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 #define ISER_MIN_POSTED_RX		(ISER_DEF_XMIT_CMDS_MAX >> 2)
123 
124 /* the max TX (send) WR supported by the iSER QP is defined by                 *
125  * max_send_wr = T * (1 + D) + C ; D is how many inflight dataouts we expect   *
126  * to have at max for SCSI command. The tx posting & completion handling code  *
127  * supports -EAGAIN scheme where tx is suspended till the QP has room for more *
128  * send WR. D=8 comes from 64K/8K                                              */
129 
130 #define ISER_INFLIGHT_DATAOUTS		8
131 
132 #define ISER_QP_MAX_REQ_DTOS		(ISER_DEF_XMIT_CMDS_MAX *    \
133 					(1 + ISER_INFLIGHT_DATAOUTS) + \
134 					ISER_MAX_TX_MISC_PDUS        + \
135 					ISER_MAX_RX_MISC_PDUS)
136 
137 /* Max registration work requests per command */
138 #define ISER_MAX_REG_WR_PER_CMD		5
139 
140 /* For Signature we don't support DATAOUTs so no need to make room for them */
141 #define ISER_QP_SIG_MAX_REQ_DTOS	(ISER_DEF_XMIT_CMDS_MAX	*       \
142 					(1 + ISER_MAX_REG_WR_PER_CMD) + \
143 					ISER_MAX_TX_MISC_PDUS         + \
144 					ISER_MAX_RX_MISC_PDUS)
145 
146 #define ISER_GET_MAX_XMIT_CMDS(send_wr) ((send_wr			\
147 					 - ISER_MAX_TX_MISC_PDUS	\
148 					 - ISER_MAX_RX_MISC_PDUS) /	\
149 					 (1 + ISER_INFLIGHT_DATAOUTS))
150 
151 #define ISER_SIGNAL_CMD_COUNT 32
152 
153 /* Constant PDU lengths calculations */
154 #define ISER_HEADERS_LEN	(sizeof(struct iser_ctrl) + sizeof(struct iscsi_hdr))
155 
156 #define ISER_RECV_DATA_SEG_LEN	128
157 #define ISER_RX_PAYLOAD_SIZE	(ISER_HEADERS_LEN + ISER_RECV_DATA_SEG_LEN)
158 #define ISER_RX_LOGIN_SIZE	(ISER_HEADERS_LEN + ISCSI_DEF_MAX_RECV_SEG_LEN)
159 
160 /* Length of an object name string */
161 #define ISER_OBJECT_NAME_SIZE		    64
162 
163 enum iser_conn_state {
164 	ISER_CONN_INIT,		   /* descriptor allocd, no conn          */
165 	ISER_CONN_PENDING,	   /* in the process of being established */
166 	ISER_CONN_UP,		   /* up and running                      */
167 	ISER_CONN_TERMINATING,	   /* in the process of being terminated  */
168 	ISER_CONN_DOWN,		   /* shut down                           */
169 	ISER_CONN_STATES_NUM
170 };
171 
172 enum iser_task_status {
173 	ISER_TASK_STATUS_INIT = 0,
174 	ISER_TASK_STATUS_STARTED,
175 	ISER_TASK_STATUS_COMPLETED
176 };
177 
178 enum iser_data_dir {
179 	ISER_DIR_IN = 0,	   /* to initiator */
180 	ISER_DIR_OUT,		   /* from initiator */
181 	ISER_DIRS_NUM
182 };
183 
184 /**
185  * struct iser_data_buf - iSER data buffer
186  *
187  * @sg:           pointer to the sg list
188  * @size:         num entries of this sg
189  * @data_len:     total beffer byte len
190  * @dma_nents:    returned by dma_map_sg
191  */
192 struct iser_data_buf {
193 	struct scatterlist *sg;
194 	int                size;
195 	unsigned long      data_len;
196 	int                dma_nents;
197 };
198 
199 /* fwd declarations */
200 struct iser_device;
201 struct iscsi_iser_task;
202 struct iscsi_endpoint;
203 struct iser_reg_resources;
204 
205 /**
206  * struct iser_mem_reg - iSER memory registration info
207  *
208  * @sge:          memory region sg element
209  * @rkey:         memory region remote key
210  * @mem_h:        pointer to registration context (FMR/Fastreg)
211  */
212 struct iser_mem_reg {
213 	struct ib_sge	 sge;
214 	u32		 rkey;
215 	void		*mem_h;
216 };
217 
218 enum iser_desc_type {
219 	ISCSI_TX_CONTROL ,
220 	ISCSI_TX_SCSI_COMMAND,
221 	ISCSI_TX_DATAOUT
222 };
223 
224 /**
225  * struct iser_tx_desc - iSER TX descriptor
226  *
227  * @iser_header:   iser header
228  * @iscsi_header:  iscsi header
229  * @type:          command/control/dataout
230  * @dma_addr:      header buffer dma_address
231  * @tx_sg:         sg[0] points to iser/iscsi headers
232  *                 sg[1] optionally points to either of immediate data
233  *                 unsolicited data-out or control
234  * @num_sge:       number sges used on this TX task
235  * @cqe:           completion handler
236  * @mapped:        Is the task header mapped
237  * @reg_wr:        registration WR
238  * @send_wr:       send WR
239  * @inv_wr:        invalidate WR
240  */
241 struct iser_tx_desc {
242 	struct iser_ctrl             iser_header;
243 	struct iscsi_hdr             iscsi_header;
244 	enum   iser_desc_type        type;
245 	u64		             dma_addr;
246 	struct ib_sge		     tx_sg[2];
247 	int                          num_sge;
248 	struct ib_cqe		     cqe;
249 	bool			     mapped;
250 	struct ib_reg_wr	     reg_wr;
251 	struct ib_send_wr	     send_wr;
252 	struct ib_send_wr	     inv_wr;
253 };
254 
255 #define ISER_RX_PAD_SIZE	(256 - (ISER_RX_PAYLOAD_SIZE + \
256 				 sizeof(u64) + sizeof(struct ib_sge) + \
257 				 sizeof(struct ib_cqe)))
258 /**
259  * struct iser_rx_desc - iSER RX descriptor
260  *
261  * @iser_header:   iser header
262  * @iscsi_header:  iscsi header
263  * @data:          received data segment
264  * @dma_addr:      receive buffer dma address
265  * @rx_sg:         ib_sge of receive buffer
266  * @cqe:           completion handler
267  * @pad:           for sense data TODO: Modify to maximum sense length supported
268  */
269 struct iser_rx_desc {
270 	struct iser_ctrl             iser_header;
271 	struct iscsi_hdr             iscsi_header;
272 	char		             data[ISER_RECV_DATA_SEG_LEN];
273 	u64		             dma_addr;
274 	struct ib_sge		     rx_sg;
275 	struct ib_cqe		     cqe;
276 	char		             pad[ISER_RX_PAD_SIZE];
277 } __packed;
278 
279 /**
280  * struct iser_login_desc - iSER login descriptor
281  *
282  * @req:           pointer to login request buffer
283  * @rsp:           pointer to login response buffer
284  * @req_dma:       DMA address of login request buffer
285  * @rsp_dma:       DMA address of login response buffer
286  * @sge:           IB sge for login post recv
287  * @cqe:           completion handler
288  */
289 struct iser_login_desc {
290 	void                         *req;
291 	void                         *rsp;
292 	u64                          req_dma;
293 	u64                          rsp_dma;
294 	struct ib_sge                sge;
295 	struct ib_cqe		     cqe;
296 } __packed;
297 
298 struct iser_conn;
299 struct ib_conn;
300 struct iscsi_iser_task;
301 
302 /**
303  * struct iser_device - iSER device handle
304  *
305  * @ib_device:     RDMA device
306  * @pd:            Protection Domain for this device
307  * @mr:            Global DMA memory region
308  * @event_handler: IB events handle routine
309  * @ig_list:	   entry in devices list
310  * @refcount:      Reference counter, dominated by open iser connections
311  */
312 struct iser_device {
313 	struct ib_device             *ib_device;
314 	struct ib_pd	             *pd;
315 	struct ib_event_handler      event_handler;
316 	struct list_head             ig_list;
317 	int                          refcount;
318 };
319 
320 /**
321  * struct iser_reg_resources - Fast registration resources
322  *
323  * @mr:         memory region
324  * @sig_mr:     signature memory region
325  * @mr_valid:   is mr valid indicator
326  */
327 struct iser_reg_resources {
328 	struct ib_mr                     *mr;
329 	struct ib_mr                     *sig_mr;
330 	u8				  mr_valid:1;
331 };
332 
333 /**
334  * struct iser_fr_desc - Fast registration descriptor
335  *
336  * @list:           entry in connection fastreg pool
337  * @rsc:            data buffer registration resources
338  * @sig_protected:  is region protected indicator
339  * @all_list:       first and last list members
340  */
341 struct iser_fr_desc {
342 	struct list_head		  list;
343 	struct iser_reg_resources	  rsc;
344 	bool				  sig_protected;
345 	struct list_head                  all_list;
346 };
347 
348 /**
349  * struct iser_fr_pool - connection fast registration pool
350  *
351  * @list:                list of fastreg descriptors
352  * @lock:                protects fastreg pool
353  * @size:                size of the pool
354  * @all_list:            first and last list members
355  */
356 struct iser_fr_pool {
357 	struct list_head        list;
358 	spinlock_t              lock;
359 	int                     size;
360 	struct list_head        all_list;
361 };
362 
363 /**
364  * struct ib_conn - Infiniband related objects
365  *
366  * @cma_id:              rdma_cm connection maneger handle
367  * @qp:                  Connection Queue-pair
368  * @cq:                  Connection completion queue
369  * @cq_size:             The number of max outstanding completions
370  * @post_recv_buf_count: post receive counter
371  * @sig_count:           send work request signal count
372  * @rx_wr:               receive work request for batch posts
373  * @device:              reference to iser device
374  * @fr_pool:             connection fast registration poool
375  * @pi_support:          Indicate device T10-PI support
376  * @reg_cqe:             completion handler
377  */
378 struct ib_conn {
379 	struct rdma_cm_id           *cma_id;
380 	struct ib_qp	            *qp;
381 	struct ib_cq		    *cq;
382 	u32			    cq_size;
383 	int                          post_recv_buf_count;
384 	u8                           sig_count;
385 	struct ib_recv_wr	     rx_wr[ISER_MIN_POSTED_RX];
386 	struct iser_device          *device;
387 	struct iser_fr_pool          fr_pool;
388 	bool			     pi_support;
389 	struct ib_cqe		     reg_cqe;
390 };
391 
392 /**
393  * struct iser_conn - iSER connection context
394  *
395  * @ib_conn:          connection RDMA resources
396  * @iscsi_conn:       link to matching iscsi connection
397  * @ep:               transport handle
398  * @state:            connection logical state
399  * @qp_max_recv_dtos: maximum number of data outs, corresponds
400  *                    to max number of post recvs
401  * @qp_max_recv_dtos_mask: (qp_max_recv_dtos - 1)
402  * @min_posted_rx:    (qp_max_recv_dtos >> 2)
403  * @max_cmds:         maximum cmds allowed for this connection
404  * @name:             connection peer portal
405  * @release_work:     deffered work for release job
406  * @state_mutex:      protects iser onnection state
407  * @stop_completion:  conn_stop completion
408  * @ib_completion:    RDMA cleanup completion
409  * @up_completion:    connection establishment completed
410  *                    (state is ISER_CONN_UP)
411  * @conn_list:        entry in ig conn list
412  * @login_desc:       login descriptor
413  * @rx_desc_head:     head of rx_descs cyclic buffer
414  * @rx_descs:         rx buffers array (cyclic buffer)
415  * @num_rx_descs:     number of rx descriptors
416  * @scsi_sg_tablesize: scsi host sg_tablesize
417  * @pages_per_mr:     maximum pages available for registration
418  * @snd_w_inv:        connection uses remote invalidation
419  */
420 struct iser_conn {
421 	struct ib_conn		     ib_conn;
422 	struct iscsi_conn	     *iscsi_conn;
423 	struct iscsi_endpoint	     *ep;
424 	enum iser_conn_state	     state;
425 	unsigned		     qp_max_recv_dtos;
426 	unsigned		     qp_max_recv_dtos_mask;
427 	unsigned		     min_posted_rx;
428 	u16                          max_cmds;
429 	char 			     name[ISER_OBJECT_NAME_SIZE];
430 	struct work_struct	     release_work;
431 	struct mutex		     state_mutex;
432 	struct completion	     stop_completion;
433 	struct completion	     ib_completion;
434 	struct completion	     up_completion;
435 	struct list_head	     conn_list;
436 	struct iser_login_desc       login_desc;
437 	unsigned int 		     rx_desc_head;
438 	struct iser_rx_desc	     *rx_descs;
439 	u32                          num_rx_descs;
440 	unsigned short               scsi_sg_tablesize;
441 	unsigned short               pages_per_mr;
442 	bool			     snd_w_inv;
443 };
444 
445 /**
446  * struct iscsi_iser_task - iser task context
447  *
448  * @desc:     TX descriptor
449  * @iser_conn:        link to iser connection
450  * @status:           current task status
451  * @sc:               link to scsi command
452  * @command_sent:     indicate if command was sent
453  * @dir:              iser data direction
454  * @rdma_reg:         task rdma registration desc
455  * @data:             iser data buffer desc
456  * @prot:             iser protection buffer desc
457  */
458 struct iscsi_iser_task {
459 	struct iser_tx_desc          desc;
460 	struct iser_conn	     *iser_conn;
461 	enum iser_task_status 	     status;
462 	struct scsi_cmnd	     *sc;
463 	int                          command_sent;
464 	int                          dir[ISER_DIRS_NUM];
465 	struct iser_mem_reg          rdma_reg[ISER_DIRS_NUM];
466 	struct iser_data_buf         data[ISER_DIRS_NUM];
467 	struct iser_data_buf         prot[ISER_DIRS_NUM];
468 };
469 
470 /**
471  * struct iser_global - iSER global context
472  *
473  * @device_list_mutex:    protects device_list
474  * @device_list:          iser devices global list
475  * @connlist_mutex:       protects connlist
476  * @connlist:             iser connections global list
477  * @desc_cache:           kmem cache for tx dataout
478  */
479 struct iser_global {
480 	struct mutex      device_list_mutex;
481 	struct list_head  device_list;
482 	struct mutex      connlist_mutex;
483 	struct list_head  connlist;
484 	struct kmem_cache *desc_cache;
485 };
486 
487 extern struct iser_global ig;
488 extern int iser_debug_level;
489 extern bool iser_pi_enable;
490 extern int iser_pi_guard;
491 extern unsigned int iser_max_sectors;
492 extern bool iser_always_reg;
493 
494 int iser_send_control(struct iscsi_conn *conn,
495 		      struct iscsi_task *task);
496 
497 int iser_send_command(struct iscsi_conn *conn,
498 		      struct iscsi_task *task);
499 
500 int iser_send_data_out(struct iscsi_conn *conn,
501 		       struct iscsi_task *task,
502 		       struct iscsi_data *hdr);
503 
504 void iscsi_iser_recv(struct iscsi_conn *conn,
505 		     struct iscsi_hdr *hdr,
506 		     char *rx_data,
507 		     int rx_data_len);
508 
509 void iser_conn_init(struct iser_conn *iser_conn);
510 
511 void iser_conn_release(struct iser_conn *iser_conn);
512 
513 int iser_conn_terminate(struct iser_conn *iser_conn);
514 
515 void iser_release_work(struct work_struct *work);
516 
517 void iser_err_comp(struct ib_wc *wc, const char *type);
518 void iser_login_rsp(struct ib_cq *cq, struct ib_wc *wc);
519 void iser_task_rsp(struct ib_cq *cq, struct ib_wc *wc);
520 void iser_cmd_comp(struct ib_cq *cq, struct ib_wc *wc);
521 void iser_ctrl_comp(struct ib_cq *cq, struct ib_wc *wc);
522 void iser_dataout_comp(struct ib_cq *cq, struct ib_wc *wc);
523 void iser_reg_comp(struct ib_cq *cq, struct ib_wc *wc);
524 
525 void iser_task_rdma_init(struct iscsi_iser_task *task);
526 
527 void iser_task_rdma_finalize(struct iscsi_iser_task *task);
528 
529 void iser_free_rx_descriptors(struct iser_conn *iser_conn);
530 
531 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
532 				     struct iser_data_buf *mem,
533 				     enum iser_data_dir cmd_dir);
534 
535 int iser_reg_mem_fastreg(struct iscsi_iser_task *task,
536 			 enum iser_data_dir dir,
537 			 bool all_imm);
538 void iser_unreg_mem_fastreg(struct iscsi_iser_task *task,
539 			    enum iser_data_dir dir);
540 
541 int  iser_connect(struct iser_conn *iser_conn,
542 		  struct sockaddr *src_addr,
543 		  struct sockaddr *dst_addr,
544 		  int non_blocking);
545 
546 int  iser_post_recvl(struct iser_conn *iser_conn);
547 int  iser_post_recvm(struct iser_conn *iser_conn, int count);
548 int  iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
549 		    bool signal);
550 
551 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
552 			   struct iser_data_buf *data,
553 			   enum iser_data_dir iser_dir,
554 			   enum dma_data_direction dma_dir);
555 
556 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task,
557 			      struct iser_data_buf *data,
558 			      enum dma_data_direction dir);
559 
560 int  iser_initialize_task_headers(struct iscsi_task *task,
561 			struct iser_tx_desc *tx_desc);
562 int iser_alloc_rx_descriptors(struct iser_conn *iser_conn,
563 			      struct iscsi_session *session);
564 int iser_alloc_fastreg_pool(struct ib_conn *ib_conn,
565 			    unsigned cmds_max,
566 			    unsigned int size);
567 void iser_free_fastreg_pool(struct ib_conn *ib_conn);
568 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
569 			     enum iser_data_dir cmd_dir, sector_t *sector);
570 
571 static inline struct iser_conn *
572 to_iser_conn(struct ib_conn *ib_conn)
573 {
574 	return container_of(ib_conn, struct iser_conn, ib_conn);
575 }
576 
577 static inline struct iser_rx_desc *
578 iser_rx(struct ib_cqe *cqe)
579 {
580 	return container_of(cqe, struct iser_rx_desc, cqe);
581 }
582 
583 static inline struct iser_tx_desc *
584 iser_tx(struct ib_cqe *cqe)
585 {
586 	return container_of(cqe, struct iser_tx_desc, cqe);
587 }
588 
589 static inline struct iser_login_desc *
590 iser_login(struct ib_cqe *cqe)
591 {
592 	return container_of(cqe, struct iser_login_desc, cqe);
593 }
594 
595 #endif
596