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 ": "
74 #define iser_dbg(fmt, arg...) \
76 if (unlikely(iser_debug_level > 2)) \
77 printk(KERN_DEBUG PFX "%s: " fmt,\
81 #define iser_warn(fmt, arg...) \
83 if (unlikely(iser_debug_level > 0)) \
84 pr_warn(PFX "%s: " fmt, \
88 #define iser_info(fmt, arg...) \
90 if (unlikely(iser_debug_level > 1)) \
91 pr_info(PFX "%s: " fmt, \
95 #define iser_err(fmt, arg...) \
96 pr_err(PFX "%s: " fmt, __func__ , ## arg)
99 #define SIZE_4K (1ULL << SHIFT_4K)
100 #define MASK_4K (~(SIZE_4K-1))
101 /* support up to 512KB in one RDMA */
102 #define ISCSI_ISER_SG_TABLESIZE (0x80000 >> SHIFT_4K)
103 #define ISER_DEF_XMIT_CMDS_DEFAULT 512
104 #if ISCSI_DEF_XMIT_CMDS_MAX > ISER_DEF_XMIT_CMDS_DEFAULT
105 #define ISER_DEF_XMIT_CMDS_MAX ISCSI_DEF_XMIT_CMDS_MAX
107 #define ISER_DEF_XMIT_CMDS_MAX ISER_DEF_XMIT_CMDS_DEFAULT
109 #define ISER_DEF_CMD_PER_LUN ISER_DEF_XMIT_CMDS_MAX
112 /* Maximal bounds on received asynchronous PDUs */
113 #define ISER_MAX_RX_MISC_PDUS 4 /* NOOP_IN(2) , ASYNC_EVENT(2) */
115 #define ISER_MAX_TX_MISC_PDUS 6 /* NOOP_OUT(2), TEXT(1), *
116 * SCSI_TMFUNC(2), LOGOUT(1) */
118 #define ISER_QP_MAX_RECV_DTOS (ISER_DEF_XMIT_CMDS_MAX)
120 #define ISER_MIN_POSTED_RX (ISER_DEF_XMIT_CMDS_MAX >> 2)
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 */
128 #define ISER_INFLIGHT_DATAOUTS 8
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)
135 /* Max registration work requests per command */
136 #define ISER_MAX_REG_WR_PER_CMD 5
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)
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))
149 #define ISER_WC_BATCH_COUNT 16
150 #define ISER_SIGNAL_CMD_COUNT 32
152 #define ISER_VER 0x10
153 #define ISER_WSV 0x08
154 #define ISER_RSV 0x04
156 #define ISER_FASTREG_LI_WRID 0xffffffffffffffffULL
157 #define ISER_BEACON_WRID 0xfffffffffffffffeULL
160 * struct iser_hdr - iSER header
162 * @flags: flags support (zbva, remote_inv)
164 * @write_stag: write rkey
165 * @write_va: write virtual address
166 * @reaf_stag: read rkey
167 * @read_va: read virtual address
176 } __attribute__((packed));
179 #define ISER_ZBVA_NOT_SUPPORTED 0x80
180 #define ISER_SEND_W_INV_NOT_SUPPORTED 0x40
187 /* Constant PDU lengths calculations */
188 #define ISER_HEADERS_LEN (sizeof(struct iser_hdr) + sizeof(struct iscsi_hdr))
190 #define ISER_RECV_DATA_SEG_LEN 128
191 #define ISER_RX_PAYLOAD_SIZE (ISER_HEADERS_LEN + ISER_RECV_DATA_SEG_LEN)
192 #define ISER_RX_LOGIN_SIZE (ISER_HEADERS_LEN + ISCSI_DEF_MAX_RECV_SEG_LEN)
194 /* Length of an object name string */
195 #define ISER_OBJECT_NAME_SIZE 64
197 enum iser_conn_state {
198 ISER_CONN_INIT, /* descriptor allocd, no conn */
199 ISER_CONN_PENDING, /* in the process of being established */
200 ISER_CONN_UP, /* up and running */
201 ISER_CONN_TERMINATING, /* in the process of being terminated */
202 ISER_CONN_DOWN, /* shut down */
206 enum iser_task_status {
207 ISER_TASK_STATUS_INIT = 0,
208 ISER_TASK_STATUS_STARTED,
209 ISER_TASK_STATUS_COMPLETED
213 ISER_DIR_IN = 0, /* to initiator */
214 ISER_DIR_OUT, /* from initiator */
219 * struct iser_data_buf - iSER data buffer
221 * @sg: pointer to the sg list
222 * @size: num entries of this sg
223 * @data_len: total beffer byte len
224 * @dma_nents: returned by dma_map_sg
225 * @orig_sg: pointer to the original sg list (in case
227 * @orig_size: num entris of orig sg list
229 struct iser_data_buf {
230 struct scatterlist *sg;
232 unsigned long data_len;
233 unsigned int dma_nents;
234 struct scatterlist *orig_sg;
235 unsigned int orig_size;
238 /* fwd declarations */
240 struct iscsi_iser_task;
241 struct iscsi_endpoint;
244 * struct iser_mem_reg - iSER memory registration info
246 * @sge: memory region sg element
247 * @rkey: memory region remote key
248 * @mem_h: pointer to registration context (FMR/Fastreg)
250 struct iser_mem_reg {
256 enum iser_desc_type {
258 ISCSI_TX_SCSI_COMMAND,
263 * struct iser_tx_desc - iSER TX descriptor (for send wr_id)
265 * @iser_header: iser header
266 * @iscsi_header: iscsi header
267 * @type: command/control/dataout
268 * @dam_addr: header buffer dma_address
269 * @tx_sg: sg[0] points to iser/iscsi headers
270 * sg[1] optionally points to either of immediate data
271 * unsolicited data-out or control
272 * @num_sge: number sges used on this TX task
273 * @mapped: Is the task header mapped
275 struct iser_tx_desc {
276 struct iser_hdr iser_header;
277 struct iscsi_hdr iscsi_header;
278 enum iser_desc_type type;
280 struct ib_sge tx_sg[2];
285 #define ISER_RX_PAD_SIZE (256 - (ISER_RX_PAYLOAD_SIZE + \
286 sizeof(u64) + sizeof(struct ib_sge)))
288 * struct iser_rx_desc - iSER RX descriptor (for recv wr_id)
290 * @iser_header: iser header
291 * @iscsi_header: iscsi header
292 * @data: received data segment
293 * @dma_addr: receive buffer dma address
294 * @rx_sg: ib_sge of receive buffer
295 * @pad: for sense data TODO: Modify to maximum sense length supported
297 struct iser_rx_desc {
298 struct iser_hdr iser_header;
299 struct iscsi_hdr iscsi_header;
300 char data[ISER_RECV_DATA_SEG_LEN];
303 char pad[ISER_RX_PAD_SIZE];
304 } __attribute__((packed));
308 struct iscsi_iser_task;
311 * struct iser_comp - iSER completion context
313 * @device: pointer to device handle
314 * @cq: completion queue
315 * @wcs: work completion array
316 * @tasklet: Tasklet handle
317 * @active_qps: Number of active QPs attached
318 * to completion context
321 struct iser_device *device;
323 struct ib_wc wcs[ISER_WC_BATCH_COUNT];
324 struct tasklet_struct tasklet;
329 * struct iser_device - iSER device handle
331 * @ib_device: RDMA device
332 * @pd: Protection Domain for this device
333 * @dev_attr: Device attributes container
334 * @mr: Global DMA memory region
335 * @event_handler: IB events handle routine
336 * @ig_list: entry in devices list
337 * @refcount: Reference counter, dominated by open iser connections
338 * @comps_used: Number of completion contexts used, Min between online
339 * cpus and device max completion vectors
340 * @comps: Dinamically allocated array of completion handlers
341 * Memory registration pool Function pointers (FMR or Fastreg):
342 * @iser_alloc_rdma_reg_res: Allocation of memory regions pool
343 * @iser_free_rdma_reg_res: Free of memory regions pool
344 * @iser_reg_rdma_mem: Memory registration routine
345 * @iser_unreg_rdma_mem: Memory deregistration routine
348 struct ib_device *ib_device;
350 struct ib_device_attr dev_attr;
352 struct ib_event_handler event_handler;
353 struct list_head ig_list;
356 struct iser_comp *comps;
357 int (*iser_alloc_rdma_reg_res)(struct ib_conn *ib_conn,
359 void (*iser_free_rdma_reg_res)(struct ib_conn *ib_conn);
360 int (*iser_reg_rdma_mem)(struct iscsi_iser_task *iser_task,
361 enum iser_data_dir cmd_dir);
362 void (*iser_unreg_rdma_mem)(struct iscsi_iser_task *iser_task,
363 enum iser_data_dir cmd_dir);
366 #define ISER_CHECK_GUARD 0xc0
367 #define ISER_CHECK_REFTAG 0x0f
368 #define ISER_CHECK_APPTAG 0x30
371 * struct iser_reg_resources - Fast registration recources
374 * @frpl: fast reg page list
375 * @mr_valid: is mr valid indicator
377 struct iser_reg_resources {
379 struct ib_fast_reg_page_list *frpl;
384 * struct iser_pi_context - Protection information context
386 * @rsc: protection buffer registration resources
387 * @sig_mr: signature enable memory region
388 * @sig_mr_valid: is sig_mr valid indicator
389 * @sig_protected: is region protected indicator
391 struct iser_pi_context {
392 struct iser_reg_resources rsc;
393 struct ib_mr *sig_mr;
399 * struct fast_reg_descriptor - Fast registration descriptor
401 * @list: entry in connection fastreg pool
402 * @rsc: data buffer registration resources
403 * @pi_ctx: protection information context
405 struct fast_reg_descriptor {
406 struct list_head list;
407 struct iser_reg_resources rsc;
408 struct iser_pi_context *pi_ctx;
412 * struct ib_conn - Infiniband related objects
414 * @cma_id: rdma_cm connection maneger handle
415 * @qp: Connection Queue-pair
416 * @post_recv_buf_count: post receive counter
417 * @sig_count: send work request signal count
418 * @rx_wr: receive work request for batch posts
419 * @device: reference to iser device
420 * @comp: iser completion context
421 * @pi_support: Indicate device T10-PI support
422 * @beacon: beacon send wr to signal all flush errors were drained
423 * @flush_comp: completes when all connection completions consumed
424 * @lock: protects fmr/fastreg pool
426 * @pool: FMR pool for fast registrations
427 * @page_vec: page vector to hold mapped commands pages
428 * used for registration
430 * @pool: Fast registration descriptors pool for fast
432 * @pool_size: Size of pool
435 struct rdma_cm_id *cma_id;
437 int post_recv_buf_count;
439 struct ib_recv_wr rx_wr[ISER_MIN_POSTED_RX];
440 struct iser_device *device;
441 struct iser_comp *comp;
443 struct ib_send_wr beacon;
444 struct completion flush_comp;
448 struct ib_fmr_pool *pool;
449 struct iser_page_vec *page_vec;
452 struct list_head pool;
459 * struct iser_conn - iSER connection context
461 * @ib_conn: connection RDMA resources
462 * @iscsi_conn: link to matching iscsi connection
463 * @ep: transport handle
464 * @state: connection logical state
465 * @qp_max_recv_dtos: maximum number of data outs, corresponds
466 * to max number of post recvs
467 * @qp_max_recv_dtos_mask: (qp_max_recv_dtos - 1)
468 * @min_posted_rx: (qp_max_recv_dtos >> 2)
469 * @max_cmds: maximum cmds allowed for this connection
470 * @name: connection peer portal
471 * @release_work: deffered work for release job
472 * @state_mutex: protects iser onnection state
473 * @stop_completion: conn_stop completion
474 * @ib_completion: RDMA cleanup completion
475 * @up_completion: connection establishment completed
476 * (state is ISER_CONN_UP)
477 * @conn_list: entry in ig conn list
478 * @login_buf: login data buffer (stores login parameters)
479 * @login_req_buf: login request buffer
480 * @login_req_dma: login request buffer dma address
481 * @login_resp_buf: login response buffer
482 * @login_resp_dma: login response buffer dma address
483 * @rx_desc_head: head of rx_descs cyclic buffer
484 * @rx_descs: rx buffers array (cyclic buffer)
485 * @num_rx_descs: number of rx descriptors
488 struct ib_conn ib_conn;
489 struct iscsi_conn *iscsi_conn;
490 struct iscsi_endpoint *ep;
491 enum iser_conn_state state;
492 unsigned qp_max_recv_dtos;
493 unsigned qp_max_recv_dtos_mask;
494 unsigned min_posted_rx;
496 char name[ISER_OBJECT_NAME_SIZE];
497 struct work_struct release_work;
498 struct mutex state_mutex;
499 struct completion stop_completion;
500 struct completion ib_completion;
501 struct completion up_completion;
502 struct list_head conn_list;
505 char *login_req_buf, *login_resp_buf;
506 u64 login_req_dma, login_resp_dma;
507 unsigned int rx_desc_head;
508 struct iser_rx_desc *rx_descs;
513 * struct iscsi_iser_task - iser task context
515 * @desc: TX descriptor
516 * @iser_conn: link to iser connection
517 * @status: current task status
518 * @sc: link to scsi command
519 * @command_sent: indicate if command was sent
520 * @dir: iser data direction
521 * @rdma_reg: task rdma registration desc
522 * @data: iser data buffer desc
523 * @prot: iser protection buffer desc
525 struct iscsi_iser_task {
526 struct iser_tx_desc desc;
527 struct iser_conn *iser_conn;
528 enum iser_task_status status;
529 struct scsi_cmnd *sc;
531 int dir[ISER_DIRS_NUM];
532 struct iser_mem_reg rdma_reg[ISER_DIRS_NUM];
533 struct iser_data_buf data[ISER_DIRS_NUM];
534 struct iser_data_buf prot[ISER_DIRS_NUM];
537 struct iser_page_vec {
545 * struct iser_global: iSER global context
547 * @device_list_mutex: protects device_list
548 * @device_list: iser devices global list
549 * @connlist_mutex: protects connlist
550 * @connlist: iser connections global list
551 * @desc_cache: kmem cache for tx dataout
554 struct mutex device_list_mutex;
555 struct list_head device_list;
556 struct mutex connlist_mutex;
557 struct list_head connlist;
558 struct kmem_cache *desc_cache;
561 extern struct iser_global ig;
562 extern int iser_debug_level;
563 extern bool iser_pi_enable;
564 extern int iser_pi_guard;
566 int iser_send_control(struct iscsi_conn *conn,
567 struct iscsi_task *task);
569 int iser_send_command(struct iscsi_conn *conn,
570 struct iscsi_task *task);
572 int iser_send_data_out(struct iscsi_conn *conn,
573 struct iscsi_task *task,
574 struct iscsi_data *hdr);
576 void iscsi_iser_recv(struct iscsi_conn *conn,
577 struct iscsi_hdr *hdr,
581 void iser_conn_init(struct iser_conn *iser_conn);
583 void iser_conn_release(struct iser_conn *iser_conn);
585 int iser_conn_terminate(struct iser_conn *iser_conn);
587 void iser_release_work(struct work_struct *work);
589 void iser_rcv_completion(struct iser_rx_desc *desc,
590 unsigned long dto_xfer_len,
591 struct ib_conn *ib_conn);
593 void iser_snd_completion(struct iser_tx_desc *desc,
594 struct ib_conn *ib_conn);
596 void iser_task_rdma_init(struct iscsi_iser_task *task);
598 void iser_task_rdma_finalize(struct iscsi_iser_task *task);
600 void iser_free_rx_descriptors(struct iser_conn *iser_conn);
602 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
603 struct iser_data_buf *mem,
604 enum iser_data_dir cmd_dir);
606 int iser_reg_rdma_mem_fmr(struct iscsi_iser_task *task,
607 enum iser_data_dir cmd_dir);
608 int iser_reg_rdma_mem_fastreg(struct iscsi_iser_task *task,
609 enum iser_data_dir cmd_dir);
611 int iser_connect(struct iser_conn *iser_conn,
612 struct sockaddr *src_addr,
613 struct sockaddr *dst_addr,
616 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
617 enum iser_data_dir cmd_dir);
618 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
619 enum iser_data_dir cmd_dir);
621 int iser_post_recvl(struct iser_conn *iser_conn);
622 int iser_post_recvm(struct iser_conn *iser_conn, int count);
623 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
626 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
627 struct iser_data_buf *data,
628 enum iser_data_dir iser_dir,
629 enum dma_data_direction dma_dir);
631 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task,
632 struct iser_data_buf *data,
633 enum dma_data_direction dir);
635 int iser_initialize_task_headers(struct iscsi_task *task,
636 struct iser_tx_desc *tx_desc);
637 int iser_alloc_rx_descriptors(struct iser_conn *iser_conn,
638 struct iscsi_session *session);
639 int iser_create_fmr_pool(struct ib_conn *ib_conn, unsigned cmds_max);
640 void iser_free_fmr_pool(struct ib_conn *ib_conn);
641 int iser_create_fastreg_pool(struct ib_conn *ib_conn, unsigned cmds_max);
642 void iser_free_fastreg_pool(struct ib_conn *ib_conn);
643 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
644 enum iser_data_dir cmd_dir, sector_t *sector);
645 struct fast_reg_descriptor *
646 iser_reg_desc_get(struct ib_conn *ib_conn);
648 iser_reg_desc_put(struct ib_conn *ib_conn,
649 struct fast_reg_descriptor *desc);