1 /*
2 * iSER transport for the Open iSCSI Initiator & iSER transport internals
3 *
4 * Copyright (C) 2004 Dmitry Yusupov
5 * Copyright (C) 2004 Alex Aizman
6 * Copyright (C) 2005 Mike Christie
7 * based on code maintained by open-iscsi@googlegroups.com
8 *
9 * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
10 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved.
11 * Copyright (c) 2013-2014 Mellanox Technologies. All rights reserved.
12 *
13 * This software is available to you under a choice of one of two
14 * licenses. You may choose to be licensed under the terms of the GNU
15 * General Public License (GPL) Version 2, available from the file
16 * COPYING in the main directory of this source tree, or the
17 * OpenIB.org BSD license below:
18 *
19 * Redistribution and use in source and binary forms, with or
20 * without modification, are permitted provided that the following
21 * conditions are met:
22 *
23 * - Redistributions of source code must retain the above
24 * copyright notice, this list of conditions and the following
25 * disclaimer.
26 *
27 * - Redistributions in binary form must reproduce the above
28 * copyright notice, this list of conditions and the following
29 * disclaimer in the documentation and/or other materials
30 * provided with the distribution.
31 *
32 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
33 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
34 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
35 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
36 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
37 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
38 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
39 * SOFTWARE.
40 */
41 #ifndef __ISCSI_ISER_H__
42 #define __ISCSI_ISER_H__
43
44 #include <linux/types.h>
45 #include <linux/net.h>
46 #include <linux/printk.h>
47 #include <scsi/libiscsi.h>
48 #include <scsi/scsi_transport_iscsi.h>
49 #include <scsi/scsi_cmnd.h>
50 #include <scsi/scsi_device.h>
51 #include <scsi/iser.h>
52
53 #include <linux/interrupt.h>
54 #include <linux/wait.h>
55 #include <linux/sched.h>
56 #include <linux/list.h>
57 #include <linux/slab.h>
58 #include <linux/dma-mapping.h>
59 #include <linux/mutex.h>
60 #include <linux/mempool.h>
61 #include <linux/uio.h>
62
63 #include <linux/socket.h>
64 #include <linux/in.h>
65 #include <linux/in6.h>
66
67 #include <rdma/ib_verbs.h>
68 #include <rdma/rdma_cm.h>
69
70 #define DRV_NAME "iser"
71 #define PFX DRV_NAME ": "
72 #define DRV_VER "1.6"
73
74 #define iser_dbg(fmt, arg...) \
75 do { \
76 if (unlikely(iser_debug_level > 2)) \
77 printk(KERN_DEBUG PFX "%s: " fmt,\
78 __func__ , ## arg); \
79 } while (0)
80
81 #define iser_warn(fmt, arg...) \
82 do { \
83 if (unlikely(iser_debug_level > 0)) \
84 pr_warn(PFX "%s: " fmt, \
85 __func__ , ## arg); \
86 } while (0)
87
88 #define iser_info(fmt, arg...) \
89 do { \
90 if (unlikely(iser_debug_level > 1)) \
91 pr_info(PFX "%s: " fmt, \
92 __func__ , ## arg); \
93 } while (0)
94
95 #define iser_err(fmt, arg...) \
96 pr_err(PFX "%s: " fmt, __func__ , ## arg)
97
98 /* Default support is 512KB I/O size */
99 #define ISER_DEF_MAX_SECTORS 1024
100 #define ISCSI_ISER_DEF_SG_TABLESIZE \
101 ((ISER_DEF_MAX_SECTORS * SECTOR_SIZE) >> ilog2(SZ_4K))
102 /* Maximum support is 16MB I/O size */
103 #define ISCSI_ISER_MAX_SG_TABLESIZE ((32768 * SECTOR_SIZE) >> ilog2(SZ_4K))
104
105 #define ISER_DEF_XMIT_CMDS_DEFAULT 512
106 #if ISCSI_DEF_XMIT_CMDS_MAX > ISER_DEF_XMIT_CMDS_DEFAULT
107 #define ISER_DEF_XMIT_CMDS_MAX ISCSI_DEF_XMIT_CMDS_MAX
108 #else
109 #define ISER_DEF_XMIT_CMDS_MAX ISER_DEF_XMIT_CMDS_DEFAULT
110 #endif
111 #define ISER_DEF_CMD_PER_LUN ISER_DEF_XMIT_CMDS_MAX
112
113 /* QP settings */
114 /* Maximal bounds on received asynchronous PDUs */
115 #define ISER_MAX_RX_MISC_PDUS 4 /* NOOP_IN(2) , ASYNC_EVENT(2) */
116
117 #define ISER_MAX_TX_MISC_PDUS 6 /* NOOP_OUT(2), TEXT(1), *
118 * SCSI_TMFUNC(2), LOGOUT(1) */
119
120 #define ISER_QP_MAX_RECV_DTOS (ISER_DEF_XMIT_CMDS_MAX)
121
122 /* the max TX (send) WR supported by the iSER QP is defined by *
123 * max_send_wr = T * (1 + D) + C ; D is how many inflight dataouts we expect *
124 * to have at max for SCSI command. The tx posting & completion handling code *
125 * supports -EAGAIN scheme where tx is suspended till the QP has room for more *
126 * send WR. D=8 comes from 64K/8K */
127
128 #define ISER_INFLIGHT_DATAOUTS 8
129
130 #define ISER_QP_MAX_REQ_DTOS (ISER_DEF_XMIT_CMDS_MAX * \
131 (1 + ISER_INFLIGHT_DATAOUTS) + \
132 ISER_MAX_TX_MISC_PDUS + \
133 ISER_MAX_RX_MISC_PDUS)
134
135 /* Max registration work requests per command */
136 #define ISER_MAX_REG_WR_PER_CMD 5
137
138 /* For Signature we don't support DATAOUTs so no need to make room for them */
139 #define ISER_QP_SIG_MAX_REQ_DTOS (ISER_DEF_XMIT_CMDS_MAX * \
140 (1 + ISER_MAX_REG_WR_PER_CMD) + \
141 ISER_MAX_TX_MISC_PDUS + \
142 ISER_MAX_RX_MISC_PDUS)
143
144 #define ISER_GET_MAX_XMIT_CMDS(send_wr) ((send_wr \
145 - ISER_MAX_TX_MISC_PDUS \
146 - ISER_MAX_RX_MISC_PDUS) / \
147 (1 + ISER_INFLIGHT_DATAOUTS))
148
149 /* Constant PDU lengths calculations */
150 #define ISER_HEADERS_LEN (sizeof(struct iser_ctrl) + sizeof(struct iscsi_hdr))
151
152 #define ISER_RECV_DATA_SEG_LEN 128
153 #define ISER_RX_PAYLOAD_SIZE (ISER_HEADERS_LEN + ISER_RECV_DATA_SEG_LEN)
154 #define ISER_RX_LOGIN_SIZE (ISER_HEADERS_LEN + ISCSI_DEF_MAX_RECV_SEG_LEN)
155
156 /* Length of an object name string */
157 #define ISER_OBJECT_NAME_SIZE 64
158
159 enum iser_conn_state {
160 ISER_CONN_INIT, /* descriptor allocd, no conn */
161 ISER_CONN_PENDING, /* in the process of being established */
162 ISER_CONN_UP, /* up and running */
163 ISER_CONN_TERMINATING, /* in the process of being terminated */
164 ISER_CONN_DOWN, /* shut down */
165 ISER_CONN_STATES_NUM
166 };
167
168 enum iser_task_status {
169 ISER_TASK_STATUS_INIT = 0,
170 ISER_TASK_STATUS_STARTED,
171 ISER_TASK_STATUS_COMPLETED
172 };
173
174 enum iser_data_dir {
175 ISER_DIR_IN = 0, /* to initiator */
176 ISER_DIR_OUT, /* from initiator */
177 ISER_DIRS_NUM
178 };
179
180 /**
181 * struct iser_data_buf - iSER data buffer
182 *
183 * @sg: pointer to the sg list
184 * @size: num entries of this sg
185 * @data_len: total beffer byte len
186 * @dma_nents: returned by dma_map_sg
187 */
188 struct iser_data_buf {
189 struct scatterlist *sg;
190 int size;
191 unsigned long data_len;
192 int dma_nents;
193 };
194
195 /* fwd declarations */
196 struct iser_device;
197 struct iscsi_iser_task;
198 struct iscsi_endpoint;
199 struct iser_reg_resources;
200
201 /**
202 * struct iser_mem_reg - iSER memory registration info
203 *
204 * @sge: memory region sg element
205 * @rkey: memory region remote key
206 * @desc: pointer to fast registration context
207 */
208 struct iser_mem_reg {
209 struct ib_sge sge;
210 u32 rkey;
211 struct iser_fr_desc *desc;
212 };
213
214 enum iser_desc_type {
215 ISCSI_TX_CONTROL ,
216 ISCSI_TX_SCSI_COMMAND,
217 ISCSI_TX_DATAOUT
218 };
219
220 /**
221 * struct iser_tx_desc - iSER TX descriptor
222 *
223 * @iser_header: iser header
224 * @iscsi_header: iscsi header
225 * @type: command/control/dataout
226 * @dma_addr: header buffer dma_address
227 * @tx_sg: sg[0] points to iser/iscsi headers
228 * sg[1] optionally points to either of immediate data
229 * unsolicited data-out or control
230 * @num_sge: number sges used on this TX task
231 * @cqe: completion handler
232 * @mapped: Is the task header mapped
233 * @reg_wr: registration WR
234 * @send_wr: send WR
235 * @inv_wr: invalidate WR
236 */
237 struct iser_tx_desc {
238 struct iser_ctrl iser_header;
239 struct iscsi_hdr iscsi_header;
240 enum iser_desc_type type;
241 u64 dma_addr;
242 struct ib_sge tx_sg[2];
243 int num_sge;
244 struct ib_cqe cqe;
245 bool mapped;
246 struct ib_reg_wr reg_wr;
247 struct ib_send_wr send_wr;
248 struct ib_send_wr inv_wr;
249 };
250
251 #define ISER_RX_PAD_SIZE (256 - (ISER_RX_PAYLOAD_SIZE + \
252 sizeof(u64) + sizeof(struct ib_sge) + \
253 sizeof(struct ib_cqe)))
254 /**
255 * struct iser_rx_desc - iSER RX descriptor
256 *
257 * @iser_header: iser header
258 * @iscsi_header: iscsi header
259 * @data: received data segment
260 * @dma_addr: receive buffer dma address
261 * @rx_sg: ib_sge of receive buffer
262 * @cqe: completion handler
263 * @pad: for sense data TODO: Modify to maximum sense length supported
264 */
265 struct iser_rx_desc {
266 struct iser_ctrl iser_header;
267 struct iscsi_hdr iscsi_header;
268 char data[ISER_RECV_DATA_SEG_LEN];
269 u64 dma_addr;
270 struct ib_sge rx_sg;
271 struct ib_cqe cqe;
272 char pad[ISER_RX_PAD_SIZE];
273 } __packed;
274
275 /**
276 * struct iser_login_desc - iSER login descriptor
277 *
278 * @req: pointer to login request buffer
279 * @rsp: pointer to login response buffer
280 * @req_dma: DMA address of login request buffer
281 * @rsp_dma: DMA address of login response buffer
282 * @sge: IB sge for login post recv
283 * @cqe: completion handler
284 */
285 struct iser_login_desc {
286 void *req;
287 void *rsp;
288 u64 req_dma;
289 u64 rsp_dma;
290 struct ib_sge sge;
291 struct ib_cqe cqe;
292 } __packed;
293
294 struct iser_conn;
295 struct ib_conn;
296
297 /**
298 * struct iser_device - iSER device handle
299 *
300 * @ib_device: RDMA device
301 * @pd: Protection Domain for this device
302 * @mr: Global DMA memory region
303 * @event_handler: IB events handle routine
304 * @ig_list: entry in devices list
305 * @refcount: Reference counter, dominated by open iser connections
306 */
307 struct iser_device {
308 struct ib_device *ib_device;
309 struct ib_pd *pd;
310 struct ib_event_handler event_handler;
311 struct list_head ig_list;
312 int refcount;
313 };
314
315 /**
316 * struct iser_reg_resources - Fast registration resources
317 *
318 * @mr: memory region
319 * @sig_mr: signature memory region
320 * @mr_valid: is mr valid indicator
321 */
322 struct iser_reg_resources {
323 struct ib_mr *mr;
324 struct ib_mr *sig_mr;
325 u8 mr_valid:1;
326 };
327
328 /**
329 * struct iser_fr_desc - Fast registration descriptor
330 *
331 * @list: entry in connection fastreg pool
332 * @rsc: data buffer registration resources
333 * @sig_protected: is region protected indicator
334 * @all_list: first and last list members
335 */
336 struct iser_fr_desc {
337 struct list_head list;
338 struct iser_reg_resources rsc;
339 bool sig_protected;
340 struct list_head all_list;
341 };
342
343 /**
344 * struct iser_fr_pool - connection fast registration pool
345 *
346 * @list: list of fastreg descriptors
347 * @lock: protects fastreg pool
348 * @size: size of the pool
349 * @all_list: first and last list members
350 */
351 struct iser_fr_pool {
352 struct list_head list;
353 spinlock_t lock;
354 int size;
355 struct list_head all_list;
356 };
357
358 /**
359 * struct ib_conn - Infiniband related objects
360 *
361 * @cma_id: rdma_cm connection maneger handle
362 * @qp: Connection Queue-pair
363 * @cq: Connection completion queue
364 * @cq_size: The number of max outstanding completions
365 * @device: reference to iser device
366 * @fr_pool: connection fast registration pool
367 * @pi_support: Indicate device T10-PI support
368 * @reg_cqe: completion handler
369 */
370 struct ib_conn {
371 struct rdma_cm_id *cma_id;
372 struct ib_qp *qp;
373 struct ib_cq *cq;
374 u32 cq_size;
375 struct iser_device *device;
376 struct iser_fr_pool fr_pool;
377 bool pi_support;
378 struct ib_cqe reg_cqe;
379 };
380
381 /**
382 * struct iser_conn - iSER connection context
383 *
384 * @ib_conn: connection RDMA resources
385 * @iscsi_conn: link to matching iscsi connection
386 * @ep: transport handle
387 * @state: connection logical state
388 * @qp_max_recv_dtos: maximum number of data outs, corresponds
389 * to max number of post recvs
390 * @max_cmds: maximum cmds allowed for this connection
391 * @name: connection peer portal
392 * @release_work: deffered work for release job
393 * @state_mutex: protects iser onnection state
394 * @stop_completion: conn_stop completion
395 * @ib_completion: RDMA cleanup completion
396 * @up_completion: connection establishment completed
397 * (state is ISER_CONN_UP)
398 * @conn_list: entry in ig conn list
399 * @login_desc: login descriptor
400 * @rx_descs: rx buffers array (cyclic buffer)
401 * @num_rx_descs: number of rx descriptors
402 * @scsi_sg_tablesize: scsi host sg_tablesize
403 * @pages_per_mr: maximum pages available for registration
404 * @snd_w_inv: connection uses remote invalidation
405 */
406 struct iser_conn {
407 struct ib_conn ib_conn;
408 struct iscsi_conn *iscsi_conn;
409 struct iscsi_endpoint *ep;
410 enum iser_conn_state state;
411 unsigned qp_max_recv_dtos;
412 u16 max_cmds;
413 char name[ISER_OBJECT_NAME_SIZE];
414 struct work_struct release_work;
415 struct mutex state_mutex;
416 struct completion stop_completion;
417 struct completion ib_completion;
418 struct completion up_completion;
419 struct list_head conn_list;
420 struct iser_login_desc login_desc;
421 struct iser_rx_desc *rx_descs;
422 u32 num_rx_descs;
423 unsigned short scsi_sg_tablesize;
424 unsigned short pages_per_mr;
425 bool snd_w_inv;
426 };
427
428 /**
429 * struct iscsi_iser_task - iser task context
430 *
431 * @desc: TX descriptor
432 * @iser_conn: link to iser connection
433 * @status: current task status
434 * @sc: link to scsi command
435 * @command_sent: indicate if command was sent
436 * @dir: iser data direction
437 * @rdma_reg: task rdma registration desc
438 * @data: iser data buffer desc
439 * @prot: iser protection buffer desc
440 */
441 struct iscsi_iser_task {
442 struct iser_tx_desc desc;
443 struct iser_conn *iser_conn;
444 enum iser_task_status status;
445 struct scsi_cmnd *sc;
446 int command_sent;
447 int dir[ISER_DIRS_NUM];
448 struct iser_mem_reg rdma_reg[ISER_DIRS_NUM];
449 struct iser_data_buf data[ISER_DIRS_NUM];
450 struct iser_data_buf prot[ISER_DIRS_NUM];
451 };
452
453 /**
454 * struct iser_global - iSER global context
455 *
456 * @device_list_mutex: protects device_list
457 * @device_list: iser devices global list
458 * @connlist_mutex: protects connlist
459 * @connlist: iser connections global list
460 * @desc_cache: kmem cache for tx dataout
461 */
462 struct iser_global {
463 struct mutex device_list_mutex;
464 struct list_head device_list;
465 struct mutex connlist_mutex;
466 struct list_head connlist;
467 struct kmem_cache *desc_cache;
468 };
469
470 extern struct iser_global ig;
471 extern int iser_debug_level;
472 extern bool iser_pi_enable;
473 extern unsigned int iser_max_sectors;
474 extern bool iser_always_reg;
475
476 int iser_send_control(struct iscsi_conn *conn,
477 struct iscsi_task *task);
478
479 int iser_send_command(struct iscsi_conn *conn,
480 struct iscsi_task *task);
481
482 int iser_send_data_out(struct iscsi_conn *conn,
483 struct iscsi_task *task,
484 struct iscsi_data *hdr);
485
486 void iscsi_iser_recv(struct iscsi_conn *conn,
487 struct iscsi_hdr *hdr,
488 char *rx_data,
489 int rx_data_len);
490
491 void iser_conn_init(struct iser_conn *iser_conn);
492
493 void iser_conn_release(struct iser_conn *iser_conn);
494
495 int iser_conn_terminate(struct iser_conn *iser_conn);
496
497 void iser_release_work(struct work_struct *work);
498
499 void iser_err_comp(struct ib_wc *wc, const char *type);
500 void iser_login_rsp(struct ib_cq *cq, struct ib_wc *wc);
501 void iser_task_rsp(struct ib_cq *cq, struct ib_wc *wc);
502 void iser_cmd_comp(struct ib_cq *cq, struct ib_wc *wc);
503 void iser_ctrl_comp(struct ib_cq *cq, struct ib_wc *wc);
504 void iser_dataout_comp(struct ib_cq *cq, struct ib_wc *wc);
505 void iser_reg_comp(struct ib_cq *cq, struct ib_wc *wc);
506
507 void iser_task_rdma_init(struct iscsi_iser_task *task);
508
509 void iser_task_rdma_finalize(struct iscsi_iser_task *task);
510
511 void iser_free_rx_descriptors(struct iser_conn *iser_conn);
512
513 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
514 struct iser_data_buf *mem,
515 enum iser_data_dir cmd_dir);
516
517 int iser_reg_mem_fastreg(struct iscsi_iser_task *task,
518 enum iser_data_dir dir,
519 bool all_imm);
520 void iser_unreg_mem_fastreg(struct iscsi_iser_task *task,
521 enum iser_data_dir dir);
522
523 int iser_connect(struct iser_conn *iser_conn,
524 struct sockaddr *src_addr,
525 struct sockaddr *dst_addr,
526 int non_blocking);
527
528 int iser_post_recvl(struct iser_conn *iser_conn);
529 int iser_post_recvm(struct iser_conn *iser_conn,
530 struct iser_rx_desc *rx_desc);
531 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc);
532
533 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
534 enum iser_data_dir iser_dir,
535 enum dma_data_direction dma_dir);
536
537 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task,
538 enum iser_data_dir iser_dir,
539 enum dma_data_direction dma_dir);
540
541 int iser_initialize_task_headers(struct iscsi_task *task,
542 struct iser_tx_desc *tx_desc);
543 int iser_alloc_rx_descriptors(struct iser_conn *iser_conn,
544 struct iscsi_session *session);
545 int iser_alloc_fastreg_pool(struct ib_conn *ib_conn,
546 unsigned cmds_max,
547 unsigned int size);
548 void iser_free_fastreg_pool(struct ib_conn *ib_conn);
549 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
550 enum iser_data_dir cmd_dir, sector_t *sector);
551
552 static inline struct iser_conn *
to_iser_conn(struct ib_conn * ib_conn)553 to_iser_conn(struct ib_conn *ib_conn)
554 {
555 return container_of(ib_conn, struct iser_conn, ib_conn);
556 }
557
558 static inline struct iser_rx_desc *
iser_rx(struct ib_cqe * cqe)559 iser_rx(struct ib_cqe *cqe)
560 {
561 return container_of(cqe, struct iser_rx_desc, cqe);
562 }
563
564 static inline struct iser_tx_desc *
iser_tx(struct ib_cqe * cqe)565 iser_tx(struct ib_cqe *cqe)
566 {
567 return container_of(cqe, struct iser_tx_desc, cqe);
568 }
569
570 static inline struct iser_login_desc *
iser_login(struct ib_cqe * cqe)571 iser_login(struct ib_cqe *cqe)
572 {
573 return container_of(cqe, struct iser_login_desc, cqe);
574 }
575
576 #endif
577