2 * Copyright (C) 2017, Microsoft Corporation.
4 * Author(s): Long Li <longli@microsoft.com>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See
14 * the GNU General Public License for more details.
16 #include <linux/module.h>
17 #include <linux/highmem.h>
18 #include "smbdirect.h"
19 #include "cifs_debug.h"
21 static struct smbd_response *get_empty_queue_buffer(
22 struct smbd_connection *info);
23 static struct smbd_response *get_receive_buffer(
24 struct smbd_connection *info);
25 static void put_receive_buffer(
26 struct smbd_connection *info,
27 struct smbd_response *response);
28 static int allocate_receive_buffers(struct smbd_connection *info, int num_buf);
29 static void destroy_receive_buffers(struct smbd_connection *info);
31 static void put_empty_packet(
32 struct smbd_connection *info, struct smbd_response *response);
33 static void enqueue_reassembly(
34 struct smbd_connection *info,
35 struct smbd_response *response, int data_length);
36 static struct smbd_response *_get_first_reassembly(
37 struct smbd_connection *info);
39 static int smbd_post_recv(
40 struct smbd_connection *info,
41 struct smbd_response *response);
43 static int smbd_post_send_empty(struct smbd_connection *info);
44 static int smbd_post_send_data(
45 struct smbd_connection *info,
46 struct kvec *iov, int n_vec, int remaining_data_length);
47 static int smbd_post_send_page(struct smbd_connection *info,
48 struct page *page, unsigned long offset,
49 size_t size, int remaining_data_length);
51 static void destroy_mr_list(struct smbd_connection *info);
52 static int allocate_mr_list(struct smbd_connection *info);
54 /* SMBD version number */
55 #define SMBD_V1 0x0100
57 /* Port numbers for SMBD transport */
59 #define SMBD_PORT 5445
61 /* Address lookup and resolve timeout in ms */
62 #define RDMA_RESOLVE_TIMEOUT 5000
64 /* SMBD negotiation timeout in seconds */
65 #define SMBD_NEGOTIATE_TIMEOUT 120
67 /* SMBD minimum receive size and fragmented sized defined in [MS-SMBD] */
68 #define SMBD_MIN_RECEIVE_SIZE 128
69 #define SMBD_MIN_FRAGMENTED_SIZE 131072
72 * Default maximum number of RDMA read/write outstanding on this connection
73 * This value is possibly decreased during QP creation on hardware limit
75 #define SMBD_CM_RESPONDER_RESOURCES 32
77 /* Maximum number of retries on data transfer operations */
78 #define SMBD_CM_RETRY 6
79 /* No need to retry on Receiver Not Ready since SMBD manages credits */
80 #define SMBD_CM_RNR_RETRY 0
83 * User configurable initial values per SMBD transport connection
84 * as defined in [MS-SMBD] 3.1.1.1
85 * Those may change after a SMBD negotiation
87 /* The local peer's maximum number of credits to grant to the peer */
88 int smbd_receive_credit_max = 255;
90 /* The remote peer's credit request of local peer */
91 int smbd_send_credit_target = 255;
93 /* The maximum single message size can be sent to remote peer */
94 int smbd_max_send_size = 1364;
96 /* The maximum fragmented upper-layer payload receive size supported */
97 int smbd_max_fragmented_recv_size = 1024 * 1024;
99 /* The maximum single-message size which can be received */
100 int smbd_max_receive_size = 8192;
102 /* The timeout to initiate send of a keepalive message on idle */
103 int smbd_keep_alive_interval = 120;
106 * User configurable initial values for RDMA transport
107 * The actual values used may be lower and are limited to hardware capabilities
109 /* Default maximum number of SGEs in a RDMA write/read */
110 int smbd_max_frmr_depth = 2048;
112 /* If payload is less than this byte, use RDMA send/recv not read/write */
113 int rdma_readwrite_threshold = 4096;
115 /* Transport logging functions
116 * Logging are defined as classes. They can be OR'ed to define the actual
117 * logging level via module parameter smbd_logging_class
118 * e.g. cifs.smbd_logging_class=0xa0 will log all log_rdma_recv() and
121 #define LOG_OUTGOING 0x1
122 #define LOG_INCOMING 0x2
124 #define LOG_WRITE 0x8
125 #define LOG_RDMA_SEND 0x10
126 #define LOG_RDMA_RECV 0x20
127 #define LOG_KEEP_ALIVE 0x40
128 #define LOG_RDMA_EVENT 0x80
129 #define LOG_RDMA_MR 0x100
130 static unsigned int smbd_logging_class;
131 module_param(smbd_logging_class, uint, 0644);
132 MODULE_PARM_DESC(smbd_logging_class,
133 "Logging class for SMBD transport 0x0 to 0x100");
137 static unsigned int smbd_logging_level = ERR;
138 module_param(smbd_logging_level, uint, 0644);
139 MODULE_PARM_DESC(smbd_logging_level,
140 "Logging level for SMBD transport, 0 (default): error, 1: info");
142 #define log_rdma(level, class, fmt, args...) \
144 if (level <= smbd_logging_level || class & smbd_logging_class) \
145 cifs_dbg(VFS, "%s:%d " fmt, __func__, __LINE__, ##args);\
148 #define log_outgoing(level, fmt, args...) \
149 log_rdma(level, LOG_OUTGOING, fmt, ##args)
150 #define log_incoming(level, fmt, args...) \
151 log_rdma(level, LOG_INCOMING, fmt, ##args)
152 #define log_read(level, fmt, args...) log_rdma(level, LOG_READ, fmt, ##args)
153 #define log_write(level, fmt, args...) log_rdma(level, LOG_WRITE, fmt, ##args)
154 #define log_rdma_send(level, fmt, args...) \
155 log_rdma(level, LOG_RDMA_SEND, fmt, ##args)
156 #define log_rdma_recv(level, fmt, args...) \
157 log_rdma(level, LOG_RDMA_RECV, fmt, ##args)
158 #define log_keep_alive(level, fmt, args...) \
159 log_rdma(level, LOG_KEEP_ALIVE, fmt, ##args)
160 #define log_rdma_event(level, fmt, args...) \
161 log_rdma(level, LOG_RDMA_EVENT, fmt, ##args)
162 #define log_rdma_mr(level, fmt, args...) \
163 log_rdma(level, LOG_RDMA_MR, fmt, ##args)
166 * Destroy the transport and related RDMA and memory resources
167 * Need to go through all the pending counters and make sure on one is using
168 * the transport while it is destroyed
170 static void smbd_destroy_rdma_work(struct work_struct *work)
172 struct smbd_response *response;
173 struct smbd_connection *info =
174 container_of(work, struct smbd_connection, destroy_work);
177 log_rdma_event(INFO, "destroying qp\n");
178 ib_drain_qp(info->id->qp);
179 rdma_destroy_qp(info->id);
181 /* Unblock all I/O waiting on the send queue */
182 wake_up_interruptible_all(&info->wait_send_queue);
184 log_rdma_event(INFO, "cancelling idle timer\n");
185 cancel_delayed_work_sync(&info->idle_timer_work);
186 log_rdma_event(INFO, "cancelling send immediate work\n");
187 cancel_delayed_work_sync(&info->send_immediate_work);
189 log_rdma_event(INFO, "wait for all send to finish\n");
190 wait_event(info->wait_smbd_send_pending,
191 info->smbd_send_pending == 0);
193 log_rdma_event(INFO, "wait for all recv to finish\n");
194 wake_up_interruptible(&info->wait_reassembly_queue);
195 wait_event(info->wait_smbd_recv_pending,
196 info->smbd_recv_pending == 0);
198 log_rdma_event(INFO, "wait for all send posted to IB to finish\n");
199 wait_event(info->wait_send_pending,
200 atomic_read(&info->send_pending) == 0);
201 wait_event(info->wait_send_payload_pending,
202 atomic_read(&info->send_payload_pending) == 0);
204 log_rdma_event(INFO, "freeing mr list\n");
205 wake_up_interruptible_all(&info->wait_mr);
206 wait_event(info->wait_for_mr_cleanup,
207 atomic_read(&info->mr_used_count) == 0);
208 destroy_mr_list(info);
210 /* It's not posssible for upper layer to get to reassembly */
211 log_rdma_event(INFO, "drain the reassembly queue\n");
213 spin_lock_irqsave(&info->reassembly_queue_lock, flags);
214 response = _get_first_reassembly(info);
216 list_del(&response->list);
217 spin_unlock_irqrestore(
218 &info->reassembly_queue_lock, flags);
219 put_receive_buffer(info, response);
221 spin_unlock_irqrestore(&info->reassembly_queue_lock, flags);
224 info->reassembly_data_length = 0;
226 log_rdma_event(INFO, "free receive buffers\n");
227 wait_event(info->wait_receive_queues,
228 info->count_receive_queue + info->count_empty_packet_queue
229 == info->receive_credit_max);
230 destroy_receive_buffers(info);
232 ib_free_cq(info->send_cq);
233 ib_free_cq(info->recv_cq);
234 ib_dealloc_pd(info->pd);
235 rdma_destroy_id(info->id);
238 mempool_destroy(info->request_mempool);
239 kmem_cache_destroy(info->request_cache);
241 mempool_destroy(info->response_mempool);
242 kmem_cache_destroy(info->response_cache);
244 info->transport_status = SMBD_DESTROYED;
245 wake_up_all(&info->wait_destroy);
248 static int smbd_process_disconnected(struct smbd_connection *info)
250 schedule_work(&info->destroy_work);
254 static void smbd_disconnect_rdma_work(struct work_struct *work)
256 struct smbd_connection *info =
257 container_of(work, struct smbd_connection, disconnect_work);
259 if (info->transport_status == SMBD_CONNECTED) {
260 info->transport_status = SMBD_DISCONNECTING;
261 rdma_disconnect(info->id);
265 static void smbd_disconnect_rdma_connection(struct smbd_connection *info)
267 queue_work(info->workqueue, &info->disconnect_work);
270 /* Upcall from RDMA CM */
271 static int smbd_conn_upcall(
272 struct rdma_cm_id *id, struct rdma_cm_event *event)
274 struct smbd_connection *info = id->context;
276 log_rdma_event(INFO, "event=%d status=%d\n",
277 event->event, event->status);
279 switch (event->event) {
280 case RDMA_CM_EVENT_ADDR_RESOLVED:
281 case RDMA_CM_EVENT_ROUTE_RESOLVED:
283 complete(&info->ri_done);
286 case RDMA_CM_EVENT_ADDR_ERROR:
287 info->ri_rc = -EHOSTUNREACH;
288 complete(&info->ri_done);
291 case RDMA_CM_EVENT_ROUTE_ERROR:
292 info->ri_rc = -ENETUNREACH;
293 complete(&info->ri_done);
296 case RDMA_CM_EVENT_ESTABLISHED:
297 log_rdma_event(INFO, "connected event=%d\n", event->event);
298 info->transport_status = SMBD_CONNECTED;
299 wake_up_interruptible(&info->conn_wait);
302 case RDMA_CM_EVENT_CONNECT_ERROR:
303 case RDMA_CM_EVENT_UNREACHABLE:
304 case RDMA_CM_EVENT_REJECTED:
305 log_rdma_event(INFO, "connecting failed event=%d\n", event->event);
306 info->transport_status = SMBD_DISCONNECTED;
307 wake_up_interruptible(&info->conn_wait);
310 case RDMA_CM_EVENT_DEVICE_REMOVAL:
311 case RDMA_CM_EVENT_DISCONNECTED:
312 /* This happenes when we fail the negotiation */
313 if (info->transport_status == SMBD_NEGOTIATE_FAILED) {
314 info->transport_status = SMBD_DISCONNECTED;
315 wake_up(&info->conn_wait);
319 info->transport_status = SMBD_DISCONNECTED;
320 smbd_process_disconnected(info);
330 /* Upcall from RDMA QP */
332 smbd_qp_async_error_upcall(struct ib_event *event, void *context)
334 struct smbd_connection *info = context;
336 log_rdma_event(ERR, "%s on device %s info %p\n",
337 ib_event_msg(event->event), event->device->name, info);
339 switch (event->event) {
340 case IB_EVENT_CQ_ERR:
341 case IB_EVENT_QP_FATAL:
342 smbd_disconnect_rdma_connection(info);
349 static inline void *smbd_request_payload(struct smbd_request *request)
351 return (void *)request->packet;
354 static inline void *smbd_response_payload(struct smbd_response *response)
356 return (void *)response->packet;
359 /* Called when a RDMA send is done */
360 static void send_done(struct ib_cq *cq, struct ib_wc *wc)
363 struct smbd_request *request =
364 container_of(wc->wr_cqe, struct smbd_request, cqe);
366 log_rdma_send(INFO, "smbd_request %p completed wc->status=%d\n",
367 request, wc->status);
369 if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_SEND) {
370 log_rdma_send(ERR, "wc->status=%d wc->opcode=%d\n",
371 wc->status, wc->opcode);
372 smbd_disconnect_rdma_connection(request->info);
375 for (i = 0; i < request->num_sge; i++)
376 ib_dma_unmap_single(request->info->id->device,
377 request->sge[i].addr,
378 request->sge[i].length,
381 if (request->has_payload) {
382 if (atomic_dec_and_test(&request->info->send_payload_pending))
383 wake_up(&request->info->wait_send_payload_pending);
385 if (atomic_dec_and_test(&request->info->send_pending))
386 wake_up(&request->info->wait_send_pending);
389 mempool_free(request, request->info->request_mempool);
392 static void dump_smbd_negotiate_resp(struct smbd_negotiate_resp *resp)
394 log_rdma_event(INFO, "resp message min_version %u max_version %u "
395 "negotiated_version %u credits_requested %u "
396 "credits_granted %u status %u max_readwrite_size %u "
397 "preferred_send_size %u max_receive_size %u "
398 "max_fragmented_size %u\n",
399 resp->min_version, resp->max_version, resp->negotiated_version,
400 resp->credits_requested, resp->credits_granted, resp->status,
401 resp->max_readwrite_size, resp->preferred_send_size,
402 resp->max_receive_size, resp->max_fragmented_size);
406 * Process a negotiation response message, according to [MS-SMBD]3.1.5.7
407 * response, packet_length: the negotiation response message
408 * return value: true if negotiation is a success, false if failed
410 static bool process_negotiation_response(
411 struct smbd_response *response, int packet_length)
413 struct smbd_connection *info = response->info;
414 struct smbd_negotiate_resp *packet = smbd_response_payload(response);
416 if (packet_length < sizeof(struct smbd_negotiate_resp)) {
418 "error: packet_length=%d\n", packet_length);
422 if (le16_to_cpu(packet->negotiated_version) != SMBD_V1) {
423 log_rdma_event(ERR, "error: negotiated_version=%x\n",
424 le16_to_cpu(packet->negotiated_version));
427 info->protocol = le16_to_cpu(packet->negotiated_version);
429 if (packet->credits_requested == 0) {
430 log_rdma_event(ERR, "error: credits_requested==0\n");
433 info->receive_credit_target = le16_to_cpu(packet->credits_requested);
435 if (packet->credits_granted == 0) {
436 log_rdma_event(ERR, "error: credits_granted==0\n");
439 atomic_set(&info->send_credits, le16_to_cpu(packet->credits_granted));
441 atomic_set(&info->receive_credits, 0);
443 if (le32_to_cpu(packet->preferred_send_size) > info->max_receive_size) {
444 log_rdma_event(ERR, "error: preferred_send_size=%d\n",
445 le32_to_cpu(packet->preferred_send_size));
448 info->max_receive_size = le32_to_cpu(packet->preferred_send_size);
450 if (le32_to_cpu(packet->max_receive_size) < SMBD_MIN_RECEIVE_SIZE) {
451 log_rdma_event(ERR, "error: max_receive_size=%d\n",
452 le32_to_cpu(packet->max_receive_size));
455 info->max_send_size = min_t(int, info->max_send_size,
456 le32_to_cpu(packet->max_receive_size));
458 if (le32_to_cpu(packet->max_fragmented_size) <
459 SMBD_MIN_FRAGMENTED_SIZE) {
460 log_rdma_event(ERR, "error: max_fragmented_size=%d\n",
461 le32_to_cpu(packet->max_fragmented_size));
464 info->max_fragmented_send_size =
465 le32_to_cpu(packet->max_fragmented_size);
466 info->rdma_readwrite_threshold =
467 rdma_readwrite_threshold > info->max_fragmented_send_size ?
468 info->max_fragmented_send_size :
469 rdma_readwrite_threshold;
472 info->max_readwrite_size = min_t(u32,
473 le32_to_cpu(packet->max_readwrite_size),
474 info->max_frmr_depth * PAGE_SIZE);
475 info->max_frmr_depth = info->max_readwrite_size / PAGE_SIZE;
481 * Check and schedule to send an immediate packet
482 * This is used to extend credtis to remote peer to keep the transport busy
484 static void check_and_send_immediate(struct smbd_connection *info)
486 if (info->transport_status != SMBD_CONNECTED)
489 info->send_immediate = true;
492 * Promptly send a packet if our peer is running low on receive
495 if (atomic_read(&info->receive_credits) <
496 info->receive_credit_target - 1)
498 info->workqueue, &info->send_immediate_work, 0);
501 static void smbd_post_send_credits(struct work_struct *work)
504 int use_receive_queue = 1;
506 struct smbd_response *response;
507 struct smbd_connection *info =
508 container_of(work, struct smbd_connection,
509 post_send_credits_work);
511 if (info->transport_status != SMBD_CONNECTED) {
512 wake_up(&info->wait_receive_queues);
516 if (info->receive_credit_target >
517 atomic_read(&info->receive_credits)) {
519 if (use_receive_queue)
520 response = get_receive_buffer(info);
522 response = get_empty_queue_buffer(info);
524 /* now switch to emtpy packet queue */
525 if (use_receive_queue) {
526 use_receive_queue = 0;
532 response->type = SMBD_TRANSFER_DATA;
533 response->first_segment = false;
534 rc = smbd_post_recv(info, response);
537 "post_recv failed rc=%d\n", rc);
538 put_receive_buffer(info, response);
546 spin_lock(&info->lock_new_credits_offered);
547 info->new_credits_offered += ret;
548 spin_unlock(&info->lock_new_credits_offered);
550 atomic_add(ret, &info->receive_credits);
552 /* Check if we can post new receive and grant credits to peer */
553 check_and_send_immediate(info);
556 static void smbd_recv_done_work(struct work_struct *work)
558 struct smbd_connection *info =
559 container_of(work, struct smbd_connection, recv_done_work);
562 * We may have new send credits granted from remote peer
563 * If any sender is blcoked on lack of credets, unblock it
565 if (atomic_read(&info->send_credits))
566 wake_up_interruptible(&info->wait_send_queue);
569 * Check if we need to send something to remote peer to
570 * grant more credits or respond to KEEP_ALIVE packet
572 check_and_send_immediate(info);
575 /* Called from softirq, when recv is done */
576 static void recv_done(struct ib_cq *cq, struct ib_wc *wc)
578 struct smbd_data_transfer *data_transfer;
579 struct smbd_response *response =
580 container_of(wc->wr_cqe, struct smbd_response, cqe);
581 struct smbd_connection *info = response->info;
584 log_rdma_recv(INFO, "response=%p type=%d wc status=%d wc opcode %d "
585 "byte_len=%d pkey_index=%x\n",
586 response, response->type, wc->status, wc->opcode,
587 wc->byte_len, wc->pkey_index);
589 if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_RECV) {
590 log_rdma_recv(INFO, "wc->status=%d opcode=%d\n",
591 wc->status, wc->opcode);
592 smbd_disconnect_rdma_connection(info);
596 ib_dma_sync_single_for_cpu(
599 response->sge.length,
602 switch (response->type) {
603 /* SMBD negotiation response */
604 case SMBD_NEGOTIATE_RESP:
605 dump_smbd_negotiate_resp(smbd_response_payload(response));
606 info->full_packet_received = true;
607 info->negotiate_done =
608 process_negotiation_response(response, wc->byte_len);
609 complete(&info->negotiate_completion);
612 /* SMBD data transfer packet */
613 case SMBD_TRANSFER_DATA:
614 data_transfer = smbd_response_payload(response);
615 data_length = le32_to_cpu(data_transfer->data_length);
618 * If this is a packet with data playload place the data in
619 * reassembly queue and wake up the reading thread
622 if (info->full_packet_received)
623 response->first_segment = true;
625 if (le32_to_cpu(data_transfer->remaining_data_length))
626 info->full_packet_received = false;
628 info->full_packet_received = true;
635 put_empty_packet(info, response);
638 wake_up_interruptible(&info->wait_reassembly_queue);
640 atomic_dec(&info->receive_credits);
641 info->receive_credit_target =
642 le16_to_cpu(data_transfer->credits_requested);
643 atomic_add(le16_to_cpu(data_transfer->credits_granted),
644 &info->send_credits);
646 log_incoming(INFO, "data flags %d data_offset %d "
647 "data_length %d remaining_data_length %d\n",
648 le16_to_cpu(data_transfer->flags),
649 le32_to_cpu(data_transfer->data_offset),
650 le32_to_cpu(data_transfer->data_length),
651 le32_to_cpu(data_transfer->remaining_data_length));
653 /* Send a KEEP_ALIVE response right away if requested */
654 info->keep_alive_requested = KEEP_ALIVE_NONE;
655 if (le16_to_cpu(data_transfer->flags) &
656 SMB_DIRECT_RESPONSE_REQUESTED) {
657 info->keep_alive_requested = KEEP_ALIVE_PENDING;
660 queue_work(info->workqueue, &info->recv_done_work);
665 "unexpected response type=%d\n", response->type);
669 put_receive_buffer(info, response);
672 static struct rdma_cm_id *smbd_create_id(
673 struct smbd_connection *info,
674 struct sockaddr *dstaddr, int port)
676 struct rdma_cm_id *id;
680 id = rdma_create_id(&init_net, smbd_conn_upcall, info,
681 RDMA_PS_TCP, IB_QPT_RC);
684 log_rdma_event(ERR, "rdma_create_id() failed %i\n", rc);
688 if (dstaddr->sa_family == AF_INET6)
689 sport = &((struct sockaddr_in6 *)dstaddr)->sin6_port;
691 sport = &((struct sockaddr_in *)dstaddr)->sin_port;
693 *sport = htons(port);
695 init_completion(&info->ri_done);
696 info->ri_rc = -ETIMEDOUT;
698 rc = rdma_resolve_addr(id, NULL, (struct sockaddr *)dstaddr,
699 RDMA_RESOLVE_TIMEOUT);
701 log_rdma_event(ERR, "rdma_resolve_addr() failed %i\n", rc);
704 wait_for_completion_interruptible_timeout(
705 &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT));
708 log_rdma_event(ERR, "rdma_resolve_addr() completed %i\n", rc);
712 info->ri_rc = -ETIMEDOUT;
713 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
715 log_rdma_event(ERR, "rdma_resolve_route() failed %i\n", rc);
718 wait_for_completion_interruptible_timeout(
719 &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT));
722 log_rdma_event(ERR, "rdma_resolve_route() completed %i\n", rc);
734 * Test if FRWR (Fast Registration Work Requests) is supported on the device
735 * This implementation requries FRWR on RDMA read/write
736 * return value: true if it is supported
738 static bool frwr_is_supported(struct ib_device_attr *attrs)
740 if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
742 if (attrs->max_fast_reg_page_list_len == 0)
747 static int smbd_ia_open(
748 struct smbd_connection *info,
749 struct sockaddr *dstaddr, int port)
753 info->id = smbd_create_id(info, dstaddr, port);
754 if (IS_ERR(info->id)) {
755 rc = PTR_ERR(info->id);
759 if (!frwr_is_supported(&info->id->device->attrs)) {
761 "Fast Registration Work Requests "
762 "(FRWR) is not supported\n");
764 "Device capability flags = %llx "
765 "max_fast_reg_page_list_len = %u\n",
766 info->id->device->attrs.device_cap_flags,
767 info->id->device->attrs.max_fast_reg_page_list_len);
768 rc = -EPROTONOSUPPORT;
771 info->max_frmr_depth = min_t(int,
773 info->id->device->attrs.max_fast_reg_page_list_len);
774 info->mr_type = IB_MR_TYPE_MEM_REG;
775 if (info->id->device->attrs.device_cap_flags & IB_DEVICE_SG_GAPS_REG)
776 info->mr_type = IB_MR_TYPE_SG_GAPS;
778 info->pd = ib_alloc_pd(info->id->device, 0);
779 if (IS_ERR(info->pd)) {
780 rc = PTR_ERR(info->pd);
781 log_rdma_event(ERR, "ib_alloc_pd() returned %d\n", rc);
788 rdma_destroy_id(info->id);
796 * Send a negotiation request message to the peer
797 * The negotiation procedure is in [MS-SMBD] 3.1.5.2 and 3.1.5.3
798 * After negotiation, the transport is connected and ready for
799 * carrying upper layer SMB payload
801 static int smbd_post_send_negotiate_req(struct smbd_connection *info)
803 struct ib_send_wr send_wr, *send_wr_fail;
805 struct smbd_request *request;
806 struct smbd_negotiate_req *packet;
808 request = mempool_alloc(info->request_mempool, GFP_KERNEL);
812 request->info = info;
814 packet = smbd_request_payload(request);
815 packet->min_version = cpu_to_le16(SMBD_V1);
816 packet->max_version = cpu_to_le16(SMBD_V1);
817 packet->reserved = 0;
818 packet->credits_requested = cpu_to_le16(info->send_credit_target);
819 packet->preferred_send_size = cpu_to_le32(info->max_send_size);
820 packet->max_receive_size = cpu_to_le32(info->max_receive_size);
821 packet->max_fragmented_size =
822 cpu_to_le32(info->max_fragmented_recv_size);
824 request->num_sge = 1;
825 request->sge[0].addr = ib_dma_map_single(
826 info->id->device, (void *)packet,
827 sizeof(*packet), DMA_TO_DEVICE);
828 if (ib_dma_mapping_error(info->id->device, request->sge[0].addr)) {
830 goto dma_mapping_failed;
833 request->sge[0].length = sizeof(*packet);
834 request->sge[0].lkey = info->pd->local_dma_lkey;
836 ib_dma_sync_single_for_device(
837 info->id->device, request->sge[0].addr,
838 request->sge[0].length, DMA_TO_DEVICE);
840 request->cqe.done = send_done;
843 send_wr.wr_cqe = &request->cqe;
844 send_wr.sg_list = request->sge;
845 send_wr.num_sge = request->num_sge;
846 send_wr.opcode = IB_WR_SEND;
847 send_wr.send_flags = IB_SEND_SIGNALED;
849 log_rdma_send(INFO, "sge addr=%llx length=%x lkey=%x\n",
850 request->sge[0].addr,
851 request->sge[0].length, request->sge[0].lkey);
853 request->has_payload = false;
854 atomic_inc(&info->send_pending);
855 rc = ib_post_send(info->id->qp, &send_wr, &send_wr_fail);
859 /* if we reach here, post send failed */
860 log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc);
861 atomic_dec(&info->send_pending);
862 ib_dma_unmap_single(info->id->device, request->sge[0].addr,
863 request->sge[0].length, DMA_TO_DEVICE);
865 smbd_disconnect_rdma_connection(info);
868 mempool_free(request, info->request_mempool);
873 * Extend the credits to remote peer
874 * This implements [MS-SMBD] 3.1.5.9
875 * The idea is that we should extend credits to remote peer as quickly as
876 * it's allowed, to maintain data flow. We allocate as much receive
877 * buffer as possible, and extend the receive credits to remote peer
878 * return value: the new credtis being granted.
880 static int manage_credits_prior_sending(struct smbd_connection *info)
884 spin_lock(&info->lock_new_credits_offered);
885 new_credits = info->new_credits_offered;
886 info->new_credits_offered = 0;
887 spin_unlock(&info->lock_new_credits_offered);
893 * Check if we need to send a KEEP_ALIVE message
894 * The idle connection timer triggers a KEEP_ALIVE message when expires
895 * SMB_DIRECT_RESPONSE_REQUESTED is set in the message flag to have peer send
898 * 1 if SMB_DIRECT_RESPONSE_REQUESTED needs to be set
901 static int manage_keep_alive_before_sending(struct smbd_connection *info)
903 if (info->keep_alive_requested == KEEP_ALIVE_PENDING) {
904 info->keep_alive_requested = KEEP_ALIVE_SENT;
911 * Build and prepare the SMBD packet header
912 * This function waits for avaialbe send credits and build a SMBD packet
913 * header. The caller then optional append payload to the packet after
916 * size: the size of the payload
917 * remaining_data_length: remaining data to send if this is part of a
920 * request_out: the request allocated from this function
921 * return values: 0 on success, otherwise actual error code returned
923 static int smbd_create_header(struct smbd_connection *info,
924 int size, int remaining_data_length,
925 struct smbd_request **request_out)
927 struct smbd_request *request;
928 struct smbd_data_transfer *packet;
932 /* Wait for send credits. A SMBD packet needs one credit */
933 rc = wait_event_interruptible(info->wait_send_queue,
934 atomic_read(&info->send_credits) > 0 ||
935 info->transport_status != SMBD_CONNECTED);
939 if (info->transport_status != SMBD_CONNECTED) {
940 log_outgoing(ERR, "disconnected not sending\n");
943 atomic_dec(&info->send_credits);
945 request = mempool_alloc(info->request_mempool, GFP_KERNEL);
951 request->info = info;
953 /* Fill in the packet header */
954 packet = smbd_request_payload(request);
955 packet->credits_requested = cpu_to_le16(info->send_credit_target);
956 packet->credits_granted =
957 cpu_to_le16(manage_credits_prior_sending(info));
958 info->send_immediate = false;
961 if (manage_keep_alive_before_sending(info))
962 packet->flags |= cpu_to_le16(SMB_DIRECT_RESPONSE_REQUESTED);
964 packet->reserved = 0;
966 packet->data_offset = 0;
968 packet->data_offset = cpu_to_le32(24);
969 packet->data_length = cpu_to_le32(size);
970 packet->remaining_data_length = cpu_to_le32(remaining_data_length);
973 log_outgoing(INFO, "credits_requested=%d credits_granted=%d "
974 "data_offset=%d data_length=%d remaining_data_length=%d\n",
975 le16_to_cpu(packet->credits_requested),
976 le16_to_cpu(packet->credits_granted),
977 le32_to_cpu(packet->data_offset),
978 le32_to_cpu(packet->data_length),
979 le32_to_cpu(packet->remaining_data_length));
981 /* Map the packet to DMA */
982 header_length = sizeof(struct smbd_data_transfer);
983 /* If this is a packet without payload, don't send padding */
985 header_length = offsetof(struct smbd_data_transfer, padding);
987 request->num_sge = 1;
988 request->sge[0].addr = ib_dma_map_single(info->id->device,
992 if (ib_dma_mapping_error(info->id->device, request->sge[0].addr)) {
993 mempool_free(request, info->request_mempool);
998 request->sge[0].length = header_length;
999 request->sge[0].lkey = info->pd->local_dma_lkey;
1001 *request_out = request;
1005 atomic_inc(&info->send_credits);
1009 static void smbd_destroy_header(struct smbd_connection *info,
1010 struct smbd_request *request)
1013 ib_dma_unmap_single(info->id->device,
1014 request->sge[0].addr,
1015 request->sge[0].length,
1017 mempool_free(request, info->request_mempool);
1018 atomic_inc(&info->send_credits);
1021 /* Post the send request */
1022 static int smbd_post_send(struct smbd_connection *info,
1023 struct smbd_request *request, bool has_payload)
1025 struct ib_send_wr send_wr, *send_wr_fail;
1028 for (i = 0; i < request->num_sge; i++) {
1030 "rdma_request sge[%d] addr=%llu length=%u\n",
1031 i, request->sge[i].addr, request->sge[i].length);
1032 ib_dma_sync_single_for_device(
1034 request->sge[i].addr,
1035 request->sge[i].length,
1039 request->cqe.done = send_done;
1041 send_wr.next = NULL;
1042 send_wr.wr_cqe = &request->cqe;
1043 send_wr.sg_list = request->sge;
1044 send_wr.num_sge = request->num_sge;
1045 send_wr.opcode = IB_WR_SEND;
1046 send_wr.send_flags = IB_SEND_SIGNALED;
1049 request->has_payload = true;
1050 atomic_inc(&info->send_payload_pending);
1052 request->has_payload = false;
1053 atomic_inc(&info->send_pending);
1056 rc = ib_post_send(info->id->qp, &send_wr, &send_wr_fail);
1058 log_rdma_send(ERR, "ib_post_send failed rc=%d\n", rc);
1060 if (atomic_dec_and_test(&info->send_payload_pending))
1061 wake_up(&info->wait_send_payload_pending);
1063 if (atomic_dec_and_test(&info->send_pending))
1064 wake_up(&info->wait_send_pending);
1066 smbd_disconnect_rdma_connection(info);
1068 /* Reset timer for idle connection after packet is sent */
1069 mod_delayed_work(info->workqueue, &info->idle_timer_work,
1070 info->keep_alive_interval*HZ);
1075 static int smbd_post_send_sgl(struct smbd_connection *info,
1076 struct scatterlist *sgl, int data_length, int remaining_data_length)
1080 struct smbd_request *request;
1081 struct scatterlist *sg;
1083 rc = smbd_create_header(
1084 info, data_length, remaining_data_length, &request);
1088 num_sgs = sgl ? sg_nents(sgl) : 0;
1089 for_each_sg(sgl, sg, num_sgs, i) {
1090 request->sge[i+1].addr =
1091 ib_dma_map_page(info->id->device, sg_page(sg),
1092 sg->offset, sg->length, DMA_BIDIRECTIONAL);
1093 if (ib_dma_mapping_error(
1094 info->id->device, request->sge[i+1].addr)) {
1096 request->sge[i+1].addr = 0;
1097 goto dma_mapping_failure;
1099 request->sge[i+1].length = sg->length;
1100 request->sge[i+1].lkey = info->pd->local_dma_lkey;
1104 rc = smbd_post_send(info, request, data_length);
1108 dma_mapping_failure:
1109 for (i = 1; i < request->num_sge; i++)
1110 if (request->sge[i].addr)
1111 ib_dma_unmap_single(info->id->device,
1112 request->sge[i].addr,
1113 request->sge[i].length,
1115 smbd_destroy_header(info, request);
1121 * page: the page to send
1122 * offset: offset in the page to send
1123 * size: length in the page to send
1124 * remaining_data_length: remaining data to send in this payload
1126 static int smbd_post_send_page(struct smbd_connection *info, struct page *page,
1127 unsigned long offset, size_t size, int remaining_data_length)
1129 struct scatterlist sgl;
1131 sg_init_table(&sgl, 1);
1132 sg_set_page(&sgl, page, size, offset);
1134 return smbd_post_send_sgl(info, &sgl, size, remaining_data_length);
1138 * Send an empty message
1139 * Empty message is used to extend credits to peer to for keep live
1140 * while there is no upper layer payload to send at the time
1142 static int smbd_post_send_empty(struct smbd_connection *info)
1144 info->count_send_empty++;
1145 return smbd_post_send_sgl(info, NULL, 0, 0);
1149 * Send a data buffer
1150 * iov: the iov array describing the data buffers
1151 * n_vec: number of iov array
1152 * remaining_data_length: remaining data to send following this packet
1153 * in segmented SMBD packet
1155 static int smbd_post_send_data(
1156 struct smbd_connection *info, struct kvec *iov, int n_vec,
1157 int remaining_data_length)
1160 u32 data_length = 0;
1161 struct scatterlist sgl[SMBDIRECT_MAX_SGE];
1163 if (n_vec > SMBDIRECT_MAX_SGE) {
1164 cifs_dbg(VFS, "Can't fit data to SGL, n_vec=%d\n", n_vec);
1168 sg_init_table(sgl, n_vec);
1169 for (i = 0; i < n_vec; i++) {
1170 data_length += iov[i].iov_len;
1171 sg_set_buf(&sgl[i], iov[i].iov_base, iov[i].iov_len);
1174 return smbd_post_send_sgl(info, sgl, data_length, remaining_data_length);
1178 * Post a receive request to the transport
1179 * The remote peer can only send data when a receive request is posted
1180 * The interaction is controlled by send/receive credit system
1182 static int smbd_post_recv(
1183 struct smbd_connection *info, struct smbd_response *response)
1185 struct ib_recv_wr recv_wr, *recv_wr_fail = NULL;
1188 response->sge.addr = ib_dma_map_single(
1189 info->id->device, response->packet,
1190 info->max_receive_size, DMA_FROM_DEVICE);
1191 if (ib_dma_mapping_error(info->id->device, response->sge.addr))
1194 response->sge.length = info->max_receive_size;
1195 response->sge.lkey = info->pd->local_dma_lkey;
1197 response->cqe.done = recv_done;
1199 recv_wr.wr_cqe = &response->cqe;
1200 recv_wr.next = NULL;
1201 recv_wr.sg_list = &response->sge;
1202 recv_wr.num_sge = 1;
1204 rc = ib_post_recv(info->id->qp, &recv_wr, &recv_wr_fail);
1206 ib_dma_unmap_single(info->id->device, response->sge.addr,
1207 response->sge.length, DMA_FROM_DEVICE);
1208 smbd_disconnect_rdma_connection(info);
1209 log_rdma_recv(ERR, "ib_post_recv failed rc=%d\n", rc);
1215 /* Perform SMBD negotiate according to [MS-SMBD] 3.1.5.2 */
1216 static int smbd_negotiate(struct smbd_connection *info)
1219 struct smbd_response *response = get_receive_buffer(info);
1221 response->type = SMBD_NEGOTIATE_RESP;
1222 rc = smbd_post_recv(info, response);
1223 log_rdma_event(INFO,
1224 "smbd_post_recv rc=%d iov.addr=%llx iov.length=%x "
1226 rc, response->sge.addr,
1227 response->sge.length, response->sge.lkey);
1231 init_completion(&info->negotiate_completion);
1232 info->negotiate_done = false;
1233 rc = smbd_post_send_negotiate_req(info);
1237 rc = wait_for_completion_interruptible_timeout(
1238 &info->negotiate_completion, SMBD_NEGOTIATE_TIMEOUT * HZ);
1239 log_rdma_event(INFO, "wait_for_completion_timeout rc=%d\n", rc);
1241 if (info->negotiate_done)
1246 else if (rc == -ERESTARTSYS)
1254 static void put_empty_packet(
1255 struct smbd_connection *info, struct smbd_response *response)
1257 spin_lock(&info->empty_packet_queue_lock);
1258 list_add_tail(&response->list, &info->empty_packet_queue);
1259 info->count_empty_packet_queue++;
1260 spin_unlock(&info->empty_packet_queue_lock);
1262 queue_work(info->workqueue, &info->post_send_credits_work);
1266 * Implement Connection.FragmentReassemblyBuffer defined in [MS-SMBD] 3.1.1.1
1267 * This is a queue for reassembling upper layer payload and present to upper
1268 * layer. All the inncoming payload go to the reassembly queue, regardless of
1269 * if reassembly is required. The uuper layer code reads from the queue for all
1270 * incoming payloads.
1271 * Put a received packet to the reassembly queue
1272 * response: the packet received
1273 * data_length: the size of payload in this packet
1275 static void enqueue_reassembly(
1276 struct smbd_connection *info,
1277 struct smbd_response *response,
1280 spin_lock(&info->reassembly_queue_lock);
1281 list_add_tail(&response->list, &info->reassembly_queue);
1282 info->reassembly_queue_length++;
1284 * Make sure reassembly_data_length is updated after list and
1285 * reassembly_queue_length are updated. On the dequeue side
1286 * reassembly_data_length is checked without a lock to determine
1287 * if reassembly_queue_length and list is up to date
1290 info->reassembly_data_length += data_length;
1291 spin_unlock(&info->reassembly_queue_lock);
1292 info->count_reassembly_queue++;
1293 info->count_enqueue_reassembly_queue++;
1297 * Get the first entry at the front of reassembly queue
1298 * Caller is responsible for locking
1299 * return value: the first entry if any, NULL if queue is empty
1301 static struct smbd_response *_get_first_reassembly(struct smbd_connection *info)
1303 struct smbd_response *ret = NULL;
1305 if (!list_empty(&info->reassembly_queue)) {
1306 ret = list_first_entry(
1307 &info->reassembly_queue,
1308 struct smbd_response, list);
1313 static struct smbd_response *get_empty_queue_buffer(
1314 struct smbd_connection *info)
1316 struct smbd_response *ret = NULL;
1317 unsigned long flags;
1319 spin_lock_irqsave(&info->empty_packet_queue_lock, flags);
1320 if (!list_empty(&info->empty_packet_queue)) {
1321 ret = list_first_entry(
1322 &info->empty_packet_queue,
1323 struct smbd_response, list);
1324 list_del(&ret->list);
1325 info->count_empty_packet_queue--;
1327 spin_unlock_irqrestore(&info->empty_packet_queue_lock, flags);
1333 * Get a receive buffer
1334 * For each remote send, we need to post a receive. The receive buffers are
1335 * pre-allocated in advance.
1336 * return value: the receive buffer, NULL if none is available
1338 static struct smbd_response *get_receive_buffer(struct smbd_connection *info)
1340 struct smbd_response *ret = NULL;
1341 unsigned long flags;
1343 spin_lock_irqsave(&info->receive_queue_lock, flags);
1344 if (!list_empty(&info->receive_queue)) {
1345 ret = list_first_entry(
1346 &info->receive_queue,
1347 struct smbd_response, list);
1348 list_del(&ret->list);
1349 info->count_receive_queue--;
1350 info->count_get_receive_buffer++;
1352 spin_unlock_irqrestore(&info->receive_queue_lock, flags);
1358 * Return a receive buffer
1359 * Upon returning of a receive buffer, we can post new receive and extend
1360 * more receive credits to remote peer. This is done immediately after a
1361 * receive buffer is returned.
1363 static void put_receive_buffer(
1364 struct smbd_connection *info, struct smbd_response *response)
1366 unsigned long flags;
1368 ib_dma_unmap_single(info->id->device, response->sge.addr,
1369 response->sge.length, DMA_FROM_DEVICE);
1371 spin_lock_irqsave(&info->receive_queue_lock, flags);
1372 list_add_tail(&response->list, &info->receive_queue);
1373 info->count_receive_queue++;
1374 info->count_put_receive_buffer++;
1375 spin_unlock_irqrestore(&info->receive_queue_lock, flags);
1377 queue_work(info->workqueue, &info->post_send_credits_work);
1380 /* Preallocate all receive buffer on transport establishment */
1381 static int allocate_receive_buffers(struct smbd_connection *info, int num_buf)
1384 struct smbd_response *response;
1386 INIT_LIST_HEAD(&info->reassembly_queue);
1387 spin_lock_init(&info->reassembly_queue_lock);
1388 info->reassembly_data_length = 0;
1389 info->reassembly_queue_length = 0;
1391 INIT_LIST_HEAD(&info->receive_queue);
1392 spin_lock_init(&info->receive_queue_lock);
1393 info->count_receive_queue = 0;
1395 INIT_LIST_HEAD(&info->empty_packet_queue);
1396 spin_lock_init(&info->empty_packet_queue_lock);
1397 info->count_empty_packet_queue = 0;
1399 init_waitqueue_head(&info->wait_receive_queues);
1401 for (i = 0; i < num_buf; i++) {
1402 response = mempool_alloc(info->response_mempool, GFP_KERNEL);
1404 goto allocate_failed;
1406 response->info = info;
1407 list_add_tail(&response->list, &info->receive_queue);
1408 info->count_receive_queue++;
1414 while (!list_empty(&info->receive_queue)) {
1415 response = list_first_entry(
1416 &info->receive_queue,
1417 struct smbd_response, list);
1418 list_del(&response->list);
1419 info->count_receive_queue--;
1421 mempool_free(response, info->response_mempool);
1426 static void destroy_receive_buffers(struct smbd_connection *info)
1428 struct smbd_response *response;
1430 while ((response = get_receive_buffer(info)))
1431 mempool_free(response, info->response_mempool);
1433 while ((response = get_empty_queue_buffer(info)))
1434 mempool_free(response, info->response_mempool);
1438 * Check and send an immediate or keep alive packet
1439 * The condition to send those packets are defined in [MS-SMBD] 3.1.1.1
1440 * Connection.KeepaliveRequested and Connection.SendImmediate
1441 * The idea is to extend credits to server as soon as it becomes available
1443 static void send_immediate_work(struct work_struct *work)
1445 struct smbd_connection *info = container_of(
1446 work, struct smbd_connection,
1447 send_immediate_work.work);
1449 if (info->keep_alive_requested == KEEP_ALIVE_PENDING ||
1450 info->send_immediate) {
1451 log_keep_alive(INFO, "send an empty message\n");
1452 smbd_post_send_empty(info);
1456 /* Implement idle connection timer [MS-SMBD] 3.1.6.2 */
1457 static void idle_connection_timer(struct work_struct *work)
1459 struct smbd_connection *info = container_of(
1460 work, struct smbd_connection,
1461 idle_timer_work.work);
1463 if (info->keep_alive_requested != KEEP_ALIVE_NONE) {
1465 "error status info->keep_alive_requested=%d\n",
1466 info->keep_alive_requested);
1467 smbd_disconnect_rdma_connection(info);
1471 log_keep_alive(INFO, "about to send an empty idle message\n");
1472 smbd_post_send_empty(info);
1474 /* Setup the next idle timeout work */
1475 queue_delayed_work(info->workqueue, &info->idle_timer_work,
1476 info->keep_alive_interval*HZ);
1479 /* Destroy this SMBD connection, called from upper layer */
1480 void smbd_destroy(struct smbd_connection *info)
1482 log_rdma_event(INFO, "destroying rdma session\n");
1484 /* Kick off the disconnection process */
1485 smbd_disconnect_rdma_connection(info);
1487 log_rdma_event(INFO, "wait for transport being destroyed\n");
1488 wait_event(info->wait_destroy,
1489 info->transport_status == SMBD_DESTROYED);
1491 destroy_workqueue(info->workqueue);
1496 * Reconnect this SMBD connection, called from upper layer
1497 * return value: 0 on success, or actual error code
1499 int smbd_reconnect(struct TCP_Server_Info *server)
1501 log_rdma_event(INFO, "reconnecting rdma session\n");
1503 if (!server->smbd_conn) {
1504 log_rdma_event(INFO, "rdma session already destroyed\n");
1509 * This is possible if transport is disconnected and we haven't received
1510 * notification from RDMA, but upper layer has detected timeout
1512 if (server->smbd_conn->transport_status == SMBD_CONNECTED) {
1513 log_rdma_event(INFO, "disconnecting transport\n");
1514 smbd_disconnect_rdma_connection(server->smbd_conn);
1517 /* wait until the transport is destroyed */
1518 if (!wait_event_timeout(server->smbd_conn->wait_destroy,
1519 server->smbd_conn->transport_status == SMBD_DESTROYED, 5*HZ))
1522 destroy_workqueue(server->smbd_conn->workqueue);
1523 kfree(server->smbd_conn);
1526 log_rdma_event(INFO, "creating rdma session\n");
1527 server->smbd_conn = smbd_get_connection(
1528 server, (struct sockaddr *) &server->dstaddr);
1529 log_rdma_event(INFO, "created rdma session info=%p\n",
1532 return server->smbd_conn ? 0 : -ENOENT;
1535 static void destroy_caches_and_workqueue(struct smbd_connection *info)
1537 destroy_receive_buffers(info);
1538 destroy_workqueue(info->workqueue);
1539 mempool_destroy(info->response_mempool);
1540 kmem_cache_destroy(info->response_cache);
1541 mempool_destroy(info->request_mempool);
1542 kmem_cache_destroy(info->request_cache);
1545 #define MAX_NAME_LEN 80
1546 static int allocate_caches_and_workqueue(struct smbd_connection *info)
1548 char name[MAX_NAME_LEN];
1551 snprintf(name, MAX_NAME_LEN, "smbd_request_%p", info);
1552 info->request_cache =
1555 sizeof(struct smbd_request) +
1556 sizeof(struct smbd_data_transfer),
1557 0, SLAB_HWCACHE_ALIGN, NULL);
1558 if (!info->request_cache)
1561 info->request_mempool =
1562 mempool_create(info->send_credit_target, mempool_alloc_slab,
1563 mempool_free_slab, info->request_cache);
1564 if (!info->request_mempool)
1567 snprintf(name, MAX_NAME_LEN, "smbd_response_%p", info);
1568 info->response_cache =
1571 sizeof(struct smbd_response) +
1572 info->max_receive_size,
1573 0, SLAB_HWCACHE_ALIGN, NULL);
1574 if (!info->response_cache)
1577 info->response_mempool =
1578 mempool_create(info->receive_credit_max, mempool_alloc_slab,
1579 mempool_free_slab, info->response_cache);
1580 if (!info->response_mempool)
1583 snprintf(name, MAX_NAME_LEN, "smbd_%p", info);
1584 info->workqueue = create_workqueue(name);
1585 if (!info->workqueue)
1588 rc = allocate_receive_buffers(info, info->receive_credit_max);
1590 log_rdma_event(ERR, "failed to allocate receive buffers\n");
1597 destroy_workqueue(info->workqueue);
1599 mempool_destroy(info->response_mempool);
1601 kmem_cache_destroy(info->response_cache);
1603 mempool_destroy(info->request_mempool);
1605 kmem_cache_destroy(info->request_cache);
1609 /* Create a SMBD connection, called by upper layer */
1610 static struct smbd_connection *_smbd_get_connection(
1611 struct TCP_Server_Info *server, struct sockaddr *dstaddr, int port)
1614 struct smbd_connection *info;
1615 struct rdma_conn_param conn_param;
1616 struct ib_qp_init_attr qp_attr;
1617 struct sockaddr_in *addr_in = (struct sockaddr_in *) dstaddr;
1618 struct ib_port_immutable port_immutable;
1621 info = kzalloc(sizeof(struct smbd_connection), GFP_KERNEL);
1625 info->transport_status = SMBD_CONNECTING;
1626 rc = smbd_ia_open(info, dstaddr, port);
1628 log_rdma_event(INFO, "smbd_ia_open rc=%d\n", rc);
1629 goto create_id_failed;
1632 if (smbd_send_credit_target > info->id->device->attrs.max_cqe ||
1633 smbd_send_credit_target > info->id->device->attrs.max_qp_wr) {
1635 "consider lowering send_credit_target = %d. "
1636 "Possible CQE overrun, device "
1637 "reporting max_cpe %d max_qp_wr %d\n",
1638 smbd_send_credit_target,
1639 info->id->device->attrs.max_cqe,
1640 info->id->device->attrs.max_qp_wr);
1644 if (smbd_receive_credit_max > info->id->device->attrs.max_cqe ||
1645 smbd_receive_credit_max > info->id->device->attrs.max_qp_wr) {
1647 "consider lowering receive_credit_max = %d. "
1648 "Possible CQE overrun, device "
1649 "reporting max_cpe %d max_qp_wr %d\n",
1650 smbd_receive_credit_max,
1651 info->id->device->attrs.max_cqe,
1652 info->id->device->attrs.max_qp_wr);
1656 info->receive_credit_max = smbd_receive_credit_max;
1657 info->send_credit_target = smbd_send_credit_target;
1658 info->max_send_size = smbd_max_send_size;
1659 info->max_fragmented_recv_size = smbd_max_fragmented_recv_size;
1660 info->max_receive_size = smbd_max_receive_size;
1661 info->keep_alive_interval = smbd_keep_alive_interval;
1663 if (info->id->device->attrs.max_sge < SMBDIRECT_MAX_SGE) {
1664 log_rdma_event(ERR, "warning: device max_sge = %d too small\n",
1665 info->id->device->attrs.max_sge);
1666 log_rdma_event(ERR, "Queue Pair creation may fail\n");
1669 info->send_cq = NULL;
1670 info->recv_cq = NULL;
1671 info->send_cq = ib_alloc_cq(info->id->device, info,
1672 info->send_credit_target, 0, IB_POLL_SOFTIRQ);
1673 if (IS_ERR(info->send_cq)) {
1674 info->send_cq = NULL;
1675 goto alloc_cq_failed;
1678 info->recv_cq = ib_alloc_cq(info->id->device, info,
1679 info->receive_credit_max, 0, IB_POLL_SOFTIRQ);
1680 if (IS_ERR(info->recv_cq)) {
1681 info->recv_cq = NULL;
1682 goto alloc_cq_failed;
1685 memset(&qp_attr, 0, sizeof(qp_attr));
1686 qp_attr.event_handler = smbd_qp_async_error_upcall;
1687 qp_attr.qp_context = info;
1688 qp_attr.cap.max_send_wr = info->send_credit_target;
1689 qp_attr.cap.max_recv_wr = info->receive_credit_max;
1690 qp_attr.cap.max_send_sge = SMBDIRECT_MAX_SGE;
1691 qp_attr.cap.max_recv_sge = SMBDIRECT_MAX_SGE;
1692 qp_attr.cap.max_inline_data = 0;
1693 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1694 qp_attr.qp_type = IB_QPT_RC;
1695 qp_attr.send_cq = info->send_cq;
1696 qp_attr.recv_cq = info->recv_cq;
1697 qp_attr.port_num = ~0;
1699 rc = rdma_create_qp(info->id, info->pd, &qp_attr);
1701 log_rdma_event(ERR, "rdma_create_qp failed %i\n", rc);
1702 goto create_qp_failed;
1705 memset(&conn_param, 0, sizeof(conn_param));
1706 conn_param.initiator_depth = 0;
1708 conn_param.responder_resources =
1709 info->id->device->attrs.max_qp_rd_atom
1710 < SMBD_CM_RESPONDER_RESOURCES ?
1711 info->id->device->attrs.max_qp_rd_atom :
1712 SMBD_CM_RESPONDER_RESOURCES;
1713 info->responder_resources = conn_param.responder_resources;
1714 log_rdma_mr(INFO, "responder_resources=%d\n",
1715 info->responder_resources);
1717 /* Need to send IRD/ORD in private data for iWARP */
1718 info->id->device->get_port_immutable(
1719 info->id->device, info->id->port_num, &port_immutable);
1720 if (port_immutable.core_cap_flags & RDMA_CORE_PORT_IWARP) {
1721 ird_ord_hdr[0] = info->responder_resources;
1723 conn_param.private_data = ird_ord_hdr;
1724 conn_param.private_data_len = sizeof(ird_ord_hdr);
1726 conn_param.private_data = NULL;
1727 conn_param.private_data_len = 0;
1730 conn_param.retry_count = SMBD_CM_RETRY;
1731 conn_param.rnr_retry_count = SMBD_CM_RNR_RETRY;
1732 conn_param.flow_control = 0;
1733 init_waitqueue_head(&info->wait_destroy);
1735 log_rdma_event(INFO, "connecting to IP %pI4 port %d\n",
1736 &addr_in->sin_addr, port);
1738 init_waitqueue_head(&info->conn_wait);
1739 rc = rdma_connect(info->id, &conn_param);
1741 log_rdma_event(ERR, "rdma_connect() failed with %i\n", rc);
1742 goto rdma_connect_failed;
1745 wait_event_interruptible(
1746 info->conn_wait, info->transport_status != SMBD_CONNECTING);
1748 if (info->transport_status != SMBD_CONNECTED) {
1749 log_rdma_event(ERR, "rdma_connect failed port=%d\n", port);
1750 goto rdma_connect_failed;
1753 log_rdma_event(INFO, "rdma_connect connected\n");
1755 rc = allocate_caches_and_workqueue(info);
1757 log_rdma_event(ERR, "cache allocation failed\n");
1758 goto allocate_cache_failed;
1761 init_waitqueue_head(&info->wait_send_queue);
1762 init_waitqueue_head(&info->wait_reassembly_queue);
1764 INIT_DELAYED_WORK(&info->idle_timer_work, idle_connection_timer);
1765 INIT_DELAYED_WORK(&info->send_immediate_work, send_immediate_work);
1766 queue_delayed_work(info->workqueue, &info->idle_timer_work,
1767 info->keep_alive_interval*HZ);
1769 init_waitqueue_head(&info->wait_smbd_send_pending);
1770 info->smbd_send_pending = 0;
1772 init_waitqueue_head(&info->wait_smbd_recv_pending);
1773 info->smbd_recv_pending = 0;
1775 init_waitqueue_head(&info->wait_send_pending);
1776 atomic_set(&info->send_pending, 0);
1778 init_waitqueue_head(&info->wait_send_payload_pending);
1779 atomic_set(&info->send_payload_pending, 0);
1781 INIT_WORK(&info->disconnect_work, smbd_disconnect_rdma_work);
1782 INIT_WORK(&info->destroy_work, smbd_destroy_rdma_work);
1783 INIT_WORK(&info->recv_done_work, smbd_recv_done_work);
1784 INIT_WORK(&info->post_send_credits_work, smbd_post_send_credits);
1785 info->new_credits_offered = 0;
1786 spin_lock_init(&info->lock_new_credits_offered);
1788 rc = smbd_negotiate(info);
1790 log_rdma_event(ERR, "smbd_negotiate rc=%d\n", rc);
1791 goto negotiation_failed;
1794 rc = allocate_mr_list(info);
1796 log_rdma_mr(ERR, "memory registration allocation failed\n");
1797 goto allocate_mr_failed;
1803 /* At this point, need to a full transport shutdown */
1808 cancel_delayed_work_sync(&info->idle_timer_work);
1809 destroy_caches_and_workqueue(info);
1810 info->transport_status = SMBD_NEGOTIATE_FAILED;
1811 init_waitqueue_head(&info->conn_wait);
1812 rdma_disconnect(info->id);
1813 wait_event(info->conn_wait,
1814 info->transport_status == SMBD_DISCONNECTED);
1816 allocate_cache_failed:
1817 rdma_connect_failed:
1818 rdma_destroy_qp(info->id);
1823 ib_free_cq(info->send_cq);
1825 ib_free_cq(info->recv_cq);
1828 ib_dealloc_pd(info->pd);
1829 rdma_destroy_id(info->id);
1836 struct smbd_connection *smbd_get_connection(
1837 struct TCP_Server_Info *server, struct sockaddr *dstaddr)
1839 struct smbd_connection *ret;
1840 int port = SMBD_PORT;
1843 ret = _smbd_get_connection(server, dstaddr, port);
1845 /* Try SMB_PORT if SMBD_PORT doesn't work */
1846 if (!ret && port == SMBD_PORT) {
1854 * Receive data from receive reassembly queue
1855 * All the incoming data packets are placed in reassembly queue
1856 * buf: the buffer to read data into
1857 * size: the length of data to read
1858 * return value: actual data read
1859 * Note: this implementation copies the data from reassebmly queue to receive
1860 * buffers used by upper layer. This is not the optimal code path. A better way
1861 * to do it is to not have upper layer allocate its receive buffers but rather
1862 * borrow the buffer from reassembly queue, and return it after data is
1863 * consumed. But this will require more changes to upper layer code, and also
1864 * need to consider packet boundaries while they still being reassembled.
1866 static int smbd_recv_buf(struct smbd_connection *info, char *buf,
1869 struct smbd_response *response;
1870 struct smbd_data_transfer *data_transfer;
1871 int to_copy, to_read, data_read, offset;
1872 u32 data_length, remaining_data_length, data_offset;
1876 if (info->transport_status != SMBD_CONNECTED) {
1877 log_read(ERR, "disconnected\n");
1882 * No need to hold the reassembly queue lock all the time as we are
1883 * the only one reading from the front of the queue. The transport
1884 * may add more entries to the back of the queue at the same time
1886 log_read(INFO, "size=%d info->reassembly_data_length=%d\n", size,
1887 info->reassembly_data_length);
1888 if (info->reassembly_data_length >= size) {
1890 int queue_removed = 0;
1893 * Need to make sure reassembly_data_length is read before
1894 * reading reassembly_queue_length and calling
1895 * _get_first_reassembly. This call is lock free
1896 * as we never read at the end of the queue which are being
1897 * updated in SOFTIRQ as more data is received
1900 queue_length = info->reassembly_queue_length;
1903 offset = info->first_entry_offset;
1904 while (data_read < size) {
1905 response = _get_first_reassembly(info);
1906 data_transfer = smbd_response_payload(response);
1907 data_length = le32_to_cpu(data_transfer->data_length);
1908 remaining_data_length =
1910 data_transfer->remaining_data_length);
1911 data_offset = le32_to_cpu(data_transfer->data_offset);
1914 * The upper layer expects RFC1002 length at the
1915 * beginning of the payload. Return it to indicate
1916 * the total length of the packet. This minimize the
1917 * change to upper layer packet processing logic. This
1918 * will be eventually remove when an intermediate
1919 * transport layer is added
1921 if (response->first_segment && size == 4) {
1922 unsigned int rfc1002_len =
1923 data_length + remaining_data_length;
1924 *((__be32 *)buf) = cpu_to_be32(rfc1002_len);
1926 response->first_segment = false;
1927 log_read(INFO, "returning rfc1002 length %d\n",
1929 goto read_rfc1002_done;
1932 to_copy = min_t(int, data_length - offset, to_read);
1935 (char *)data_transfer + data_offset + offset,
1938 /* move on to the next buffer? */
1939 if (to_copy == data_length - offset) {
1942 * No need to lock if we are not at the
1946 list_del(&response->list);
1949 &info->reassembly_queue_lock);
1950 list_del(&response->list);
1952 &info->reassembly_queue_lock);
1955 info->count_reassembly_queue--;
1956 info->count_dequeue_reassembly_queue++;
1957 put_receive_buffer(info, response);
1959 log_read(INFO, "put_receive_buffer offset=0\n");
1964 data_read += to_copy;
1966 log_read(INFO, "_get_first_reassembly memcpy %d bytes "
1967 "data_transfer_length-offset=%d after that "
1968 "to_read=%d data_read=%d offset=%d\n",
1969 to_copy, data_length - offset,
1970 to_read, data_read, offset);
1973 spin_lock_irq(&info->reassembly_queue_lock);
1974 info->reassembly_data_length -= data_read;
1975 info->reassembly_queue_length -= queue_removed;
1976 spin_unlock_irq(&info->reassembly_queue_lock);
1978 info->first_entry_offset = offset;
1979 log_read(INFO, "returning to thread data_read=%d "
1980 "reassembly_data_length=%d first_entry_offset=%d\n",
1981 data_read, info->reassembly_data_length,
1982 info->first_entry_offset);
1987 log_read(INFO, "wait_event on more data\n");
1988 rc = wait_event_interruptible(
1989 info->wait_reassembly_queue,
1990 info->reassembly_data_length >= size ||
1991 info->transport_status != SMBD_CONNECTED);
1992 /* Don't return any data if interrupted */
2000 * Receive a page from receive reassembly queue
2001 * page: the page to read data into
2002 * to_read: the length of data to read
2003 * return value: actual data read
2005 static int smbd_recv_page(struct smbd_connection *info,
2006 struct page *page, unsigned int to_read)
2011 /* make sure we have the page ready for read */
2012 ret = wait_event_interruptible(
2013 info->wait_reassembly_queue,
2014 info->reassembly_data_length >= to_read ||
2015 info->transport_status != SMBD_CONNECTED);
2019 /* now we can read from reassembly queue and not sleep */
2020 to_address = kmap_atomic(page);
2022 log_read(INFO, "reading from page=%p address=%p to_read=%d\n",
2023 page, to_address, to_read);
2025 ret = smbd_recv_buf(info, to_address, to_read);
2026 kunmap_atomic(to_address);
2032 * Receive data from transport
2033 * msg: a msghdr point to the buffer, can be ITER_KVEC or ITER_BVEC
2034 * return: total bytes read, or 0. SMB Direct will not do partial read.
2036 int smbd_recv(struct smbd_connection *info, struct msghdr *msg)
2040 unsigned int to_read;
2043 info->smbd_recv_pending++;
2045 switch (msg->msg_iter.type) {
2046 case READ | ITER_KVEC:
2047 buf = msg->msg_iter.kvec->iov_base;
2048 to_read = msg->msg_iter.kvec->iov_len;
2049 rc = smbd_recv_buf(info, buf, to_read);
2052 case READ | ITER_BVEC:
2053 page = msg->msg_iter.bvec->bv_page;
2054 to_read = msg->msg_iter.bvec->bv_len;
2055 rc = smbd_recv_page(info, page, to_read);
2059 /* It's a bug in upper layer to get there */
2060 cifs_dbg(VFS, "CIFS: invalid msg type %d\n",
2061 msg->msg_iter.type);
2065 info->smbd_recv_pending--;
2066 wake_up(&info->wait_smbd_recv_pending);
2068 /* SMBDirect will read it all or nothing */
2070 msg->msg_iter.count = 0;
2075 * Send data to transport
2076 * Each rqst is transported as a SMBDirect payload
2077 * rqst: the data to write
2078 * return value: 0 if successfully write, otherwise error code
2080 int smbd_send(struct smbd_connection *info, struct smb_rqst *rqst)
2085 int buflen = 0, remaining_data_length;
2088 info->max_send_size - sizeof(struct smbd_data_transfer);
2092 info->smbd_send_pending++;
2093 if (info->transport_status != SMBD_CONNECTED) {
2099 * Skip the RFC1002 length defined in MS-SMB2 section 2.1
2100 * It is used only for TCP transport in the iov[0]
2101 * In future we may want to add a transport layer under protocol
2102 * layer so this will only be issued to TCP transport
2105 if (rqst->rq_iov[0].iov_len != 4) {
2106 log_write(ERR, "expected the pdu length in 1st iov, but got %zu\n", rqst->rq_iov[0].iov_len);
2109 iov = &rqst->rq_iov[1];
2111 /* total up iov array first */
2112 for (i = 0; i < rqst->rq_nvec-1; i++) {
2113 buflen += iov[i].iov_len;
2116 /* add in the page array if there is one */
2117 if (rqst->rq_npages) {
2118 buflen += rqst->rq_pagesz * (rqst->rq_npages - 1);
2119 buflen += rqst->rq_tailsz;
2122 if (buflen + sizeof(struct smbd_data_transfer) >
2123 info->max_fragmented_send_size) {
2124 log_write(ERR, "payload size %d > max size %d\n",
2125 buflen, info->max_fragmented_send_size);
2130 cifs_dbg(FYI, "Sending smb (RDMA): smb_len=%u\n", buflen);
2131 for (i = 0; i < rqst->rq_nvec-1; i++)
2132 dump_smb(iov[i].iov_base, iov[i].iov_len);
2134 remaining_data_length = buflen;
2136 log_write(INFO, "rqst->rq_nvec=%d rqst->rq_npages=%d rq_pagesz=%d "
2137 "rq_tailsz=%d buflen=%d\n",
2138 rqst->rq_nvec, rqst->rq_npages, rqst->rq_pagesz,
2139 rqst->rq_tailsz, buflen);
2141 start = i = iov[0].iov_len ? 0 : 1;
2144 buflen += iov[i].iov_len;
2145 if (buflen > max_iov_size) {
2147 remaining_data_length -=
2148 (buflen-iov[i].iov_len);
2149 log_write(INFO, "sending iov[] from start=%d "
2151 "remaining_data_length=%d\n",
2153 remaining_data_length);
2154 rc = smbd_post_send_data(
2155 info, &iov[start], i-start,
2156 remaining_data_length);
2160 /* iov[start] is too big, break it */
2161 nvecs = (buflen+max_iov_size-1)/max_iov_size;
2162 log_write(INFO, "iov[%d] iov_base=%p buflen=%d"
2163 " break to %d vectors\n",
2164 start, iov[start].iov_base,
2166 for (j = 0; j < nvecs; j++) {
2168 (char *)iov[start].iov_base +
2170 vec.iov_len = max_iov_size;
2174 max_iov_size*(nvecs-1);
2175 remaining_data_length -= vec.iov_len;
2177 "sending vec j=%d iov_base=%p"
2179 "remaining_data_length=%d\n",
2180 j, vec.iov_base, vec.iov_len,
2181 remaining_data_length);
2182 rc = smbd_post_send_data(
2184 remaining_data_length);
2189 if (i == rqst->rq_nvec-1)
2196 if (i == rqst->rq_nvec-1) {
2197 /* send out all remaining vecs */
2198 remaining_data_length -= buflen;
2200 "sending iov[] from start=%d i=%d "
2201 "nvecs=%d remaining_data_length=%d\n",
2203 remaining_data_length);
2204 rc = smbd_post_send_data(info, &iov[start],
2205 i-start, remaining_data_length);
2211 log_write(INFO, "looping i=%d buflen=%d\n", i, buflen);
2214 /* now sending pages if there are any */
2215 for (i = 0; i < rqst->rq_npages; i++) {
2216 buflen = (i == rqst->rq_npages-1) ?
2217 rqst->rq_tailsz : rqst->rq_pagesz;
2218 nvecs = (buflen + max_iov_size - 1) / max_iov_size;
2219 log_write(INFO, "sending pages buflen=%d nvecs=%d\n",
2221 for (j = 0; j < nvecs; j++) {
2222 size = max_iov_size;
2224 size = buflen - j*max_iov_size;
2225 remaining_data_length -= size;
2226 log_write(INFO, "sending pages i=%d offset=%d size=%d"
2227 " remaining_data_length=%d\n",
2228 i, j*max_iov_size, size, remaining_data_length);
2229 rc = smbd_post_send_page(
2230 info, rqst->rq_pages[i], j*max_iov_size,
2231 size, remaining_data_length);
2239 * As an optimization, we don't wait for individual I/O to finish
2240 * before sending the next one.
2241 * Send them all and wait for pending send count to get to 0
2242 * that means all the I/Os have been out and we are good to return
2245 wait_event(info->wait_send_payload_pending,
2246 atomic_read(&info->send_payload_pending) == 0);
2248 info->smbd_send_pending--;
2249 wake_up(&info->wait_smbd_send_pending);
2254 static void register_mr_done(struct ib_cq *cq, struct ib_wc *wc)
2260 log_rdma_mr(ERR, "status=%d\n", wc->status);
2262 mr = container_of(cqe, struct smbd_mr, cqe);
2263 smbd_disconnect_rdma_connection(mr->conn);
2268 * The work queue function that recovers MRs
2269 * We need to call ib_dereg_mr() and ib_alloc_mr() before this MR can be used
2270 * again. Both calls are slow, so finish them in a workqueue. This will not
2272 * There is one workqueue that recovers MRs, there is no need to lock as the
2273 * I/O requests calling smbd_register_mr will never update the links in the
2276 static void smbd_mr_recovery_work(struct work_struct *work)
2278 struct smbd_connection *info =
2279 container_of(work, struct smbd_connection, mr_recovery_work);
2280 struct smbd_mr *smbdirect_mr;
2283 list_for_each_entry(smbdirect_mr, &info->mr_list, list) {
2284 if (smbdirect_mr->state == MR_INVALIDATED ||
2285 smbdirect_mr->state == MR_ERROR) {
2287 if (smbdirect_mr->state == MR_INVALIDATED) {
2289 info->id->device, smbdirect_mr->sgl,
2290 smbdirect_mr->sgl_count,
2292 smbdirect_mr->state = MR_READY;
2293 } else if (smbdirect_mr->state == MR_ERROR) {
2295 /* recover this MR entry */
2296 rc = ib_dereg_mr(smbdirect_mr->mr);
2299 "ib_dereg_mr failed rc=%x\n",
2301 smbd_disconnect_rdma_connection(info);
2304 smbdirect_mr->mr = ib_alloc_mr(
2305 info->pd, info->mr_type,
2306 info->max_frmr_depth);
2307 if (IS_ERR(smbdirect_mr->mr)) {
2309 "ib_alloc_mr failed mr_type=%x "
2310 "max_frmr_depth=%x\n",
2312 info->max_frmr_depth);
2313 smbd_disconnect_rdma_connection(info);
2316 smbdirect_mr->state = MR_READY;
2318 /* smbdirect_mr->state is updated by this function
2319 * and is read and updated by I/O issuing CPUs trying
2320 * to get a MR, the call to atomic_inc_return
2321 * implicates a memory barrier and guarantees this
2322 * value is updated before waking up any calls to
2323 * get_mr() from the I/O issuing CPUs
2325 if (atomic_inc_return(&info->mr_ready_count) == 1)
2326 wake_up_interruptible(&info->wait_mr);
2331 static void destroy_mr_list(struct smbd_connection *info)
2333 struct smbd_mr *mr, *tmp;
2335 cancel_work_sync(&info->mr_recovery_work);
2336 list_for_each_entry_safe(mr, tmp, &info->mr_list, list) {
2337 if (mr->state == MR_INVALIDATED)
2338 ib_dma_unmap_sg(info->id->device, mr->sgl,
2339 mr->sgl_count, mr->dir);
2340 ib_dereg_mr(mr->mr);
2347 * Allocate MRs used for RDMA read/write
2348 * The number of MRs will not exceed hardware capability in responder_resources
2349 * All MRs are kept in mr_list. The MR can be recovered after it's used
2350 * Recovery is done in smbd_mr_recovery_work. The content of list entry changes
2351 * as MRs are used and recovered for I/O, but the list links will not change
2353 static int allocate_mr_list(struct smbd_connection *info)
2356 struct smbd_mr *smbdirect_mr, *tmp;
2358 INIT_LIST_HEAD(&info->mr_list);
2359 init_waitqueue_head(&info->wait_mr);
2360 spin_lock_init(&info->mr_list_lock);
2361 atomic_set(&info->mr_ready_count, 0);
2362 atomic_set(&info->mr_used_count, 0);
2363 init_waitqueue_head(&info->wait_for_mr_cleanup);
2364 /* Allocate more MRs (2x) than hardware responder_resources */
2365 for (i = 0; i < info->responder_resources * 2; i++) {
2366 smbdirect_mr = kzalloc(sizeof(*smbdirect_mr), GFP_KERNEL);
2369 smbdirect_mr->mr = ib_alloc_mr(info->pd, info->mr_type,
2370 info->max_frmr_depth);
2371 if (IS_ERR(smbdirect_mr->mr)) {
2372 log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x "
2373 "max_frmr_depth=%x\n",
2374 info->mr_type, info->max_frmr_depth);
2377 smbdirect_mr->sgl = kcalloc(
2378 info->max_frmr_depth,
2379 sizeof(struct scatterlist),
2381 if (!smbdirect_mr->sgl) {
2382 log_rdma_mr(ERR, "failed to allocate sgl\n");
2383 ib_dereg_mr(smbdirect_mr->mr);
2386 smbdirect_mr->state = MR_READY;
2387 smbdirect_mr->conn = info;
2389 list_add_tail(&smbdirect_mr->list, &info->mr_list);
2390 atomic_inc(&info->mr_ready_count);
2392 INIT_WORK(&info->mr_recovery_work, smbd_mr_recovery_work);
2396 kfree(smbdirect_mr);
2398 list_for_each_entry_safe(smbdirect_mr, tmp, &info->mr_list, list) {
2399 ib_dereg_mr(smbdirect_mr->mr);
2400 kfree(smbdirect_mr->sgl);
2401 kfree(smbdirect_mr);
2407 * Get a MR from mr_list. This function waits until there is at least one
2408 * MR available in the list. It may access the list while the
2409 * smbd_mr_recovery_work is recovering the MR list. This doesn't need a lock
2410 * as they never modify the same places. However, there may be several CPUs
2411 * issueing I/O trying to get MR at the same time, mr_list_lock is used to
2412 * protect this situation.
2414 static struct smbd_mr *get_mr(struct smbd_connection *info)
2416 struct smbd_mr *ret;
2419 rc = wait_event_interruptible(info->wait_mr,
2420 atomic_read(&info->mr_ready_count) ||
2421 info->transport_status != SMBD_CONNECTED);
2423 log_rdma_mr(ERR, "wait_event_interruptible rc=%x\n", rc);
2427 if (info->transport_status != SMBD_CONNECTED) {
2428 log_rdma_mr(ERR, "info->transport_status=%x\n",
2429 info->transport_status);
2433 spin_lock(&info->mr_list_lock);
2434 list_for_each_entry(ret, &info->mr_list, list) {
2435 if (ret->state == MR_READY) {
2436 ret->state = MR_REGISTERED;
2437 spin_unlock(&info->mr_list_lock);
2438 atomic_dec(&info->mr_ready_count);
2439 atomic_inc(&info->mr_used_count);
2444 spin_unlock(&info->mr_list_lock);
2446 * It is possible that we could fail to get MR because other processes may
2447 * try to acquire a MR at the same time. If this is the case, retry it.
2453 * Register memory for RDMA read/write
2454 * pages[]: the list of pages to register memory with
2455 * num_pages: the number of pages to register
2456 * tailsz: if non-zero, the bytes to register in the last page
2457 * writing: true if this is a RDMA write (SMB read), false for RDMA read
2458 * need_invalidate: true if this MR needs to be locally invalidated after I/O
2459 * return value: the MR registered, NULL if failed.
2461 struct smbd_mr *smbd_register_mr(
2462 struct smbd_connection *info, struct page *pages[], int num_pages,
2463 int tailsz, bool writing, bool need_invalidate)
2465 struct smbd_mr *smbdirect_mr;
2467 enum dma_data_direction dir;
2468 struct ib_reg_wr *reg_wr;
2469 struct ib_send_wr *bad_wr;
2471 if (num_pages > info->max_frmr_depth) {
2472 log_rdma_mr(ERR, "num_pages=%d max_frmr_depth=%d\n",
2473 num_pages, info->max_frmr_depth);
2477 smbdirect_mr = get_mr(info);
2478 if (!smbdirect_mr) {
2479 log_rdma_mr(ERR, "get_mr returning NULL\n");
2482 smbdirect_mr->need_invalidate = need_invalidate;
2483 smbdirect_mr->sgl_count = num_pages;
2484 sg_init_table(smbdirect_mr->sgl, num_pages);
2486 for (i = 0; i < num_pages - 1; i++)
2487 sg_set_page(&smbdirect_mr->sgl[i], pages[i], PAGE_SIZE, 0);
2489 sg_set_page(&smbdirect_mr->sgl[i], pages[i],
2490 tailsz ? tailsz : PAGE_SIZE, 0);
2492 dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
2493 smbdirect_mr->dir = dir;
2494 rc = ib_dma_map_sg(info->id->device, smbdirect_mr->sgl, num_pages, dir);
2496 log_rdma_mr(INFO, "ib_dma_map_sg num_pages=%x dir=%x rc=%x\n",
2497 num_pages, dir, rc);
2501 rc = ib_map_mr_sg(smbdirect_mr->mr, smbdirect_mr->sgl, num_pages,
2503 if (rc != num_pages) {
2505 "ib_map_mr_sg failed rc = %x num_pages = %x\n",
2510 ib_update_fast_reg_key(smbdirect_mr->mr,
2511 ib_inc_rkey(smbdirect_mr->mr->rkey));
2512 reg_wr = &smbdirect_mr->wr;
2513 reg_wr->wr.opcode = IB_WR_REG_MR;
2514 smbdirect_mr->cqe.done = register_mr_done;
2515 reg_wr->wr.wr_cqe = &smbdirect_mr->cqe;
2516 reg_wr->wr.num_sge = 0;
2517 reg_wr->wr.send_flags = IB_SEND_SIGNALED;
2518 reg_wr->mr = smbdirect_mr->mr;
2519 reg_wr->key = smbdirect_mr->mr->rkey;
2520 reg_wr->access = writing ?
2521 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
2522 IB_ACCESS_REMOTE_READ;
2525 * There is no need for waiting for complemtion on ib_post_send
2526 * on IB_WR_REG_MR. Hardware enforces a barrier and order of execution
2527 * on the next ib_post_send when we actaully send I/O to remote peer
2529 rc = ib_post_send(info->id->qp, ®_wr->wr, &bad_wr);
2531 return smbdirect_mr;
2533 log_rdma_mr(ERR, "ib_post_send failed rc=%x reg_wr->key=%x\n",
2536 /* If all failed, attempt to recover this MR by setting it MR_ERROR*/
2538 ib_dma_unmap_sg(info->id->device, smbdirect_mr->sgl,
2539 smbdirect_mr->sgl_count, smbdirect_mr->dir);
2542 smbdirect_mr->state = MR_ERROR;
2543 if (atomic_dec_and_test(&info->mr_used_count))
2544 wake_up(&info->wait_for_mr_cleanup);
2546 smbd_disconnect_rdma_connection(info);
2551 static void local_inv_done(struct ib_cq *cq, struct ib_wc *wc)
2553 struct smbd_mr *smbdirect_mr;
2557 smbdirect_mr = container_of(cqe, struct smbd_mr, cqe);
2558 smbdirect_mr->state = MR_INVALIDATED;
2559 if (wc->status != IB_WC_SUCCESS) {
2560 log_rdma_mr(ERR, "invalidate failed status=%x\n", wc->status);
2561 smbdirect_mr->state = MR_ERROR;
2563 complete(&smbdirect_mr->invalidate_done);
2567 * Deregister a MR after I/O is done
2568 * This function may wait if remote invalidation is not used
2569 * and we have to locally invalidate the buffer to prevent data is being
2570 * modified by remote peer after upper layer consumes it
2572 int smbd_deregister_mr(struct smbd_mr *smbdirect_mr)
2574 struct ib_send_wr *wr, *bad_wr;
2575 struct smbd_connection *info = smbdirect_mr->conn;
2578 if (smbdirect_mr->need_invalidate) {
2579 /* Need to finish local invalidation before returning */
2580 wr = &smbdirect_mr->inv_wr;
2581 wr->opcode = IB_WR_LOCAL_INV;
2582 smbdirect_mr->cqe.done = local_inv_done;
2583 wr->wr_cqe = &smbdirect_mr->cqe;
2585 wr->ex.invalidate_rkey = smbdirect_mr->mr->rkey;
2586 wr->send_flags = IB_SEND_SIGNALED;
2588 init_completion(&smbdirect_mr->invalidate_done);
2589 rc = ib_post_send(info->id->qp, wr, &bad_wr);
2591 log_rdma_mr(ERR, "ib_post_send failed rc=%x\n", rc);
2592 smbd_disconnect_rdma_connection(info);
2595 wait_for_completion(&smbdirect_mr->invalidate_done);
2596 smbdirect_mr->need_invalidate = false;
2599 * For remote invalidation, just set it to MR_INVALIDATED
2600 * and defer to mr_recovery_work to recover the MR for next use
2602 smbdirect_mr->state = MR_INVALIDATED;
2605 * Schedule the work to do MR recovery for future I/Os
2606 * MR recovery is slow and we don't want it to block the current I/O
2608 queue_work(info->workqueue, &info->mr_recovery_work);
2611 if (atomic_dec_and_test(&info->mr_used_count))
2612 wake_up(&info->wait_for_mr_cleanup);