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 "1.4.1"
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))
105 /* support up to 512KB in one RDMA */
106 #define ISCSI_ISER_SG_TABLESIZE (0x80000 >> SHIFT_4K)
107 #define ISER_DEF_XMIT_CMDS_DEFAULT 512
108 #if ISCSI_DEF_XMIT_CMDS_MAX > ISER_DEF_XMIT_CMDS_DEFAULT
109 #define ISER_DEF_XMIT_CMDS_MAX ISCSI_DEF_XMIT_CMDS_MAX
111 #define ISER_DEF_XMIT_CMDS_MAX ISER_DEF_XMIT_CMDS_DEFAULT
113 #define ISER_DEF_CMD_PER_LUN ISER_DEF_XMIT_CMDS_MAX
116 /* Maximal bounds on received asynchronous PDUs */
117 #define ISER_MAX_RX_MISC_PDUS 4 /* NOOP_IN(2) , ASYNC_EVENT(2) */
119 #define ISER_MAX_TX_MISC_PDUS 6 /* NOOP_OUT(2), TEXT(1), *
120 * SCSI_TMFUNC(2), LOGOUT(1) */
122 #define ISER_QP_MAX_RECV_DTOS (ISER_DEF_XMIT_CMDS_MAX)
124 #define ISER_MIN_POSTED_RX (ISER_DEF_XMIT_CMDS_MAX >> 2)
126 /* the max TX (send) WR supported by the iSER QP is defined by *
127 * max_send_wr = T * (1 + D) + C ; D is how many inflight dataouts we expect *
128 * to have at max for SCSI command. The tx posting & completion handling code *
129 * supports -EAGAIN scheme where tx is suspended till the QP has room for more *
130 * send WR. D=8 comes from 64K/8K */
132 #define ISER_INFLIGHT_DATAOUTS 8
134 #define ISER_QP_MAX_REQ_DTOS (ISER_DEF_XMIT_CMDS_MAX * \
135 (1 + ISER_INFLIGHT_DATAOUTS) + \
136 ISER_MAX_TX_MISC_PDUS + \
137 ISER_MAX_RX_MISC_PDUS)
139 /* Max registration work requests per command */
140 #define ISER_MAX_REG_WR_PER_CMD 5
142 /* For Signature we don't support DATAOUTs so no need to make room for them */
143 #define ISER_QP_SIG_MAX_REQ_DTOS (ISER_DEF_XMIT_CMDS_MAX * \
144 (1 + ISER_MAX_REG_WR_PER_CMD) + \
145 ISER_MAX_TX_MISC_PDUS + \
146 ISER_MAX_RX_MISC_PDUS)
148 #define ISER_WC_BATCH_COUNT 16
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
160 __be32 write_stag; /* write rkey */
162 __be32 read_stag; /* read rkey */
164 } __attribute__((packed));
167 #define ISER_ZBVA_NOT_SUPPORTED 0x80
168 #define ISER_SEND_W_INV_NOT_SUPPORTED 0x40
175 /* Constant PDU lengths calculations */
176 #define ISER_HEADERS_LEN (sizeof(struct iser_hdr) + sizeof(struct iscsi_hdr))
178 #define ISER_RECV_DATA_SEG_LEN 128
179 #define ISER_RX_PAYLOAD_SIZE (ISER_HEADERS_LEN + ISER_RECV_DATA_SEG_LEN)
180 #define ISER_RX_LOGIN_SIZE (ISER_HEADERS_LEN + ISCSI_DEF_MAX_RECV_SEG_LEN)
182 /* Length of an object name string */
183 #define ISER_OBJECT_NAME_SIZE 64
185 enum iser_conn_state {
186 ISER_CONN_INIT, /* descriptor allocd, no conn */
187 ISER_CONN_PENDING, /* in the process of being established */
188 ISER_CONN_UP, /* up and running */
189 ISER_CONN_TERMINATING, /* in the process of being terminated */
190 ISER_CONN_DOWN, /* shut down */
194 enum iser_task_status {
195 ISER_TASK_STATUS_INIT = 0,
196 ISER_TASK_STATUS_STARTED,
197 ISER_TASK_STATUS_COMPLETED
201 ISER_DIR_IN = 0, /* to initiator */
202 ISER_DIR_OUT, /* from initiator */
206 struct iser_data_buf {
207 void *buf; /* pointer to the sg list */
208 unsigned int size; /* num entries of this sg */
209 unsigned long data_len; /* total data len */
210 unsigned int dma_nents; /* returned by dma_map_sg */
211 char *copy_buf; /* allocated copy buf for SGs unaligned *
212 * for rdma which are copied */
213 struct scatterlist sg_single; /* SG-ified clone of a non SG SC or *
217 /* fwd declarations */
219 struct iscsi_iser_task;
220 struct iscsi_endpoint;
222 struct iser_mem_reg {
231 struct iser_regd_buf {
232 struct iser_mem_reg reg; /* memory registration info */
234 struct iser_device *device; /* device->device for dma_unmap */
235 enum dma_data_direction direction; /* direction for dma_unmap */
236 unsigned int data_size;
239 enum iser_desc_type {
241 ISCSI_TX_SCSI_COMMAND,
245 struct iser_tx_desc {
246 struct iser_hdr iser_header;
247 struct iscsi_hdr iscsi_header;
248 enum iser_desc_type type;
250 /* sg[0] points to iser/iscsi headers, sg[1] optionally points to either
251 of immediate data, unsolicited data-out or control (login,text) */
252 struct ib_sge tx_sg[2];
256 #define ISER_RX_PAD_SIZE (256 - (ISER_RX_PAYLOAD_SIZE + \
257 sizeof(u64) + sizeof(struct ib_sge)))
258 struct iser_rx_desc {
259 struct iser_hdr iser_header;
260 struct iscsi_hdr iscsi_header;
261 char data[ISER_RECV_DATA_SEG_LEN];
264 char pad[ISER_RX_PAD_SIZE];
265 } __attribute__((packed));
267 #define ISER_MAX_CQ 4
271 struct iscsi_iser_task;
274 * struct iser_comp - iSER completion context
276 * @device: pointer to device handle
277 * @cq: completion queue
278 * @wcs: work completion array
279 * @tasklet: Tasklet handle
280 * @active_qps: Number of active QPs attached
281 * to completion context
284 struct iser_device *device;
286 struct ib_wc wcs[ISER_WC_BATCH_COUNT];
287 struct tasklet_struct tasklet;
292 struct ib_device *ib_device;
294 struct ib_device_attr dev_attr;
296 struct ib_event_handler event_handler;
297 struct list_head ig_list; /* entry in ig devices list */
300 struct iser_comp comps[ISER_MAX_CQ];
301 int (*iser_alloc_rdma_reg_res)(struct ib_conn *ib_conn,
303 void (*iser_free_rdma_reg_res)(struct ib_conn *ib_conn);
304 int (*iser_reg_rdma_mem)(struct iscsi_iser_task *iser_task,
305 enum iser_data_dir cmd_dir);
306 void (*iser_unreg_rdma_mem)(struct iscsi_iser_task *iser_task,
307 enum iser_data_dir cmd_dir);
310 #define ISER_CHECK_GUARD 0xc0
311 #define ISER_CHECK_REFTAG 0x0f
312 #define ISER_CHECK_APPTAG 0x30
314 enum iser_reg_indicator {
315 ISER_DATA_KEY_VALID = 1 << 0,
316 ISER_PROT_KEY_VALID = 1 << 1,
317 ISER_SIG_KEY_VALID = 1 << 2,
318 ISER_FASTREG_PROTECTED = 1 << 3,
321 struct iser_pi_context {
322 struct ib_mr *prot_mr;
323 struct ib_fast_reg_page_list *prot_frpl;
324 struct ib_mr *sig_mr;
327 struct fast_reg_descriptor {
328 struct list_head list;
329 /* For fast registration - FRWR */
330 struct ib_mr *data_mr;
331 struct ib_fast_reg_page_list *data_frpl;
332 struct iser_pi_context *pi_ctx;
333 /* registration indicators container */
338 * struct ib_conn - Infiniband related objects
340 * @cma_id: rdma_cm connection maneger handle
341 * @qp: Connection Queue-pair
342 * @post_recv_buf_count: post receive counter
343 * @rx_wr: receive work request for batch posts
344 * @device: reference to iser device
345 * @comp: iser completion context
346 * @pi_support: Indicate device T10-PI support
347 * @beacon: beacon send wr to signal all flush errors were drained
348 * @flush_comp: completes when all connection completions consumed
349 * @lock: protects fmr/fastreg pool
351 * @pool: FMR pool for fast registrations
352 * @page_vec: page vector to hold mapped commands pages
353 * used for registration
355 * @pool: Fast registration descriptors pool for fast
357 * @pool_size: Size of pool
360 struct rdma_cm_id *cma_id;
362 int post_recv_buf_count;
363 struct ib_recv_wr rx_wr[ISER_MIN_POSTED_RX];
364 struct iser_device *device;
365 struct iser_comp *comp;
367 struct ib_send_wr beacon;
368 struct completion flush_comp;
372 struct ib_fmr_pool *pool;
373 struct iser_page_vec *page_vec;
376 struct list_head pool;
383 struct ib_conn ib_conn;
384 struct iscsi_conn *iscsi_conn;
385 struct iscsi_endpoint *ep;
386 enum iser_conn_state state; /* rdma connection state */
387 unsigned qp_max_recv_dtos; /* num of rx buffers */
388 unsigned qp_max_recv_dtos_mask; /* above minus 1 */
389 unsigned min_posted_rx; /* qp_max_recv_dtos >> 2 */
390 char name[ISER_OBJECT_NAME_SIZE];
391 struct work_struct release_work;
392 struct mutex state_mutex;
393 struct completion stop_completion;
394 struct completion ib_completion;
395 struct completion up_completion;
396 struct list_head conn_list; /* entry in ig conn list */
399 char *login_req_buf, *login_resp_buf;
400 u64 login_req_dma, login_resp_dma;
401 unsigned int rx_desc_head;
402 struct iser_rx_desc *rx_descs;
406 struct iscsi_iser_task {
407 struct iser_tx_desc desc;
408 struct iser_conn *iser_conn;
409 enum iser_task_status status;
410 struct scsi_cmnd *sc;
411 int command_sent; /* set if command sent */
412 int dir[ISER_DIRS_NUM]; /* set if dir use*/
413 struct iser_regd_buf rdma_regd[ISER_DIRS_NUM];/* regd rdma buf */
414 struct iser_data_buf data[ISER_DIRS_NUM]; /* orig. data des*/
415 struct iser_data_buf data_copy[ISER_DIRS_NUM];/* contig. copy */
416 struct iser_data_buf prot[ISER_DIRS_NUM]; /* prot desc */
417 struct iser_data_buf prot_copy[ISER_DIRS_NUM];/* prot copy */
420 struct iser_page_vec {
428 struct mutex device_list_mutex;/* */
429 struct list_head device_list; /* all iSER devices */
430 struct mutex connlist_mutex;
431 struct list_head connlist; /* all iSER IB connections */
433 struct kmem_cache *desc_cache;
436 extern struct iser_global ig;
437 extern int iser_debug_level;
438 extern bool iser_pi_enable;
439 extern int iser_pi_guard;
441 int iser_send_control(struct iscsi_conn *conn,
442 struct iscsi_task *task);
444 int iser_send_command(struct iscsi_conn *conn,
445 struct iscsi_task *task);
447 int iser_send_data_out(struct iscsi_conn *conn,
448 struct iscsi_task *task,
449 struct iscsi_data *hdr);
451 void iscsi_iser_recv(struct iscsi_conn *conn,
452 struct iscsi_hdr *hdr,
456 void iser_conn_init(struct iser_conn *iser_conn);
458 void iser_conn_release(struct iser_conn *iser_conn);
460 int iser_conn_terminate(struct iser_conn *iser_conn);
462 void iser_release_work(struct work_struct *work);
464 void iser_rcv_completion(struct iser_rx_desc *desc,
465 unsigned long dto_xfer_len,
466 struct ib_conn *ib_conn);
468 void iser_snd_completion(struct iser_tx_desc *desc,
469 struct ib_conn *ib_conn);
471 void iser_task_rdma_init(struct iscsi_iser_task *task);
473 void iser_task_rdma_finalize(struct iscsi_iser_task *task);
475 void iser_free_rx_descriptors(struct iser_conn *iser_conn);
477 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
478 struct iser_data_buf *mem,
479 struct iser_data_buf *mem_copy,
480 enum iser_data_dir cmd_dir);
482 int iser_reg_rdma_mem_fmr(struct iscsi_iser_task *task,
483 enum iser_data_dir cmd_dir);
484 int iser_reg_rdma_mem_fastreg(struct iscsi_iser_task *task,
485 enum iser_data_dir cmd_dir);
487 int iser_connect(struct iser_conn *iser_conn,
488 struct sockaddr *src_addr,
489 struct sockaddr *dst_addr,
492 int iser_reg_page_vec(struct ib_conn *ib_conn,
493 struct iser_page_vec *page_vec,
494 struct iser_mem_reg *mem_reg);
496 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
497 enum iser_data_dir cmd_dir);
498 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
499 enum iser_data_dir cmd_dir);
501 int iser_post_recvl(struct iser_conn *iser_conn);
502 int iser_post_recvm(struct iser_conn *iser_conn, int count);
503 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc);
505 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
506 struct iser_data_buf *data,
507 enum iser_data_dir iser_dir,
508 enum dma_data_direction dma_dir);
510 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task,
511 struct iser_data_buf *data);
512 int iser_initialize_task_headers(struct iscsi_task *task,
513 struct iser_tx_desc *tx_desc);
514 int iser_alloc_rx_descriptors(struct iser_conn *iser_conn,
515 struct iscsi_session *session);
516 int iser_create_fmr_pool(struct ib_conn *ib_conn, unsigned cmds_max);
517 void iser_free_fmr_pool(struct ib_conn *ib_conn);
518 int iser_create_fastreg_pool(struct ib_conn *ib_conn, unsigned cmds_max);
519 void iser_free_fastreg_pool(struct ib_conn *ib_conn);
520 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
521 enum iser_data_dir cmd_dir, sector_t *sector);