Commit | Line | Data |
---|---|---|
c942fddf | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
03bee01d LL |
2 | /* |
3 | * Copyright (C) 2017, Microsoft Corporation. | |
4 | * | |
5 | * Author(s): Long Li <longli@microsoft.com> | |
03bee01d | 6 | */ |
f198186a | 7 | #include <linux/module.h> |
f64b78fd | 8 | #include <linux/highmem.h> |
03bee01d | 9 | #include "smbdirect.h" |
f198186a | 10 | #include "cifs_debug.h" |
b6903bcf | 11 | #include "cifsproto.h" |
35e2cc1b | 12 | #include "smb2proto.h" |
f198186a LL |
13 | |
14 | static struct smbd_response *get_empty_queue_buffer( | |
15 | struct smbd_connection *info); | |
16 | static struct smbd_response *get_receive_buffer( | |
17 | struct smbd_connection *info); | |
18 | static void put_receive_buffer( | |
19 | struct smbd_connection *info, | |
20 | struct smbd_response *response); | |
21 | static int allocate_receive_buffers(struct smbd_connection *info, int num_buf); | |
22 | static void destroy_receive_buffers(struct smbd_connection *info); | |
23 | ||
24 | static void put_empty_packet( | |
25 | struct smbd_connection *info, struct smbd_response *response); | |
26 | static void enqueue_reassembly( | |
27 | struct smbd_connection *info, | |
28 | struct smbd_response *response, int data_length); | |
29 | static struct smbd_response *_get_first_reassembly( | |
30 | struct smbd_connection *info); | |
31 | ||
32 | static int smbd_post_recv( | |
33 | struct smbd_connection *info, | |
34 | struct smbd_response *response); | |
35 | ||
36 | static int smbd_post_send_empty(struct smbd_connection *info); | |
d649e1bb LL |
37 | static int smbd_post_send_data( |
38 | struct smbd_connection *info, | |
39 | struct kvec *iov, int n_vec, int remaining_data_length); | |
40 | static int smbd_post_send_page(struct smbd_connection *info, | |
41 | struct page *page, unsigned long offset, | |
42 | size_t size, int remaining_data_length); | |
03bee01d | 43 | |
c7398583 LL |
44 | static void destroy_mr_list(struct smbd_connection *info); |
45 | static int allocate_mr_list(struct smbd_connection *info); | |
46 | ||
03bee01d LL |
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 */ | |
81 | int smbd_receive_credit_max = 255; | |
82 | ||
83 | /* The remote peer's credit request of local peer */ | |
84 | int smbd_send_credit_target = 255; | |
85 | ||
86 | /* The maximum single message size can be sent to remote peer */ | |
87 | int smbd_max_send_size = 1364; | |
88 | ||
89 | /* The maximum fragmented upper-layer payload receive size supported */ | |
90 | int smbd_max_fragmented_recv_size = 1024 * 1024; | |
91 | ||
92 | /* The maximum single-message size which can be received */ | |
3c62df55 | 93 | int smbd_max_receive_size = 1364; |
03bee01d LL |
94 | |
95 | /* The timeout to initiate send of a keepalive message on idle */ | |
96 | int 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 | */ | |
d2e81f92 | 102 | /* Default maximum number of pages in a single RDMA write/read */ |
03bee01d LL |
103 | int smbd_max_frmr_depth = 2048; |
104 | ||
105 | /* If payload is less than this byte, use RDMA send/recv not read/write */ | |
106 | int rdma_readwrite_threshold = 4096; | |
f198186a LL |
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 | |
123 | static unsigned int smbd_logging_class; | |
124 | module_param(smbd_logging_class, uint, 0644); | |
125 | MODULE_PARM_DESC(smbd_logging_class, | |
126 | "Logging class for SMBD transport 0x0 to 0x100"); | |
127 | ||
128 | #define ERR 0x0 | |
129 | #define INFO 0x1 | |
130 | static unsigned int smbd_logging_level = ERR; | |
131 | module_param(smbd_logging_level, uint, 0644); | |
132 | MODULE_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...) \ | |
136 | do { \ | |
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 | ||
f198186a LL |
158 | static 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 | ||
169 | static 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 */ | |
175 | static 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; | |
e8b3bfe9 | 224 | wake_up_interruptible(&info->disconn_wait); |
050b8c37 LL |
225 | wake_up_interruptible(&info->wait_reassembly_queue); |
226 | wake_up_interruptible_all(&info->wait_send_queue); | |
f198186a LL |
227 | break; |
228 | ||
229 | default: | |
230 | break; | |
231 | } | |
232 | ||
233 | return 0; | |
234 | } | |
235 | ||
236 | /* Upcall from RDMA QP */ | |
237 | static void | |
238 | smbd_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); | |
21ac58f4 | 249 | break; |
f198186a LL |
250 | |
251 | default: | |
252 | break; | |
253 | } | |
254 | } | |
255 | ||
256 | static inline void *smbd_request_payload(struct smbd_request *request) | |
257 | { | |
258 | return (void *)request->packet; | |
259 | } | |
260 | ||
261 | static 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 */ | |
267 | static 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 | ||
0350d7a3 | 273 | log_rdma_send(INFO, "smbd_request 0x%p completed wc->status=%d\n", |
f198186a LL |
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 | ||
072a14ec LL |
288 | if (atomic_dec_and_test(&request->info->send_pending)) |
289 | wake_up(&request->info->wait_send_pending); | |
290 | ||
3ffbe78a | 291 | wake_up(&request->info->wait_post_send); |
f198186a LL |
292 | |
293 | mempool_free(request, request->info->request_mempool); | |
294 | } | |
295 | ||
296 | static void dump_smbd_negotiate_resp(struct smbd_negotiate_resp *resp) | |
297 | { | |
a0a3036b JP |
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); | |
f198186a LL |
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 | */ | |
311 | static 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); | |
c7398583 LL |
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; | |
f198186a LL |
377 | |
378 | return true; | |
379 | } | |
380 | ||
f198186a LL |
381 | static 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 | ||
044b541c LL |
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 | } | |
f198186a LL |
440 | } |
441 | ||
f198186a LL |
442 | /* Called from softirq, when recv is done */ |
443 | static 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 | ||
0350d7a3 | 451 | log_rdma_recv(INFO, "response=0x%p type=%d wc status=%d wc opcode %d byte_len=%d pkey_index=%u\n", |
a0a3036b JP |
452 | response, response->type, wc->status, wc->opcode, |
453 | wc->byte_len, wc->pkey_index); | |
f198186a LL |
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); | |
4ebb8795 LL |
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 | } | |
f198186a | 518 | |
a0a3036b JP |
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)); | |
f198186a LL |
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 | ||
f198186a LL |
532 | return; |
533 | ||
534 | default: | |
535 | log_rdma_recv(ERR, | |
536 | "unexpected response type=%d\n", response->type); | |
537 | } | |
538 | ||
539 | error: | |
540 | put_receive_buffer(info, response); | |
541 | } | |
542 | ||
543 | static 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 | } | |
0555b221 | 575 | rc = wait_for_completion_interruptible_timeout( |
f198186a | 576 | &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT)); |
0555b221 SF |
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 | } | |
f198186a LL |
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 | } | |
0555b221 | 594 | rc = wait_for_completion_interruptible_timeout( |
f198186a | 595 | &info->ri_done, msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT)); |
0555b221 SF |
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 | } | |
f198186a LL |
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 | ||
609 | out: | |
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 | */ | |
619 | static 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 | ||
628 | static 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)) { | |
a0a3036b JP |
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); | |
f198186a LL |
645 | rc = -EPROTONOSUPPORT; |
646 | goto out2; | |
647 | } | |
c7398583 LL |
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; | |
e945c653 | 652 | if (info->id->device->attrs.kernel_cap_flags & IBK_SG_GAPS_REG) |
c7398583 | 653 | info->mr_type = IB_MR_TYPE_SG_GAPS; |
f198186a LL |
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 | ||
664 | out2: | |
665 | rdma_destroy_id(info->id); | |
666 | info->id = NULL; | |
667 | ||
668 | out1: | |
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 | */ | |
678 | static int smbd_post_send_negotiate_req(struct smbd_connection *info) | |
679 | { | |
73930595 | 680 | struct ib_send_wr send_wr; |
f198186a LL |
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 | ||
0350d7a3 | 726 | log_rdma_send(INFO, "sge addr=0x%llx length=%u lkey=0x%x\n", |
f198186a LL |
727 | request->sge[0].addr, |
728 | request->sge[0].length, request->sge[0].lkey); | |
729 | ||
f198186a | 730 | atomic_inc(&info->send_pending); |
73930595 | 731 | rc = ib_post_send(info->id->qp, &send_wr, NULL); |
f198186a LL |
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 | ||
21a4e14a LL |
741 | smbd_disconnect_rdma_connection(info); |
742 | ||
f198186a LL |
743 | dma_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 | */ | |
756 | static 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 | */ | |
777 | static 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 | ||
f1b7b862 LL |
786 | /* Post the send request */ |
787 | static 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, | |
0350d7a3 | 795 | "rdma_request sge[%d] addr=0x%llx length=%u\n", |
f1b7b862 LL |
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 | ||
826 | static int smbd_post_send_sgl(struct smbd_connection *info, | |
827 | struct scatterlist *sgl, int data_length, int remaining_data_length) | |
f198186a | 828 | { |
f1b7b862 LL |
829 | int num_sgs; |
830 | int i, rc; | |
831 | int header_length; | |
f198186a LL |
832 | struct smbd_request *request; |
833 | struct smbd_data_transfer *packet; | |
d4e5160d | 834 | int new_credits; |
f1b7b862 | 835 | struct scatterlist *sg; |
f198186a | 836 | |
f1b7b862 | 837 | wait_credit: |
f198186a LL |
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) | |
f1b7b862 | 843 | goto err_wait_credit; |
f198186a LL |
844 | |
845 | if (info->transport_status != SMBD_CONNECTED) { | |
f1b7b862 LL |
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 | ||
855 | wait_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; | |
f198186a | 870 | } |
f198186a LL |
871 | |
872 | request = mempool_alloc(info->request_mempool, GFP_KERNEL); | |
873 | if (!request) { | |
874 | rc = -ENOMEM; | |
d4e5160d | 875 | goto err_alloc; |
f198186a LL |
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); | |
d4e5160d LL |
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 | ||
f198186a LL |
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; | |
f1b7b862 | 895 | if (!data_length) |
f198186a LL |
896 | packet->data_offset = 0; |
897 | else | |
898 | packet->data_offset = cpu_to_le32(24); | |
f1b7b862 | 899 | packet->data_length = cpu_to_le32(data_length); |
f198186a LL |
900 | packet->remaining_data_length = cpu_to_le32(remaining_data_length); |
901 | packet->padding = 0; | |
902 | ||
a0a3036b JP |
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)); | |
f198186a LL |
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 */ | |
f1b7b862 | 913 | if (!data_length) |
f198186a LL |
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, | |
7f46d23e | 920 | DMA_TO_DEVICE); |
f198186a | 921 | if (ib_dma_mapping_error(info->id->device, request->sge[0].addr)) { |
f198186a | 922 | rc = -EIO; |
f1b7b862 | 923 | request->sge[0].addr = 0; |
d4e5160d | 924 | goto err_dma; |
f198186a LL |
925 | } |
926 | ||
927 | request->sge[0].length = header_length; | |
928 | request->sge[0].lkey = info->pd->local_dma_lkey; | |
929 | ||
f1b7b862 | 930 | /* Fill in the packet data payload */ |
f198186a LL |
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), | |
7f46d23e | 935 | sg->offset, sg->length, DMA_TO_DEVICE); |
f198186a LL |
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; | |
f1b7b862 | 940 | goto err_dma; |
f198186a LL |
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 | ||
072a14ec | 947 | rc = smbd_post_send(info, request); |
f198186a LL |
948 | if (!rc) |
949 | return 0; | |
950 | ||
f1b7b862 LL |
951 | err_dma: |
952 | for (i = 0; i < request->num_sge; i++) | |
f198186a LL |
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); | |
f1b7b862 LL |
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 | ||
966 | err_alloc: | |
967 | if (atomic_dec_and_test(&info->send_pending)) | |
968 | wake_up(&info->wait_send_pending); | |
969 | ||
970 | err_wait_send_queue: | |
971 | /* roll back send credits and pending */ | |
972 | atomic_inc(&info->send_credits); | |
973 | ||
974 | err_wait_credit: | |
f198186a LL |
975 | return rc; |
976 | } | |
977 | ||
d649e1bb LL |
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 | */ | |
985 | static 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 | ||
f198186a LL |
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 | */ | |
1001 | static 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 | ||
d649e1bb LL |
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 | */ | |
1014 | static 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; | |
d2e81f92 | 1020 | struct scatterlist sgl[SMBDIRECT_MAX_SEND_SGE - 1]; |
d649e1bb | 1021 | |
d2e81f92 | 1022 | if (n_vec > SMBDIRECT_MAX_SEND_SGE - 1) { |
d649e1bb | 1023 | cifs_dbg(VFS, "Can't fit data to SGL, n_vec=%d\n", n_vec); |
37941ea1 | 1024 | return -EINVAL; |
d649e1bb LL |
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 | ||
f198186a LL |
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 | */ | |
1041 | static int smbd_post_recv( | |
1042 | struct smbd_connection *info, struct smbd_response *response) | |
1043 | { | |
73930595 | 1044 | struct ib_recv_wr recv_wr; |
f198186a LL |
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 | ||
73930595 | 1063 | rc = ib_post_recv(info->id->qp, &recv_wr, NULL); |
f198186a LL |
1064 | if (rc) { |
1065 | ib_dma_unmap_single(info->id->device, response->sge.addr, | |
1066 | response->sge.length, DMA_FROM_DEVICE); | |
21a4e14a | 1067 | smbd_disconnect_rdma_connection(info); |
f198186a LL |
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 */ | |
1075 | static 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); | |
0350d7a3 | 1082 | log_rdma_event(INFO, "smbd_post_recv rc=%d iov.addr=0x%llx iov.length=%u iov.lkey=0x%x\n", |
a0a3036b JP |
1083 | rc, response->sge.addr, |
1084 | response->sge.length, response->sge.lkey); | |
f198186a LL |
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 | ||
1111 | static 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 | */ | |
1132 | static 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 | */ | |
1158 | static 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 | ||
1170 | static 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 | */ | |
1195 | static 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 | */ | |
1220 | static 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 */ | |
1238 | static 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 | ||
1270 | allocate_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 | ||
1283 | static 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 | ||
f198186a LL |
1294 | /* Implement idle connection timer [MS-SMBD] 3.1.6.2 */ |
1295 | static 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 | ||
050b8c37 LL |
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 | */ | |
1322 | void smbd_destroy(struct TCP_Server_Info *server) | |
8ef130f9 | 1323 | { |
050b8c37 LL |
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 | ||
8ef130f9 | 1333 | log_rdma_event(INFO, "destroying rdma session\n"); |
050b8c37 LL |
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"); | |
e8b3bfe9 | 1337 | wait_event_interruptible( |
050b8c37 LL |
1338 | info->disconn_wait, |
1339 | info->transport_status == SMBD_DISCONNECTED); | |
1340 | } | |
8ef130f9 | 1341 | |
050b8c37 LL |
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); | |
8ef130f9 | 1348 | |
050b8c37 LL |
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); | |
050b8c37 | 1352 | |
fb64f7f1 | 1353 | /* It's not possible for upper layer to get to reassembly */ |
050b8c37 LL |
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)) { | |
cc391b69 | 1385 | cifs_server_unlock(server); |
050b8c37 | 1386 | msleep(1000); |
cc391b69 | 1387 | cifs_server_lock(server); |
050b8c37 LL |
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; | |
8ef130f9 LL |
1404 | |
1405 | destroy_workqueue(info->workqueue); | |
d63cdbae | 1406 | log_rdma_event(INFO, "rdma session destroyed\n"); |
8ef130f9 LL |
1407 | kfree(info); |
1408 | } | |
1409 | ||
ad57b8e1 LL |
1410 | /* |
1411 | * Reconnect this SMBD connection, called from upper layer | |
1412 | * return value: 0 on success, or actual error code | |
1413 | */ | |
1414 | int smbd_reconnect(struct TCP_Server_Info *server) | |
1415 | { | |
1416 | log_rdma_event(INFO, "reconnecting rdma session\n"); | |
1417 | ||
1418 | if (!server->smbd_conn) { | |
48f238a7 LL |
1419 | log_rdma_event(INFO, "rdma session already destroyed\n"); |
1420 | goto create_conn; | |
ad57b8e1 LL |
1421 | } |
1422 | ||
1423 | /* | |
1424 | * This is possible if transport is disconnected and we haven't received | |
1425 | * notification from RDMA, but upper layer has detected timeout | |
1426 | */ | |
1427 | if (server->smbd_conn->transport_status == SMBD_CONNECTED) { | |
1428 | log_rdma_event(INFO, "disconnecting transport\n"); | |
050b8c37 | 1429 | smbd_destroy(server); |
ad57b8e1 LL |
1430 | } |
1431 | ||
48f238a7 | 1432 | create_conn: |
ad57b8e1 LL |
1433 | log_rdma_event(INFO, "creating rdma session\n"); |
1434 | server->smbd_conn = smbd_get_connection( | |
1435 | server, (struct sockaddr *) &server->dstaddr); | |
d63cdbae LL |
1436 | |
1437 | if (server->smbd_conn) | |
1438 | cifs_dbg(VFS, "RDMA transport re-established\n"); | |
ad57b8e1 LL |
1439 | |
1440 | return server->smbd_conn ? 0 : -ENOENT; | |
1441 | } | |
1442 | ||
f198186a LL |
1443 | static void destroy_caches_and_workqueue(struct smbd_connection *info) |
1444 | { | |
1445 | destroy_receive_buffers(info); | |
1446 | destroy_workqueue(info->workqueue); | |
1447 | mempool_destroy(info->response_mempool); | |
1448 | kmem_cache_destroy(info->response_cache); | |
1449 | mempool_destroy(info->request_mempool); | |
1450 | kmem_cache_destroy(info->request_cache); | |
1451 | } | |
1452 | ||
1453 | #define MAX_NAME_LEN 80 | |
1454 | static int allocate_caches_and_workqueue(struct smbd_connection *info) | |
1455 | { | |
1456 | char name[MAX_NAME_LEN]; | |
1457 | int rc; | |
1458 | ||
74ea5f98 | 1459 | scnprintf(name, MAX_NAME_LEN, "smbd_request_%p", info); |
f198186a LL |
1460 | info->request_cache = |
1461 | kmem_cache_create( | |
1462 | name, | |
1463 | sizeof(struct smbd_request) + | |
1464 | sizeof(struct smbd_data_transfer), | |
1465 | 0, SLAB_HWCACHE_ALIGN, NULL); | |
1466 | if (!info->request_cache) | |
1467 | return -ENOMEM; | |
1468 | ||
1469 | info->request_mempool = | |
1470 | mempool_create(info->send_credit_target, mempool_alloc_slab, | |
1471 | mempool_free_slab, info->request_cache); | |
1472 | if (!info->request_mempool) | |
1473 | goto out1; | |
1474 | ||
74ea5f98 | 1475 | scnprintf(name, MAX_NAME_LEN, "smbd_response_%p", info); |
f198186a LL |
1476 | info->response_cache = |
1477 | kmem_cache_create( | |
1478 | name, | |
1479 | sizeof(struct smbd_response) + | |
1480 | info->max_receive_size, | |
1481 | 0, SLAB_HWCACHE_ALIGN, NULL); | |
1482 | if (!info->response_cache) | |
1483 | goto out2; | |
1484 | ||
1485 | info->response_mempool = | |
1486 | mempool_create(info->receive_credit_max, mempool_alloc_slab, | |
1487 | mempool_free_slab, info->response_cache); | |
1488 | if (!info->response_mempool) | |
1489 | goto out3; | |
1490 | ||
74ea5f98 | 1491 | scnprintf(name, MAX_NAME_LEN, "smbd_%p", info); |
f198186a LL |
1492 | info->workqueue = create_workqueue(name); |
1493 | if (!info->workqueue) | |
1494 | goto out4; | |
1495 | ||
1496 | rc = allocate_receive_buffers(info, info->receive_credit_max); | |
1497 | if (rc) { | |
1498 | log_rdma_event(ERR, "failed to allocate receive buffers\n"); | |
1499 | goto out5; | |
1500 | } | |
1501 | ||
1502 | return 0; | |
1503 | ||
1504 | out5: | |
1505 | destroy_workqueue(info->workqueue); | |
1506 | out4: | |
1507 | mempool_destroy(info->response_mempool); | |
1508 | out3: | |
1509 | kmem_cache_destroy(info->response_cache); | |
1510 | out2: | |
1511 | mempool_destroy(info->request_mempool); | |
1512 | out1: | |
1513 | kmem_cache_destroy(info->request_cache); | |
1514 | return -ENOMEM; | |
1515 | } | |
1516 | ||
1517 | /* Create a SMBD connection, called by upper layer */ | |
9084432c | 1518 | static struct smbd_connection *_smbd_get_connection( |
f198186a LL |
1519 | struct TCP_Server_Info *server, struct sockaddr *dstaddr, int port) |
1520 | { | |
1521 | int rc; | |
1522 | struct smbd_connection *info; | |
1523 | struct rdma_conn_param conn_param; | |
1524 | struct ib_qp_init_attr qp_attr; | |
1525 | struct sockaddr_in *addr_in = (struct sockaddr_in *) dstaddr; | |
c7398583 LL |
1526 | struct ib_port_immutable port_immutable; |
1527 | u32 ird_ord_hdr[2]; | |
f198186a LL |
1528 | |
1529 | info = kzalloc(sizeof(struct smbd_connection), GFP_KERNEL); | |
1530 | if (!info) | |
1531 | return NULL; | |
1532 | ||
1533 | info->transport_status = SMBD_CONNECTING; | |
1534 | rc = smbd_ia_open(info, dstaddr, port); | |
1535 | if (rc) { | |
1536 | log_rdma_event(INFO, "smbd_ia_open rc=%d\n", rc); | |
1537 | goto create_id_failed; | |
1538 | } | |
1539 | ||
1540 | if (smbd_send_credit_target > info->id->device->attrs.max_cqe || | |
1541 | smbd_send_credit_target > info->id->device->attrs.max_qp_wr) { | |
0350d7a3 | 1542 | log_rdma_event(ERR, "consider lowering send_credit_target = %d. Possible CQE overrun, device reporting max_cqe %d max_qp_wr %d\n", |
a0a3036b JP |
1543 | smbd_send_credit_target, |
1544 | info->id->device->attrs.max_cqe, | |
1545 | info->id->device->attrs.max_qp_wr); | |
f198186a LL |
1546 | goto config_failed; |
1547 | } | |
1548 | ||
1549 | if (smbd_receive_credit_max > info->id->device->attrs.max_cqe || | |
1550 | smbd_receive_credit_max > info->id->device->attrs.max_qp_wr) { | |
0350d7a3 | 1551 | log_rdma_event(ERR, "consider lowering receive_credit_max = %d. Possible CQE overrun, device reporting max_cqe %d max_qp_wr %d\n", |
a0a3036b JP |
1552 | smbd_receive_credit_max, |
1553 | info->id->device->attrs.max_cqe, | |
1554 | info->id->device->attrs.max_qp_wr); | |
f198186a LL |
1555 | goto config_failed; |
1556 | } | |
1557 | ||
1558 | info->receive_credit_max = smbd_receive_credit_max; | |
1559 | info->send_credit_target = smbd_send_credit_target; | |
1560 | info->max_send_size = smbd_max_send_size; | |
1561 | info->max_fragmented_recv_size = smbd_max_fragmented_recv_size; | |
1562 | info->max_receive_size = smbd_max_receive_size; | |
1563 | info->keep_alive_interval = smbd_keep_alive_interval; | |
1564 | ||
d2e81f92 TT |
1565 | if (info->id->device->attrs.max_send_sge < SMBDIRECT_MAX_SEND_SGE || |
1566 | info->id->device->attrs.max_recv_sge < SMBDIRECT_MAX_RECV_SGE) { | |
33023fb8 | 1567 | log_rdma_event(ERR, |
d2e81f92 TT |
1568 | "device %.*s max_send_sge/max_recv_sge = %d/%d too small\n", |
1569 | IB_DEVICE_NAME_MAX, | |
1570 | info->id->device->name, | |
1571 | info->id->device->attrs.max_send_sge, | |
33023fb8 | 1572 | info->id->device->attrs.max_recv_sge); |
d2e81f92 | 1573 | goto config_failed; |
f198186a LL |
1574 | } |
1575 | ||
1576 | info->send_cq = NULL; | |
1577 | info->recv_cq = NULL; | |
20cf4e02 CL |
1578 | info->send_cq = |
1579 | ib_alloc_cq_any(info->id->device, info, | |
1580 | info->send_credit_target, IB_POLL_SOFTIRQ); | |
f198186a LL |
1581 | if (IS_ERR(info->send_cq)) { |
1582 | info->send_cq = NULL; | |
1583 | goto alloc_cq_failed; | |
1584 | } | |
1585 | ||
20cf4e02 CL |
1586 | info->recv_cq = |
1587 | ib_alloc_cq_any(info->id->device, info, | |
1588 | info->receive_credit_max, IB_POLL_SOFTIRQ); | |
f198186a LL |
1589 | if (IS_ERR(info->recv_cq)) { |
1590 | info->recv_cq = NULL; | |
1591 | goto alloc_cq_failed; | |
1592 | } | |
1593 | ||
1594 | memset(&qp_attr, 0, sizeof(qp_attr)); | |
1595 | qp_attr.event_handler = smbd_qp_async_error_upcall; | |
1596 | qp_attr.qp_context = info; | |
1597 | qp_attr.cap.max_send_wr = info->send_credit_target; | |
1598 | qp_attr.cap.max_recv_wr = info->receive_credit_max; | |
d2e81f92 TT |
1599 | qp_attr.cap.max_send_sge = SMBDIRECT_MAX_SEND_SGE; |
1600 | qp_attr.cap.max_recv_sge = SMBDIRECT_MAX_RECV_SGE; | |
f198186a LL |
1601 | qp_attr.cap.max_inline_data = 0; |
1602 | qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR; | |
1603 | qp_attr.qp_type = IB_QPT_RC; | |
1604 | qp_attr.send_cq = info->send_cq; | |
1605 | qp_attr.recv_cq = info->recv_cq; | |
1606 | qp_attr.port_num = ~0; | |
1607 | ||
1608 | rc = rdma_create_qp(info->id, info->pd, &qp_attr); | |
1609 | if (rc) { | |
1610 | log_rdma_event(ERR, "rdma_create_qp failed %i\n", rc); | |
1611 | goto create_qp_failed; | |
1612 | } | |
1613 | ||
1614 | memset(&conn_param, 0, sizeof(conn_param)); | |
1615 | conn_param.initiator_depth = 0; | |
1616 | ||
c7398583 LL |
1617 | conn_param.responder_resources = |
1618 | info->id->device->attrs.max_qp_rd_atom | |
1619 | < SMBD_CM_RESPONDER_RESOURCES ? | |
1620 | info->id->device->attrs.max_qp_rd_atom : | |
1621 | SMBD_CM_RESPONDER_RESOURCES; | |
1622 | info->responder_resources = conn_param.responder_resources; | |
1623 | log_rdma_mr(INFO, "responder_resources=%d\n", | |
1624 | info->responder_resources); | |
1625 | ||
1626 | /* Need to send IRD/ORD in private data for iWARP */ | |
3023a1e9 | 1627 | info->id->device->ops.get_port_immutable( |
c7398583 LL |
1628 | info->id->device, info->id->port_num, &port_immutable); |
1629 | if (port_immutable.core_cap_flags & RDMA_CORE_PORT_IWARP) { | |
1630 | ird_ord_hdr[0] = info->responder_resources; | |
1631 | ird_ord_hdr[1] = 1; | |
1632 | conn_param.private_data = ird_ord_hdr; | |
1633 | conn_param.private_data_len = sizeof(ird_ord_hdr); | |
1634 | } else { | |
1635 | conn_param.private_data = NULL; | |
1636 | conn_param.private_data_len = 0; | |
1637 | } | |
1638 | ||
f198186a LL |
1639 | conn_param.retry_count = SMBD_CM_RETRY; |
1640 | conn_param.rnr_retry_count = SMBD_CM_RNR_RETRY; | |
1641 | conn_param.flow_control = 0; | |
f198186a LL |
1642 | |
1643 | log_rdma_event(INFO, "connecting to IP %pI4 port %d\n", | |
1644 | &addr_in->sin_addr, port); | |
1645 | ||
1646 | init_waitqueue_head(&info->conn_wait); | |
050b8c37 LL |
1647 | init_waitqueue_head(&info->disconn_wait); |
1648 | init_waitqueue_head(&info->wait_reassembly_queue); | |
f198186a LL |
1649 | rc = rdma_connect(info->id, &conn_param); |
1650 | if (rc) { | |
1651 | log_rdma_event(ERR, "rdma_connect() failed with %i\n", rc); | |
1652 | goto rdma_connect_failed; | |
1653 | } | |
1654 | ||
1655 | wait_event_interruptible( | |
1656 | info->conn_wait, info->transport_status != SMBD_CONNECTING); | |
1657 | ||
1658 | if (info->transport_status != SMBD_CONNECTED) { | |
1659 | log_rdma_event(ERR, "rdma_connect failed port=%d\n", port); | |
1660 | goto rdma_connect_failed; | |
1661 | } | |
1662 | ||
1663 | log_rdma_event(INFO, "rdma_connect connected\n"); | |
1664 | ||
1665 | rc = allocate_caches_and_workqueue(info); | |
1666 | if (rc) { | |
1667 | log_rdma_event(ERR, "cache allocation failed\n"); | |
1668 | goto allocate_cache_failed; | |
1669 | } | |
1670 | ||
1671 | init_waitqueue_head(&info->wait_send_queue); | |
f198186a | 1672 | INIT_DELAYED_WORK(&info->idle_timer_work, idle_connection_timer); |
f198186a LL |
1673 | queue_delayed_work(info->workqueue, &info->idle_timer_work, |
1674 | info->keep_alive_interval*HZ); | |
1675 | ||
1676 | init_waitqueue_head(&info->wait_send_pending); | |
1677 | atomic_set(&info->send_pending, 0); | |
1678 | ||
3ffbe78a | 1679 | init_waitqueue_head(&info->wait_post_send); |
f198186a LL |
1680 | |
1681 | INIT_WORK(&info->disconnect_work, smbd_disconnect_rdma_work); | |
f198186a LL |
1682 | INIT_WORK(&info->post_send_credits_work, smbd_post_send_credits); |
1683 | info->new_credits_offered = 0; | |
1684 | spin_lock_init(&info->lock_new_credits_offered); | |
1685 | ||
1686 | rc = smbd_negotiate(info); | |
1687 | if (rc) { | |
1688 | log_rdma_event(ERR, "smbd_negotiate rc=%d\n", rc); | |
1689 | goto negotiation_failed; | |
1690 | } | |
1691 | ||
c7398583 LL |
1692 | rc = allocate_mr_list(info); |
1693 | if (rc) { | |
1694 | log_rdma_mr(ERR, "memory registration allocation failed\n"); | |
1695 | goto allocate_mr_failed; | |
1696 | } | |
1697 | ||
f198186a LL |
1698 | return info; |
1699 | ||
c7398583 LL |
1700 | allocate_mr_failed: |
1701 | /* At this point, need to a full transport shutdown */ | |
050b8c37 | 1702 | smbd_destroy(server); |
c7398583 LL |
1703 | return NULL; |
1704 | ||
f198186a LL |
1705 | negotiation_failed: |
1706 | cancel_delayed_work_sync(&info->idle_timer_work); | |
1707 | destroy_caches_and_workqueue(info); | |
1708 | info->transport_status = SMBD_NEGOTIATE_FAILED; | |
1709 | init_waitqueue_head(&info->conn_wait); | |
1710 | rdma_disconnect(info->id); | |
1711 | wait_event(info->conn_wait, | |
1712 | info->transport_status == SMBD_DISCONNECTED); | |
1713 | ||
1714 | allocate_cache_failed: | |
1715 | rdma_connect_failed: | |
1716 | rdma_destroy_qp(info->id); | |
1717 | ||
1718 | create_qp_failed: | |
1719 | alloc_cq_failed: | |
1720 | if (info->send_cq) | |
1721 | ib_free_cq(info->send_cq); | |
1722 | if (info->recv_cq) | |
1723 | ib_free_cq(info->recv_cq); | |
1724 | ||
1725 | config_failed: | |
1726 | ib_dealloc_pd(info->pd); | |
1727 | rdma_destroy_id(info->id); | |
1728 | ||
1729 | create_id_failed: | |
1730 | kfree(info); | |
1731 | return NULL; | |
1732 | } | |
399f9539 LL |
1733 | |
1734 | struct smbd_connection *smbd_get_connection( | |
1735 | struct TCP_Server_Info *server, struct sockaddr *dstaddr) | |
1736 | { | |
1737 | struct smbd_connection *ret; | |
1738 | int port = SMBD_PORT; | |
1739 | ||
1740 | try_again: | |
1741 | ret = _smbd_get_connection(server, dstaddr, port); | |
1742 | ||
1743 | /* Try SMB_PORT if SMBD_PORT doesn't work */ | |
1744 | if (!ret && port == SMBD_PORT) { | |
1745 | port = SMB_PORT; | |
1746 | goto try_again; | |
1747 | } | |
1748 | return ret; | |
1749 | } | |
f64b78fd LL |
1750 | |
1751 | /* | |
1752 | * Receive data from receive reassembly queue | |
1753 | * All the incoming data packets are placed in reassembly queue | |
1754 | * buf: the buffer to read data into | |
1755 | * size: the length of data to read | |
1756 | * return value: actual data read | |
1757 | * Note: this implementation copies the data from reassebmly queue to receive | |
1758 | * buffers used by upper layer. This is not the optimal code path. A better way | |
1759 | * to do it is to not have upper layer allocate its receive buffers but rather | |
1760 | * borrow the buffer from reassembly queue, and return it after data is | |
1761 | * consumed. But this will require more changes to upper layer code, and also | |
1762 | * need to consider packet boundaries while they still being reassembled. | |
1763 | */ | |
2026b06e SF |
1764 | static int smbd_recv_buf(struct smbd_connection *info, char *buf, |
1765 | unsigned int size) | |
f64b78fd LL |
1766 | { |
1767 | struct smbd_response *response; | |
1768 | struct smbd_data_transfer *data_transfer; | |
1769 | int to_copy, to_read, data_read, offset; | |
1770 | u32 data_length, remaining_data_length, data_offset; | |
1771 | int rc; | |
f64b78fd LL |
1772 | |
1773 | again: | |
f64b78fd LL |
1774 | /* |
1775 | * No need to hold the reassembly queue lock all the time as we are | |
1776 | * the only one reading from the front of the queue. The transport | |
1777 | * may add more entries to the back of the queue at the same time | |
1778 | */ | |
1779 | log_read(INFO, "size=%d info->reassembly_data_length=%d\n", size, | |
1780 | info->reassembly_data_length); | |
1781 | if (info->reassembly_data_length >= size) { | |
1782 | int queue_length; | |
1783 | int queue_removed = 0; | |
1784 | ||
1785 | /* | |
1786 | * Need to make sure reassembly_data_length is read before | |
1787 | * reading reassembly_queue_length and calling | |
1788 | * _get_first_reassembly. This call is lock free | |
1789 | * as we never read at the end of the queue which are being | |
1790 | * updated in SOFTIRQ as more data is received | |
1791 | */ | |
1792 | virt_rmb(); | |
1793 | queue_length = info->reassembly_queue_length; | |
1794 | data_read = 0; | |
1795 | to_read = size; | |
1796 | offset = info->first_entry_offset; | |
1797 | while (data_read < size) { | |
1798 | response = _get_first_reassembly(info); | |
1799 | data_transfer = smbd_response_payload(response); | |
1800 | data_length = le32_to_cpu(data_transfer->data_length); | |
1801 | remaining_data_length = | |
1802 | le32_to_cpu( | |
1803 | data_transfer->remaining_data_length); | |
1804 | data_offset = le32_to_cpu(data_transfer->data_offset); | |
1805 | ||
1806 | /* | |
1807 | * The upper layer expects RFC1002 length at the | |
1808 | * beginning of the payload. Return it to indicate | |
1809 | * the total length of the packet. This minimize the | |
1810 | * change to upper layer packet processing logic. This | |
1811 | * will be eventually remove when an intermediate | |
1812 | * transport layer is added | |
1813 | */ | |
1814 | if (response->first_segment && size == 4) { | |
1815 | unsigned int rfc1002_len = | |
1816 | data_length + remaining_data_length; | |
1817 | *((__be32 *)buf) = cpu_to_be32(rfc1002_len); | |
1818 | data_read = 4; | |
1819 | response->first_segment = false; | |
1820 | log_read(INFO, "returning rfc1002 length %d\n", | |
1821 | rfc1002_len); | |
1822 | goto read_rfc1002_done; | |
1823 | } | |
1824 | ||
1825 | to_copy = min_t(int, data_length - offset, to_read); | |
1826 | memcpy( | |
1827 | buf + data_read, | |
1828 | (char *)data_transfer + data_offset + offset, | |
1829 | to_copy); | |
1830 | ||
1831 | /* move on to the next buffer? */ | |
1832 | if (to_copy == data_length - offset) { | |
1833 | queue_length--; | |
1834 | /* | |
1835 | * No need to lock if we are not at the | |
1836 | * end of the queue | |
1837 | */ | |
f9de151b SF |
1838 | if (queue_length) |
1839 | list_del(&response->list); | |
1840 | else { | |
e36c048a AB |
1841 | spin_lock_irq( |
1842 | &info->reassembly_queue_lock); | |
f9de151b | 1843 | list_del(&response->list); |
e36c048a AB |
1844 | spin_unlock_irq( |
1845 | &info->reassembly_queue_lock); | |
f9de151b SF |
1846 | } |
1847 | queue_removed++; | |
f64b78fd LL |
1848 | info->count_reassembly_queue--; |
1849 | info->count_dequeue_reassembly_queue++; | |
1850 | put_receive_buffer(info, response); | |
1851 | offset = 0; | |
1852 | log_read(INFO, "put_receive_buffer offset=0\n"); | |
1853 | } else | |
1854 | offset += to_copy; | |
1855 | ||
1856 | to_read -= to_copy; | |
1857 | data_read += to_copy; | |
1858 | ||
a0a3036b JP |
1859 | 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", |
1860 | to_copy, data_length - offset, | |
1861 | to_read, data_read, offset); | |
f64b78fd LL |
1862 | } |
1863 | ||
e36c048a | 1864 | spin_lock_irq(&info->reassembly_queue_lock); |
f64b78fd LL |
1865 | info->reassembly_data_length -= data_read; |
1866 | info->reassembly_queue_length -= queue_removed; | |
e36c048a | 1867 | spin_unlock_irq(&info->reassembly_queue_lock); |
f64b78fd LL |
1868 | |
1869 | info->first_entry_offset = offset; | |
a0a3036b JP |
1870 | log_read(INFO, "returning to thread data_read=%d reassembly_data_length=%d first_entry_offset=%d\n", |
1871 | data_read, info->reassembly_data_length, | |
1872 | info->first_entry_offset); | |
f64b78fd LL |
1873 | read_rfc1002_done: |
1874 | return data_read; | |
1875 | } | |
1876 | ||
1877 | log_read(INFO, "wait_event on more data\n"); | |
1878 | rc = wait_event_interruptible( | |
1879 | info->wait_reassembly_queue, | |
1880 | info->reassembly_data_length >= size || | |
1881 | info->transport_status != SMBD_CONNECTED); | |
1882 | /* Don't return any data if interrupted */ | |
1883 | if (rc) | |
98e0d408 | 1884 | return rc; |
f64b78fd | 1885 | |
e8b3bfe9 LL |
1886 | if (info->transport_status != SMBD_CONNECTED) { |
1887 | log_read(ERR, "disconnected\n"); | |
acd4680e | 1888 | return -ECONNABORTED; |
e8b3bfe9 LL |
1889 | } |
1890 | ||
f64b78fd LL |
1891 | goto again; |
1892 | } | |
1893 | ||
1894 | /* | |
1895 | * Receive a page from receive reassembly queue | |
1896 | * page: the page to read data into | |
1897 | * to_read: the length of data to read | |
1898 | * return value: actual data read | |
1899 | */ | |
2026b06e | 1900 | static int smbd_recv_page(struct smbd_connection *info, |
6509f50c LL |
1901 | struct page *page, unsigned int page_offset, |
1902 | unsigned int to_read) | |
f64b78fd LL |
1903 | { |
1904 | int ret; | |
1905 | char *to_address; | |
6509f50c | 1906 | void *page_address; |
f64b78fd LL |
1907 | |
1908 | /* make sure we have the page ready for read */ | |
1909 | ret = wait_event_interruptible( | |
1910 | info->wait_reassembly_queue, | |
1911 | info->reassembly_data_length >= to_read || | |
1912 | info->transport_status != SMBD_CONNECTED); | |
1913 | if (ret) | |
6509f50c | 1914 | return ret; |
f64b78fd LL |
1915 | |
1916 | /* now we can read from reassembly queue and not sleep */ | |
6509f50c LL |
1917 | page_address = kmap_atomic(page); |
1918 | to_address = (char *) page_address + page_offset; | |
f64b78fd LL |
1919 | |
1920 | log_read(INFO, "reading from page=%p address=%p to_read=%d\n", | |
1921 | page, to_address, to_read); | |
1922 | ||
1923 | ret = smbd_recv_buf(info, to_address, to_read); | |
6509f50c | 1924 | kunmap_atomic(page_address); |
f64b78fd LL |
1925 | |
1926 | return ret; | |
1927 | } | |
1928 | ||
1929 | /* | |
1930 | * Receive data from transport | |
1931 | * msg: a msghdr point to the buffer, can be ITER_KVEC or ITER_BVEC | |
1932 | * return: total bytes read, or 0. SMB Direct will not do partial read. | |
1933 | */ | |
1934 | int smbd_recv(struct smbd_connection *info, struct msghdr *msg) | |
1935 | { | |
1936 | char *buf; | |
1937 | struct page *page; | |
6509f50c | 1938 | unsigned int to_read, page_offset; |
f64b78fd LL |
1939 | int rc; |
1940 | ||
00e23707 DH |
1941 | if (iov_iter_rw(&msg->msg_iter) == WRITE) { |
1942 | /* It's a bug in upper layer to get there */ | |
a0a3036b | 1943 | cifs_dbg(VFS, "Invalid msg iter dir %u\n", |
00e23707 DH |
1944 | iov_iter_rw(&msg->msg_iter)); |
1945 | rc = -EINVAL; | |
1946 | goto out; | |
1947 | } | |
1948 | ||
1949 | switch (iov_iter_type(&msg->msg_iter)) { | |
1950 | case ITER_KVEC: | |
f64b78fd LL |
1951 | buf = msg->msg_iter.kvec->iov_base; |
1952 | to_read = msg->msg_iter.kvec->iov_len; | |
1953 | rc = smbd_recv_buf(info, buf, to_read); | |
1954 | break; | |
1955 | ||
00e23707 | 1956 | case ITER_BVEC: |
f64b78fd | 1957 | page = msg->msg_iter.bvec->bv_page; |
6509f50c | 1958 | page_offset = msg->msg_iter.bvec->bv_offset; |
f64b78fd | 1959 | to_read = msg->msg_iter.bvec->bv_len; |
6509f50c | 1960 | rc = smbd_recv_page(info, page, page_offset, to_read); |
f64b78fd LL |
1961 | break; |
1962 | ||
1963 | default: | |
1964 | /* It's a bug in upper layer to get there */ | |
a0a3036b | 1965 | cifs_dbg(VFS, "Invalid msg type %d\n", |
00e23707 | 1966 | iov_iter_type(&msg->msg_iter)); |
6509f50c | 1967 | rc = -EINVAL; |
f64b78fd LL |
1968 | } |
1969 | ||
00e23707 | 1970 | out: |
f64b78fd LL |
1971 | /* SMBDirect will read it all or nothing */ |
1972 | if (rc > 0) | |
1973 | msg->msg_iter.count = 0; | |
1974 | return rc; | |
1975 | } | |
d649e1bb LL |
1976 | |
1977 | /* | |
1978 | * Send data to transport | |
1979 | * Each rqst is transported as a SMBDirect payload | |
1980 | * rqst: the data to write | |
1981 | * return value: 0 if successfully write, otherwise error code | |
1982 | */ | |
4739f232 LL |
1983 | int smbd_send(struct TCP_Server_Info *server, |
1984 | int num_rqst, struct smb_rqst *rqst_array) | |
d649e1bb | 1985 | { |
81f39f95 | 1986 | struct smbd_connection *info = server->smbd_conn; |
adeb964d | 1987 | struct kvec vecs[SMBDIRECT_MAX_SEND_SGE - 1]; |
d649e1bb LL |
1988 | int nvecs; |
1989 | int size; | |
35e2cc1b | 1990 | unsigned int buflen, remaining_data_length; |
adeb964d | 1991 | unsigned int offset, remaining_vec_data_length; |
d649e1bb LL |
1992 | int start, i, j; |
1993 | int max_iov_size = | |
1994 | info->max_send_size - sizeof(struct smbd_data_transfer); | |
8bcda1d2 | 1995 | struct kvec *iov; |
d649e1bb | 1996 | int rc; |
4739f232 LL |
1997 | struct smb_rqst *rqst; |
1998 | int rqst_idx; | |
d649e1bb | 1999 | |
adeb964d TT |
2000 | if (info->transport_status != SMBD_CONNECTED) |
2001 | return -EAGAIN; | |
d649e1bb | 2002 | |
b6903bcf LL |
2003 | /* |
2004 | * Add in the page array if there is one. The caller needs to set | |
2005 | * rq_tailsz to PAGE_SIZE when the buffer has multiple pages and | |
2006 | * ends at page boundary | |
2007 | */ | |
4739f232 LL |
2008 | remaining_data_length = 0; |
2009 | for (i = 0; i < num_rqst; i++) | |
2010 | remaining_data_length += smb_rqst_len(server, &rqst_array[i]); | |
d649e1bb | 2011 | |
adeb964d TT |
2012 | if (unlikely(remaining_data_length > info->max_fragmented_send_size)) { |
2013 | /* assertion: payload never exceeds negotiated maximum */ | |
d649e1bb | 2014 | log_write(ERR, "payload size %d > max size %d\n", |
4739f232 | 2015 | remaining_data_length, info->max_fragmented_send_size); |
adeb964d | 2016 | return -EINVAL; |
d649e1bb LL |
2017 | } |
2018 | ||
7f46d23e LL |
2019 | log_write(INFO, "num_rqst=%d total length=%u\n", |
2020 | num_rqst, remaining_data_length); | |
4739f232 | 2021 | |
7f46d23e | 2022 | rqst_idx = 0; |
adeb964d TT |
2023 | do { |
2024 | rqst = &rqst_array[rqst_idx]; | |
2025 | iov = rqst->rq_iov; | |
2026 | ||
2027 | cifs_dbg(FYI, "Sending smb (RDMA): idx=%d smb_len=%lu\n", | |
2028 | rqst_idx, smb_rqst_len(server, rqst)); | |
2029 | remaining_vec_data_length = 0; | |
2030 | for (i = 0; i < rqst->rq_nvec; i++) { | |
2031 | remaining_vec_data_length += iov[i].iov_len; | |
2032 | dump_smb(iov[i].iov_base, iov[i].iov_len); | |
2033 | } | |
2034 | ||
2035 | log_write(INFO, "rqst_idx=%d nvec=%d rqst->rq_npages=%d rq_pagesz=%d rq_tailsz=%d buflen=%lu\n", | |
2036 | rqst_idx, rqst->rq_nvec, | |
2037 | rqst->rq_npages, rqst->rq_pagesz, | |
2038 | rqst->rq_tailsz, smb_rqst_len(server, rqst)); | |
2039 | ||
2040 | start = 0; | |
2041 | offset = 0; | |
2042 | do { | |
2043 | buflen = 0; | |
2044 | i = start; | |
2045 | j = 0; | |
2046 | while (i < rqst->rq_nvec && | |
2047 | j < SMBDIRECT_MAX_SEND_SGE - 1 && | |
2048 | buflen < max_iov_size) { | |
2049 | ||
2050 | vecs[j].iov_base = iov[i].iov_base + offset; | |
2051 | if (buflen + iov[i].iov_len > max_iov_size) { | |
2052 | vecs[j].iov_len = | |
2053 | max_iov_size - iov[i].iov_len; | |
2054 | buflen = max_iov_size; | |
2055 | offset = vecs[j].iov_len; | |
2056 | } else { | |
2057 | vecs[j].iov_len = | |
2058 | iov[i].iov_len - offset; | |
2059 | buflen += vecs[j].iov_len; | |
2060 | offset = 0; | |
2061 | ++i; | |
d649e1bb | 2062 | } |
adeb964d | 2063 | ++j; |
d649e1bb | 2064 | } |
adeb964d TT |
2065 | |
2066 | remaining_vec_data_length -= buflen; | |
2067 | remaining_data_length -= buflen; | |
2068 | log_write(INFO, "sending %s iov[%d] from start=%d nvecs=%d remaining_data_length=%d\n", | |
2069 | remaining_vec_data_length > 0 ? | |
2070 | "partial" : "complete", | |
2071 | rqst->rq_nvec, start, j, | |
2072 | remaining_data_length); | |
2073 | ||
d649e1bb | 2074 | start = i; |
adeb964d TT |
2075 | rc = smbd_post_send_data(info, vecs, j, remaining_data_length); |
2076 | if (rc) | |
2077 | goto done; | |
2078 | } while (remaining_vec_data_length > 0); | |
2079 | ||
2080 | /* now sending pages if there are any */ | |
2081 | for (i = 0; i < rqst->rq_npages; i++) { | |
2082 | rqst_page_get_length(rqst, i, &buflen, &offset); | |
2083 | nvecs = (buflen + max_iov_size - 1) / max_iov_size; | |
2084 | log_write(INFO, "sending pages buflen=%d nvecs=%d\n", | |
2085 | buflen, nvecs); | |
2086 | for (j = 0; j < nvecs; j++) { | |
2087 | size = min_t(unsigned int, max_iov_size, remaining_data_length); | |
2088 | remaining_data_length -= size; | |
2089 | log_write(INFO, "sending pages i=%d offset=%d size=%d remaining_data_length=%d\n", | |
2090 | i, j * max_iov_size + offset, size, | |
a0a3036b | 2091 | remaining_data_length); |
adeb964d TT |
2092 | rc = smbd_post_send_page( |
2093 | info, rqst->rq_pages[i], | |
2094 | j*max_iov_size + offset, | |
2095 | size, remaining_data_length); | |
d649e1bb LL |
2096 | if (rc) |
2097 | goto done; | |
d649e1bb LL |
2098 | } |
2099 | } | |
adeb964d | 2100 | } while (++rqst_idx < num_rqst); |
4739f232 | 2101 | |
d649e1bb LL |
2102 | done: |
2103 | /* | |
2104 | * As an optimization, we don't wait for individual I/O to finish | |
2105 | * before sending the next one. | |
2106 | * Send them all and wait for pending send count to get to 0 | |
2107 | * that means all the I/Os have been out and we are good to return | |
2108 | */ | |
2109 | ||
072a14ec LL |
2110 | wait_event(info->wait_send_pending, |
2111 | atomic_read(&info->send_pending) == 0); | |
d649e1bb | 2112 | |
d649e1bb LL |
2113 | return rc; |
2114 | } | |
c7398583 LL |
2115 | |
2116 | static void register_mr_done(struct ib_cq *cq, struct ib_wc *wc) | |
2117 | { | |
2118 | struct smbd_mr *mr; | |
2119 | struct ib_cqe *cqe; | |
2120 | ||
2121 | if (wc->status) { | |
2122 | log_rdma_mr(ERR, "status=%d\n", wc->status); | |
2123 | cqe = wc->wr_cqe; | |
2124 | mr = container_of(cqe, struct smbd_mr, cqe); | |
2125 | smbd_disconnect_rdma_connection(mr->conn); | |
2126 | } | |
2127 | } | |
2128 | ||
2129 | /* | |
2130 | * The work queue function that recovers MRs | |
2131 | * We need to call ib_dereg_mr() and ib_alloc_mr() before this MR can be used | |
2132 | * again. Both calls are slow, so finish them in a workqueue. This will not | |
2133 | * block I/O path. | |
2134 | * There is one workqueue that recovers MRs, there is no need to lock as the | |
2135 | * I/O requests calling smbd_register_mr will never update the links in the | |
2136 | * mr_list. | |
2137 | */ | |
2138 | static void smbd_mr_recovery_work(struct work_struct *work) | |
2139 | { | |
2140 | struct smbd_connection *info = | |
2141 | container_of(work, struct smbd_connection, mr_recovery_work); | |
2142 | struct smbd_mr *smbdirect_mr; | |
2143 | int rc; | |
2144 | ||
2145 | list_for_each_entry(smbdirect_mr, &info->mr_list, list) { | |
c21ce58e | 2146 | if (smbdirect_mr->state == MR_ERROR) { |
c7398583 | 2147 | |
7cf20bce LL |
2148 | /* recover this MR entry */ |
2149 | rc = ib_dereg_mr(smbdirect_mr->mr); | |
2150 | if (rc) { | |
2151 | log_rdma_mr(ERR, | |
2152 | "ib_dereg_mr failed rc=%x\n", | |
2153 | rc); | |
2154 | smbd_disconnect_rdma_connection(info); | |
2155 | continue; | |
2156 | } | |
2157 | ||
2158 | smbdirect_mr->mr = ib_alloc_mr( | |
2159 | info->pd, info->mr_type, | |
2160 | info->max_frmr_depth); | |
2161 | if (IS_ERR(smbdirect_mr->mr)) { | |
a0a3036b JP |
2162 | log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x max_frmr_depth=%x\n", |
2163 | info->mr_type, | |
2164 | info->max_frmr_depth); | |
7cf20bce LL |
2165 | smbd_disconnect_rdma_connection(info); |
2166 | continue; | |
2167 | } | |
ff526d86 LL |
2168 | } else |
2169 | /* This MR is being used, don't recover it */ | |
2170 | continue; | |
7cf20bce | 2171 | |
ff526d86 | 2172 | smbdirect_mr->state = MR_READY; |
c7398583 | 2173 | |
ff526d86 LL |
2174 | /* smbdirect_mr->state is updated by this function |
2175 | * and is read and updated by I/O issuing CPUs trying | |
2176 | * to get a MR, the call to atomic_inc_return | |
2177 | * implicates a memory barrier and guarantees this | |
2178 | * value is updated before waking up any calls to | |
2179 | * get_mr() from the I/O issuing CPUs | |
2180 | */ | |
2181 | if (atomic_inc_return(&info->mr_ready_count) == 1) | |
2182 | wake_up_interruptible(&info->wait_mr); | |
c7398583 LL |
2183 | } |
2184 | } | |
2185 | ||
2186 | static void destroy_mr_list(struct smbd_connection *info) | |
2187 | { | |
2188 | struct smbd_mr *mr, *tmp; | |
2189 | ||
2190 | cancel_work_sync(&info->mr_recovery_work); | |
2191 | list_for_each_entry_safe(mr, tmp, &info->mr_list, list) { | |
2192 | if (mr->state == MR_INVALIDATED) | |
2193 | ib_dma_unmap_sg(info->id->device, mr->sgl, | |
2194 | mr->sgl_count, mr->dir); | |
2195 | ib_dereg_mr(mr->mr); | |
2196 | kfree(mr->sgl); | |
2197 | kfree(mr); | |
2198 | } | |
2199 | } | |
2200 | ||
2201 | /* | |
2202 | * Allocate MRs used for RDMA read/write | |
2203 | * The number of MRs will not exceed hardware capability in responder_resources | |
2204 | * All MRs are kept in mr_list. The MR can be recovered after it's used | |
2205 | * Recovery is done in smbd_mr_recovery_work. The content of list entry changes | |
2206 | * as MRs are used and recovered for I/O, but the list links will not change | |
2207 | */ | |
2208 | static int allocate_mr_list(struct smbd_connection *info) | |
2209 | { | |
2210 | int i; | |
2211 | struct smbd_mr *smbdirect_mr, *tmp; | |
2212 | ||
2213 | INIT_LIST_HEAD(&info->mr_list); | |
2214 | init_waitqueue_head(&info->wait_mr); | |
2215 | spin_lock_init(&info->mr_list_lock); | |
2216 | atomic_set(&info->mr_ready_count, 0); | |
2217 | atomic_set(&info->mr_used_count, 0); | |
2218 | init_waitqueue_head(&info->wait_for_mr_cleanup); | |
2219 | /* Allocate more MRs (2x) than hardware responder_resources */ | |
2220 | for (i = 0; i < info->responder_resources * 2; i++) { | |
2221 | smbdirect_mr = kzalloc(sizeof(*smbdirect_mr), GFP_KERNEL); | |
2222 | if (!smbdirect_mr) | |
2223 | goto out; | |
2224 | smbdirect_mr->mr = ib_alloc_mr(info->pd, info->mr_type, | |
2225 | info->max_frmr_depth); | |
2226 | if (IS_ERR(smbdirect_mr->mr)) { | |
a0a3036b JP |
2227 | log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x max_frmr_depth=%x\n", |
2228 | info->mr_type, info->max_frmr_depth); | |
c7398583 LL |
2229 | goto out; |
2230 | } | |
2231 | smbdirect_mr->sgl = kcalloc( | |
2232 | info->max_frmr_depth, | |
2233 | sizeof(struct scatterlist), | |
2234 | GFP_KERNEL); | |
2235 | if (!smbdirect_mr->sgl) { | |
2236 | log_rdma_mr(ERR, "failed to allocate sgl\n"); | |
2237 | ib_dereg_mr(smbdirect_mr->mr); | |
2238 | goto out; | |
2239 | } | |
2240 | smbdirect_mr->state = MR_READY; | |
2241 | smbdirect_mr->conn = info; | |
2242 | ||
2243 | list_add_tail(&smbdirect_mr->list, &info->mr_list); | |
2244 | atomic_inc(&info->mr_ready_count); | |
2245 | } | |
2246 | INIT_WORK(&info->mr_recovery_work, smbd_mr_recovery_work); | |
2247 | return 0; | |
2248 | ||
2249 | out: | |
2250 | kfree(smbdirect_mr); | |
2251 | ||
2252 | list_for_each_entry_safe(smbdirect_mr, tmp, &info->mr_list, list) { | |
2253 | ib_dereg_mr(smbdirect_mr->mr); | |
2254 | kfree(smbdirect_mr->sgl); | |
2255 | kfree(smbdirect_mr); | |
2256 | } | |
2257 | return -ENOMEM; | |
2258 | } | |
2259 | ||
2260 | /* | |
2261 | * Get a MR from mr_list. This function waits until there is at least one | |
2262 | * MR available in the list. It may access the list while the | |
2263 | * smbd_mr_recovery_work is recovering the MR list. This doesn't need a lock | |
2264 | * as they never modify the same places. However, there may be several CPUs | |
2265 | * issueing I/O trying to get MR at the same time, mr_list_lock is used to | |
2266 | * protect this situation. | |
2267 | */ | |
2268 | static struct smbd_mr *get_mr(struct smbd_connection *info) | |
2269 | { | |
2270 | struct smbd_mr *ret; | |
2271 | int rc; | |
2272 | again: | |
2273 | rc = wait_event_interruptible(info->wait_mr, | |
2274 | atomic_read(&info->mr_ready_count) || | |
2275 | info->transport_status != SMBD_CONNECTED); | |
2276 | if (rc) { | |
2277 | log_rdma_mr(ERR, "wait_event_interruptible rc=%x\n", rc); | |
2278 | return NULL; | |
2279 | } | |
2280 | ||
2281 | if (info->transport_status != SMBD_CONNECTED) { | |
2282 | log_rdma_mr(ERR, "info->transport_status=%x\n", | |
2283 | info->transport_status); | |
2284 | return NULL; | |
2285 | } | |
2286 | ||
2287 | spin_lock(&info->mr_list_lock); | |
2288 | list_for_each_entry(ret, &info->mr_list, list) { | |
2289 | if (ret->state == MR_READY) { | |
2290 | ret->state = MR_REGISTERED; | |
2291 | spin_unlock(&info->mr_list_lock); | |
2292 | atomic_dec(&info->mr_ready_count); | |
2293 | atomic_inc(&info->mr_used_count); | |
2294 | return ret; | |
2295 | } | |
2296 | } | |
2297 | ||
2298 | spin_unlock(&info->mr_list_lock); | |
2299 | /* | |
2300 | * It is possible that we could fail to get MR because other processes may | |
2301 | * try to acquire a MR at the same time. If this is the case, retry it. | |
2302 | */ | |
2303 | goto again; | |
2304 | } | |
2305 | ||
2306 | /* | |
2307 | * Register memory for RDMA read/write | |
2308 | * pages[]: the list of pages to register memory with | |
2309 | * num_pages: the number of pages to register | |
2310 | * tailsz: if non-zero, the bytes to register in the last page | |
2311 | * writing: true if this is a RDMA write (SMB read), false for RDMA read | |
2312 | * need_invalidate: true if this MR needs to be locally invalidated after I/O | |
2313 | * return value: the MR registered, NULL if failed. | |
2314 | */ | |
2315 | struct smbd_mr *smbd_register_mr( | |
2316 | struct smbd_connection *info, struct page *pages[], int num_pages, | |
7cf20bce | 2317 | int offset, int tailsz, bool writing, bool need_invalidate) |
c7398583 LL |
2318 | { |
2319 | struct smbd_mr *smbdirect_mr; | |
2320 | int rc, i; | |
2321 | enum dma_data_direction dir; | |
2322 | struct ib_reg_wr *reg_wr; | |
c7398583 LL |
2323 | |
2324 | if (num_pages > info->max_frmr_depth) { | |
2325 | log_rdma_mr(ERR, "num_pages=%d max_frmr_depth=%d\n", | |
2326 | num_pages, info->max_frmr_depth); | |
2327 | return NULL; | |
2328 | } | |
2329 | ||
2330 | smbdirect_mr = get_mr(info); | |
2331 | if (!smbdirect_mr) { | |
2332 | log_rdma_mr(ERR, "get_mr returning NULL\n"); | |
2333 | return NULL; | |
2334 | } | |
2335 | smbdirect_mr->need_invalidate = need_invalidate; | |
2336 | smbdirect_mr->sgl_count = num_pages; | |
2337 | sg_init_table(smbdirect_mr->sgl, num_pages); | |
2338 | ||
7cf20bce LL |
2339 | log_rdma_mr(INFO, "num_pages=0x%x offset=0x%x tailsz=0x%x\n", |
2340 | num_pages, offset, tailsz); | |
c7398583 | 2341 | |
7cf20bce LL |
2342 | if (num_pages == 1) { |
2343 | sg_set_page(&smbdirect_mr->sgl[0], pages[0], tailsz, offset); | |
2344 | goto skip_multiple_pages; | |
2345 | } | |
2346 | ||
2347 | /* We have at least two pages to register */ | |
2348 | sg_set_page( | |
2349 | &smbdirect_mr->sgl[0], pages[0], PAGE_SIZE - offset, offset); | |
2350 | i = 1; | |
2351 | while (i < num_pages - 1) { | |
2352 | sg_set_page(&smbdirect_mr->sgl[i], pages[i], PAGE_SIZE, 0); | |
2353 | i++; | |
2354 | } | |
c7398583 LL |
2355 | sg_set_page(&smbdirect_mr->sgl[i], pages[i], |
2356 | tailsz ? tailsz : PAGE_SIZE, 0); | |
2357 | ||
7cf20bce | 2358 | skip_multiple_pages: |
c7398583 LL |
2359 | dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE; |
2360 | smbdirect_mr->dir = dir; | |
2361 | rc = ib_dma_map_sg(info->id->device, smbdirect_mr->sgl, num_pages, dir); | |
2362 | if (!rc) { | |
7cf20bce | 2363 | log_rdma_mr(ERR, "ib_dma_map_sg num_pages=%x dir=%x rc=%x\n", |
c7398583 LL |
2364 | num_pages, dir, rc); |
2365 | goto dma_map_error; | |
2366 | } | |
2367 | ||
2368 | rc = ib_map_mr_sg(smbdirect_mr->mr, smbdirect_mr->sgl, num_pages, | |
2369 | NULL, PAGE_SIZE); | |
2370 | if (rc != num_pages) { | |
7cf20bce LL |
2371 | log_rdma_mr(ERR, |
2372 | "ib_map_mr_sg failed rc = %d num_pages = %x\n", | |
c7398583 LL |
2373 | rc, num_pages); |
2374 | goto map_mr_error; | |
2375 | } | |
2376 | ||
2377 | ib_update_fast_reg_key(smbdirect_mr->mr, | |
2378 | ib_inc_rkey(smbdirect_mr->mr->rkey)); | |
2379 | reg_wr = &smbdirect_mr->wr; | |
2380 | reg_wr->wr.opcode = IB_WR_REG_MR; | |
2381 | smbdirect_mr->cqe.done = register_mr_done; | |
2382 | reg_wr->wr.wr_cqe = &smbdirect_mr->cqe; | |
2383 | reg_wr->wr.num_sge = 0; | |
2384 | reg_wr->wr.send_flags = IB_SEND_SIGNALED; | |
2385 | reg_wr->mr = smbdirect_mr->mr; | |
2386 | reg_wr->key = smbdirect_mr->mr->rkey; | |
2387 | reg_wr->access = writing ? | |
2388 | IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE : | |
2389 | IB_ACCESS_REMOTE_READ; | |
2390 | ||
2391 | /* | |
2392 | * There is no need for waiting for complemtion on ib_post_send | |
2393 | * on IB_WR_REG_MR. Hardware enforces a barrier and order of execution | |
2394 | * on the next ib_post_send when we actaully send I/O to remote peer | |
2395 | */ | |
73930595 | 2396 | rc = ib_post_send(info->id->qp, ®_wr->wr, NULL); |
c7398583 LL |
2397 | if (!rc) |
2398 | return smbdirect_mr; | |
2399 | ||
2400 | log_rdma_mr(ERR, "ib_post_send failed rc=%x reg_wr->key=%x\n", | |
2401 | rc, reg_wr->key); | |
2402 | ||
2403 | /* If all failed, attempt to recover this MR by setting it MR_ERROR*/ | |
2404 | map_mr_error: | |
2405 | ib_dma_unmap_sg(info->id->device, smbdirect_mr->sgl, | |
2406 | smbdirect_mr->sgl_count, smbdirect_mr->dir); | |
2407 | ||
2408 | dma_map_error: | |
2409 | smbdirect_mr->state = MR_ERROR; | |
2410 | if (atomic_dec_and_test(&info->mr_used_count)) | |
2411 | wake_up(&info->wait_for_mr_cleanup); | |
2412 | ||
21a4e14a LL |
2413 | smbd_disconnect_rdma_connection(info); |
2414 | ||
c7398583 LL |
2415 | return NULL; |
2416 | } | |
2417 | ||
2418 | static void local_inv_done(struct ib_cq *cq, struct ib_wc *wc) | |
2419 | { | |
2420 | struct smbd_mr *smbdirect_mr; | |
2421 | struct ib_cqe *cqe; | |
2422 | ||
2423 | cqe = wc->wr_cqe; | |
2424 | smbdirect_mr = container_of(cqe, struct smbd_mr, cqe); | |
2425 | smbdirect_mr->state = MR_INVALIDATED; | |
2426 | if (wc->status != IB_WC_SUCCESS) { | |
2427 | log_rdma_mr(ERR, "invalidate failed status=%x\n", wc->status); | |
2428 | smbdirect_mr->state = MR_ERROR; | |
2429 | } | |
2430 | complete(&smbdirect_mr->invalidate_done); | |
2431 | } | |
2432 | ||
2433 | /* | |
2434 | * Deregister a MR after I/O is done | |
2435 | * This function may wait if remote invalidation is not used | |
2436 | * and we have to locally invalidate the buffer to prevent data is being | |
2437 | * modified by remote peer after upper layer consumes it | |
2438 | */ | |
2439 | int smbd_deregister_mr(struct smbd_mr *smbdirect_mr) | |
2440 | { | |
73930595 | 2441 | struct ib_send_wr *wr; |
c7398583 LL |
2442 | struct smbd_connection *info = smbdirect_mr->conn; |
2443 | int rc = 0; | |
2444 | ||
2445 | if (smbdirect_mr->need_invalidate) { | |
2446 | /* Need to finish local invalidation before returning */ | |
2447 | wr = &smbdirect_mr->inv_wr; | |
2448 | wr->opcode = IB_WR_LOCAL_INV; | |
2449 | smbdirect_mr->cqe.done = local_inv_done; | |
2450 | wr->wr_cqe = &smbdirect_mr->cqe; | |
2451 | wr->num_sge = 0; | |
2452 | wr->ex.invalidate_rkey = smbdirect_mr->mr->rkey; | |
2453 | wr->send_flags = IB_SEND_SIGNALED; | |
2454 | ||
2455 | init_completion(&smbdirect_mr->invalidate_done); | |
73930595 | 2456 | rc = ib_post_send(info->id->qp, wr, NULL); |
c7398583 LL |
2457 | if (rc) { |
2458 | log_rdma_mr(ERR, "ib_post_send failed rc=%x\n", rc); | |
2459 | smbd_disconnect_rdma_connection(info); | |
2460 | goto done; | |
2461 | } | |
2462 | wait_for_completion(&smbdirect_mr->invalidate_done); | |
2463 | smbdirect_mr->need_invalidate = false; | |
2464 | } else | |
2465 | /* | |
2466 | * For remote invalidation, just set it to MR_INVALIDATED | |
2467 | * and defer to mr_recovery_work to recover the MR for next use | |
2468 | */ | |
2469 | smbdirect_mr->state = MR_INVALIDATED; | |
2470 | ||
c21ce58e LL |
2471 | if (smbdirect_mr->state == MR_INVALIDATED) { |
2472 | ib_dma_unmap_sg( | |
2473 | info->id->device, smbdirect_mr->sgl, | |
2474 | smbdirect_mr->sgl_count, | |
2475 | smbdirect_mr->dir); | |
2476 | smbdirect_mr->state = MR_READY; | |
2477 | if (atomic_inc_return(&info->mr_ready_count) == 1) | |
2478 | wake_up_interruptible(&info->wait_mr); | |
2479 | } else | |
2480 | /* | |
2481 | * Schedule the work to do MR recovery for future I/Os MR | |
2482 | * recovery is slow and don't want it to block current I/O | |
2483 | */ | |
2484 | queue_work(info->workqueue, &info->mr_recovery_work); | |
c7398583 LL |
2485 | |
2486 | done: | |
2487 | if (atomic_dec_and_test(&info->mr_used_count)) | |
2488 | wake_up(&info->wait_for_mr_cleanup); | |
2489 | ||
2490 | return rc; | |
2491 | } |