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>
52 #include <linux/interrupt.h>
53 #include <linux/wait.h>
54 #include <linux/sched.h>
55 #include <linux/list.h>
56 #include <linux/slab.h>
57 #include <linux/dma-mapping.h>
58 #include <linux/mutex.h>
59 #include <linux/mempool.h>
60 #include <linux/uio.h>
62 #include <linux/socket.h>
64 #include <linux/in6.h>
66 #include <rdma/ib_verbs.h>
67 #include <rdma/ib_fmr_pool.h>
68 #include <rdma/rdma_cm.h>
70 #define DRV_NAME "iser"
71 #define PFX DRV_NAME ": "
72 #define DRV_VER "1.4.8"
74 #define iser_dbg(fmt, arg...) \
76 if (iser_debug_level > 2) \
77 printk(KERN_DEBUG PFX "%s: " fmt,\
81 #define iser_warn(fmt, arg...) \
83 if (iser_debug_level > 0) \
84 pr_warn(PFX "%s: " fmt, \
88 #define iser_info(fmt, arg...) \
90 if (iser_debug_level > 1) \
91 pr_info(PFX "%s: " fmt, \
95 #define iser_err(fmt, arg...) \
97 printk(KERN_ERR PFX "%s: " fmt, \
102 #define SIZE_4K (1ULL << SHIFT_4K)
103 #define MASK_4K (~(SIZE_4K-1))
104 /* support up to 512KB in one RDMA */
105 #define ISCSI_ISER_SG_TABLESIZE (0x80000 >> SHIFT_4K)
106 #define ISER_DEF_XMIT_CMDS_DEFAULT 512
107 #if ISCSI_DEF_XMIT_CMDS_MAX > ISER_DEF_XMIT_CMDS_DEFAULT
108 #define ISER_DEF_XMIT_CMDS_MAX ISCSI_DEF_XMIT_CMDS_MAX
110 #define ISER_DEF_XMIT_CMDS_MAX ISER_DEF_XMIT_CMDS_DEFAULT
112 #define ISER_DEF_CMD_PER_LUN ISER_DEF_XMIT_CMDS_MAX
115 /* Maximal bounds on received asynchronous PDUs */
116 #define ISER_MAX_RX_MISC_PDUS 4 /* NOOP_IN(2) , ASYNC_EVENT(2) */
118 #define ISER_MAX_TX_MISC_PDUS 6 /* NOOP_OUT(2), TEXT(1), *
119 * SCSI_TMFUNC(2), LOGOUT(1) */
121 #define ISER_QP_MAX_RECV_DTOS (ISER_DEF_XMIT_CMDS_MAX)
123 #define ISER_MIN_POSTED_RX (ISER_DEF_XMIT_CMDS_MAX >> 2)
125 /* the max TX (send) WR supported by the iSER QP is defined by *
126 * max_send_wr = T * (1 + D) + C ; D is how many inflight dataouts we expect *
127 * to have at max for SCSI command. The tx posting & completion handling code *
128 * supports -EAGAIN scheme where tx is suspended till the QP has room for more *
129 * send WR. D=8 comes from 64K/8K */
131 #define ISER_INFLIGHT_DATAOUTS 8
133 #define ISER_QP_MAX_REQ_DTOS (ISER_DEF_XMIT_CMDS_MAX * \
134 (1 + ISER_INFLIGHT_DATAOUTS) + \
135 ISER_MAX_TX_MISC_PDUS + \
136 ISER_MAX_RX_MISC_PDUS)
138 /* Max registration work requests per command */
139 #define ISER_MAX_REG_WR_PER_CMD 5
141 /* For Signature we don't support DATAOUTs so no need to make room for them */
142 #define ISER_QP_SIG_MAX_REQ_DTOS (ISER_DEF_XMIT_CMDS_MAX * \
143 (1 + ISER_MAX_REG_WR_PER_CMD) + \
144 ISER_MAX_TX_MISC_PDUS + \
145 ISER_MAX_RX_MISC_PDUS)
147 #define ISER_WC_BATCH_COUNT 16
148 #define ISER_SIGNAL_CMD_COUNT 32
150 #define ISER_VER 0x10
151 #define ISER_WSV 0x08
152 #define ISER_RSV 0x04
154 #define ISER_FASTREG_LI_WRID 0xffffffffffffffffULL
155 #define ISER_BEACON_WRID 0xfffffffffffffffeULL
158 * struct iser_hdr - iSER header
160 * @flags: flags support (zbva, remote_inv)
162 * @write_stag: write rkey
163 * @write_va: write virtual address
164 * @reaf_stag: read rkey
165 * @read_va: read virtual address
174 } __attribute__((packed));
177 #define ISER_ZBVA_NOT_SUPPORTED 0x80
178 #define ISER_SEND_W_INV_NOT_SUPPORTED 0x40
185 /* Constant PDU lengths calculations */
186 #define ISER_HEADERS_LEN (sizeof(struct iser_hdr) + sizeof(struct iscsi_hdr))
188 #define ISER_RECV_DATA_SEG_LEN 128
189 #define ISER_RX_PAYLOAD_SIZE (ISER_HEADERS_LEN + ISER_RECV_DATA_SEG_LEN)
190 #define ISER_RX_LOGIN_SIZE (ISER_HEADERS_LEN + ISCSI_DEF_MAX_RECV_SEG_LEN)
192 /* Length of an object name string */
193 #define ISER_OBJECT_NAME_SIZE 64
195 enum iser_conn_state {
196 ISER_CONN_INIT, /* descriptor allocd, no conn */
197 ISER_CONN_PENDING, /* in the process of being established */
198 ISER_CONN_UP, /* up and running */
199 ISER_CONN_TERMINATING, /* in the process of being terminated */
200 ISER_CONN_DOWN, /* shut down */
204 enum iser_task_status {
205 ISER_TASK_STATUS_INIT = 0,
206 ISER_TASK_STATUS_STARTED,
207 ISER_TASK_STATUS_COMPLETED
211 ISER_DIR_IN = 0, /* to initiator */
212 ISER_DIR_OUT, /* from initiator */
217 * struct iser_data_buf - iSER data buffer
219 * @buf: pointer to the sg list
220 * @size: num entries of this sg
221 * @data_len: total beffer byte len
222 * @dma_nents: returned by dma_map_sg
223 * @copy_buf: allocated copy buf for SGs unaligned
224 * for rdma which are copied
225 * @sg_single: SG-ified clone of a non SG SC or
228 struct iser_data_buf {
231 unsigned long data_len;
232 unsigned int dma_nents;
234 struct scatterlist sg_single;
237 /* fwd declarations */
239 struct iscsi_iser_task;
240 struct iscsi_endpoint;
243 * struct iser_mem_reg - iSER memory registration info
245 * @lkey: MR local key
246 * @rkey: MR remote key
247 * @va: MR start address (buffer va)
249 * @mem_h: pointer to registration context (FMR/Fastreg)
250 * @is_mr: indicates weather we registered the buffer
252 struct iser_mem_reg {
262 * struct iser_regd_buf - iSER buffer registration desc
264 * @reg: memory registration info
265 * @virt_addr: virtual address of buffer
266 * @device: reference to iser device
267 * @direction: dma direction (for dma_unmap)
268 * @data_size: data buffer size in bytes
270 struct iser_regd_buf {
271 struct iser_mem_reg reg;
273 struct iser_device *device;
274 enum dma_data_direction direction;
275 unsigned int data_size;
278 enum iser_desc_type {
280 ISCSI_TX_SCSI_COMMAND,
285 * struct iser_tx_desc - iSER TX descriptor (for send wr_id)
287 * @iser_header: iser header
288 * @iscsi_header: iscsi header
289 * @type: command/control/dataout
290 * @dam_addr: header buffer dma_address
291 * @tx_sg: sg[0] points to iser/iscsi headers
292 * sg[1] optionally points to either of immediate data
293 * unsolicited data-out or control
294 * @num_sge: number sges used on this TX task
296 struct iser_tx_desc {
297 struct iser_hdr iser_header;
298 struct iscsi_hdr iscsi_header;
299 enum iser_desc_type type;
301 struct ib_sge tx_sg[2];
305 #define ISER_RX_PAD_SIZE (256 - (ISER_RX_PAYLOAD_SIZE + \
306 sizeof(u64) + sizeof(struct ib_sge)))
308 * struct iser_rx_desc - iSER RX descriptor (for recv wr_id)
310 * @iser_header: iser header
311 * @iscsi_header: iscsi header
312 * @data: received data segment
313 * @dma_addr: receive buffer dma address
314 * @rx_sg: ib_sge of receive buffer
315 * @pad: for sense data TODO: Modify to maximum sense length supported
317 struct iser_rx_desc {
318 struct iser_hdr iser_header;
319 struct iscsi_hdr iscsi_header;
320 char data[ISER_RECV_DATA_SEG_LEN];
323 char pad[ISER_RX_PAD_SIZE];
324 } __attribute__((packed));
326 #define ISER_MAX_CQ 4
330 struct iscsi_iser_task;
333 * struct iser_comp - iSER completion context
335 * @device: pointer to device handle
336 * @cq: completion queue
337 * @wcs: work completion array
338 * @tasklet: Tasklet handle
339 * @active_qps: Number of active QPs attached
340 * to completion context
343 struct iser_device *device;
345 struct ib_wc wcs[ISER_WC_BATCH_COUNT];
346 struct tasklet_struct tasklet;
351 * struct iser_device - iSER device handle
353 * @ib_device: RDMA device
354 * @pd: Protection Domain for this device
355 * @dev_attr: Device attributes container
356 * @mr: Global DMA memory region
357 * @event_handler: IB events handle routine
358 * @ig_list: entry in devices list
359 * @refcount: Reference counter, dominated by open iser connections
360 * @comps_used: Number of completion contexts used, Min between online
361 * cpus and device max completion vectors
362 * @comps: Dinamically allocated array of completion handlers
363 * Memory registration pool Function pointers (FMR or Fastreg):
364 * @iser_alloc_rdma_reg_res: Allocation of memory regions pool
365 * @iser_free_rdma_reg_res: Free of memory regions pool
366 * @iser_reg_rdma_mem: Memory registration routine
367 * @iser_unreg_rdma_mem: Memory deregistration routine
370 struct ib_device *ib_device;
372 struct ib_device_attr dev_attr;
374 struct ib_event_handler event_handler;
375 struct list_head ig_list;
378 struct iser_comp comps[ISER_MAX_CQ];
379 int (*iser_alloc_rdma_reg_res)(struct ib_conn *ib_conn,
381 void (*iser_free_rdma_reg_res)(struct ib_conn *ib_conn);
382 int (*iser_reg_rdma_mem)(struct iscsi_iser_task *iser_task,
383 enum iser_data_dir cmd_dir);
384 void (*iser_unreg_rdma_mem)(struct iscsi_iser_task *iser_task,
385 enum iser_data_dir cmd_dir);
388 #define ISER_CHECK_GUARD 0xc0
389 #define ISER_CHECK_REFTAG 0x0f
390 #define ISER_CHECK_APPTAG 0x30
392 enum iser_reg_indicator {
393 ISER_DATA_KEY_VALID = 1 << 0,
394 ISER_PROT_KEY_VALID = 1 << 1,
395 ISER_SIG_KEY_VALID = 1 << 2,
396 ISER_FASTREG_PROTECTED = 1 << 3,
400 * struct iser_pi_context - Protection information context
402 * @prot_mr: protection memory region
403 * @prot_frpl: protection fastreg page list
404 * @sig_mr: signature feature enabled memory region
406 struct iser_pi_context {
407 struct ib_mr *prot_mr;
408 struct ib_fast_reg_page_list *prot_frpl;
409 struct ib_mr *sig_mr;
413 * struct fast_reg_descriptor - Fast registration descriptor
415 * @list: entry in connection fastreg pool
416 * @data_mr: data memory region
417 * @data_frpl: data fastreg page list
418 * @pi_ctx: protection information context
419 * @reg_indicators: fast registration indicators
421 struct fast_reg_descriptor {
422 struct list_head list;
423 struct ib_mr *data_mr;
424 struct ib_fast_reg_page_list *data_frpl;
425 struct iser_pi_context *pi_ctx;
430 * struct ib_conn - Infiniband related objects
432 * @cma_id: rdma_cm connection maneger handle
433 * @qp: Connection Queue-pair
434 * @post_recv_buf_count: post receive counter
435 * @rx_wr: receive work request for batch posts
436 * @device: reference to iser device
437 * @comp: iser completion context
438 * @pi_support: Indicate device T10-PI support
439 * @beacon: beacon send wr to signal all flush errors were drained
440 * @flush_comp: completes when all connection completions consumed
441 * @lock: protects fmr/fastreg pool
443 * @pool: FMR pool for fast registrations
444 * @page_vec: page vector to hold mapped commands pages
445 * used for registration
447 * @pool: Fast registration descriptors pool for fast
449 * @pool_size: Size of pool
452 struct rdma_cm_id *cma_id;
454 int post_recv_buf_count;
455 struct ib_recv_wr rx_wr[ISER_MIN_POSTED_RX];
456 struct iser_device *device;
457 struct iser_comp *comp;
459 struct ib_send_wr beacon;
460 struct completion flush_comp;
464 struct ib_fmr_pool *pool;
465 struct iser_page_vec *page_vec;
468 struct list_head pool;
475 * struct iser_conn - iSER connection context
477 * @ib_conn: connection RDMA resources
478 * @iscsi_conn: link to matching iscsi connection
479 * @ep: transport handle
480 * @state: connection logical state
481 * @qp_max_recv_dtos: maximum number of data outs, corresponds
482 * to max number of post recvs
483 * @qp_max_recv_dtos_mask: (qp_max_recv_dtos - 1)
484 * @min_posted_rx: (qp_max_recv_dtos >> 2)
485 * @name: connection peer portal
486 * @release_work: deffered work for release job
487 * @state_mutex: protects iser onnection state
488 * @stop_completion: conn_stop completion
489 * @ib_completion: RDMA cleanup completion
490 * @up_completion: connection establishment completed
491 * (state is ISER_CONN_UP)
492 * @conn_list: entry in ig conn list
493 * @login_buf: login data buffer (stores login parameters)
494 * @login_req_buf: login request buffer
495 * @login_req_dma: login request buffer dma address
496 * @login_resp_buf: login response buffer
497 * @login_resp_dma: login response buffer dma address
498 * @rx_desc_head: head of rx_descs cyclic buffer
499 * @rx_descs: rx buffers array (cyclic buffer)
500 * @num_rx_descs: number of rx descriptors
503 struct ib_conn ib_conn;
504 struct iscsi_conn *iscsi_conn;
505 struct iscsi_endpoint *ep;
506 enum iser_conn_state state;
507 unsigned qp_max_recv_dtos;
508 unsigned qp_max_recv_dtos_mask;
509 unsigned min_posted_rx;
510 char name[ISER_OBJECT_NAME_SIZE];
511 struct work_struct release_work;
512 struct mutex state_mutex;
513 struct completion stop_completion;
514 struct completion ib_completion;
515 struct completion up_completion;
516 struct list_head conn_list;
519 char *login_req_buf, *login_resp_buf;
520 u64 login_req_dma, login_resp_dma;
521 unsigned int rx_desc_head;
522 struct iser_rx_desc *rx_descs;
527 * struct iscsi_iser_task - iser task context
529 * @desc: TX descriptor
530 * @iser_conn: link to iser connection
531 * @status: current task status
532 * @sc: link to scsi command
533 * @command_sent: indicate if command was sent
534 * @dir: iser data direction
535 * @rdma_regd: task rdma registration desc
536 * @data: iser data buffer desc
537 * @data_copy: iser data copy buffer desc (bounce buffer)
538 * @prot: iser protection buffer desc
539 * @prot_copy: iser protection copy buffer desc (bounce buffer)
541 struct iscsi_iser_task {
542 struct iser_tx_desc desc;
543 struct iser_conn *iser_conn;
544 enum iser_task_status status;
545 struct scsi_cmnd *sc;
547 int dir[ISER_DIRS_NUM];
548 struct iser_regd_buf rdma_regd[ISER_DIRS_NUM];
549 struct iser_data_buf data[ISER_DIRS_NUM];
550 struct iser_data_buf data_copy[ISER_DIRS_NUM];
551 struct iser_data_buf prot[ISER_DIRS_NUM];
552 struct iser_data_buf prot_copy[ISER_DIRS_NUM];
555 struct iser_page_vec {
563 * struct iser_global: iSER global context
565 * @device_list_mutex: protects device_list
566 * @device_list: iser devices global list
567 * @connlist_mutex: protects connlist
568 * @connlist: iser connections global list
569 * @desc_cache: kmem cache for tx dataout
572 struct mutex device_list_mutex;
573 struct list_head device_list;
574 struct mutex connlist_mutex;
575 struct list_head connlist;
576 struct kmem_cache *desc_cache;
579 extern struct iser_global ig;
580 extern int iser_debug_level;
581 extern bool iser_pi_enable;
582 extern int iser_pi_guard;
584 int iser_send_control(struct iscsi_conn *conn,
585 struct iscsi_task *task);
587 int iser_send_command(struct iscsi_conn *conn,
588 struct iscsi_task *task);
590 int iser_send_data_out(struct iscsi_conn *conn,
591 struct iscsi_task *task,
592 struct iscsi_data *hdr);
594 void iscsi_iser_recv(struct iscsi_conn *conn,
595 struct iscsi_hdr *hdr,
599 void iser_conn_init(struct iser_conn *iser_conn);
601 void iser_conn_release(struct iser_conn *iser_conn);
603 int iser_conn_terminate(struct iser_conn *iser_conn);
605 void iser_release_work(struct work_struct *work);
607 void iser_rcv_completion(struct iser_rx_desc *desc,
608 unsigned long dto_xfer_len,
609 struct ib_conn *ib_conn);
611 void iser_snd_completion(struct iser_tx_desc *desc,
612 struct ib_conn *ib_conn);
614 void iser_task_rdma_init(struct iscsi_iser_task *task);
616 void iser_task_rdma_finalize(struct iscsi_iser_task *task);
618 void iser_free_rx_descriptors(struct iser_conn *iser_conn);
620 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
621 struct iser_data_buf *mem,
622 struct iser_data_buf *mem_copy,
623 enum iser_data_dir cmd_dir);
625 int iser_reg_rdma_mem_fmr(struct iscsi_iser_task *task,
626 enum iser_data_dir cmd_dir);
627 int iser_reg_rdma_mem_fastreg(struct iscsi_iser_task *task,
628 enum iser_data_dir cmd_dir);
630 int iser_connect(struct iser_conn *iser_conn,
631 struct sockaddr *src_addr,
632 struct sockaddr *dst_addr,
635 int iser_reg_page_vec(struct ib_conn *ib_conn,
636 struct iser_page_vec *page_vec,
637 struct iser_mem_reg *mem_reg);
639 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
640 enum iser_data_dir cmd_dir);
641 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
642 enum iser_data_dir cmd_dir);
644 int iser_post_recvl(struct iser_conn *iser_conn);
645 int iser_post_recvm(struct iser_conn *iser_conn, int count);
646 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
649 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
650 struct iser_data_buf *data,
651 enum iser_data_dir iser_dir,
652 enum dma_data_direction dma_dir);
654 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task,
655 struct iser_data_buf *data);
656 int iser_initialize_task_headers(struct iscsi_task *task,
657 struct iser_tx_desc *tx_desc);
658 int iser_alloc_rx_descriptors(struct iser_conn *iser_conn,
659 struct iscsi_session *session);
660 int iser_create_fmr_pool(struct ib_conn *ib_conn, unsigned cmds_max);
661 void iser_free_fmr_pool(struct ib_conn *ib_conn);
662 int iser_create_fastreg_pool(struct ib_conn *ib_conn, unsigned cmds_max);
663 void iser_free_fastreg_pool(struct ib_conn *ib_conn);
664 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
665 enum iser_data_dir cmd_dir, sector_t *sector);