2164 info->mr_type,
2165 info->max_frmr_depth);
2166 smbd_disconnect_rdma_connection(info);
2167 continue;
2168 }
2169 } else
2170 /* This MR is being used, don't recover it */
2171 continue;
2172
2173 smbdirect_mr->state = MR_READY;
2174
2175 /* smbdirect_mr->state is updated by this function
2176 * and is read and updated by I/O issuing CPUs trying
2177 * to get a MR, the call to atomic_inc_return
2178 * implicates a memory barrier and guarantees this
2179 * value is updated before waking up any calls to
2180 * get_mr() from the I/O issuing CPUs
2181 */
2182 if (atomic_inc_return(&info->mr_ready_count) == 1)
2183 wake_up_interruptible(&info->wait_mr);
2184 }
2185}
2186
2187static void destroy_mr_list(struct smbd_connection *info)
2188{
2189 struct smbd_mr *mr, *tmp;
2190
2191 cancel_work_sync(&info->mr_recovery_work);
2192 list_for_each_entry_safe(mr, tmp, &info->mr_list, list) {
2193 if (mr->state == MR_INVALIDATED)
2194 ib_dma_unmap_sg(info->id->device, mr->sgl,
2195 mr->sgl_count, mr->dir);
2196 ib_dereg_mr(mr->mr);
2197 kfree(mr->sgl);
2198 kfree(mr);
2199 }
2200}
2201
2202/*
2203 * Allocate MRs used for RDMA read/write
2204 * The number of MRs will not exceed hardware capability in responder_resources
2205 * All MRs are kept in mr_list. The MR can be recovered after it's used
2206 * Recovery is done in smbd_mr_recovery_work. The content of list entry changes
2207 * as MRs are used and recovered for I/O, but the list links will not change
2208 */
2209static int allocate_mr_list(struct smbd_connection *info)
2210{
2211 int i;
2212 struct smbd_mr *smbdirect_mr, *tmp;
2213
2214 INIT_LIST_HEAD(&info->mr_list);
2215 init_waitqueue_head(&info->wait_mr);
2216 spin_lock_init(&info->mr_list_lock);
2217 atomic_set(&info->mr_ready_count, 0);
2218 atomic_set(&info->mr_used_count, 0);
2219 init_waitqueue_head(&info->wait_for_mr_cleanup);
2220 /* Allocate more MRs (2x) than hardware responder_resources */
2221 for (i = 0; i < info->responder_resources * 2; i++) {
2222 smbdirect_mr = kzalloc(sizeof(*smbdirect_mr), GFP_KERNEL);
2223 if (!smbdirect_mr)
2224 goto out;
2225 smbdirect_mr->mr = ib_alloc_mr(info->pd, info->mr_type,
2226 info->max_frmr_depth);
2227 if (IS_ERR(smbdirect_mr->mr)) {
2228 log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x max_frmr_depth=%x\n",
2229 info->mr_type, info->max_frmr_depth);
2230 goto out;
2231 }
2232 smbdirect_mr->sgl = kcalloc(
2233 info->max_frmr_depth,
2234 sizeof(struct scatterlist),
2235 GFP_KERNEL);
2236 if (!smbdirect_mr->sgl) {
2237 log_rdma_mr(ERR, "failed to allocate sgl\n");
2238 ib_dereg_mr(smbdirect_mr->mr);
2239 goto out;
2240 }
2241 smbdirect_mr->state = MR_READY;
2242 smbdirect_mr->conn = info;
2243
2244 list_add_tail(&smbdirect_mr->list, &info->mr_list);
2245 atomic_inc(&info->mr_ready_count);
2246 }
2247 INIT_WORK(&info->mr_recovery_work, smbd_mr_recovery_work);
2248 return 0;
2249
2250out:
2251 kfree(smbdirect_mr);
2252
2253 list_for_each_entry_safe(smbdirect_mr, tmp, &info->mr_list, list) {
2254 ib_dereg_mr(smbdirect_mr->mr);
2255 kfree(smbdirect_mr->sgl);
2256 kfree(smbdirect_mr);
2257 }
2258 return -ENOMEM;
2259}
2260
2261/*
2262 * Get a MR from mr_list. This function waits until there is at least one
2263 * MR available in the list. It may access the list while the
2264 * smbd_mr_recovery_work is recovering the MR list. This doesn't need a lock
2265 * as they never modify the same places. However, there may be several CPUs
2266 * issueing I/O trying to get MR at the same time, mr_list_lock is used to
2267 * protect this situation.
2268 */
2269static struct smbd_mr *get_mr(struct smbd_connection *info)
2270{
2271 struct smbd_mr *ret;
2272 int rc;
2273again:
2274 rc = wait_event_interruptible(info->wait_mr,
2275 atomic_read(&info->mr_ready_count) ||
2276 info->transport_status != SMBD_CONNECTED);
2277 if (rc) {
2278 log_rdma_mr(ERR, "wait_event_interruptible rc=%x\n", rc);
2279 return NULL;
2280 }
2281
2282 if (info->transport_status != SMBD_CONNECTED) {
2283 log_rdma_mr(ERR, "info->transport_status=%x\n",
2284 info->transport_status);
2285 return NULL;
2286 }
2287
2288 spin_lock(&info->mr_list_lock);
2289 list_for_each_entry(ret, &info->mr_list, list) {
2290 if (ret->state == MR_READY) {
2291 ret->state = MR_REGISTERED;
2292 spin_unlock(&info->mr_list_lock);
2293 atomic_dec(&info->mr_ready_count);
2294 atomic_inc(&info->mr_used_count);
2295 return ret;
2296 }
2297 }
2298
2299 spin_unlock(&info->mr_list_lock);
2300 /*
2301 * It is possible that we could fail to get MR because other processes may
2302 * try to acquire a MR at the same time. If this is the case, retry it.
2303 */
2304 goto again;
2305}
2306
2307/*
2308 * Register memory for RDMA read/write
2309 * pages[]: the list of pages to register memory with
2310 * num_pages: the number of pages to register
2311 * tailsz: if non-zero, the bytes to register in the last page
2312 * writing: true if this is a RDMA write (SMB read), false for RDMA read
2313 * need_invalidate: true if this MR needs to be locally invalidated after I/O
2314 * return value: the MR registered, NULL if failed.
2315 */
2316struct smbd_mr *smbd_register_mr(
2317 struct smbd_connection *info, struct page *pages[], int num_pages,
2318 int offset, int tailsz, bool writing, bool need_invalidate)
2319{
2320 struct smbd_mr *smbdirect_mr;
2321 int rc, i;
2322 enum dma_data_direction dir;
2323 struct ib_reg_wr *reg_wr;
2324
2325 if (num_pages > info->max_frmr_depth) {
2326 log_rdma_mr(ERR, "num_pages=%d max_frmr_depth=%d\n",
2327 num_pages, info->max_frmr_depth);
2328 return NULL;
2329 }
2330
2331 smbdirect_mr = get_mr(info);
2332 if (!smbdirect_mr) {
2333 log_rdma_mr(ERR, "get_mr returning NULL\n");
2334 return NULL;
2335 }
2336 smbdirect_mr->need_invalidate = need_invalidate;
2337 smbdirect_mr->sgl_count = num_pages;
2338 sg_init_table(smbdirect_mr->sgl, num_pages);
2339
2340 log_rdma_mr(INFO, "num_pages=0x%x offset=0x%x tailsz=0x%x\n",
2341 num_pages, offset, tailsz);
2342
2343 if (num_pages == 1) {
2344 sg_set_page(&smbdirect_mr->sgl[0], pages[0], tailsz, offset);
2345 goto skip_multiple_pages;
2346 }
2347
2348 /* We have at least two pages to register */
2349 sg_set_page(
2350 &smbdirect_mr->sgl[0], pages[0], PAGE_SIZE - offset, offset);
2351 i = 1;
2352 while (i < num_pages - 1) {
2353 sg_set_page(&smbdirect_mr->sgl[i], pages[i], PAGE_SIZE, 0);
2354 i++;
2355 }
2356 sg_set_page(&smbdirect_mr->sgl[i], pages[i],
2357 tailsz ? tailsz : PAGE_SIZE, 0);
2358
2359skip_multiple_pages:
2360 dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
2361 smbdirect_mr->dir = dir;
2362 rc = ib_dma_map_sg(info->id->device, smbdirect_mr->sgl, num_pages, dir);
2363 if (!rc) {
2364 log_rdma_mr(ERR, "ib_dma_map_sg num_pages=%x dir=%x rc=%x\n",
2365 num_pages, dir, rc);
2366 goto dma_map_error;
2367 }
2368
2369 rc = ib_map_mr_sg(smbdirect_mr->mr, smbdirect_mr->sgl, num_pages,
2370 NULL, PAGE_SIZE);
2371 if (rc != num_pages) {
2372 log_rdma_mr(ERR,
2373 "ib_map_mr_sg failed rc = %d num_pages = %x\n",
2374 rc, num_pages);
2375 goto map_mr_error;
2376 }
2377
2378 ib_update_fast_reg_key(smbdirect_mr->mr,
2379 ib_inc_rkey(smbdirect_mr->mr->rkey));
2380 reg_wr = &smbdirect_mr->wr;
2381 reg_wr->wr.opcode = IB_WR_REG_MR;
2382 smbdirect_mr->cqe.done = register_mr_done;
2383 reg_wr->wr.wr_cqe = &smbdirect_mr->cqe;
2384 reg_wr->wr.num_sge = 0;
2385 reg_wr->wr.send_flags = IB_SEND_SIGNALED;
2386 reg_wr->mr = smbdirect_mr->mr;
2387 reg_wr->key = smbdirect_mr->mr->rkey;
2388 reg_wr->access = writing ?
2389 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2390 IB_ACCESS_REMOTE_READ;
2391
2392 /*
2393 * There is no need for waiting for complemtion on ib_post_send
2394 * on IB_WR_REG_MR. Hardware enforces a barrier and order of execution
2395 * on the next ib_post_send when we actaully send I/O to remote peer
2396 */
2397 rc = ib_post_send(info->id->qp, ®_wr->wr, NULL);
2398 if (!rc)
2399 return smbdirect_mr;
2400
2401 log_rdma_mr(ERR, "ib_post_send failed rc=%x reg_wr->key=%x\n",
2402 rc, reg_wr->key);
2403
2404 /* If all failed, attempt to recover this MR by setting it MR_ERROR*/
2405map_mr_error:
2406 ib_dma_unmap_sg(info->id->device, smbdirect_mr->sgl,
2407 smbdirect_mr->sgl_count, smbdirect_mr->dir);
2408
2409dma_map_error:
2410 smbdirect_mr->state = MR_ERROR;
2411 if (atomic_dec_and_test(&info->mr_used_count))
2412 wake_up(&info->wait_for_mr_cleanup);
2413
2414 smbd_disconnect_rdma_connection(info);
2415
2416 return NULL;
2417}
2418
2419static void local_inv_done(struct ib_cq *cq, struct ib_wc *wc)
2420{
2421 struct smbd_mr *smbdirect_mr;
2422 struct ib_cqe *cqe;
2423
2424 cqe = wc->wr_cqe;
2425 smbdirect_mr = container_of(cqe, struct smbd_mr, cqe);
2426 smbdirect_mr->state = MR_INVALIDATED;
2427 if (wc->status != IB_WC_SUCCESS) {
2428 log_rdma_mr(ERR, "invalidate failed status=%x\n", wc->status);
2429 smbdirect_mr->state = MR_ERROR;
2430 }
2431 complete(&smbdirect_mr->invalidate_done);
2432}
2433
2434/*
2435 * Deregister a MR after I/O is done
2436 * This function may wait if remote invalidation is not used
2437 * and we have to locally invalidate the buffer to prevent data is being
2438 * modified by remote peer after upper layer consumes it
2439 */
2440int smbd_deregister_mr(struct smbd_mr *smbdirect_mr)
2441{
2442 struct ib_send_wr *wr;
2443 struct smbd_connection *info = smbdirect_mr->conn;
2444 int rc = 0;
2445
2446 if (smbdirect_mr->need_invalidate) {
2447 /* Need to finish local invalidation before returning */
2448 wr = &smbdirect_mr->inv_wr;
2449 wr->opcode = IB_WR_LOCAL_INV;
2450 smbdirect_mr->cqe.done = local_inv_done;
2451 wr->wr_cqe = &smbdirect_mr->cqe;
2452 wr->num_sge = 0;
2453 wr->ex.invalidate_rkey = smbdirect_mr->mr->rkey;
2454 wr->send_flags = IB_SEND_SIGNALED;
2455
2456 init_completion(&smbdirect_mr->invalidate_done);
2457 rc = ib_post_send(info->id->qp, wr, NULL);
2458 if (rc) {
2459 log_rdma_mr(ERR, "ib_post_send failed rc=%x\n", rc);
2460 smbd_disconnect_rdma_connection(info);
2461 goto done;
2462 }
2463 wait_for_completion(&smbdirect_mr->invalidate_done);
2464 smbdirect_mr->need_invalidate = false;
2465 } else
2466 /*
2467 * For remote invalidation, just set it to MR_INVALIDATED
2468 * and defer to mr_recovery_work to recover the MR for next use
2469 */
2470 smbdirect_mr->state = MR_INVALIDATED;
2471
2472 if (smbdirect_mr->state == MR_INVALIDATED) {
2473 ib_dma_unmap_sg(
2474 info->id->device, smbdirect_mr->sgl,
2475 smbdirect_mr->sgl_count,
2476 smbdirect_mr->dir);
2477 smbdirect_mr->state = MR_READY;
2478 if (atomic_inc_return(&info->mr_ready_count) == 1)
2479 wake_up_interruptible(&info->wait_mr);
2480 } else
2481 /*
2482 * Schedule the work to do MR recovery for future I/Os MR
2483 * recovery is slow and don't want it to block current I/O
2484 */
2485 queue_work(info->workqueue, &info->mr_recovery_work);
2486
2487done:
2488 if (atomic_dec_and_test(&info->mr_used_count))
2489 wake_up(&info->wait_for_mr_cleanup);
2490
2491 return rc;
2492}
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (C) 2017, Microsoft Corporation.
4 *
5 * Author(s): Long Li <longli@microsoft.com>
6 */
7#include <linux/module.h>
8#include <linux/highmem.h>
9#include "smbdirect.h"
10#include "cifs_debug.h"
11#include "cifsproto.h"
12#include "smb2proto.h"
13
14static struct smbd_response *get_empty_queue_buffer(
15 struct smbd_connection *info);
16static struct smbd_response *get_receive_buffer(
17 struct smbd_connection *info);
18static void put_receive_buffer(
19 struct smbd_connection *info,
20 struct smbd_response *response);
21static int allocate_receive_buffers(struct smbd_connection *info, int num_buf);
22static void destroy_receive_buffers(struct smbd_connection *info);
23
24static void put_empty_packet(
25 struct smbd_connection *info, struct smbd_response *response);
26static void enqueue_reassembly(
27 struct smbd_connection *info,
28 struct smbd_response *response, int data_length);
29static struct smbd_response *_get_first_reassembly(
30 struct smbd_connection *info);
31
32static int smbd_post_recv(
33 struct smbd_connection *info,
34 struct smbd_response *response);
35
36static int smbd_post_send_empty(struct smbd_connection *info);
37static int smbd_post_send_data(
38 struct smbd_connection *info,
39 struct kvec *iov, int n_vec, int remaining_data_length);
40static int smbd_post_send_page(struct smbd_connection *info,
41 struct page *page, unsigned long offset,
42 size_t size, int remaining_data_length);
43
44static void destroy_mr_list(struct smbd_connection *info);
45static int allocate_mr_list(struct smbd_connection *info);
46
47/* SMBD version number */
48#define SMBD_V1 0x0100
49
50/* Port numbers for SMBD transport */
51#define SMB_PORT 445
52#define SMBD_PORT 5445
53
54/* Address lookup and resolve timeout in ms */
55#define RDMA_RESOLVE_TIMEOUT 5000
56
57/* SMBD negotiation timeout in seconds */
58#define SMBD_NEGOTIATE_TIMEOUT 120
59
60/* SMBD minimum receive size and fragmented sized defined in [MS-SMBD] */
61#define SMBD_MIN_RECEIVE_SIZE 128
62#define SMBD_MIN_FRAGMENTED_SIZE 131072
63
64/*
65 * Default maximum number of RDMA read/write outstanding on this connection
66 * This value is possibly decreased during QP creation on hardware limit
67 */
68#define SMBD_CM_RESPONDER_RESOURCES 32
69
70/* Maximum number of retries on data transfer operations */
71#define SMBD_CM_RETRY 6
72/* No need to retry on Receiver Not Ready since SMBD manages credits */
73#define SMBD_CM_RNR_RETRY 0
74
75/*
76 * User configurable initial values per SMBD transport connection
77 * as defined in [MS-SMBD] 3.1.1.1
78 * Those may change after a SMBD negotiation
79 */
80/* The local peer's maximum number of credits to grant to the peer */
81int smbd_receive_credit_max = 255;
82
83/* The remote peer's credit request of local peer */
84int smbd_send_credit_target = 255;
85
86/* The maximum single message size can be sent to remote peer */
87int smbd_max_send_size = 1364;
88
89/* The maximum fragmented upper-layer payload receive size supported */
90int smbd_max_fragmented_recv_size = 1024 * 1024;
91
92/* The maximum single-message size which can be received */
93int smbd_max_receive_size = 8192;
94
95/* The timeout to initiate send of a keepalive message on idle */
96int smbd_keep_alive_interval = 120;
97
98/*
99 * User configurable initial values for RDMA transport
100 * The actual values used may be lower and are limited to hardware capabilities
101 */
102/* Default maximum number of SGEs in a RDMA write/read */
103int smbd_max_frmr_depth = 2048;
104
105/* If payload is less than this byte, use RDMA send/recv not read/write */
106int rdma_readwrite_threshold = 4096;
107
108/* Transport logging functions
109 * Logging are defined as classes. They can be OR'ed to define the actual
110 * logging level via module parameter smbd_logging_class
111 * e.g. cifs.smbd_logging_class=0xa0 will log all log_rdma_recv() and
112 * log_rdma_event()
113 */
114#define LOG_OUTGOING 0x1
115#define LOG_INCOMING 0x2
116#define LOG_READ 0x4
117#define LOG_WRITE 0x8
118#define LOG_RDMA_SEND 0x10
119#define LOG_RDMA_RECV 0x20
120#define LOG_KEEP_ALIVE 0x40
121#define LOG_RDMA_EVENT 0x80
122#define LOG_RDMA_MR 0x100
123static unsigned int smbd_logging_class;
124module_param(smbd_logging_class, uint, 0644);
125MODULE_PARM_DESC(smbd_logging_class,
126 "Logging class for SMBD transport 0x0 to 0x100");
127
128#define ERR 0x0
129#define INFO 0x1
130static unsigned int smbd_logging_level = ERR;
131module_param(smbd_logging_level, uint, 0644);
132MODULE_PARM_DESC(smbd_logging_level,
133 "Logging level for SMBD transport, 0 (default): error, 1: info");
134
135#define log_rdma(level, class, fmt, args...) \
136do { \
137 if (level <= smbd_logging_level || class & smbd_logging_class) \
138 cifs_dbg(VFS, "%s:%d " fmt, __func__, __LINE__, ##args);\
139} while (0)
140
141#define log_outgoing(level, fmt, args...) \
142 log_rdma(level, LOG_OUTGOING, fmt, ##args)
143#define log_incoming(level, fmt, args...) \
144 log_rdma(level, LOG_INCOMING, fmt, ##args)
145#define log_read(level, fmt, args...) log_rdma(level, LOG_READ, fmt, ##args)
146#define log_write(level, fmt, args...) log_rdma(level, LOG_WRITE, fmt, ##args)
147#define log_rdma_send(level, fmt, args...) \
148 log_rdma(level, LOG_RDMA_SEND, fmt, ##args)
149#define log_rdma_recv(level, fmt, args...) \
150 log_rdma(level, LOG_RDMA_RECV, fmt, ##args)
151#define log_keep_alive(level, fmt, args...) \
152 log_rdma(level, LOG_KEEP_ALIVE, fmt, ##args)
153#define log_rdma_event(level, fmt, args...) \
154 log_rdma(level, LOG_RDMA_EVENT, fmt, ##args)
155#define log_rdma_mr(level, fmt, args...) \
156 log_rdma(level, LOG_RDMA_MR, fmt, ##args)
157
158static void smbd_disconnect_rdma_work(struct work_struct *work)
159{
160 struct smbd_connection *info =
161 container_of(work, struct smbd_connection, disconnect_work);
162
163 if (info->transport_status == SMBD_CONNECTED) {
164 info->transport_status = SMBD_DISCONNECTING;
165 rdma_disconnect(info->id);
166 }
167}
168
169static void smbd_disconnect_rdma_connection(struct smbd_connection *info)
170{
171 queue_work(info->workqueue, &info->disconnect_work);
172}
173
174/* Upcall from RDMA CM */
175static int smbd_conn_upcall(
176 struct rdma_cm_id *id, struct rdma_cm_event *event)
177{
178 struct smbd_connection *info = id->context;
179
180 log_rdma_event(INFO, "event=%d status=%d\n",
181 event->event, event->status);
182
183 switch (event->event) {
184 case RDMA_CM_EVENT_ADDR_RESOLVED:
185 case RDMA_CM_EVENT_ROUTE_RESOLVED:
186 info->ri_rc = 0;
187 complete(&info->ri_done);
188 break;
189
190 case RDMA_CM_EVENT_ADDR_ERROR:
191 info->ri_rc = -EHOSTUNREACH;
192 complete(&info->ri_done);
193 break;
194
195 case RDMA_CM_EVENT_ROUTE_ERROR:
196 info->ri_rc = -ENETUNREACH;
197 complete(&info->ri_done);
198 break;
199
200 case RDMA_CM_EVENT_ESTABLISHED:
201 log_rdma_event(INFO, "connected event=%d\n", event->event);
202 info->transport_status = SMBD_CONNECTED;
203 wake_up_interruptible(&info->conn_wait);
204 break;
205
206 case RDMA_CM_EVENT_CONNECT_ERROR:
207 case RDMA_CM_EVENT_UNREACHABLE:
208 case RDMA_CM_EVENT_REJECTED:
209 log_rdma_event(INFO, "connecting failed event=%d\n", event->event);
210 info->transport_status = SMBD_DISCONNECTED;
211 wake_up_interruptible(&info->conn_wait);
212 break;
213
214 case RDMA_CM_EVENT_DEVICE_REMOVAL:
215 case RDMA_CM_EVENT_DISCONNECTED:
216 /* This happenes when we fail the negotiation */
217 if (info->transport_status == SMBD_NEGOTIATE_FAILED) {
218 info->transport_status = SMBD_DISCONNECTED;
219 wake_up(&info->conn_wait);
220 break;
221 }
222
223 info->transport_status = SMBD_DISCONNECTED;
224 wake_up_interruptible(&info->disconn_wait);
225 wake_up_interruptible(&info->wait_reassembly_queue);
226 wake_up_interruptible_all(&info->wait_send_queue);
227 break;
228
229 default:
230 break;
231 }
232
233 return 0;
234}
235
236/* Upcall from RDMA QP */
237static void
238smbd_qp_async_error_upcall(struct ib_event *event, void *context)
239{
240 struct smbd_connection *info = context;
241
242 log_rdma_event(ERR, "%s on device %s info %p\n",
243 ib_event_msg(event->event), event->device->name, info);
244
245 switch (event->event) {
246 case IB_EVENT_CQ_ERR:
247 case IB_EVENT_QP_FATAL:
248 smbd_disconnect_rdma_connection(info);
249
250 default:
251 break;
252 }
253}
254
255static inline void *smbd_request_payload(struct smbd_request *request)
256{
257 return (void *)request->packet;
258}
259
260static inline void *smbd_response_payload(struct smbd_response *response)
261{
262 return (void *)response->packet;
263}
264
265/* Called when a RDMA send is done */
266static void send_done(struct ib_cq *cq, struct ib_wc *wc)
267{
268 int i;
269 struct smbd_request *request =
270 container_of(wc->wr_cqe, struct smbd_request, cqe);
271
272 log_rdma_send(INFO, "smbd_request %p completed wc->status=%d\n",
273 request, wc->status);
274
275 if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_SEND) {
276 log_rdma_send(ERR, "wc->status=%d wc->opcode=%d\n",
277 wc->status, wc->opcode);
278 smbd_disconnect_rdma_connection(request->info);
279 }
280
281 for (i = 0; i < request->num_sge; i++)
282 ib_dma_unmap_single(request->info->id->device,
283 request->sge[i].addr,
284 request->sge[i].length,
285 DMA_TO_DEVICE);
286
287 if (request->has_payload) {
288 if (atomic_dec_and_test(&request->info->send_payload_pending))
289 wake_up(&request->info->wait_send_payload_pending);
290 } else {
291 if (atomic_dec_and_test(&request->info->send_pending))
292 wake_up(&request->info->wait_send_pending);
293 }
294
295 mempool_free(request, request->info->request_mempool);
296}
297
298static void dump_smbd_negotiate_resp(struct smbd_negotiate_resp *resp)
299{
300 log_rdma_event(INFO, "resp message min_version %u max_version %u "
301 "negotiated_version %u credits_requested %u "
302 "credits_granted %u status %u max_readwrite_size %u "
303 "preferred_send_size %u max_receive_size %u "
304 "max_fragmented_size %u\n",
305 resp->min_version, resp->max_version, resp->negotiated_version,
306 resp->credits_requested, resp->credits_granted, resp->status,
307 resp->max_readwrite_size, resp->preferred_send_size,
308 resp->max_receive_size, resp->max_fragmented_size);
309}
310
311/*
312 * Process a negotiation response message, according to [MS-SMBD]3.1.5.7
313 * response, packet_length: the negotiation response message
314 * return value: true if negotiation is a success, false if failed
315 */
316static bool process_negotiation_response(
317 struct smbd_response *response, int packet_length)
318{
319 struct smbd_connection *info = response->info;
320 struct smbd_negotiate_resp *packet = smbd_response_payload(response);
321
322 if (packet_length < sizeof(struct smbd_negotiate_resp)) {
323 log_rdma_event(ERR,
324 "error: packet_length=%d\n", packet_length);
325 return false;
326 }
327
328 if (le16_to_cpu(packet->negotiated_version) != SMBD_V1) {
329 log_rdma_event(ERR, "error: negotiated_version=%x\n",
330 le16_to_cpu(packet->negotiated_version));
331 return false;
332 }
333 info->protocol = le16_to_cpu(packet->negotiated_version);
334
335 if (packet->credits_requested == 0) {
336 log_rdma_event(ERR, "error: credits_requested==0\n");
337 return false;
338 }
339 info->receive_credit_target = le16_to_cpu(packet->credits_requested);
340
341 if (packet->credits_granted == 0) {
342 log_rdma_event(ERR, "error: credits_granted==0\n");
343 return false;
344 }
345 atomic_set(&info->send_credits, le16_to_cpu(packet->credits_granted));
346
347 atomic_set(&info->receive_credits, 0);
348
349 if (le32_to_cpu(packet->preferred_send_size) > info->max_receive_size) {
350 log_rdma_event(ERR, "error: preferred_send_size=%d\n",
351 le32_to_cpu(packet->preferred_send_size));
352 return false;
353 }
354 info->max_receive_size = le32_to_cpu(packet->preferred_send_size);
355
356 if (le32_to_cpu(packet->max_receive_size) < SMBD_MIN_RECEIVE_SIZE) {
357 log_rdma_event(ERR, "error: max_receive_size=%d\n",
358 le32_to_cpu(packet->max_receive_size));
359 return false;
360 }
361 info->max_send_size = min_t(int, info->max_send_size,
362 le32_to_cpu(packet->max_receive_size));
363
364 if (le32_to_cpu(packet->max_fragmented_size) <
365 SMBD_MIN_FRAGMENTED_SIZE) {
366 log_rdma_event(ERR, "error: max_fragmented_size=%d\n",
367 le32_to_cpu(packet->max_fragmented_size));
368 return false;
369 }
370 info->max_fragmented_send_size =
371 le32_to_cpu(packet->max_fragmented_size);
372 info->rdma_readwrite_threshold =
373 rdma_readwrite_threshold > info->max_fragmented_send_size ?
374 info->max_fragmented_send_size :
375 rdma_readwrite_threshold;
376
377
378 info->max_readwrite_size = min_t(u32,
379 le32_to_cpu(packet->max_readwrite_size),
380 info->max_frmr_depth * PAGE_SIZE);
381 info->max_frmr_depth = info->max_readwrite_size / PAGE_SIZE;
382
383 return true;
384}
385
386/*
387 * Check and schedule to send an immediate packet
388 * This is used to extend credtis to remote peer to keep the transport busy
389 */
390static void check_and_send_immediate(struct smbd_connection *info)
391{
392 if (info->transport_status != SMBD_CONNECTED)
393 return;
394
395 info->send_immediate = true;
396
397 /*
398 * Promptly send a packet if our peer is running low on receive
399 * credits
400 */
401 if (atomic_read(&info->receive_credits) <
402 info->receive_credit_target - 1)
403 queue_delayed_work(
404 info->workqueue, &info->send_immediate_work, 0);
405}
406
407static void smbd_post_send_credits(struct work_struct *work)
408{
409 int ret = 0;
410 int use_receive_queue = 1;
411 int rc;
412 struct smbd_response *response;
413 struct smbd_connection *info =
414 container_of(work, struct smbd_connection,
415 post_send_credits_work);
416
417 if (info->transport_status != SMBD_CONNECTED) {
418 wake_up(&info->wait_receive_queues);
419 return;
420 }
421
422 if (info->receive_credit_target >
423 atomic_read(&info->receive_credits)) {
424 while (true) {
425 if (use_receive_queue)
426 response = get_receive_buffer(info);
427 else
428 response = get_empty_queue_buffer(info);
429 if (!response) {
430 /* now switch to emtpy packet queue */
431 if (use_receive_queue) {
432 use_receive_queue = 0;
433 continue;
434 } else
435 break;
436 }
437
438 response->type = SMBD_TRANSFER_DATA;
439 response->first_segment = false;
440 rc = smbd_post_recv(info, response);
441 if (rc) {
442 log_rdma_recv(ERR,
443 "post_recv failed rc=%d\n", rc);
444 put_receive_buffer(info, response);
445 break;
446 }
447
448 ret++;
449 }
450 }
451
452 spin_lock(&info->lock_new_credits_offered);
453 info->new_credits_offered += ret;
454 spin_unlock(&info->lock_new_credits_offered);
455
456 atomic_add(ret, &info->receive_credits);
457
458 /* Check if we can post new receive and grant credits to peer */
459 check_and_send_immediate(info);
460}
461
462static void smbd_recv_done_work(struct work_struct *work)
463{
464 struct smbd_connection *info =
465 container_of(work, struct smbd_connection, recv_done_work);
466
467 /*
468 * We may have new send credits granted from remote peer
469 * If any sender is blcoked on lack of credets, unblock it
470 */
471 if (atomic_read(&info->send_credits))
472 wake_up_interruptible(&info->wait_send_queue);
473
474 /*
475 * Check if we need to send something to remote peer to
476 * grant more credits or respond to KEEP_ALIVE packet
477 */
478 check_and_send_immediate(info);
479}
480
481/* Called from softirq, when recv is done */
482static void recv_done(struct ib_cq *cq, struct ib_wc *wc)
483{
484 struct smbd_data_transfer *data_transfer;
485 struct smbd_response *response =
486 container_of(wc->wr_cqe, struct smbd_response, cqe);
487 struct smbd_connection *info = response->info;
488 int data_length = 0;
489
490 log_rdma_recv(INFO, "response=%p type=%d wc status=%d wc opcode %d "
491 "byte_len=%d pkey_index=%x\n",
492 response, response->type, wc->status, wc->opcode,
493 wc->byte_len, wc->pkey_index);
494
495 if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_RECV) {
496 log_rdma_recv(INFO, "wc->status=%d opcode=%d\n",
497 wc->status, wc->opcode);
498 smbd_disconnect_rdma_connection(info);
499 goto error;
500 }
501
502 ib_dma_sync_single_for_cpu(
503 wc->qp->device,
504 response->sge.addr,
505 response->sge.length,
506 DMA_FROM_DEVICE);
507
508 switch (response->type) {
509 /* SMBD negotiation response */
510 case SMBD_NEGOTIATE_RESP:
511 dump_smbd_negotiate_resp(smbd_response_payload(response));
512 info->full_packet_received = true;
513 info->negotiate_done =
514 process_negotiation_response(response, wc->byte_len);
515 complete(&info->negotiate_completion);
516 break;
517
518 /* SMBD data transfer packet */
519 case SMBD_TRANSFER_DATA:
520 data_transfer = smbd_response_payload(response);
521 data_length = le32_to_cpu(data_transfer->data_length);
522
523 /*
524 * If this is a packet with data playload place the data in
525 * reassembly queue and wake up the reading thread
526 */
527 if (data_length) {
528 if (info->full_packet_received)
529 response->first_segment = true;
530
531 if (le32_to_cpu(data_transfer->remaining_data_length))
532 info->full_packet_received = false;
533 else
534 info->full_packet_received = true;
535
536 enqueue_reassembly(
537 info,
538 response,
539 data_length);
540 } else
541 put_empty_packet(info, response);
542
543 if (data_length)
544 wake_up_interruptible(&info->wait_reassembly_queue);
545
546 atomic_dec(&info->receive_credits);
547 info->receive_credit_target =
548 le16_to_cpu(data_transfer->credits_requested);
549 atomic_add(le16_to_cpu(data_transfer->credits_granted),
550 &info->send_credits);
551
552 log_incoming(INFO, "data flags %d data_offset %d "
553 "data_length %d remaining_data_length %d\n",
554 le16_to_cpu(data_transfer->flags),
555 le32_to_cpu(data_transfer->data_offset),
556 le32_to_cpu(data_transfer->data_length),
557 le32_to_cpu(data_transfer->remaining_data_length));
558
559 /* Send a KEEP_ALIVE response right away if requested */
560 info->keep_alive_requested = KEEP_ALIVE_NONE;
561 if (le16_to_cpu(data_transfer->flags) &
562 SMB_DIRECT_RESPONSE_REQUESTED) {
563 info->keep_alive_requested = KEEP_ALIVE_PENDING;
564 }
565
566 queue_work(info->workqueue, &info->recv_done_work);
567 return;
568
569 default:
570 log_rdma_recv(ERR,
571 "unexpected response type=%d\n", response->type);
572 }
573
574error:
575 put_receive_buffer(info, response);
576}
577
578static struct rdma_cm_id *smbd_create_id(
579 struct smbd_connection *info,
580 struct sockaddr *dstaddr, int port)
581{
582 struct rdma_cm_id *id;
583 int rc;
584 __be16 *sport;
585
586 id = rdma_create_id(&init_net, smbd_conn_upcall, info,
587 RDMA_PS_TCP, IB_QPT_RC);
588 if (IS_ERR(id)) {
589 rc = PTR_ERR(id);
590 log_rdma_event(ERR, "rdma_create_id() failed %i\n", rc);
591 return id;
592 }
593
594 if (dstaddr->sa_family == AF_INET6)
595 sport = &((struct sockaddr_in6 *)dstaddr)->sin6_port;
596 else
597 sport = &((struct sockaddr_in *)dstaddr)->sin_port;
598
599 *sport = htons(port);
600
601 init_completion(&info->ri_done);
602 info->ri_rc = -ETIMEDOUT;
603
604 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)dstaddr,
605 RDMA_RESOLVE_TIMEOUT);
606 if (rc) {
607 log_rdma_event(ERR, "rdma_resolve_addr() failed %i\n", rc);
608 goto out;
609 }
610 wait_for_completion_interruptible_timeout(
611 &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT));
612 rc = info->ri_rc;
613 if (rc) {
614 log_rdma_event(ERR, "rdma_resolve_addr() completed %i\n", rc);
615 goto out;
616 }
617
618 info->ri_rc = -ETIMEDOUT;
619 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
620 if (rc) {
621 log_rdma_event(ERR, "rdma_resolve_route() failed %i\n", rc);
622 goto out;
623 }
624 wait_for_completion_interruptible_timeout(
625 &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT));
626 rc = info->ri_rc;
627 if (rc) {
628 log_rdma_event(ERR, "rdma_resolve_route() completed %i\n", rc);
629 goto out;
630 }
631
632 return id;
633
634out:
635 rdma_destroy_id(id);
636 return ERR_PTR(rc);
637}
638
639/*
640 * Test if FRWR (Fast Registration Work Requests) is supported on the device
641 * This implementation requries FRWR on RDMA read/write
642 * return value: true if it is supported
643 */
644static bool frwr_is_supported(struct ib_device_attr *attrs)
645{
646 if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
647 return false;
648 if (attrs->max_fast_reg_page_list_len == 0)
649 return false;
650 return true;
651}
652
653static int smbd_ia_open(
654 struct smbd_connection *info,
655 struct sockaddr *dstaddr, int port)
656{
657 int rc;
658
659 info->id = smbd_create_id(info, dstaddr, port);
660 if (IS_ERR(info->id)) {
661 rc = PTR_ERR(info->id);
662 goto out1;
663 }
664
665 if (!frwr_is_supported(&info->id->device->attrs)) {
666 log_rdma_event(ERR,
667 "Fast Registration Work Requests "
668 "(FRWR) is not supported\n");
669 log_rdma_event(ERR,
670 "Device capability flags = %llx "
671 "max_fast_reg_page_list_len = %u\n",
672 info->id->device->attrs.device_cap_flags,
673 info->id->device->attrs.max_fast_reg_page_list_len);
674 rc = -EPROTONOSUPPORT;
675 goto out2;
676 }
677 info->max_frmr_depth = min_t(int,
678 smbd_max_frmr_depth,
679 info->id->device->attrs.max_fast_reg_page_list_len);
680 info->mr_type = IB_MR_TYPE_MEM_REG;
681 if (info->id->device->attrs.device_cap_flags & IB_DEVICE_SG_GAPS_REG)
682 info->mr_type = IB_MR_TYPE_SG_GAPS;
683
684 info->pd = ib_alloc_pd(info->id->device, 0);
685 if (IS_ERR(info->pd)) {
686 rc = PTR_ERR(info->pd);
687 log_rdma_event(ERR, "ib_alloc_pd() returned %d\n", rc);
688 goto out2;
689 }
690
691 return 0;
692
693out2:
694 rdma_destroy_id(info->id);
695 info->id = NULL;
696
697out1:
698 return rc;
699}
700
701/*
702 * Send a negotiation request message to the peer
703 * The negotiation procedure is in [MS-SMBD] 3.1.5.2 and 3.1.5.3
704 * After negotiation, the transport is connected and ready for
705 * carrying upper layer SMB payload
706 */
707static int smbd_post_send_negotiate_req(struct smbd_connection *info)
708{
709 struct ib_send_wr send_wr;
710 int rc = -ENOMEM;
711 struct smbd_request *request;
712 struct smbd_negotiate_req *packet;
713
714 request = mempool_alloc(info->request_mempool, GFP_KERNEL);
715 if (!request)
716 return rc;
717
718 request->info = info;
719
720 packet = smbd_request_payload(request);
721 packet->min_version = cpu_to_le16(SMBD_V1);
722 packet->max_version = cpu_to_le16(SMBD_V1);
723 packet->reserved = 0;
724 packet->credits_requested = cpu_to_le16(info->send_credit_target);
725 packet->preferred_send_size = cpu_to_le32(info->max_send_size);
726 packet->max_receive_size = cpu_to_le32(info->max_receive_size);
727 packet->max_fragmented_size =
728 cpu_to_le32(info->max_fragmented_recv_size);
729
730 request->num_sge = 1;
731 request->sge[0].addr = ib_dma_map_single(
732 info->id->device, (void *)packet,
733 sizeof(*packet), DMA_TO_DEVICE);
734 if (ib_dma_mapping_error(info->id->device, request->sge[0].addr)) {
735 rc = -EIO;
736 goto dma_mapping_failed;
737 }
738
739 request->sge[0].length = sizeof(*packet);
740 request->sge[0].lkey = info->pd->local_dma_lkey;
741
742 ib_dma_sync_single_for_device(
743 info->id->device, request->sge[0].addr,
744 request->sge[0].length, DMA_TO_DEVICE);
745
746 request->cqe.done = send_done;
747
748 send_wr.next = NULL;
749 send_wr.wr_cqe = &request->cqe;
750 send_wr.sg_list = request->sge;
751 send_wr.num_sge = request->num_sge;
752 send_wr.opcode = IB_WR_SEND;
753 send_wr.send_flags = IB_SEND_SIGNALED;
754
755 log_rdma_send(INFO, "sge addr=%llx length=%x lkey=%x\n",
756 request->sge[0].addr,
757 request->sge[0].length, request->sge[0].lkey);
758
759 request->has_payload = false;
760 atomic_inc(&info->send_pending);
761 rc = ib_post_send(info->id->qp, &send_wr, NULL);
762 if (!rc)
763 return 0;
764
765 /* if we reach here, post send failed */
766 log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc);
767 atomic_dec(&info->send_pending);
768 ib_dma_unmap_single(info->id->device, request->sge[0].addr,
769 request->sge[0].length, DMA_TO_DEVICE);
770
771 smbd_disconnect_rdma_connection(info);
772
773dma_mapping_failed:
774 mempool_free(request, info->request_mempool);
775 return rc;
776}
777
778/*
779 * Extend the credits to remote peer
780 * This implements [MS-SMBD] 3.1.5.9
781 * The idea is that we should extend credits to remote peer as quickly as
782 * it's allowed, to maintain data flow. We allocate as much receive
783 * buffer as possible, and extend the receive credits to remote peer
784 * return value: the new credtis being granted.
785 */
786static int manage_credits_prior_sending(struct smbd_connection *info)
787{
788 int new_credits;
789
790 spin_lock(&info->lock_new_credits_offered);
791 new_credits = info->new_credits_offered;
792 info->new_credits_offered = 0;
793 spin_unlock(&info->lock_new_credits_offered);
794
795 return new_credits;
796}
797
798/*
799 * Check if we need to send a KEEP_ALIVE message
800 * The idle connection timer triggers a KEEP_ALIVE message when expires
801 * SMB_DIRECT_RESPONSE_REQUESTED is set in the message flag to have peer send
802 * back a response.
803 * return value:
804 * 1 if SMB_DIRECT_RESPONSE_REQUESTED needs to be set
805 * 0: otherwise
806 */
807static int manage_keep_alive_before_sending(struct smbd_connection *info)
808{
809 if (info->keep_alive_requested == KEEP_ALIVE_PENDING) {
810 info->keep_alive_requested = KEEP_ALIVE_SENT;
811 return 1;
812 }
813 return 0;
814}
815
816/*
817 * Build and prepare the SMBD packet header
818 * This function waits for avaialbe send credits and build a SMBD packet
819 * header. The caller then optional append payload to the packet after
820 * the header
821 * intput values
822 * size: the size of the payload
823 * remaining_data_length: remaining data to send if this is part of a
824 * fragmented packet
825 * output values
826 * request_out: the request allocated from this function
827 * return values: 0 on success, otherwise actual error code returned
828 */
829static int smbd_create_header(struct smbd_connection *info,
830 int size, int remaining_data_length,
831 struct smbd_request **request_out)
832{
833 struct smbd_request *request;
834 struct smbd_data_transfer *packet;
835 int header_length;
836 int rc;
837
838 /* Wait for send credits. A SMBD packet needs one credit */
839 rc = wait_event_interruptible(info->wait_send_queue,
840 atomic_read(&info->send_credits) > 0 ||
841 info->transport_status != SMBD_CONNECTED);
842 if (rc)
843 return rc;
844
845 if (info->transport_status != SMBD_CONNECTED) {
846 log_outgoing(ERR, "disconnected not sending\n");
847 return -EAGAIN;
848 }
849 atomic_dec(&info->send_credits);
850
851 request = mempool_alloc(info->request_mempool, GFP_KERNEL);
852 if (!request) {
853 rc = -ENOMEM;
854 goto err;
855 }
856
857 request->info = info;
858
859 /* Fill in the packet header */
860 packet = smbd_request_payload(request);
861 packet->credits_requested = cpu_to_le16(info->send_credit_target);
862 packet->credits_granted =
863 cpu_to_le16(manage_credits_prior_sending(info));
864 info->send_immediate = false;
865
866 packet->flags = 0;
867 if (manage_keep_alive_before_sending(info))
868 packet->flags |= cpu_to_le16(SMB_DIRECT_RESPONSE_REQUESTED);
869
870 packet->reserved = 0;
871 if (!size)
872 packet->data_offset = 0;
873 else
874 packet->data_offset = cpu_to_le32(24);
875 packet->data_length = cpu_to_le32(size);
876 packet->remaining_data_length = cpu_to_le32(remaining_data_length);
877 packet->padding = 0;
878
879 log_outgoing(INFO, "credits_requested=%d credits_granted=%d "
880 "data_offset=%d data_length=%d remaining_data_length=%d\n",
881 le16_to_cpu(packet->credits_requested),
882 le16_to_cpu(packet->credits_granted),
883 le32_to_cpu(packet->data_offset),
884 le32_to_cpu(packet->data_length),
885 le32_to_cpu(packet->remaining_data_length));
886
887 /* Map the packet to DMA */
888 header_length = sizeof(struct smbd_data_transfer);
889 /* If this is a packet without payload, don't send padding */
890 if (!size)
891 header_length = offsetof(struct smbd_data_transfer, padding);
892
893 request->num_sge = 1;
894 request->sge[0].addr = ib_dma_map_single(info->id->device,
895 (void *)packet,
896 header_length,
897 DMA_TO_DEVICE);
898 if (ib_dma_mapping_error(info->id->device, request->sge[0].addr)) {
899 mempool_free(request, info->request_mempool);
900 rc = -EIO;
901 goto err;
902 }
903
904 request->sge[0].length = header_length;
905 request->sge[0].lkey = info->pd->local_dma_lkey;
906
907 *request_out = request;
908 return 0;
909
910err:
911 atomic_inc(&info->send_credits);
912 return rc;
913}
914
915static void smbd_destroy_header(struct smbd_connection *info,
916 struct smbd_request *request)
917{
918
919 ib_dma_unmap_single(info->id->device,
920 request->sge[0].addr,
921 request->sge[0].length,
922 DMA_TO_DEVICE);
923 mempool_free(request, info->request_mempool);
924 atomic_inc(&info->send_credits);
925}
926
927/* Post the send request */
928static int smbd_post_send(struct smbd_connection *info,
929 struct smbd_request *request, bool has_payload)
930{
931 struct ib_send_wr send_wr;
932 int rc, i;
933
934 for (i = 0; i < request->num_sge; i++) {
935 log_rdma_send(INFO,
936 "rdma_request sge[%d] addr=%llu length=%u\n",
937 i, request->sge[i].addr, request->sge[i].length);
938 ib_dma_sync_single_for_device(
939 info->id->device,
940 request->sge[i].addr,
941 request->sge[i].length,
942 DMA_TO_DEVICE);
943 }
944
945 request->cqe.done = send_done;
946
947 send_wr.next = NULL;
948 send_wr.wr_cqe = &request->cqe;
949 send_wr.sg_list = request->sge;
950 send_wr.num_sge = request->num_sge;
951 send_wr.opcode = IB_WR_SEND;
952 send_wr.send_flags = IB_SEND_SIGNALED;
953
954 if (has_payload) {
955 request->has_payload = true;
956 atomic_inc(&info->send_payload_pending);
957 } else {
958 request->has_payload = false;
959 atomic_inc(&info->send_pending);
960 }
961
962 rc = ib_post_send(info->id->qp, &send_wr, NULL);
963 if (rc) {
964 log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc);
965 if (has_payload) {
966 if (atomic_dec_and_test(&info->send_payload_pending))
967 wake_up(&info->wait_send_payload_pending);
968 } else {
969 if (atomic_dec_and_test(&info->send_pending))
970 wake_up(&info->wait_send_pending);
971 }
972 smbd_disconnect_rdma_connection(info);
973 rc = -EAGAIN;
974 } else
975 /* Reset timer for idle connection after packet is sent */
976 mod_delayed_work(info->workqueue, &info->idle_timer_work,
977 info->keep_alive_interval*HZ);
978
979 return rc;
980}
981
982static int smbd_post_send_sgl(struct smbd_connection *info,
983 struct scatterlist *sgl, int data_length, int remaining_data_length)
984{
985 int num_sgs;
986 int i, rc;
987 struct smbd_request *request;
988 struct scatterlist *sg;
989
990 rc = smbd_create_header(
991 info, data_length, remaining_data_length, &request);
992 if (rc)
993 return rc;
994
995 num_sgs = sgl ? sg_nents(sgl) : 0;
996 for_each_sg(sgl, sg, num_sgs, i) {
997 request->sge[i+1].addr =
998 ib_dma_map_page(info->id->device, sg_page(sg),
999 sg->offset, sg->length, DMA_TO_DEVICE);
1000 if (ib_dma_mapping_error(
1001 info->id->device, request->sge[i+1].addr)) {
1002 rc = -EIO;
1003 request->sge[i+1].addr = 0;
1004 goto dma_mapping_failure;
1005 }
1006 request->sge[i+1].length = sg->length;
1007 request->sge[i+1].lkey = info->pd->local_dma_lkey;
1008 request->num_sge++;
1009 }
1010
1011 rc = smbd_post_send(info, request, data_length);
1012 if (!rc)
1013 return 0;
1014
1015dma_mapping_failure:
1016 for (i = 1; i < request->num_sge; i++)
1017 if (request->sge[i].addr)
1018 ib_dma_unmap_single(info->id->device,
1019 request->sge[i].addr,
1020 request->sge[i].length,
1021 DMA_TO_DEVICE);
1022 smbd_destroy_header(info, request);
1023 return rc;
1024}
1025
1026/*
1027 * Send a page
1028 * page: the page to send
1029 * offset: offset in the page to send
1030 * size: length in the page to send
1031 * remaining_data_length: remaining data to send in this payload
1032 */
1033static int smbd_post_send_page(struct smbd_connection *info, struct page *page,
1034 unsigned long offset, size_t size, int remaining_data_length)
1035{
1036 struct scatterlist sgl;
1037
1038 sg_init_table(&sgl, 1);
1039 sg_set_page(&sgl, page, size, offset);
1040
1041 return smbd_post_send_sgl(info, &sgl, size, remaining_data_length);
1042}
1043
1044/*
1045 * Send an empty message
1046 * Empty message is used to extend credits to peer to for keep live
1047 * while there is no upper layer payload to send at the time
1048 */
1049static int smbd_post_send_empty(struct smbd_connection *info)
1050{
1051 info->count_send_empty++;
1052 return smbd_post_send_sgl(info, NULL, 0, 0);
1053}
1054
1055/*
1056 * Send a data buffer
1057 * iov: the iov array describing the data buffers
1058 * n_vec: number of iov array
1059 * remaining_data_length: remaining data to send following this packet
1060 * in segmented SMBD packet
1061 */
1062static int smbd_post_send_data(
1063 struct smbd_connection *info, struct kvec *iov, int n_vec,
1064 int remaining_data_length)
1065{
1066 int i;
1067 u32 data_length = 0;
1068 struct scatterlist sgl[SMBDIRECT_MAX_SGE];
1069
1070 if (n_vec > SMBDIRECT_MAX_SGE) {
1071 cifs_dbg(VFS, "Can't fit data to SGL, n_vec=%d\n", n_vec);
1072 return -ENOMEM;
1073 }
1074
1075 sg_init_table(sgl, n_vec);
1076 for (i = 0; i < n_vec; i++) {
1077 data_length += iov[i].iov_len;
1078 sg_set_buf(&sgl[i], iov[i].iov_base, iov[i].iov_len);
1079 }
1080
1081 return smbd_post_send_sgl(info, sgl, data_length, remaining_data_length);
1082}
1083
1084/*
1085 * Post a receive request to the transport
1086 * The remote peer can only send data when a receive request is posted
1087 * The interaction is controlled by send/receive credit system
1088 */
1089static int smbd_post_recv(
1090 struct smbd_connection *info, struct smbd_response *response)
1091{
1092 struct ib_recv_wr recv_wr;
1093 int rc = -EIO;
1094
1095 response->sge.addr = ib_dma_map_single(
1096 info->id->device, response->packet,
1097 info->max_receive_size, DMA_FROM_DEVICE);
1098 if (ib_dma_mapping_error(info->id->device, response->sge.addr))
1099 return rc;
1100
1101 response->sge.length = info->max_receive_size;
1102 response->sge.lkey = info->pd->local_dma_lkey;
1103
1104 response->cqe.done = recv_done;
1105
1106 recv_wr.wr_cqe = &response->cqe;
1107 recv_wr.next = NULL;
1108 recv_wr.sg_list = &response->sge;
1109 recv_wr.num_sge = 1;
1110
1111 rc = ib_post_recv(info->id->qp, &recv_wr, NULL);
1112 if (rc) {
1113 ib_dma_unmap_single(info->id->device, response->sge.addr,
1114 response->sge.length, DMA_FROM_DEVICE);
1115 smbd_disconnect_rdma_connection(info);
1116 log_rdma_recv(ERR, "ib_post_recv failed rc=%d\n", rc);
1117 }
1118
1119 return rc;
1120}
1121
1122/* Perform SMBD negotiate according to [MS-SMBD] 3.1.5.2 */
1123static int smbd_negotiate(struct smbd_connection *info)
1124{
1125 int rc;
1126 struct smbd_response *response = get_receive_buffer(info);
1127
1128 response->type = SMBD_NEGOTIATE_RESP;
1129 rc = smbd_post_recv(info, response);
1130 log_rdma_event(INFO,
1131 "smbd_post_recv rc=%d iov.addr=%llx iov.length=%x "
1132 "iov.lkey=%x\n",
1133 rc, response->sge.addr,
1134 response->sge.length, response->sge.lkey);
1135 if (rc)
1136 return rc;
1137
1138 init_completion(&info->negotiate_completion);
1139 info->negotiate_done = false;
1140 rc = smbd_post_send_negotiate_req(info);
1141 if (rc)
1142 return rc;
1143
1144 rc = wait_for_completion_interruptible_timeout(
1145 &info->negotiate_completion, SMBD_NEGOTIATE_TIMEOUT * HZ);
1146 log_rdma_event(INFO, "wait_for_completion_timeout rc=%d\n", rc);
1147
1148 if (info->negotiate_done)
1149 return 0;
1150
1151 if (rc == 0)
1152 rc = -ETIMEDOUT;
1153 else if (rc == -ERESTARTSYS)
1154 rc = -EINTR;
1155 else
1156 rc = -ENOTCONN;
1157
1158 return rc;
1159}
1160
1161static void put_empty_packet(
1162 struct smbd_connection *info, struct smbd_response *response)
1163{
1164 spin_lock(&info->empty_packet_queue_lock);
1165 list_add_tail(&response->list, &info->empty_packet_queue);
1166 info->count_empty_packet_queue++;
1167 spin_unlock(&info->empty_packet_queue_lock);
1168
1169 queue_work(info->workqueue, &info->post_send_credits_work);
1170}
1171
1172/*
1173 * Implement Connection.FragmentReassemblyBuffer defined in [MS-SMBD] 3.1.1.1
1174 * This is a queue for reassembling upper layer payload and present to upper
1175 * layer. All the inncoming payload go to the reassembly queue, regardless of
1176 * if reassembly is required. The uuper layer code reads from the queue for all
1177 * incoming payloads.
1178 * Put a received packet to the reassembly queue
1179 * response: the packet received
1180 * data_length: the size of payload in this packet
1181 */
1182static void enqueue_reassembly(
1183 struct smbd_connection *info,
1184 struct smbd_response *response,
1185 int data_length)
1186{
1187 spin_lock(&info->reassembly_queue_lock);
1188 list_add_tail(&response->list, &info->reassembly_queue);
1189 info->reassembly_queue_length++;
1190 /*
1191 * Make sure reassembly_data_length is updated after list and
1192 * reassembly_queue_length are updated. On the dequeue side
1193 * reassembly_data_length is checked without a lock to determine
1194 * if reassembly_queue_length and list is up to date
1195 */
1196 virt_wmb();
1197 info->reassembly_data_length += data_length;
1198 spin_unlock(&info->reassembly_queue_lock);
1199 info->count_reassembly_queue++;
1200 info->count_enqueue_reassembly_queue++;
1201}
1202
1203/*
1204 * Get the first entry at the front of reassembly queue
1205 * Caller is responsible for locking
1206 * return value: the first entry if any, NULL if queue is empty
1207 */
1208static struct smbd_response *_get_first_reassembly(struct smbd_connection *info)
1209{
1210 struct smbd_response *ret = NULL;
1211
1212 if (!list_empty(&info->reassembly_queue)) {
1213 ret = list_first_entry(
1214 &info->reassembly_queue,
1215 struct smbd_response, list);
1216 }
1217 return ret;
1218}
1219
1220static struct smbd_response *get_empty_queue_buffer(
1221 struct smbd_connection *info)
1222{
1223 struct smbd_response *ret = NULL;
1224 unsigned long flags;
1225
1226 spin_lock_irqsave(&info->empty_packet_queue_lock, flags);
1227 if (!list_empty(&info->empty_packet_queue)) {
1228 ret = list_first_entry(
1229 &info->empty_packet_queue,
1230 struct smbd_response, list);
1231 list_del(&ret->list);
1232 info->count_empty_packet_queue--;
1233 }
1234 spin_unlock_irqrestore(&info->empty_packet_queue_lock, flags);
1235
1236 return ret;
1237}
1238
1239/*
1240 * Get a receive buffer
1241 * For each remote send, we need to post a receive. The receive buffers are
1242 * pre-allocated in advance.
1243 * return value: the receive buffer, NULL if none is available
1244 */
1245static struct smbd_response *get_receive_buffer(struct smbd_connection *info)
1246{
1247 struct smbd_response *ret = NULL;
1248 unsigned long flags;
1249
1250 spin_lock_irqsave(&info->receive_queue_lock, flags);
1251 if (!list_empty(&info->receive_queue)) {
1252 ret = list_first_entry(
1253 &info->receive_queue,
1254 struct smbd_response, list);
1255 list_del(&ret->list);
1256 info->count_receive_queue--;
1257 info->count_get_receive_buffer++;
1258 }
1259 spin_unlock_irqrestore(&info->receive_queue_lock, flags);
1260
1261 return ret;
1262}
1263
1264/*
1265 * Return a receive buffer
1266 * Upon returning of a receive buffer, we can post new receive and extend
1267 * more receive credits to remote peer. This is done immediately after a
1268 * receive buffer is returned.
1269 */
1270static void put_receive_buffer(
1271 struct smbd_connection *info, struct smbd_response *response)
1272{
1273 unsigned long flags;
1274
1275 ib_dma_unmap_single(info->id->device, response->sge.addr,
1276 response->sge.length, DMA_FROM_DEVICE);
1277
1278 spin_lock_irqsave(&info->receive_queue_lock, flags);
1279 list_add_tail(&response->list, &info->receive_queue);
1280 info->count_receive_queue++;
1281 info->count_put_receive_buffer++;
1282 spin_unlock_irqrestore(&info->receive_queue_lock, flags);
1283
1284 queue_work(info->workqueue, &info->post_send_credits_work);
1285}
1286
1287/* Preallocate all receive buffer on transport establishment */
1288static int allocate_receive_buffers(struct smbd_connection *info, int num_buf)
1289{
1290 int i;
1291 struct smbd_response *response;
1292
1293 INIT_LIST_HEAD(&info->reassembly_queue);
1294 spin_lock_init(&info->reassembly_queue_lock);
1295 info->reassembly_data_length = 0;
1296 info->reassembly_queue_length = 0;
1297
1298 INIT_LIST_HEAD(&info->receive_queue);
1299 spin_lock_init(&info->receive_queue_lock);
1300 info->count_receive_queue = 0;
1301
1302 INIT_LIST_HEAD(&info->empty_packet_queue);
1303 spin_lock_init(&info->empty_packet_queue_lock);
1304 info->count_empty_packet_queue = 0;
1305
1306 init_waitqueue_head(&info->wait_receive_queues);
1307
1308 for (i = 0; i < num_buf; i++) {
1309 response = mempool_alloc(info->response_mempool, GFP_KERNEL);
1310 if (!response)
1311 goto allocate_failed;
1312
1313 response->info = info;
1314 list_add_tail(&response->list, &info->receive_queue);
1315 info->count_receive_queue++;
1316 }
1317
1318 return 0;
1319
1320allocate_failed:
1321 while (!list_empty(&info->receive_queue)) {
1322 response = list_first_entry(
1323 &info->receive_queue,
1324 struct smbd_response, list);
1325 list_del(&response->list);
1326 info->count_receive_queue--;
1327
1328 mempool_free(response, info->response_mempool);
1329 }
1330 return -ENOMEM;
1331}
1332
1333static void destroy_receive_buffers(struct smbd_connection *info)
1334{
1335 struct smbd_response *response;
1336
1337 while ((response = get_receive_buffer(info)))
1338 mempool_free(response, info->response_mempool);
1339
1340 while ((response = get_empty_queue_buffer(info)))
1341 mempool_free(response, info->response_mempool);
1342}
1343
1344/*
1345 * Check and send an immediate or keep alive packet
1346 * The condition to send those packets are defined in [MS-SMBD] 3.1.1.1
1347 * Connection.KeepaliveRequested and Connection.SendImmediate
1348 * The idea is to extend credits to server as soon as it becomes available
1349 */
1350static void send_immediate_work(struct work_struct *work)
1351{
1352 struct smbd_connection *info = container_of(
1353 work, struct smbd_connection,
1354 send_immediate_work.work);
1355
1356 if (info->keep_alive_requested == KEEP_ALIVE_PENDING ||
1357 info->send_immediate) {
1358 log_keep_alive(INFO, "send an empty message\n");
1359 smbd_post_send_empty(info);
1360 }
1361}
1362
1363/* Implement idle connection timer [MS-SMBD] 3.1.6.2 */
1364static void idle_connection_timer(struct work_struct *work)
1365{
1366 struct smbd_connection *info = container_of(
1367 work, struct smbd_connection,
1368 idle_timer_work.work);
1369
1370 if (info->keep_alive_requested != KEEP_ALIVE_NONE) {
1371 log_keep_alive(ERR,
1372 "error status info->keep_alive_requested=%d\n",
1373 info->keep_alive_requested);
1374 smbd_disconnect_rdma_connection(info);
1375 return;
1376 }
1377
1378 log_keep_alive(INFO, "about to send an empty idle message\n");
1379 smbd_post_send_empty(info);
1380
1381 /* Setup the next idle timeout work */
1382 queue_delayed_work(info->workqueue, &info->idle_timer_work,
1383 info->keep_alive_interval*HZ);
1384}
1385
1386/*
1387 * Destroy the transport and related RDMA and memory resources
1388 * Need to go through all the pending counters and make sure on one is using
1389 * the transport while it is destroyed
1390 */
1391void smbd_destroy(struct TCP_Server_Info *server)
1392{
1393 struct smbd_connection *info = server->smbd_conn;
1394 struct smbd_response *response;
1395 unsigned long flags;
1396
1397 if (!info) {
1398 log_rdma_event(INFO, "rdma session already destroyed\n");
1399 return;
1400 }
1401
1402 log_rdma_event(INFO, "destroying rdma session\n");
1403 if (info->transport_status != SMBD_DISCONNECTED) {
1404 rdma_disconnect(server->smbd_conn->id);
1405 log_rdma_event(INFO, "wait for transport being disconnected\n");
1406 wait_event_interruptible(
1407 info->disconn_wait,
1408 info->transport_status == SMBD_DISCONNECTED);
1409 }
1410
1411 log_rdma_event(INFO, "destroying qp\n");
1412 ib_drain_qp(info->id->qp);
1413 rdma_destroy_qp(info->id);
1414
1415 log_rdma_event(INFO, "cancelling idle timer\n");
1416 cancel_delayed_work_sync(&info->idle_timer_work);
1417 log_rdma_event(INFO, "cancelling send immediate work\n");
1418 cancel_delayed_work_sync(&info->send_immediate_work);
1419
1420 log_rdma_event(INFO, "wait for all send posted to IB to finish\n");
1421 wait_event(info->wait_send_pending,
1422 atomic_read(&info->send_pending) == 0);
1423 wait_event(info->wait_send_payload_pending,
1424 atomic_read(&info->send_payload_pending) == 0);
1425
1426 /* It's not posssible for upper layer to get to reassembly */
1427 log_rdma_event(INFO, "drain the reassembly queue\n");
1428 do {
1429 spin_lock_irqsave(&info->reassembly_queue_lock, flags);
1430 response = _get_first_reassembly(info);
1431 if (response) {
1432 list_del(&response->list);
1433 spin_unlock_irqrestore(
1434 &info->reassembly_queue_lock, flags);
1435 put_receive_buffer(info, response);
1436 } else
1437 spin_unlock_irqrestore(
1438 &info->reassembly_queue_lock, flags);
1439 } while (response);
1440 info->reassembly_data_length = 0;
1441
1442 log_rdma_event(INFO, "free receive buffers\n");
1443 wait_event(info->wait_receive_queues,
1444 info->count_receive_queue + info->count_empty_packet_queue
1445 == info->receive_credit_max);
1446 destroy_receive_buffers(info);
1447
1448 /*
1449 * For performance reasons, memory registration and deregistration
1450 * are not locked by srv_mutex. It is possible some processes are
1451 * blocked on transport srv_mutex while holding memory registration.
1452 * Release the transport srv_mutex to allow them to hit the failure
1453 * path when sending data, and then release memory registartions.
1454 */
1455 log_rdma_event(INFO, "freeing mr list\n");
1456 wake_up_interruptible_all(&info->wait_mr);
1457 while (atomic_read(&info->mr_used_count)) {
1458 mutex_unlock(&server->srv_mutex);
1459 msleep(1000);
1460 mutex_lock(&server->srv_mutex);
1461 }
1462 destroy_mr_list(info);
1463
1464 ib_free_cq(info->send_cq);
1465 ib_free_cq(info->recv_cq);
1466 ib_dealloc_pd(info->pd);
1467 rdma_destroy_id(info->id);
1468
1469 /* free mempools */
1470 mempool_destroy(info->request_mempool);
1471 kmem_cache_destroy(info->request_cache);
1472
1473 mempool_destroy(info->response_mempool);
1474 kmem_cache_destroy(info->response_cache);
1475
1476 info->transport_status = SMBD_DESTROYED;
1477
1478 destroy_workqueue(info->workqueue);
1479 kfree(info);
1480}
1481
1482/*
1483 * Reconnect this SMBD connection, called from upper layer
1484 * return value: 0 on success, or actual error code
1485 */
1486int smbd_reconnect(struct TCP_Server_Info *server)
1487{
1488 log_rdma_event(INFO, "reconnecting rdma session\n");
1489
1490 if (!server->smbd_conn) {
1491 log_rdma_event(INFO, "rdma session already destroyed\n");
1492 goto create_conn;
1493 }
1494
1495 /*
1496 * This is possible if transport is disconnected and we haven't received
1497 * notification from RDMA, but upper layer has detected timeout
1498 */
1499 if (server->smbd_conn->transport_status == SMBD_CONNECTED) {
1500 log_rdma_event(INFO, "disconnecting transport\n");
1501 smbd_destroy(server);
1502 }
1503
1504create_conn:
1505 log_rdma_event(INFO, "creating rdma session\n");
1506 server->smbd_conn = smbd_get_connection(
1507 server, (struct sockaddr *) &server->dstaddr);
1508 log_rdma_event(INFO, "created rdma session info=%p\n",
1509 server->smbd_conn);
1510
1511 return server->smbd_conn ? 0 : -ENOENT;
1512}
1513
1514static void destroy_caches_and_workqueue(struct smbd_connection *info)
1515{
1516 destroy_receive_buffers(info);
1517 destroy_workqueue(info->workqueue);
1518 mempool_destroy(info->response_mempool);
1519 kmem_cache_destroy(info->response_cache);
1520 mempool_destroy(info->request_mempool);
1521 kmem_cache_destroy(info->request_cache);
1522}
1523
1524#define MAX_NAME_LEN 80
1525static int allocate_caches_and_workqueue(struct smbd_connection *info)
1526{
1527 char name[MAX_NAME_LEN];
1528 int rc;
1529
1530 scnprintf(name, MAX_NAME_LEN, "smbd_request_%p", info);
1531 info->request_cache =
1532 kmem_cache_create(
1533 name,
1534 sizeof(struct smbd_request) +
1535 sizeof(struct smbd_data_transfer),
1536 0, SLAB_HWCACHE_ALIGN, NULL);
1537 if (!info->request_cache)
1538 return -ENOMEM;
1539
1540 info->request_mempool =
1541 mempool_create(info->send_credit_target, mempool_alloc_slab,
1542 mempool_free_slab, info->request_cache);
1543 if (!info->request_mempool)
1544 goto out1;
1545
1546 scnprintf(name, MAX_NAME_LEN, "smbd_response_%p", info);
1547 info->response_cache =
1548 kmem_cache_create(
1549 name,
1550 sizeof(struct smbd_response) +
1551 info->max_receive_size,
1552 0, SLAB_HWCACHE_ALIGN, NULL);
1553 if (!info->response_cache)
1554 goto out2;
1555
1556 info->response_mempool =
1557 mempool_create(info->receive_credit_max, mempool_alloc_slab,
1558 mempool_free_slab, info->response_cache);
1559 if (!info->response_mempool)
1560 goto out3;
1561
1562 scnprintf(name, MAX_NAME_LEN, "smbd_%p", info);
1563 info->workqueue = create_workqueue(name);
1564 if (!info->workqueue)
1565 goto out4;
1566
1567 rc = allocate_receive_buffers(info, info->receive_credit_max);
1568 if (rc) {
1569 log_rdma_event(ERR, "failed to allocate receive buffers\n");
1570 goto out5;
1571 }
1572
1573 return 0;
1574
1575out5:
1576 destroy_workqueue(info->workqueue);
1577out4:
1578 mempool_destroy(info->response_mempool);
1579out3:
1580 kmem_cache_destroy(info->response_cache);
1581out2:
1582 mempool_destroy(info->request_mempool);
1583out1:
1584 kmem_cache_destroy(info->request_cache);
1585 return -ENOMEM;
1586}
1587
1588/* Create a SMBD connection, called by upper layer */
1589static struct smbd_connection *_smbd_get_connection(
1590 struct TCP_Server_Info *server, struct sockaddr *dstaddr, int port)
1591{
1592 int rc;
1593 struct smbd_connection *info;
1594 struct rdma_conn_param conn_param;
1595 struct ib_qp_init_attr qp_attr;
1596 struct sockaddr_in *addr_in = (struct sockaddr_in *) dstaddr;
1597 struct ib_port_immutable port_immutable;
1598 u32 ird_ord_hdr[2];
1599
1600 info = kzalloc(sizeof(struct smbd_connection), GFP_KERNEL);
1601 if (!info)
1602 return NULL;
1603
1604 info->transport_status = SMBD_CONNECTING;
1605 rc = smbd_ia_open(info, dstaddr, port);
1606 if (rc) {
1607 log_rdma_event(INFO, "smbd_ia_open rc=%d\n", rc);
1608 goto create_id_failed;
1609 }
1610
1611 if (smbd_send_credit_target > info->id->device->attrs.max_cqe ||
1612 smbd_send_credit_target > info->id->device->attrs.max_qp_wr) {
1613 log_rdma_event(ERR,
1614 "consider lowering send_credit_target = %d. "
1615 "Possible CQE overrun, device "
1616 "reporting max_cpe %d max_qp_wr %d\n",
1617 smbd_send_credit_target,
1618 info->id->device->attrs.max_cqe,
1619 info->id->device->attrs.max_qp_wr);
1620 goto config_failed;
1621 }
1622
1623 if (smbd_receive_credit_max > info->id->device->attrs.max_cqe ||
1624 smbd_receive_credit_max > info->id->device->attrs.max_qp_wr) {
1625 log_rdma_event(ERR,
1626 "consider lowering receive_credit_max = %d. "
1627 "Possible CQE overrun, device "
1628 "reporting max_cpe %d max_qp_wr %d\n",
1629 smbd_receive_credit_max,
1630 info->id->device->attrs.max_cqe,
1631 info->id->device->attrs.max_qp_wr);
1632 goto config_failed;
1633 }
1634
1635 info->receive_credit_max = smbd_receive_credit_max;
1636 info->send_credit_target = smbd_send_credit_target;
1637 info->max_send_size = smbd_max_send_size;
1638 info->max_fragmented_recv_size = smbd_max_fragmented_recv_size;
1639 info->max_receive_size = smbd_max_receive_size;
1640 info->keep_alive_interval = smbd_keep_alive_interval;
1641
1642 if (info->id->device->attrs.max_send_sge < SMBDIRECT_MAX_SGE) {
1643 log_rdma_event(ERR,
1644 "warning: device max_send_sge = %d too small\n",
1645 info->id->device->attrs.max_send_sge);
1646 log_rdma_event(ERR, "Queue Pair creation may fail\n");
1647 }
1648 if (info->id->device->attrs.max_recv_sge < SMBDIRECT_MAX_SGE) {
1649 log_rdma_event(ERR,
1650 "warning: device max_recv_sge = %d too small\n",
1651 info->id->device->attrs.max_recv_sge);
1652 log_rdma_event(ERR, "Queue Pair creation may fail\n");
1653 }
1654
1655 info->send_cq = NULL;
1656 info->recv_cq = NULL;
1657 info->send_cq =
1658 ib_alloc_cq_any(info->id->device, info,
1659 info->send_credit_target, IB_POLL_SOFTIRQ);
1660 if (IS_ERR(info->send_cq)) {
1661 info->send_cq = NULL;
1662 goto alloc_cq_failed;
1663 }
1664
1665 info->recv_cq =
1666 ib_alloc_cq_any(info->id->device, info,
1667 info->receive_credit_max, IB_POLL_SOFTIRQ);
1668 if (IS_ERR(info->recv_cq)) {
1669 info->recv_cq = NULL;
1670 goto alloc_cq_failed;
1671 }
1672
1673 memset(&qp_attr, 0, sizeof(qp_attr));
1674 qp_attr.event_handler = smbd_qp_async_error_upcall;
1675 qp_attr.qp_context = info;
1676 qp_attr.cap.max_send_wr = info->send_credit_target;
1677 qp_attr.cap.max_recv_wr = info->receive_credit_max;
1678 qp_attr.cap.max_send_sge = SMBDIRECT_MAX_SGE;
1679 qp_attr.cap.max_recv_sge = SMBDIRECT_MAX_SGE;
1680 qp_attr.cap.max_inline_data = 0;
1681 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1682 qp_attr.qp_type = IB_QPT_RC;
1683 qp_attr.send_cq = info->send_cq;
1684 qp_attr.recv_cq = info->recv_cq;
1685 qp_attr.port_num = ~0;
1686
1687 rc = rdma_create_qp(info->id, info->pd, &qp_attr);
1688 if (rc) {
1689 log_rdma_event(ERR, "rdma_create_qp failed %i\n", rc);
1690 goto create_qp_failed;
1691 }
1692
1693 memset(&conn_param, 0, sizeof(conn_param));
1694 conn_param.initiator_depth = 0;
1695
1696 conn_param.responder_resources =
1697 info->id->device->attrs.max_qp_rd_atom
1698 < SMBD_CM_RESPONDER_RESOURCES ?
1699 info->id->device->attrs.max_qp_rd_atom :
1700 SMBD_CM_RESPONDER_RESOURCES;
1701 info->responder_resources = conn_param.responder_resources;
1702 log_rdma_mr(INFO, "responder_resources=%d\n",
1703 info->responder_resources);
1704
1705 /* Need to send IRD/ORD in private data for iWARP */
1706 info->id->device->ops.get_port_immutable(
1707 info->id->device, info->id->port_num, &port_immutable);
1708 if (port_immutable.core_cap_flags & RDMA_CORE_PORT_IWARP) {
1709 ird_ord_hdr[0] = info->responder_resources;
1710 ird_ord_hdr[1] = 1;
1711 conn_param.private_data = ird_ord_hdr;
1712 conn_param.private_data_len = sizeof(ird_ord_hdr);
1713 } else {
1714 conn_param.private_data = NULL;
1715 conn_param.private_data_len = 0;
1716 }
1717
1718 conn_param.retry_count = SMBD_CM_RETRY;
1719 conn_param.rnr_retry_count = SMBD_CM_RNR_RETRY;
1720 conn_param.flow_control = 0;
1721
1722 log_rdma_event(INFO, "connecting to IP %pI4 port %d\n",
1723 &addr_in->sin_addr, port);
1724
1725 init_waitqueue_head(&info->conn_wait);
1726 init_waitqueue_head(&info->disconn_wait);
1727 init_waitqueue_head(&info->wait_reassembly_queue);
1728 rc = rdma_connect(info->id, &conn_param);
1729 if (rc) {
1730 log_rdma_event(ERR, "rdma_connect() failed with %i\n", rc);
1731 goto rdma_connect_failed;
1732 }
1733
1734 wait_event_interruptible(
1735 info->conn_wait, info->transport_status != SMBD_CONNECTING);
1736
1737 if (info->transport_status != SMBD_CONNECTED) {
1738 log_rdma_event(ERR, "rdma_connect failed port=%d\n", port);
1739 goto rdma_connect_failed;
1740 }
1741
1742 log_rdma_event(INFO, "rdma_connect connected\n");
1743
1744 rc = allocate_caches_and_workqueue(info);
1745 if (rc) {
1746 log_rdma_event(ERR, "cache allocation failed\n");
1747 goto allocate_cache_failed;
1748 }
1749
1750 init_waitqueue_head(&info->wait_send_queue);
1751 INIT_DELAYED_WORK(&info->idle_timer_work, idle_connection_timer);
1752 INIT_DELAYED_WORK(&info->send_immediate_work, send_immediate_work);
1753 queue_delayed_work(info->workqueue, &info->idle_timer_work,
1754 info->keep_alive_interval*HZ);
1755
1756 init_waitqueue_head(&info->wait_send_pending);
1757 atomic_set(&info->send_pending, 0);
1758
1759 init_waitqueue_head(&info->wait_send_payload_pending);
1760 atomic_set(&info->send_payload_pending, 0);
1761
1762 INIT_WORK(&info->disconnect_work, smbd_disconnect_rdma_work);
1763 INIT_WORK(&info->recv_done_work, smbd_recv_done_work);
1764 INIT_WORK(&info->post_send_credits_work, smbd_post_send_credits);
1765 info->new_credits_offered = 0;
1766 spin_lock_init(&info->lock_new_credits_offered);
1767
1768 rc = smbd_negotiate(info);
1769 if (rc) {
1770 log_rdma_event(ERR, "smbd_negotiate rc=%d\n", rc);
1771 goto negotiation_failed;
1772 }
1773
1774 rc = allocate_mr_list(info);
1775 if (rc) {
1776 log_rdma_mr(ERR, "memory registration allocation failed\n");
1777 goto allocate_mr_failed;
1778 }
1779
1780 return info;
1781
1782allocate_mr_failed:
1783 /* At this point, need to a full transport shutdown */
1784 smbd_destroy(server);
1785 return NULL;
1786
1787negotiation_failed:
1788 cancel_delayed_work_sync(&info->idle_timer_work);
1789 destroy_caches_and_workqueue(info);
1790 info->transport_status = SMBD_NEGOTIATE_FAILED;
1791 init_waitqueue_head(&info->conn_wait);
1792 rdma_disconnect(info->id);
1793 wait_event(info->conn_wait,
1794 info->transport_status == SMBD_DISCONNECTED);
1795
1796allocate_cache_failed:
1797rdma_connect_failed:
1798 rdma_destroy_qp(info->id);
1799
1800create_qp_failed:
1801alloc_cq_failed:
1802 if (info->send_cq)
1803 ib_free_cq(info->send_cq);
1804 if (info->recv_cq)
1805 ib_free_cq(info->recv_cq);
1806
1807config_failed:
1808 ib_dealloc_pd(info->pd);
1809 rdma_destroy_id(info->id);
1810
1811create_id_failed:
1812 kfree(info);
1813 return NULL;
1814}
1815
1816struct smbd_connection *smbd_get_connection(
1817 struct TCP_Server_Info *server, struct sockaddr *dstaddr)
1818{
1819 struct smbd_connection *ret;
1820 int port = SMBD_PORT;
1821
1822try_again:
1823 ret = _smbd_get_connection(server, dstaddr, port);
1824
1825 /* Try SMB_PORT if SMBD_PORT doesn't work */
1826 if (!ret && port == SMBD_PORT) {
1827 port = SMB_PORT;
1828 goto try_again;
1829 }
1830 return ret;
1831}
1832
1833/*
1834 * Receive data from receive reassembly queue
1835 * All the incoming data packets are placed in reassembly queue
1836 * buf: the buffer to read data into
1837 * size: the length of data to read
1838 * return value: actual data read
1839 * Note: this implementation copies the data from reassebmly queue to receive
1840 * buffers used by upper layer. This is not the optimal code path. A better way
1841 * to do it is to not have upper layer allocate its receive buffers but rather
1842 * borrow the buffer from reassembly queue, and return it after data is
1843 * consumed. But this will require more changes to upper layer code, and also
1844 * need to consider packet boundaries while they still being reassembled.
1845 */
1846static int smbd_recv_buf(struct smbd_connection *info, char *buf,
1847 unsigned int size)
1848{
1849 struct smbd_response *response;
1850 struct smbd_data_transfer *data_transfer;
1851 int to_copy, to_read, data_read, offset;
1852 u32 data_length, remaining_data_length, data_offset;
1853 int rc;
1854
1855again:
1856 /*
1857 * No need to hold the reassembly queue lock all the time as we are
1858 * the only one reading from the front of the queue. The transport
1859 * may add more entries to the back of the queue at the same time
1860 */
1861 log_read(INFO, "size=%d info->reassembly_data_length=%d\n", size,
1862 info->reassembly_data_length);
1863 if (info->reassembly_data_length >= size) {
1864 int queue_length;
1865 int queue_removed = 0;
1866
1867 /*
1868 * Need to make sure reassembly_data_length is read before
1869 * reading reassembly_queue_length and calling
1870 * _get_first_reassembly. This call is lock free
1871 * as we never read at the end of the queue which are being
1872 * updated in SOFTIRQ as more data is received
1873 */
1874 virt_rmb();
1875 queue_length = info->reassembly_queue_length;
1876 data_read = 0;
1877 to_read = size;
1878 offset = info->first_entry_offset;
1879 while (data_read < size) {
1880 response = _get_first_reassembly(info);
1881 data_transfer = smbd_response_payload(response);
1882 data_length = le32_to_cpu(data_transfer->data_length);
1883 remaining_data_length =
1884 le32_to_cpu(
1885 data_transfer->remaining_data_length);
1886 data_offset = le32_to_cpu(data_transfer->data_offset);
1887
1888 /*
1889 * The upper layer expects RFC1002 length at the
1890 * beginning of the payload. Return it to indicate
1891 * the total length of the packet. This minimize the
1892 * change to upper layer packet processing logic. This
1893 * will be eventually remove when an intermediate
1894 * transport layer is added
1895 */
1896 if (response->first_segment && size == 4) {
1897 unsigned int rfc1002_len =
1898 data_length + remaining_data_length;
1899 *((__be32 *)buf) = cpu_to_be32(rfc1002_len);
1900 data_read = 4;
1901 response->first_segment = false;
1902 log_read(INFO, "returning rfc1002 length %d\n",
1903 rfc1002_len);
1904 goto read_rfc1002_done;
1905 }
1906
1907 to_copy = min_t(int, data_length - offset, to_read);
1908 memcpy(
1909 buf + data_read,
1910 (char *)data_transfer + data_offset + offset,
1911 to_copy);
1912
1913 /* move on to the next buffer? */
1914 if (to_copy == data_length - offset) {
1915 queue_length--;
1916 /*
1917 * No need to lock if we are not at the
1918 * end of the queue
1919 */
1920 if (queue_length)
1921 list_del(&response->list);
1922 else {
1923 spin_lock_irq(
1924 &info->reassembly_queue_lock);
1925 list_del(&response->list);
1926 spin_unlock_irq(
1927 &info->reassembly_queue_lock);
1928 }
1929 queue_removed++;
1930 info->count_reassembly_queue--;
1931 info->count_dequeue_reassembly_queue++;
1932 put_receive_buffer(info, response);
1933 offset = 0;
1934 log_read(INFO, "put_receive_buffer offset=0\n");
1935 } else
1936 offset += to_copy;
1937
1938 to_read -= to_copy;
1939 data_read += to_copy;
1940
1941 log_read(INFO, "_get_first_reassembly memcpy %d bytes "
1942 "data_transfer_length-offset=%d after that "
1943 "to_read=%d data_read=%d offset=%d\n",
1944 to_copy, data_length - offset,
1945 to_read, data_read, offset);
1946 }
1947
1948 spin_lock_irq(&info->reassembly_queue_lock);
1949 info->reassembly_data_length -= data_read;
1950 info->reassembly_queue_length -= queue_removed;
1951 spin_unlock_irq(&info->reassembly_queue_lock);
1952
1953 info->first_entry_offset = offset;
1954 log_read(INFO, "returning to thread data_read=%d "
1955 "reassembly_data_length=%d first_entry_offset=%d\n",
1956 data_read, info->reassembly_data_length,
1957 info->first_entry_offset);
1958read_rfc1002_done:
1959 return data_read;
1960 }
1961
1962 log_read(INFO, "wait_event on more data\n");
1963 rc = wait_event_interruptible(
1964 info->wait_reassembly_queue,
1965 info->reassembly_data_length >= size ||
1966 info->transport_status != SMBD_CONNECTED);
1967 /* Don't return any data if interrupted */
1968 if (rc)
1969 return rc;
1970
1971 if (info->transport_status != SMBD_CONNECTED) {
1972 log_read(ERR, "disconnected\n");
1973 return 0;
1974 }
1975
1976 goto again;
1977}
1978
1979/*
1980 * Receive a page from receive reassembly queue
1981 * page: the page to read data into
1982 * to_read: the length of data to read
1983 * return value: actual data read
1984 */
1985static int smbd_recv_page(struct smbd_connection *info,
1986 struct page *page, unsigned int page_offset,
1987 unsigned int to_read)
1988{
1989 int ret;
1990 char *to_address;
1991 void *page_address;
1992
1993 /* make sure we have the page ready for read */
1994 ret = wait_event_interruptible(
1995 info->wait_reassembly_queue,
1996 info->reassembly_data_length >= to_read ||
1997 info->transport_status != SMBD_CONNECTED);
1998 if (ret)
1999 return ret;
2000
2001 /* now we can read from reassembly queue and not sleep */
2002 page_address = kmap_atomic(page);
2003 to_address = (char *) page_address + page_offset;
2004
2005 log_read(INFO, "reading from page=%p address=%p to_read=%d\n",
2006 page, to_address, to_read);
2007
2008 ret = smbd_recv_buf(info, to_address, to_read);
2009 kunmap_atomic(page_address);
2010
2011 return ret;
2012}
2013
2014/*
2015 * Receive data from transport
2016 * msg: a msghdr point to the buffer, can be ITER_KVEC or ITER_BVEC
2017 * return: total bytes read, or 0. SMB Direct will not do partial read.
2018 */
2019int smbd_recv(struct smbd_connection *info, struct msghdr *msg)
2020{
2021 char *buf;
2022 struct page *page;
2023 unsigned int to_read, page_offset;
2024 int rc;
2025
2026 if (iov_iter_rw(&msg->msg_iter) == WRITE) {
2027 /* It's a bug in upper layer to get there */
2028 cifs_dbg(VFS, "CIFS: invalid msg iter dir %u\n",
2029 iov_iter_rw(&msg->msg_iter));
2030 rc = -EINVAL;
2031 goto out;
2032 }
2033
2034 switch (iov_iter_type(&msg->msg_iter)) {
2035 case ITER_KVEC:
2036 buf = msg->msg_iter.kvec->iov_base;
2037 to_read = msg->msg_iter.kvec->iov_len;
2038 rc = smbd_recv_buf(info, buf, to_read);
2039 break;
2040
2041 case ITER_BVEC:
2042 page = msg->msg_iter.bvec->bv_page;
2043 page_offset = msg->msg_iter.bvec->bv_offset;
2044 to_read = msg->msg_iter.bvec->bv_len;
2045 rc = smbd_recv_page(info, page, page_offset, to_read);
2046 break;
2047
2048 default:
2049 /* It's a bug in upper layer to get there */
2050 cifs_dbg(VFS, "CIFS: invalid msg type %d\n",
2051 iov_iter_type(&msg->msg_iter));
2052 rc = -EINVAL;
2053 }
2054
2055out:
2056 /* SMBDirect will read it all or nothing */
2057 if (rc > 0)
2058 msg->msg_iter.count = 0;
2059 return rc;
2060}
2061
2062/*
2063 * Send data to transport
2064 * Each rqst is transported as a SMBDirect payload
2065 * rqst: the data to write
2066 * return value: 0 if successfully write, otherwise error code
2067 */
2068int smbd_send(struct TCP_Server_Info *server,
2069 int num_rqst, struct smb_rqst *rqst_array)
2070{
2071 struct smbd_connection *info = server->smbd_conn;
2072 struct kvec vec;
2073 int nvecs;
2074 int size;
2075 unsigned int buflen, remaining_data_length;
2076 int start, i, j;
2077 int max_iov_size =
2078 info->max_send_size - sizeof(struct smbd_data_transfer);
2079 struct kvec *iov;
2080 int rc;
2081 struct smb_rqst *rqst;
2082 int rqst_idx;
2083
2084 if (info->transport_status != SMBD_CONNECTED) {
2085 rc = -EAGAIN;
2086 goto done;
2087 }
2088
2089 /*
2090 * Add in the page array if there is one. The caller needs to set
2091 * rq_tailsz to PAGE_SIZE when the buffer has multiple pages and
2092 * ends at page boundary
2093 */
2094 remaining_data_length = 0;
2095 for (i = 0; i < num_rqst; i++)
2096 remaining_data_length += smb_rqst_len(server, &rqst_array[i]);
2097
2098 if (remaining_data_length + sizeof(struct smbd_data_transfer) >
2099 info->max_fragmented_send_size) {
2100 log_write(ERR, "payload size %d > max size %d\n",
2101 remaining_data_length, info->max_fragmented_send_size);
2102 rc = -EINVAL;
2103 goto done;
2104 }
2105
2106 log_write(INFO, "num_rqst=%d total length=%u\n",
2107 num_rqst, remaining_data_length);
2108
2109 rqst_idx = 0;
2110next_rqst:
2111 rqst = &rqst_array[rqst_idx];
2112 iov = rqst->rq_iov;
2113
2114 cifs_dbg(FYI, "Sending smb (RDMA): idx=%d smb_len=%lu\n",
2115 rqst_idx, smb_rqst_len(server, rqst));
2116 for (i = 0; i < rqst->rq_nvec; i++)
2117 dump_smb(iov[i].iov_base, iov[i].iov_len);
2118
2119
2120 log_write(INFO, "rqst_idx=%d nvec=%d rqst->rq_npages=%d rq_pagesz=%d "
2121 "rq_tailsz=%d buflen=%lu\n",
2122 rqst_idx, rqst->rq_nvec, rqst->rq_npages, rqst->rq_pagesz,
2123 rqst->rq_tailsz, smb_rqst_len(server, rqst));
2124
2125 start = i = 0;
2126 buflen = 0;
2127 while (true) {
2128 buflen += iov[i].iov_len;
2129 if (buflen > max_iov_size) {
2130 if (i > start) {
2131 remaining_data_length -=
2132 (buflen-iov[i].iov_len);
2133 log_write(INFO, "sending iov[] from start=%d "
2134 "i=%d nvecs=%d "
2135 "remaining_data_length=%d\n",
2136 start, i, i-start,
2137 remaining_data_length);
2138 rc = smbd_post_send_data(
2139 info, &iov[start], i-start,
2140 remaining_data_length);
2141 if (rc)
2142 goto done;
2143 } else {
2144 /* iov[start] is too big, break it */
2145 nvecs = (buflen+max_iov_size-1)/max_iov_size;
2146 log_write(INFO, "iov[%d] iov_base=%p buflen=%d"
2147 " break to %d vectors\n",
2148 start, iov[start].iov_base,
2149 buflen, nvecs);
2150 for (j = 0; j < nvecs; j++) {
2151 vec.iov_base =
2152 (char *)iov[start].iov_base +
2153 j*max_iov_size;
2154 vec.iov_len = max_iov_size;
2155 if (j == nvecs-1)
2156 vec.iov_len =
2157 buflen -
2158 max_iov_size*(nvecs-1);
2159 remaining_data_length -= vec.iov_len;
2160 log_write(INFO,
2161 "sending vec j=%d iov_base=%p"
2162 " iov_len=%zu "
2163 "remaining_data_length=%d\n",
2164 j, vec.iov_base, vec.iov_len,
2165 remaining_data_length);
2166 rc = smbd_post_send_data(
2167 info, &vec, 1,
2168 remaining_data_length);
2169 if (rc)
2170 goto done;
2171 }
2172 i++;
2173 if (i == rqst->rq_nvec)
2174 break;
2175 }
2176 start = i;
2177 buflen = 0;
2178 } else {
2179 i++;
2180 if (i == rqst->rq_nvec) {
2181 /* send out all remaining vecs */
2182 remaining_data_length -= buflen;
2183 log_write(INFO,
2184 "sending iov[] from start=%d i=%d "
2185 "nvecs=%d remaining_data_length=%d\n",
2186 start, i, i-start,
2187 remaining_data_length);
2188 rc = smbd_post_send_data(info, &iov[start],
2189 i-start, remaining_data_length);
2190 if (rc)
2191 goto done;
2192 break;
2193 }
2194 }
2195 log_write(INFO, "looping i=%d buflen=%d\n", i, buflen);
2196 }
2197
2198 /* now sending pages if there are any */
2199 for (i = 0; i < rqst->rq_npages; i++) {
2200 unsigned int offset;
2201
2202 rqst_page_get_length(rqst, i, &buflen, &offset);
2203 nvecs = (buflen + max_iov_size - 1) / max_iov_size;
2204 log_write(INFO, "sending pages buflen=%d nvecs=%d\n",
2205 buflen, nvecs);
2206 for (j = 0; j < nvecs; j++) {
2207 size = max_iov_size;
2208 if (j == nvecs-1)
2209 size = buflen - j*max_iov_size;
2210 remaining_data_length -= size;
2211 log_write(INFO, "sending pages i=%d offset=%d size=%d"
2212 " remaining_data_length=%d\n",
2213 i, j*max_iov_size+offset, size,
2214 remaining_data_length);
2215 rc = smbd_post_send_page(
2216 info, rqst->rq_pages[i],
2217 j*max_iov_size + offset,
2218 size, remaining_data_length);
2219 if (rc)
2220 goto done;
2221 }
2222 }
2223
2224 rqst_idx++;
2225 if (rqst_idx < num_rqst)
2226 goto next_rqst;
2227
2228done:
2229 /*
2230 * As an optimization, we don't wait for individual I/O to finish
2231 * before sending the next one.
2232 * Send them all and wait for pending send count to get to 0
2233 * that means all the I/Os have been out and we are good to return
2234 */
2235
2236 wait_event(info->wait_send_payload_pending,
2237 atomic_read(&info->send_payload_pending) == 0);
2238
2239 return rc;
2240}
2241
2242static void register_mr_done(struct ib_cq *cq, struct ib_wc *wc)
2243{
2244 struct smbd_mr *mr;
2245 struct ib_cqe *cqe;
2246
2247 if (wc->status) {
2248 log_rdma_mr(ERR, "status=%d\n", wc->status);
2249 cqe = wc->wr_cqe;
2250 mr = container_of(cqe, struct smbd_mr, cqe);
2251 smbd_disconnect_rdma_connection(mr->conn);
2252 }
2253}
2254
2255/*
2256 * The work queue function that recovers MRs
2257 * We need to call ib_dereg_mr() and ib_alloc_mr() before this MR can be used
2258 * again. Both calls are slow, so finish them in a workqueue. This will not
2259 * block I/O path.
2260 * There is one workqueue that recovers MRs, there is no need to lock as the
2261 * I/O requests calling smbd_register_mr will never update the links in the
2262 * mr_list.
2263 */
2264static void smbd_mr_recovery_work(struct work_struct *work)
2265{
2266 struct smbd_connection *info =
2267 container_of(work, struct smbd_connection, mr_recovery_work);
2268 struct smbd_mr *smbdirect_mr;
2269 int rc;
2270
2271 list_for_each_entry(smbdirect_mr, &info->mr_list, list) {
2272 if (smbdirect_mr->state == MR_INVALIDATED)
2273 ib_dma_unmap_sg(
2274 info->id->device, smbdirect_mr->sgl,
2275 smbdirect_mr->sgl_count,
2276 smbdirect_mr->dir);
2277 else if (smbdirect_mr->state == MR_ERROR) {
2278
2279 /* recover this MR entry */
2280 rc = ib_dereg_mr(smbdirect_mr->mr);
2281 if (rc) {
2282 log_rdma_mr(ERR,
2283 "ib_dereg_mr failed rc=%x\n",
2284 rc);
2285 smbd_disconnect_rdma_connection(info);
2286 continue;
2287 }
2288
2289 smbdirect_mr->mr = ib_alloc_mr(
2290 info->pd, info->mr_type,
2291 info->max_frmr_depth);
2292 if (IS_ERR(smbdirect_mr->mr)) {
2293 log_rdma_mr(ERR,
2294 "ib_alloc_mr failed mr_type=%x "
2295 "max_frmr_depth=%x\n",
2296 info->mr_type,
2297 info->max_frmr_depth);
2298 smbd_disconnect_rdma_connection(info);
2299 continue;
2300 }
2301 } else
2302 /* This MR is being used, don't recover it */
2303 continue;
2304
2305 smbdirect_mr->state = MR_READY;
2306
2307 /* smbdirect_mr->state is updated by this function
2308 * and is read and updated by I/O issuing CPUs trying
2309 * to get a MR, the call to atomic_inc_return
2310 * implicates a memory barrier and guarantees this
2311 * value is updated before waking up any calls to
2312 * get_mr() from the I/O issuing CPUs
2313 */
2314 if (atomic_inc_return(&info->mr_ready_count) == 1)
2315 wake_up_interruptible(&info->wait_mr);
2316 }
2317}
2318
2319static void destroy_mr_list(struct smbd_connection *info)
2320{
2321 struct smbd_mr *mr, *tmp;
2322
2323 cancel_work_sync(&info->mr_recovery_work);
2324 list_for_each_entry_safe(mr, tmp, &info->mr_list, list) {
2325 if (mr->state == MR_INVALIDATED)
2326 ib_dma_unmap_sg(info->id->device, mr->sgl,
2327 mr->sgl_count, mr->dir);
2328 ib_dereg_mr(mr->mr);
2329 kfree(mr->sgl);
2330 kfree(mr);
2331 }
2332}
2333
2334/*
2335 * Allocate MRs used for RDMA read/write
2336 * The number of MRs will not exceed hardware capability in responder_resources
2337 * All MRs are kept in mr_list. The MR can be recovered after it's used
2338 * Recovery is done in smbd_mr_recovery_work. The content of list entry changes
2339 * as MRs are used and recovered for I/O, but the list links will not change
2340 */
2341static int allocate_mr_list(struct smbd_connection *info)
2342{
2343 int i;
2344 struct smbd_mr *smbdirect_mr, *tmp;
2345
2346 INIT_LIST_HEAD(&info->mr_list);
2347 init_waitqueue_head(&info->wait_mr);
2348 spin_lock_init(&info->mr_list_lock);
2349 atomic_set(&info->mr_ready_count, 0);
2350 atomic_set(&info->mr_used_count, 0);
2351 init_waitqueue_head(&info->wait_for_mr_cleanup);
2352 /* Allocate more MRs (2x) than hardware responder_resources */
2353 for (i = 0; i < info->responder_resources * 2; i++) {
2354 smbdirect_mr = kzalloc(sizeof(*smbdirect_mr), GFP_KERNEL);
2355 if (!smbdirect_mr)
2356 goto out;
2357 smbdirect_mr->mr = ib_alloc_mr(info->pd, info->mr_type,
2358 info->max_frmr_depth);
2359 if (IS_ERR(smbdirect_mr->mr)) {
2360 log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x "
2361 "max_frmr_depth=%x\n",
2362 info->mr_type, info->max_frmr_depth);
2363 goto out;
2364 }
2365 smbdirect_mr->sgl = kcalloc(
2366 info->max_frmr_depth,
2367 sizeof(struct scatterlist),
2368 GFP_KERNEL);
2369 if (!smbdirect_mr->sgl) {
2370 log_rdma_mr(ERR, "failed to allocate sgl\n");
2371 ib_dereg_mr(smbdirect_mr->mr);
2372 goto out;
2373 }
2374 smbdirect_mr->state = MR_READY;
2375 smbdirect_mr->conn = info;
2376
2377 list_add_tail(&smbdirect_mr->list, &info->mr_list);
2378 atomic_inc(&info->mr_ready_count);
2379 }
2380 INIT_WORK(&info->mr_recovery_work, smbd_mr_recovery_work);
2381 return 0;
2382
2383out:
2384 kfree(smbdirect_mr);
2385
2386 list_for_each_entry_safe(smbdirect_mr, tmp, &info->mr_list, list) {
2387 ib_dereg_mr(smbdirect_mr->mr);
2388 kfree(smbdirect_mr->sgl);
2389 kfree(smbdirect_mr);
2390 }
2391 return -ENOMEM;
2392}
2393
2394/*
2395 * Get a MR from mr_list. This function waits until there is at least one
2396 * MR available in the list. It may access the list while the
2397 * smbd_mr_recovery_work is recovering the MR list. This doesn't need a lock
2398 * as they never modify the same places. However, there may be several CPUs
2399 * issueing I/O trying to get MR at the same time, mr_list_lock is used to
2400 * protect this situation.
2401 */
2402static struct smbd_mr *get_mr(struct smbd_connection *info)
2403{
2404 struct smbd_mr *ret;
2405 int rc;
2406again:
2407 rc = wait_event_interruptible(info->wait_mr,
2408 atomic_read(&info->mr_ready_count) ||
2409 info->transport_status != SMBD_CONNECTED);
2410 if (rc) {
2411 log_rdma_mr(ERR, "wait_event_interruptible rc=%x\n", rc);
2412 return NULL;
2413 }
2414
2415 if (info->transport_status != SMBD_CONNECTED) {
2416 log_rdma_mr(ERR, "info->transport_status=%x\n",
2417 info->transport_status);
2418 return NULL;
2419 }
2420
2421 spin_lock(&info->mr_list_lock);
2422 list_for_each_entry(ret, &info->mr_list, list) {
2423 if (ret->state == MR_READY) {
2424 ret->state = MR_REGISTERED;
2425 spin_unlock(&info->mr_list_lock);
2426 atomic_dec(&info->mr_ready_count);
2427 atomic_inc(&info->mr_used_count);
2428 return ret;
2429 }
2430 }
2431
2432 spin_unlock(&info->mr_list_lock);
2433 /*
2434 * It is possible that we could fail to get MR because other processes may
2435 * try to acquire a MR at the same time. If this is the case, retry it.
2436 */
2437 goto again;
2438}
2439
2440/*
2441 * Register memory for RDMA read/write
2442 * pages[]: the list of pages to register memory with
2443 * num_pages: the number of pages to register
2444 * tailsz: if non-zero, the bytes to register in the last page
2445 * writing: true if this is a RDMA write (SMB read), false for RDMA read
2446 * need_invalidate: true if this MR needs to be locally invalidated after I/O
2447 * return value: the MR registered, NULL if failed.
2448 */
2449struct smbd_mr *smbd_register_mr(
2450 struct smbd_connection *info, struct page *pages[], int num_pages,
2451 int offset, int tailsz, bool writing, bool need_invalidate)
2452{
2453 struct smbd_mr *smbdirect_mr;
2454 int rc, i;
2455 enum dma_data_direction dir;
2456 struct ib_reg_wr *reg_wr;
2457
2458 if (num_pages > info->max_frmr_depth) {
2459 log_rdma_mr(ERR, "num_pages=%d max_frmr_depth=%d\n",
2460 num_pages, info->max_frmr_depth);
2461 return NULL;
2462 }
2463
2464 smbdirect_mr = get_mr(info);
2465 if (!smbdirect_mr) {
2466 log_rdma_mr(ERR, "get_mr returning NULL\n");
2467 return NULL;
2468 }
2469 smbdirect_mr->need_invalidate = need_invalidate;
2470 smbdirect_mr->sgl_count = num_pages;
2471 sg_init_table(smbdirect_mr->sgl, num_pages);
2472
2473 log_rdma_mr(INFO, "num_pages=0x%x offset=0x%x tailsz=0x%x\n",
2474 num_pages, offset, tailsz);
2475
2476 if (num_pages == 1) {
2477 sg_set_page(&smbdirect_mr->sgl[0], pages[0], tailsz, offset);
2478 goto skip_multiple_pages;
2479 }
2480
2481 /* We have at least two pages to register */
2482 sg_set_page(
2483 &smbdirect_mr->sgl[0], pages[0], PAGE_SIZE - offset, offset);
2484 i = 1;
2485 while (i < num_pages - 1) {
2486 sg_set_page(&smbdirect_mr->sgl[i], pages[i], PAGE_SIZE, 0);
2487 i++;
2488 }
2489 sg_set_page(&smbdirect_mr->sgl[i], pages[i],
2490 tailsz ? tailsz : PAGE_SIZE, 0);
2491
2492skip_multiple_pages:
2493 dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
2494 smbdirect_mr->dir = dir;
2495 rc = ib_dma_map_sg(info->id->device, smbdirect_mr->sgl, num_pages, dir);
2496 if (!rc) {
2497 log_rdma_mr(ERR, "ib_dma_map_sg num_pages=%x dir=%x rc=%x\n",
2498 num_pages, dir, rc);
2499 goto dma_map_error;
2500 }
2501
2502 rc = ib_map_mr_sg(smbdirect_mr->mr, smbdirect_mr->sgl, num_pages,
2503 NULL, PAGE_SIZE);
2504 if (rc != num_pages) {
2505 log_rdma_mr(ERR,
2506 "ib_map_mr_sg failed rc = %d num_pages = %x\n",
2507 rc, num_pages);
2508 goto map_mr_error;
2509 }
2510
2511 ib_update_fast_reg_key(smbdirect_mr->mr,
2512 ib_inc_rkey(smbdirect_mr->mr->rkey));
2513 reg_wr = &smbdirect_mr->wr;
2514 reg_wr->wr.opcode = IB_WR_REG_MR;
2515 smbdirect_mr->cqe.done = register_mr_done;
2516 reg_wr->wr.wr_cqe = &smbdirect_mr->cqe;
2517 reg_wr->wr.num_sge = 0;
2518 reg_wr->wr.send_flags = IB_SEND_SIGNALED;
2519 reg_wr->mr = smbdirect_mr->mr;
2520 reg_wr->key = smbdirect_mr->mr->rkey;
2521 reg_wr->access = writing ?
2522 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2523 IB_ACCESS_REMOTE_READ;
2524
2525 /*
2526 * There is no need for waiting for complemtion on ib_post_send
2527 * on IB_WR_REG_MR. Hardware enforces a barrier and order of execution
2528 * on the next ib_post_send when we actaully send I/O to remote peer
2529 */
2530 rc = ib_post_send(info->id->qp, ®_wr->wr, NULL);
2531 if (!rc)
2532 return smbdirect_mr;
2533
2534 log_rdma_mr(ERR, "ib_post_send failed rc=%x reg_wr->key=%x\n",
2535 rc, reg_wr->key);
2536
2537 /* If all failed, attempt to recover this MR by setting it MR_ERROR*/
2538map_mr_error:
2539 ib_dma_unmap_sg(info->id->device, smbdirect_mr->sgl,
2540 smbdirect_mr->sgl_count, smbdirect_mr->dir);
2541
2542dma_map_error:
2543 smbdirect_mr->state = MR_ERROR;
2544 if (atomic_dec_and_test(&info->mr_used_count))
2545 wake_up(&info->wait_for_mr_cleanup);
2546
2547 smbd_disconnect_rdma_connection(info);
2548
2549 return NULL;
2550}
2551
2552static void local_inv_done(struct ib_cq *cq, struct ib_wc *wc)
2553{
2554 struct smbd_mr *smbdirect_mr;
2555 struct ib_cqe *cqe;
2556
2557 cqe = wc->wr_cqe;
2558 smbdirect_mr = container_of(cqe, struct smbd_mr, cqe);
2559 smbdirect_mr->state = MR_INVALIDATED;
2560 if (wc->status != IB_WC_SUCCESS) {
2561 log_rdma_mr(ERR, "invalidate failed status=%x\n", wc->status);
2562 smbdirect_mr->state = MR_ERROR;
2563 }
2564 complete(&smbdirect_mr->invalidate_done);
2565}
2566
2567/*
2568 * Deregister a MR after I/O is done
2569 * This function may wait if remote invalidation is not used
2570 * and we have to locally invalidate the buffer to prevent data is being
2571 * modified by remote peer after upper layer consumes it
2572 */
2573int smbd_deregister_mr(struct smbd_mr *smbdirect_mr)
2574{
2575 struct ib_send_wr *wr;
2576 struct smbd_connection *info = smbdirect_mr->conn;
2577 int rc = 0;
2578
2579 if (smbdirect_mr->need_invalidate) {
2580 /* Need to finish local invalidation before returning */
2581 wr = &smbdirect_mr->inv_wr;
2582 wr->opcode = IB_WR_LOCAL_INV;
2583 smbdirect_mr->cqe.done = local_inv_done;
2584 wr->wr_cqe = &smbdirect_mr->cqe;
2585 wr->num_sge = 0;
2586 wr->ex.invalidate_rkey = smbdirect_mr->mr->rkey;
2587 wr->send_flags = IB_SEND_SIGNALED;
2588
2589 init_completion(&smbdirect_mr->invalidate_done);
2590 rc = ib_post_send(info->id->qp, wr, NULL);
2591 if (rc) {
2592 log_rdma_mr(ERR, "ib_post_send failed rc=%x\n", rc);
2593 smbd_disconnect_rdma_connection(info);
2594 goto done;
2595 }
2596 wait_for_completion(&smbdirect_mr->invalidate_done);
2597 smbdirect_mr->need_invalidate = false;
2598 } else
2599 /*
2600 * For remote invalidation, just set it to MR_INVALIDATED
2601 * and defer to mr_recovery_work to recover the MR for next use
2602 */
2603 smbdirect_mr->state = MR_INVALIDATED;
2604
2605 /*
2606 * Schedule the work to do MR recovery for future I/Os
2607 * MR recovery is slow and we don't want it to block the current I/O
2608 */
2609 queue_work(info->workqueue, &info->mr_recovery_work);
2610
2611done:
2612 if (atomic_dec_and_test(&info->mr_used_count))
2613 wake_up(&info->wait_for_mr_cleanup);
2614
2615 return rc;
2616}