2 * iSER transport for the Open iSCSI Initiator & iSER transport internals
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
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.
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:
19 * Redistribution and use in source and binary forms, with or
20 * without modification, are permitted provided that the following
23 * - Redistributions of source code must retain the above
24 * copyright notice, this list of conditions and the following
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.
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
41 #ifndef __ISCSI_ISER_H__
42 #define __ISCSI_ISER_H__
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>
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>
63 #include <linux/socket.h>
65 #include <linux/in6.h>
67 #include <rdma/ib_verbs.h>
68 #include <rdma/ib_fmr_pool.h>
69 #include <rdma/rdma_cm.h>
71 #define DRV_NAME "iser"
72 #define PFX DRV_NAME ": "
75 #define iser_dbg(fmt, arg...) \
77 if (unlikely(iser_debug_level > 2)) \
78 printk(KERN_DEBUG PFX "%s: " fmt,\
82 #define iser_warn(fmt, arg...) \
84 if (unlikely(iser_debug_level > 0)) \
85 pr_warn(PFX "%s: " fmt, \
89 #define iser_info(fmt, arg...) \
91 if (unlikely(iser_debug_level > 1)) \
92 pr_info(PFX "%s: " fmt, \
96 #define iser_err(fmt, arg...) \
97 pr_err(PFX "%s: " fmt, __func__ , ## arg)
100 #define SIZE_4K (1ULL << SHIFT_4K)
101 #define MASK_4K (~(SIZE_4K-1))
103 /* Default support is 512KB I/O size */
104 #define ISER_DEF_MAX_SECTORS 1024
105 #define ISCSI_ISER_DEF_SG_TABLESIZE \
106 ((ISER_DEF_MAX_SECTORS * SECTOR_SIZE) >> SHIFT_4K)
107 /* Maximum support is 16MB I/O size */
108 #define ISCSI_ISER_MAX_SG_TABLESIZE ((32768 * SECTOR_SIZE) >> SHIFT_4K)
110 #define ISER_DEF_XMIT_CMDS_DEFAULT 512
111 #if ISCSI_DEF_XMIT_CMDS_MAX > ISER_DEF_XMIT_CMDS_DEFAULT
112 #define ISER_DEF_XMIT_CMDS_MAX ISCSI_DEF_XMIT_CMDS_MAX
114 #define ISER_DEF_XMIT_CMDS_MAX ISER_DEF_XMIT_CMDS_DEFAULT
116 #define ISER_DEF_CMD_PER_LUN ISER_DEF_XMIT_CMDS_MAX
119 /* Maximal bounds on received asynchronous PDUs */
120 #define ISER_MAX_RX_MISC_PDUS 4 /* NOOP_IN(2) , ASYNC_EVENT(2) */
122 #define ISER_MAX_TX_MISC_PDUS 6 /* NOOP_OUT(2), TEXT(1), *
123 * SCSI_TMFUNC(2), LOGOUT(1) */
125 #define ISER_QP_MAX_RECV_DTOS (ISER_DEF_XMIT_CMDS_MAX)
127 #define ISER_MIN_POSTED_RX (ISER_DEF_XMIT_CMDS_MAX >> 2)
129 /* the max TX (send) WR supported by the iSER QP is defined by *
130 * max_send_wr = T * (1 + D) + C ; D is how many inflight dataouts we expect *
131 * to have at max for SCSI command. The tx posting & completion handling code *
132 * supports -EAGAIN scheme where tx is suspended till the QP has room for more *
133 * send WR. D=8 comes from 64K/8K */
135 #define ISER_INFLIGHT_DATAOUTS 8
137 #define ISER_QP_MAX_REQ_DTOS (ISER_DEF_XMIT_CMDS_MAX * \
138 (1 + ISER_INFLIGHT_DATAOUTS) + \
139 ISER_MAX_TX_MISC_PDUS + \
140 ISER_MAX_RX_MISC_PDUS)
142 /* Max registration work requests per command */
143 #define ISER_MAX_REG_WR_PER_CMD 5
145 /* For Signature we don't support DATAOUTs so no need to make room for them */
146 #define ISER_QP_SIG_MAX_REQ_DTOS (ISER_DEF_XMIT_CMDS_MAX * \
147 (1 + ISER_MAX_REG_WR_PER_CMD) + \
148 ISER_MAX_TX_MISC_PDUS + \
149 ISER_MAX_RX_MISC_PDUS)
151 #define ISER_GET_MAX_XMIT_CMDS(send_wr) ((send_wr \
152 - ISER_MAX_TX_MISC_PDUS \
153 - ISER_MAX_RX_MISC_PDUS) / \
154 (1 + ISER_INFLIGHT_DATAOUTS))
156 #define ISER_SIGNAL_CMD_COUNT 32
158 /* Constant PDU lengths calculations */
159 #define ISER_HEADERS_LEN (sizeof(struct iser_ctrl) + sizeof(struct iscsi_hdr))
161 #define ISER_RECV_DATA_SEG_LEN 128
162 #define ISER_RX_PAYLOAD_SIZE (ISER_HEADERS_LEN + ISER_RECV_DATA_SEG_LEN)
163 #define ISER_RX_LOGIN_SIZE (ISER_HEADERS_LEN + ISCSI_DEF_MAX_RECV_SEG_LEN)
165 /* Length of an object name string */
166 #define ISER_OBJECT_NAME_SIZE 64
168 enum iser_conn_state {
169 ISER_CONN_INIT, /* descriptor allocd, no conn */
170 ISER_CONN_PENDING, /* in the process of being established */
171 ISER_CONN_UP, /* up and running */
172 ISER_CONN_TERMINATING, /* in the process of being terminated */
173 ISER_CONN_DOWN, /* shut down */
177 enum iser_task_status {
178 ISER_TASK_STATUS_INIT = 0,
179 ISER_TASK_STATUS_STARTED,
180 ISER_TASK_STATUS_COMPLETED
184 ISER_DIR_IN = 0, /* to initiator */
185 ISER_DIR_OUT, /* from initiator */
190 * struct iser_data_buf - iSER data buffer
192 * @sg: pointer to the sg list
193 * @size: num entries of this sg
194 * @data_len: total beffer byte len
195 * @dma_nents: returned by dma_map_sg
197 struct iser_data_buf {
198 struct scatterlist *sg;
200 unsigned long data_len;
204 /* fwd declarations */
206 struct iscsi_iser_task;
207 struct iscsi_endpoint;
208 struct iser_reg_resources;
211 * struct iser_mem_reg - iSER memory registration info
213 * @sge: memory region sg element
214 * @rkey: memory region remote key
215 * @mem_h: pointer to registration context (FMR/Fastreg)
217 struct iser_mem_reg {
223 enum iser_desc_type {
225 ISCSI_TX_SCSI_COMMAND,
230 * struct iser_tx_desc - iSER TX descriptor
232 * @iser_header: iser header
233 * @iscsi_header: iscsi header
234 * @type: command/control/dataout
235 * @dam_addr: header buffer dma_address
236 * @tx_sg: sg[0] points to iser/iscsi headers
237 * sg[1] optionally points to either of immediate data
238 * unsolicited data-out or control
239 * @num_sge: number sges used on this TX task
240 * @mapped: Is the task header mapped
241 * reg_wr: registration WR
243 * inv_wr: invalidate WR
245 struct iser_tx_desc {
246 struct iser_ctrl iser_header;
247 struct iscsi_hdr iscsi_header;
248 enum iser_desc_type type;
250 struct ib_sge tx_sg[2];
254 struct ib_reg_wr reg_wr;
255 struct ib_send_wr send_wr;
256 struct ib_send_wr inv_wr;
259 #define ISER_RX_PAD_SIZE (256 - (ISER_RX_PAYLOAD_SIZE + \
260 sizeof(u64) + sizeof(struct ib_sge) + \
261 sizeof(struct ib_cqe)))
263 * struct iser_rx_desc - iSER RX descriptor
265 * @iser_header: iser header
266 * @iscsi_header: iscsi header
267 * @data: received data segment
268 * @dma_addr: receive buffer dma address
269 * @rx_sg: ib_sge of receive buffer
270 * @pad: for sense data TODO: Modify to maximum sense length supported
272 struct iser_rx_desc {
273 struct iser_ctrl iser_header;
274 struct iscsi_hdr iscsi_header;
275 char data[ISER_RECV_DATA_SEG_LEN];
279 char pad[ISER_RX_PAD_SIZE];
283 * struct iser_login_desc - iSER login descriptor
285 * @req: pointer to login request buffer
286 * @resp: pointer to login response buffer
287 * @req_dma: DMA address of login request buffer
288 * @rsp_dma: DMA address of login response buffer
289 * @sge: IB sge for login post recv
290 * @cqe: completion handler
292 struct iser_login_desc {
303 struct iscsi_iser_task;
306 * struct iser_comp - iSER completion context
308 * @cq: completion queue
309 * @active_qps: Number of active QPs attached
310 * to completion context
318 * struct iser_device - Memory registration operations
319 * per-device registration schemes
321 * @alloc_reg_res: Allocate registration resources
322 * @free_reg_res: Free registration resources
323 * @fast_reg_mem: Register memory buffers
324 * @unreg_mem: Un-register memory buffers
325 * @reg_desc_get: Get a registration descriptor for pool
326 * @reg_desc_put: Get a registration descriptor to pool
328 struct iser_reg_ops {
329 int (*alloc_reg_res)(struct ib_conn *ib_conn,
332 void (*free_reg_res)(struct ib_conn *ib_conn);
333 int (*reg_mem)(struct iscsi_iser_task *iser_task,
334 struct iser_data_buf *mem,
335 struct iser_reg_resources *rsc,
336 struct iser_mem_reg *reg);
337 void (*unreg_mem)(struct iscsi_iser_task *iser_task,
338 enum iser_data_dir cmd_dir);
339 struct iser_fr_desc * (*reg_desc_get)(struct ib_conn *ib_conn);
340 void (*reg_desc_put)(struct ib_conn *ib_conn,
341 struct iser_fr_desc *desc);
345 * struct iser_device - iSER device handle
347 * @ib_device: RDMA device
348 * @pd: Protection Domain for this device
349 * @mr: Global DMA memory region
350 * @event_handler: IB events handle routine
351 * @ig_list: entry in devices list
352 * @refcount: Reference counter, dominated by open iser connections
353 * @comps_used: Number of completion contexts used, Min between online
354 * cpus and device max completion vectors
355 * @comps: Dinamically allocated array of completion handlers
356 * @reg_ops: Registration ops
357 * @remote_inv_sup: Remote invalidate is supported on this device
360 struct ib_device *ib_device;
362 struct ib_event_handler event_handler;
363 struct list_head ig_list;
366 struct iser_comp *comps;
367 const struct iser_reg_ops *reg_ops;
372 * struct iser_reg_resources - Fast registration recources
375 * @fmr_pool: pool of fmrs
376 * @sig_mr: signature memory region
377 * @page_vec: fast reg page list used by fmr pool
378 * @mr_valid: is mr valid indicator
380 struct iser_reg_resources {
383 struct ib_fmr_pool *fmr_pool;
385 struct ib_mr *sig_mr;
386 struct iser_page_vec *page_vec;
391 * struct iser_fr_desc - Fast registration descriptor
393 * @list: entry in connection fastreg pool
394 * @rsc: data buffer registration resources
395 * @sig_protected: is region protected indicator
397 struct iser_fr_desc {
398 struct list_head list;
399 struct iser_reg_resources rsc;
401 struct list_head all_list;
405 * struct iser_fr_pool: connection fast registration pool
407 * @list: list of fastreg descriptors
408 * @lock: protects fmr/fastreg pool
409 * @size: size of the pool
411 struct iser_fr_pool {
412 struct list_head list;
415 struct list_head all_list;
419 * struct ib_conn - Infiniband related objects
421 * @cma_id: rdma_cm connection maneger handle
422 * @qp: Connection Queue-pair
423 * @post_recv_buf_count: post receive counter
424 * @sig_count: send work request signal count
425 * @rx_wr: receive work request for batch posts
426 * @device: reference to iser device
427 * @comp: iser completion context
428 * @fr_pool: connection fast registration poool
429 * @pi_support: Indicate device T10-PI support
432 struct rdma_cm_id *cma_id;
434 int post_recv_buf_count;
436 struct ib_recv_wr rx_wr[ISER_MIN_POSTED_RX];
437 struct iser_device *device;
438 struct iser_comp *comp;
439 struct iser_fr_pool fr_pool;
441 struct ib_cqe reg_cqe;
445 * struct iser_conn - iSER connection context
447 * @ib_conn: connection RDMA resources
448 * @iscsi_conn: link to matching iscsi connection
449 * @ep: transport handle
450 * @state: connection logical state
451 * @qp_max_recv_dtos: maximum number of data outs, corresponds
452 * to max number of post recvs
453 * @qp_max_recv_dtos_mask: (qp_max_recv_dtos - 1)
454 * @min_posted_rx: (qp_max_recv_dtos >> 2)
455 * @max_cmds: maximum cmds allowed for this connection
456 * @name: connection peer portal
457 * @release_work: deffered work for release job
458 * @state_mutex: protects iser onnection state
459 * @stop_completion: conn_stop completion
460 * @ib_completion: RDMA cleanup completion
461 * @up_completion: connection establishment completed
462 * (state is ISER_CONN_UP)
463 * @conn_list: entry in ig conn list
464 * @login_desc: login descriptor
465 * @rx_desc_head: head of rx_descs cyclic buffer
466 * @rx_descs: rx buffers array (cyclic buffer)
467 * @num_rx_descs: number of rx descriptors
468 * @scsi_sg_tablesize: scsi host sg_tablesize
469 * @pages_per_mr: maximum pages available for registration
472 struct ib_conn ib_conn;
473 struct iscsi_conn *iscsi_conn;
474 struct iscsi_endpoint *ep;
475 enum iser_conn_state state;
476 unsigned qp_max_recv_dtos;
477 unsigned qp_max_recv_dtos_mask;
478 unsigned min_posted_rx;
480 char name[ISER_OBJECT_NAME_SIZE];
481 struct work_struct release_work;
482 struct mutex state_mutex;
483 struct completion stop_completion;
484 struct completion ib_completion;
485 struct completion up_completion;
486 struct list_head conn_list;
487 struct iser_login_desc login_desc;
488 unsigned int rx_desc_head;
489 struct iser_rx_desc *rx_descs;
491 unsigned short scsi_sg_tablesize;
492 unsigned short pages_per_mr;
497 * struct iscsi_iser_task - iser task context
499 * @desc: TX descriptor
500 * @iser_conn: link to iser connection
501 * @status: current task status
502 * @sc: link to scsi command
503 * @command_sent: indicate if command was sent
504 * @dir: iser data direction
505 * @rdma_reg: task rdma registration desc
506 * @data: iser data buffer desc
507 * @prot: iser protection buffer desc
509 struct iscsi_iser_task {
510 struct iser_tx_desc desc;
511 struct iser_conn *iser_conn;
512 enum iser_task_status status;
513 struct scsi_cmnd *sc;
515 int dir[ISER_DIRS_NUM];
516 struct iser_mem_reg rdma_reg[ISER_DIRS_NUM];
517 struct iser_data_buf data[ISER_DIRS_NUM];
518 struct iser_data_buf prot[ISER_DIRS_NUM];
521 struct iser_page_vec {
524 struct ib_mr fake_mr;
528 * struct iser_global: iSER global context
530 * @device_list_mutex: protects device_list
531 * @device_list: iser devices global list
532 * @connlist_mutex: protects connlist
533 * @connlist: iser connections global list
534 * @desc_cache: kmem cache for tx dataout
537 struct mutex device_list_mutex;
538 struct list_head device_list;
539 struct mutex connlist_mutex;
540 struct list_head connlist;
541 struct kmem_cache *desc_cache;
544 extern struct iser_global ig;
545 extern int iser_debug_level;
546 extern bool iser_pi_enable;
547 extern int iser_pi_guard;
548 extern unsigned int iser_max_sectors;
549 extern bool iser_always_reg;
551 int iser_assign_reg_ops(struct iser_device *device);
553 int iser_send_control(struct iscsi_conn *conn,
554 struct iscsi_task *task);
556 int iser_send_command(struct iscsi_conn *conn,
557 struct iscsi_task *task);
559 int iser_send_data_out(struct iscsi_conn *conn,
560 struct iscsi_task *task,
561 struct iscsi_data *hdr);
563 void iscsi_iser_recv(struct iscsi_conn *conn,
564 struct iscsi_hdr *hdr,
568 void iser_conn_init(struct iser_conn *iser_conn);
570 void iser_conn_release(struct iser_conn *iser_conn);
572 int iser_conn_terminate(struct iser_conn *iser_conn);
574 void iser_release_work(struct work_struct *work);
576 void iser_err_comp(struct ib_wc *wc, const char *type);
577 void iser_login_rsp(struct ib_cq *cq, struct ib_wc *wc);
578 void iser_task_rsp(struct ib_cq *cq, struct ib_wc *wc);
579 void iser_cmd_comp(struct ib_cq *cq, struct ib_wc *wc);
580 void iser_ctrl_comp(struct ib_cq *cq, struct ib_wc *wc);
581 void iser_dataout_comp(struct ib_cq *cq, struct ib_wc *wc);
582 void iser_reg_comp(struct ib_cq *cq, struct ib_wc *wc);
584 void iser_task_rdma_init(struct iscsi_iser_task *task);
586 void iser_task_rdma_finalize(struct iscsi_iser_task *task);
588 void iser_free_rx_descriptors(struct iser_conn *iser_conn);
590 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
591 struct iser_data_buf *mem,
592 enum iser_data_dir cmd_dir);
594 int iser_reg_rdma_mem(struct iscsi_iser_task *task,
595 enum iser_data_dir dir,
597 void iser_unreg_rdma_mem(struct iscsi_iser_task *task,
598 enum iser_data_dir dir);
600 int iser_connect(struct iser_conn *iser_conn,
601 struct sockaddr *src_addr,
602 struct sockaddr *dst_addr,
605 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
606 enum iser_data_dir cmd_dir);
607 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
608 enum iser_data_dir cmd_dir);
610 int iser_post_recvl(struct iser_conn *iser_conn);
611 int iser_post_recvm(struct iser_conn *iser_conn, int count);
612 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
615 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
616 struct iser_data_buf *data,
617 enum iser_data_dir iser_dir,
618 enum dma_data_direction dma_dir);
620 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task,
621 struct iser_data_buf *data,
622 enum dma_data_direction dir);
624 int iser_initialize_task_headers(struct iscsi_task *task,
625 struct iser_tx_desc *tx_desc);
626 int iser_alloc_rx_descriptors(struct iser_conn *iser_conn,
627 struct iscsi_session *session);
628 int iser_alloc_fmr_pool(struct ib_conn *ib_conn,
631 void iser_free_fmr_pool(struct ib_conn *ib_conn);
632 int iser_alloc_fastreg_pool(struct ib_conn *ib_conn,
635 void iser_free_fastreg_pool(struct ib_conn *ib_conn);
636 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
637 enum iser_data_dir cmd_dir, sector_t *sector);
638 struct iser_fr_desc *
639 iser_reg_desc_get_fr(struct ib_conn *ib_conn);
641 iser_reg_desc_put_fr(struct ib_conn *ib_conn,
642 struct iser_fr_desc *desc);
643 struct iser_fr_desc *
644 iser_reg_desc_get_fmr(struct ib_conn *ib_conn);
646 iser_reg_desc_put_fmr(struct ib_conn *ib_conn,
647 struct iser_fr_desc *desc);
649 static inline struct iser_conn *
650 to_iser_conn(struct ib_conn *ib_conn)
652 return container_of(ib_conn, struct iser_conn, ib_conn);
655 static inline struct iser_rx_desc *
656 iser_rx(struct ib_cqe *cqe)
658 return container_of(cqe, struct iser_rx_desc, cqe);
661 static inline struct iser_tx_desc *
662 iser_tx(struct ib_cqe *cqe)
664 return container_of(cqe, struct iser_tx_desc, cqe);
667 static inline struct iser_login_desc *
668 iser_login(struct ib_cqe *cqe)
670 return container_of(cqe, struct iser_login_desc, cqe);