1 /*
2  * Copyright (C) 2013 Shaohua Li <shli@kernel.org>
3  * Copyright (C) 2014 Red Hat, Inc.
4  * Copyright (C) 2015 Arrikto, Inc.
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms and conditions of the GNU General Public License,
8  * version 2, as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope it will be useful, but WITHOUT
11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
13  * more details.
14  *
15  * You should have received a copy of the GNU General Public License along with
16  * this program; if not, write to the Free Software Foundation, Inc.,
17  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
18  */
19 
20 #include <linux/spinlock.h>
21 #include <linux/module.h>
22 #include <linux/idr.h>
23 #include <linux/kernel.h>
24 #include <linux/timer.h>
25 #include <linux/parser.h>
26 #include <linux/vmalloc.h>
27 #include <linux/uio_driver.h>
28 #include <linux/stringify.h>
29 #include <linux/bitops.h>
30 #include <net/genetlink.h>
31 #include <scsi/scsi_common.h>
32 #include <scsi/scsi_proto.h>
33 #include <target/target_core_base.h>
34 #include <target/target_core_fabric.h>
35 #include <target/target_core_backend.h>
36 
37 #include <linux/target_core_user.h>
38 
39 /*
40  * Define a shared-memory interface for LIO to pass SCSI commands and
41  * data to userspace for processing. This is to allow backends that
42  * are too complex for in-kernel support to be possible.
43  *
44  * It uses the UIO framework to do a lot of the device-creation and
45  * introspection work for us.
46  *
47  * See the .h file for how the ring is laid out. Note that while the
48  * command ring is defined, the particulars of the data area are
49  * not. Offset values in the command entry point to other locations
50  * internal to the mmap()ed area. There is separate space outside the
51  * command ring for data buffers. This leaves maximum flexibility for
52  * moving buffer allocations, or even page flipping or other
53  * allocation techniques, without altering the command ring layout.
54  *
55  * SECURITY:
56  * The user process must be assumed to be malicious. There's no way to
57  * prevent it breaking the command ring protocol if it wants, but in
58  * order to prevent other issues we must only ever read *data* from
59  * the shared memory area, not offsets or sizes. This applies to
60  * command ring entries as well as the mailbox. Extra code needed for
61  * this may have a 'UAM' comment.
62  */
63 
64 
65 #define TCMU_TIME_OUT (30 * MSEC_PER_SEC)
66 
67 #define DATA_BLOCK_BITS 256
68 #define DATA_BLOCK_SIZE 4096
69 
70 #define CMDR_SIZE (16 * 4096)
71 #define DATA_SIZE (DATA_BLOCK_BITS * DATA_BLOCK_SIZE)
72 
73 #define TCMU_RING_SIZE (CMDR_SIZE + DATA_SIZE)
74 
75 static struct device *tcmu_root_device;
76 
77 struct tcmu_hba {
78 	u32 host_id;
79 };
80 
81 #define TCMU_CONFIG_LEN 256
82 
83 struct tcmu_dev {
84 	struct se_device se_dev;
85 
86 	char *name;
87 	struct se_hba *hba;
88 
89 #define TCMU_DEV_BIT_OPEN 0
90 #define TCMU_DEV_BIT_BROKEN 1
91 	unsigned long flags;
92 
93 	struct uio_info uio_info;
94 
95 	struct tcmu_mailbox *mb_addr;
96 	size_t dev_size;
97 	u32 cmdr_size;
98 	u32 cmdr_last_cleaned;
99 	/* Offset of data ring from start of mb */
100 	/* Must add data_off and mb_addr to get the address */
101 	size_t data_off;
102 	size_t data_size;
103 
104 	DECLARE_BITMAP(data_bitmap, DATA_BLOCK_BITS);
105 
106 	wait_queue_head_t wait_cmdr;
107 	/* TODO should this be a mutex? */
108 	spinlock_t cmdr_lock;
109 
110 	struct idr commands;
111 	spinlock_t commands_lock;
112 
113 	struct timer_list timeout;
114 
115 	char dev_config[TCMU_CONFIG_LEN];
116 };
117 
118 #define TCMU_DEV(_se_dev) container_of(_se_dev, struct tcmu_dev, se_dev)
119 
120 #define CMDR_OFF sizeof(struct tcmu_mailbox)
121 
122 struct tcmu_cmd {
123 	struct se_cmd *se_cmd;
124 	struct tcmu_dev *tcmu_dev;
125 
126 	uint16_t cmd_id;
127 
128 	/* Can't use se_cmd when cleaning up expired cmds, because if
129 	   cmd has been completed then accessing se_cmd is off limits */
130 	DECLARE_BITMAP(data_bitmap, DATA_BLOCK_BITS);
131 
132 	unsigned long deadline;
133 
134 #define TCMU_CMD_BIT_EXPIRED 0
135 	unsigned long flags;
136 };
137 
138 static struct kmem_cache *tcmu_cmd_cache;
139 
140 /* multicast group */
141 enum tcmu_multicast_groups {
142 	TCMU_MCGRP_CONFIG,
143 };
144 
145 static const struct genl_multicast_group tcmu_mcgrps[] = {
146 	[TCMU_MCGRP_CONFIG] = { .name = "config", },
147 };
148 
149 /* Our generic netlink family */
150 static struct genl_family tcmu_genl_family = {
151 	.id = GENL_ID_GENERATE,
152 	.hdrsize = 0,
153 	.name = "TCM-USER",
154 	.version = 1,
155 	.maxattr = TCMU_ATTR_MAX,
156 	.mcgrps = tcmu_mcgrps,
157 	.n_mcgrps = ARRAY_SIZE(tcmu_mcgrps),
158 	.netnsok = true,
159 };
160 
161 static struct tcmu_cmd *tcmu_alloc_cmd(struct se_cmd *se_cmd)
162 {
163 	struct se_device *se_dev = se_cmd->se_dev;
164 	struct tcmu_dev *udev = TCMU_DEV(se_dev);
165 	struct tcmu_cmd *tcmu_cmd;
166 	int cmd_id;
167 
168 	tcmu_cmd = kmem_cache_zalloc(tcmu_cmd_cache, GFP_KERNEL);
169 	if (!tcmu_cmd)
170 		return NULL;
171 
172 	tcmu_cmd->se_cmd = se_cmd;
173 	tcmu_cmd->tcmu_dev = udev;
174 	tcmu_cmd->deadline = jiffies + msecs_to_jiffies(TCMU_TIME_OUT);
175 
176 	idr_preload(GFP_KERNEL);
177 	spin_lock_irq(&udev->commands_lock);
178 	cmd_id = idr_alloc(&udev->commands, tcmu_cmd, 0,
179 		USHRT_MAX, GFP_NOWAIT);
180 	spin_unlock_irq(&udev->commands_lock);
181 	idr_preload_end();
182 
183 	if (cmd_id < 0) {
184 		kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
185 		return NULL;
186 	}
187 	tcmu_cmd->cmd_id = cmd_id;
188 
189 	return tcmu_cmd;
190 }
191 
192 static inline void tcmu_flush_dcache_range(void *vaddr, size_t size)
193 {
194 	unsigned long offset = offset_in_page(vaddr);
195 
196 	size = round_up(size+offset, PAGE_SIZE);
197 	vaddr -= offset;
198 
199 	while (size) {
200 		flush_dcache_page(virt_to_page(vaddr));
201 		size -= PAGE_SIZE;
202 	}
203 }
204 
205 /*
206  * Some ring helper functions. We don't assume size is a power of 2 so
207  * we can't use circ_buf.h.
208  */
209 static inline size_t spc_used(size_t head, size_t tail, size_t size)
210 {
211 	int diff = head - tail;
212 
213 	if (diff >= 0)
214 		return diff;
215 	else
216 		return size + diff;
217 }
218 
219 static inline size_t spc_free(size_t head, size_t tail, size_t size)
220 {
221 	/* Keep 1 byte unused or we can't tell full from empty */
222 	return (size - spc_used(head, tail, size) - 1);
223 }
224 
225 static inline size_t head_to_end(size_t head, size_t size)
226 {
227 	return size - head;
228 }
229 
230 static inline void new_iov(struct iovec **iov, int *iov_cnt,
231 			   struct tcmu_dev *udev)
232 {
233 	struct iovec *iovec;
234 
235 	if (*iov_cnt != 0)
236 		(*iov)++;
237 	(*iov_cnt)++;
238 
239 	iovec = *iov;
240 	memset(iovec, 0, sizeof(struct iovec));
241 }
242 
243 #define UPDATE_HEAD(head, used, size) smp_store_release(&head, ((head % size) + used) % size)
244 
245 /* offset is relative to mb_addr */
246 static inline size_t get_block_offset(struct tcmu_dev *dev,
247 		int block, int remaining)
248 {
249 	return dev->data_off + block * DATA_BLOCK_SIZE +
250 		DATA_BLOCK_SIZE - remaining;
251 }
252 
253 static inline size_t iov_tail(struct tcmu_dev *udev, struct iovec *iov)
254 {
255 	return (size_t)iov->iov_base + iov->iov_len;
256 }
257 
258 static void alloc_and_scatter_data_area(struct tcmu_dev *udev,
259 	struct scatterlist *data_sg, unsigned int data_nents,
260 	struct iovec **iov, int *iov_cnt, bool copy_data)
261 {
262 	int i, block;
263 	int block_remaining = 0;
264 	void *from, *to;
265 	size_t copy_bytes, to_offset;
266 	struct scatterlist *sg;
267 
268 	for_each_sg(data_sg, sg, data_nents, i) {
269 		int sg_remaining = sg->length;
270 		from = kmap_atomic(sg_page(sg)) + sg->offset;
271 		while (sg_remaining > 0) {
272 			if (block_remaining == 0) {
273 				block = find_first_zero_bit(udev->data_bitmap,
274 						DATA_BLOCK_BITS);
275 				block_remaining = DATA_BLOCK_SIZE;
276 				set_bit(block, udev->data_bitmap);
277 			}
278 			copy_bytes = min_t(size_t, sg_remaining,
279 					block_remaining);
280 			to_offset = get_block_offset(udev, block,
281 					block_remaining);
282 			to = (void *)udev->mb_addr + to_offset;
283 			if (*iov_cnt != 0 &&
284 			    to_offset == iov_tail(udev, *iov)) {
285 				(*iov)->iov_len += copy_bytes;
286 			} else {
287 				new_iov(iov, iov_cnt, udev);
288 				(*iov)->iov_base = (void __user *) to_offset;
289 				(*iov)->iov_len = copy_bytes;
290 			}
291 			if (copy_data) {
292 				memcpy(to, from + sg->length - sg_remaining,
293 					copy_bytes);
294 				tcmu_flush_dcache_range(to, copy_bytes);
295 			}
296 			sg_remaining -= copy_bytes;
297 			block_remaining -= copy_bytes;
298 		}
299 		kunmap_atomic(from - sg->offset);
300 	}
301 }
302 
303 static void free_data_area(struct tcmu_dev *udev, struct tcmu_cmd *cmd)
304 {
305 	bitmap_xor(udev->data_bitmap, udev->data_bitmap, cmd->data_bitmap,
306 		   DATA_BLOCK_BITS);
307 }
308 
309 static void gather_data_area(struct tcmu_dev *udev, unsigned long *cmd_bitmap,
310 		struct scatterlist *data_sg, unsigned int data_nents)
311 {
312 	int i, block;
313 	int block_remaining = 0;
314 	void *from, *to;
315 	size_t copy_bytes, from_offset;
316 	struct scatterlist *sg;
317 
318 	for_each_sg(data_sg, sg, data_nents, i) {
319 		int sg_remaining = sg->length;
320 		to = kmap_atomic(sg_page(sg)) + sg->offset;
321 		while (sg_remaining > 0) {
322 			if (block_remaining == 0) {
323 				block = find_first_bit(cmd_bitmap,
324 						DATA_BLOCK_BITS);
325 				block_remaining = DATA_BLOCK_SIZE;
326 				clear_bit(block, cmd_bitmap);
327 			}
328 			copy_bytes = min_t(size_t, sg_remaining,
329 					block_remaining);
330 			from_offset = get_block_offset(udev, block,
331 					block_remaining);
332 			from = (void *) udev->mb_addr + from_offset;
333 			tcmu_flush_dcache_range(from, copy_bytes);
334 			memcpy(to + sg->length - sg_remaining, from,
335 					copy_bytes);
336 
337 			sg_remaining -= copy_bytes;
338 			block_remaining -= copy_bytes;
339 		}
340 		kunmap_atomic(to - sg->offset);
341 	}
342 }
343 
344 static inline size_t spc_bitmap_free(unsigned long *bitmap)
345 {
346 	return DATA_BLOCK_SIZE * (DATA_BLOCK_BITS -
347 			bitmap_weight(bitmap, DATA_BLOCK_BITS));
348 }
349 
350 /*
351  * We can't queue a command until we have space available on the cmd ring *and*
352  * space available on the data ring.
353  *
354  * Called with ring lock held.
355  */
356 static bool is_ring_space_avail(struct tcmu_dev *udev, size_t cmd_size, size_t data_needed)
357 {
358 	struct tcmu_mailbox *mb = udev->mb_addr;
359 	size_t space, cmd_needed;
360 	u32 cmd_head;
361 
362 	tcmu_flush_dcache_range(mb, sizeof(*mb));
363 
364 	cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
365 
366 	/*
367 	 * If cmd end-of-ring space is too small then we need space for a NOP plus
368 	 * original cmd - cmds are internally contiguous.
369 	 */
370 	if (head_to_end(cmd_head, udev->cmdr_size) >= cmd_size)
371 		cmd_needed = cmd_size;
372 	else
373 		cmd_needed = cmd_size + head_to_end(cmd_head, udev->cmdr_size);
374 
375 	space = spc_free(cmd_head, udev->cmdr_last_cleaned, udev->cmdr_size);
376 	if (space < cmd_needed) {
377 		pr_debug("no cmd space: %u %u %u\n", cmd_head,
378 		       udev->cmdr_last_cleaned, udev->cmdr_size);
379 		return false;
380 	}
381 
382 	space = spc_bitmap_free(udev->data_bitmap);
383 	if (space < data_needed) {
384 		pr_debug("no data space: only %zu available, but ask for %zu\n",
385 				space, data_needed);
386 		return false;
387 	}
388 
389 	return true;
390 }
391 
392 static int tcmu_queue_cmd_ring(struct tcmu_cmd *tcmu_cmd)
393 {
394 	struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
395 	struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
396 	size_t base_command_size, command_size;
397 	struct tcmu_mailbox *mb;
398 	struct tcmu_cmd_entry *entry;
399 	struct iovec *iov;
400 	int iov_cnt;
401 	uint32_t cmd_head;
402 	uint64_t cdb_off;
403 	bool copy_to_data_area;
404 	size_t data_length;
405 	DECLARE_BITMAP(old_bitmap, DATA_BLOCK_BITS);
406 
407 	if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags))
408 		return -EINVAL;
409 
410 	/*
411 	 * Must be a certain minimum size for response sense info, but
412 	 * also may be larger if the iov array is large.
413 	 *
414 	 * We prepare way too many iovs for potential uses here, because it's
415 	 * expensive to tell how many regions are freed in the bitmap
416 	*/
417 	base_command_size = max(offsetof(struct tcmu_cmd_entry,
418 				req.iov[se_cmd->t_bidi_data_nents +
419 					se_cmd->t_data_nents]),
420 				sizeof(struct tcmu_cmd_entry));
421 	command_size = base_command_size
422 		+ round_up(scsi_command_size(se_cmd->t_task_cdb), TCMU_OP_ALIGN_SIZE);
423 
424 	WARN_ON(command_size & (TCMU_OP_ALIGN_SIZE-1));
425 
426 	spin_lock_irq(&udev->cmdr_lock);
427 
428 	mb = udev->mb_addr;
429 	cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
430 	data_length = se_cmd->data_length;
431 	if (se_cmd->se_cmd_flags & SCF_BIDI) {
432 		BUG_ON(!(se_cmd->t_bidi_data_sg && se_cmd->t_bidi_data_nents));
433 		data_length += se_cmd->t_bidi_data_sg->length;
434 	}
435 	if ((command_size > (udev->cmdr_size / 2))
436 	    || data_length > udev->data_size)
437 		pr_warn("TCMU: Request of size %zu/%zu may be too big for %u/%zu "
438 			"cmd/data ring buffers\n", command_size, data_length,
439 			udev->cmdr_size, udev->data_size);
440 
441 	while (!is_ring_space_avail(udev, command_size, data_length)) {
442 		int ret;
443 		DEFINE_WAIT(__wait);
444 
445 		prepare_to_wait(&udev->wait_cmdr, &__wait, TASK_INTERRUPTIBLE);
446 
447 		pr_debug("sleeping for ring space\n");
448 		spin_unlock_irq(&udev->cmdr_lock);
449 		ret = schedule_timeout(msecs_to_jiffies(TCMU_TIME_OUT));
450 		finish_wait(&udev->wait_cmdr, &__wait);
451 		if (!ret) {
452 			pr_warn("tcmu: command timed out\n");
453 			return -ETIMEDOUT;
454 		}
455 
456 		spin_lock_irq(&udev->cmdr_lock);
457 
458 		/* We dropped cmdr_lock, cmd_head is stale */
459 		cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
460 	}
461 
462 	/* Insert a PAD if end-of-ring space is too small */
463 	if (head_to_end(cmd_head, udev->cmdr_size) < command_size) {
464 		size_t pad_size = head_to_end(cmd_head, udev->cmdr_size);
465 
466 		entry = (void *) mb + CMDR_OFF + cmd_head;
467 		tcmu_flush_dcache_range(entry, sizeof(*entry));
468 		tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_PAD);
469 		tcmu_hdr_set_len(&entry->hdr.len_op, pad_size);
470 		entry->hdr.cmd_id = 0; /* not used for PAD */
471 		entry->hdr.kflags = 0;
472 		entry->hdr.uflags = 0;
473 
474 		UPDATE_HEAD(mb->cmd_head, pad_size, udev->cmdr_size);
475 
476 		cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
477 		WARN_ON(cmd_head != 0);
478 	}
479 
480 	entry = (void *) mb + CMDR_OFF + cmd_head;
481 	tcmu_flush_dcache_range(entry, sizeof(*entry));
482 	tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_CMD);
483 	tcmu_hdr_set_len(&entry->hdr.len_op, command_size);
484 	entry->hdr.cmd_id = tcmu_cmd->cmd_id;
485 	entry->hdr.kflags = 0;
486 	entry->hdr.uflags = 0;
487 
488 	bitmap_copy(old_bitmap, udev->data_bitmap, DATA_BLOCK_BITS);
489 
490 	/*
491 	 * Fix up iovecs, and handle if allocation in data ring wrapped.
492 	 */
493 	iov = &entry->req.iov[0];
494 	iov_cnt = 0;
495 	copy_to_data_area = (se_cmd->data_direction == DMA_TO_DEVICE
496 		|| se_cmd->se_cmd_flags & SCF_BIDI);
497 	alloc_and_scatter_data_area(udev, se_cmd->t_data_sg,
498 		se_cmd->t_data_nents, &iov, &iov_cnt, copy_to_data_area);
499 	entry->req.iov_cnt = iov_cnt;
500 	entry->req.iov_dif_cnt = 0;
501 
502 	/* Handle BIDI commands */
503 	iov_cnt = 0;
504 	alloc_and_scatter_data_area(udev, se_cmd->t_bidi_data_sg,
505 		se_cmd->t_bidi_data_nents, &iov, &iov_cnt, false);
506 	entry->req.iov_bidi_cnt = iov_cnt;
507 
508 	/* cmd's data_bitmap is what changed in process */
509 	bitmap_xor(tcmu_cmd->data_bitmap, old_bitmap, udev->data_bitmap,
510 			DATA_BLOCK_BITS);
511 
512 	/* All offsets relative to mb_addr, not start of entry! */
513 	cdb_off = CMDR_OFF + cmd_head + base_command_size;
514 	memcpy((void *) mb + cdb_off, se_cmd->t_task_cdb, scsi_command_size(se_cmd->t_task_cdb));
515 	entry->req.cdb_off = cdb_off;
516 	tcmu_flush_dcache_range(entry, sizeof(*entry));
517 
518 	UPDATE_HEAD(mb->cmd_head, command_size, udev->cmdr_size);
519 	tcmu_flush_dcache_range(mb, sizeof(*mb));
520 
521 	spin_unlock_irq(&udev->cmdr_lock);
522 
523 	/* TODO: only if FLUSH and FUA? */
524 	uio_event_notify(&udev->uio_info);
525 
526 	mod_timer(&udev->timeout,
527 		round_jiffies_up(jiffies + msecs_to_jiffies(TCMU_TIME_OUT)));
528 
529 	return 0;
530 }
531 
532 static int tcmu_queue_cmd(struct se_cmd *se_cmd)
533 {
534 	struct se_device *se_dev = se_cmd->se_dev;
535 	struct tcmu_dev *udev = TCMU_DEV(se_dev);
536 	struct tcmu_cmd *tcmu_cmd;
537 	int ret;
538 
539 	tcmu_cmd = tcmu_alloc_cmd(se_cmd);
540 	if (!tcmu_cmd)
541 		return -ENOMEM;
542 
543 	ret = tcmu_queue_cmd_ring(tcmu_cmd);
544 	if (ret < 0) {
545 		pr_err("TCMU: Could not queue command\n");
546 		spin_lock_irq(&udev->commands_lock);
547 		idr_remove(&udev->commands, tcmu_cmd->cmd_id);
548 		spin_unlock_irq(&udev->commands_lock);
549 
550 		kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
551 	}
552 
553 	return ret;
554 }
555 
556 static void tcmu_handle_completion(struct tcmu_cmd *cmd, struct tcmu_cmd_entry *entry)
557 {
558 	struct se_cmd *se_cmd = cmd->se_cmd;
559 	struct tcmu_dev *udev = cmd->tcmu_dev;
560 
561 	if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags)) {
562 		/*
563 		 * cmd has been completed already from timeout, just reclaim
564 		 * data ring space and free cmd
565 		 */
566 		free_data_area(udev, cmd);
567 
568 		kmem_cache_free(tcmu_cmd_cache, cmd);
569 		return;
570 	}
571 
572 	if (entry->hdr.uflags & TCMU_UFLAG_UNKNOWN_OP) {
573 		free_data_area(udev, cmd);
574 		pr_warn("TCMU: Userspace set UNKNOWN_OP flag on se_cmd %p\n",
575 			cmd->se_cmd);
576 		entry->rsp.scsi_status = SAM_STAT_CHECK_CONDITION;
577 	} else if (entry->rsp.scsi_status == SAM_STAT_CHECK_CONDITION) {
578 		memcpy(se_cmd->sense_buffer, entry->rsp.sense_buffer,
579 			       se_cmd->scsi_sense_length);
580 		free_data_area(udev, cmd);
581 	} else if (se_cmd->se_cmd_flags & SCF_BIDI) {
582 		DECLARE_BITMAP(bitmap, DATA_BLOCK_BITS);
583 
584 		/* Get Data-In buffer before clean up */
585 		bitmap_copy(bitmap, cmd->data_bitmap, DATA_BLOCK_BITS);
586 		gather_data_area(udev, bitmap,
587 			se_cmd->t_bidi_data_sg, se_cmd->t_bidi_data_nents);
588 		free_data_area(udev, cmd);
589 	} else if (se_cmd->data_direction == DMA_FROM_DEVICE) {
590 		DECLARE_BITMAP(bitmap, DATA_BLOCK_BITS);
591 
592 		bitmap_copy(bitmap, cmd->data_bitmap, DATA_BLOCK_BITS);
593 		gather_data_area(udev, bitmap,
594 			se_cmd->t_data_sg, se_cmd->t_data_nents);
595 		free_data_area(udev, cmd);
596 	} else if (se_cmd->data_direction == DMA_TO_DEVICE) {
597 		free_data_area(udev, cmd);
598 	} else if (se_cmd->data_direction != DMA_NONE) {
599 		pr_warn("TCMU: data direction was %d!\n",
600 			se_cmd->data_direction);
601 	}
602 
603 	target_complete_cmd(cmd->se_cmd, entry->rsp.scsi_status);
604 	cmd->se_cmd = NULL;
605 
606 	kmem_cache_free(tcmu_cmd_cache, cmd);
607 }
608 
609 static unsigned int tcmu_handle_completions(struct tcmu_dev *udev)
610 {
611 	struct tcmu_mailbox *mb;
612 	unsigned long flags;
613 	int handled = 0;
614 
615 	if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags)) {
616 		pr_err("ring broken, not handling completions\n");
617 		return 0;
618 	}
619 
620 	spin_lock_irqsave(&udev->cmdr_lock, flags);
621 
622 	mb = udev->mb_addr;
623 	tcmu_flush_dcache_range(mb, sizeof(*mb));
624 
625 	while (udev->cmdr_last_cleaned != ACCESS_ONCE(mb->cmd_tail)) {
626 
627 		struct tcmu_cmd_entry *entry = (void *) mb + CMDR_OFF + udev->cmdr_last_cleaned;
628 		struct tcmu_cmd *cmd;
629 
630 		tcmu_flush_dcache_range(entry, sizeof(*entry));
631 
632 		if (tcmu_hdr_get_op(entry->hdr.len_op) == TCMU_OP_PAD) {
633 			UPDATE_HEAD(udev->cmdr_last_cleaned,
634 				    tcmu_hdr_get_len(entry->hdr.len_op),
635 				    udev->cmdr_size);
636 			continue;
637 		}
638 		WARN_ON(tcmu_hdr_get_op(entry->hdr.len_op) != TCMU_OP_CMD);
639 
640 		spin_lock(&udev->commands_lock);
641 		cmd = idr_find(&udev->commands, entry->hdr.cmd_id);
642 		if (cmd)
643 			idr_remove(&udev->commands, cmd->cmd_id);
644 		spin_unlock(&udev->commands_lock);
645 
646 		if (!cmd) {
647 			pr_err("cmd_id not found, ring is broken\n");
648 			set_bit(TCMU_DEV_BIT_BROKEN, &udev->flags);
649 			break;
650 		}
651 
652 		tcmu_handle_completion(cmd, entry);
653 
654 		UPDATE_HEAD(udev->cmdr_last_cleaned,
655 			    tcmu_hdr_get_len(entry->hdr.len_op),
656 			    udev->cmdr_size);
657 
658 		handled++;
659 	}
660 
661 	if (mb->cmd_tail == mb->cmd_head)
662 		del_timer(&udev->timeout); /* no more pending cmds */
663 
664 	spin_unlock_irqrestore(&udev->cmdr_lock, flags);
665 
666 	wake_up(&udev->wait_cmdr);
667 
668 	return handled;
669 }
670 
671 static int tcmu_check_expired_cmd(int id, void *p, void *data)
672 {
673 	struct tcmu_cmd *cmd = p;
674 
675 	if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
676 		return 0;
677 
678 	if (!time_after(jiffies, cmd->deadline))
679 		return 0;
680 
681 	set_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags);
682 	target_complete_cmd(cmd->se_cmd, SAM_STAT_CHECK_CONDITION);
683 	cmd->se_cmd = NULL;
684 
685 	kmem_cache_free(tcmu_cmd_cache, cmd);
686 
687 	return 0;
688 }
689 
690 static void tcmu_device_timedout(unsigned long data)
691 {
692 	struct tcmu_dev *udev = (struct tcmu_dev *)data;
693 	unsigned long flags;
694 	int handled;
695 
696 	handled = tcmu_handle_completions(udev);
697 
698 	pr_warn("%d completions handled from timeout\n", handled);
699 
700 	spin_lock_irqsave(&udev->commands_lock, flags);
701 	idr_for_each(&udev->commands, tcmu_check_expired_cmd, NULL);
702 	spin_unlock_irqrestore(&udev->commands_lock, flags);
703 
704 	/*
705 	 * We don't need to wakeup threads on wait_cmdr since they have their
706 	 * own timeout.
707 	 */
708 }
709 
710 static int tcmu_attach_hba(struct se_hba *hba, u32 host_id)
711 {
712 	struct tcmu_hba *tcmu_hba;
713 
714 	tcmu_hba = kzalloc(sizeof(struct tcmu_hba), GFP_KERNEL);
715 	if (!tcmu_hba)
716 		return -ENOMEM;
717 
718 	tcmu_hba->host_id = host_id;
719 	hba->hba_ptr = tcmu_hba;
720 
721 	return 0;
722 }
723 
724 static void tcmu_detach_hba(struct se_hba *hba)
725 {
726 	kfree(hba->hba_ptr);
727 	hba->hba_ptr = NULL;
728 }
729 
730 static struct se_device *tcmu_alloc_device(struct se_hba *hba, const char *name)
731 {
732 	struct tcmu_dev *udev;
733 
734 	udev = kzalloc(sizeof(struct tcmu_dev), GFP_KERNEL);
735 	if (!udev)
736 		return NULL;
737 
738 	udev->name = kstrdup(name, GFP_KERNEL);
739 	if (!udev->name) {
740 		kfree(udev);
741 		return NULL;
742 	}
743 
744 	udev->hba = hba;
745 
746 	init_waitqueue_head(&udev->wait_cmdr);
747 	spin_lock_init(&udev->cmdr_lock);
748 
749 	idr_init(&udev->commands);
750 	spin_lock_init(&udev->commands_lock);
751 
752 	setup_timer(&udev->timeout, tcmu_device_timedout,
753 		(unsigned long)udev);
754 
755 	return &udev->se_dev;
756 }
757 
758 static int tcmu_irqcontrol(struct uio_info *info, s32 irq_on)
759 {
760 	struct tcmu_dev *tcmu_dev = container_of(info, struct tcmu_dev, uio_info);
761 
762 	tcmu_handle_completions(tcmu_dev);
763 
764 	return 0;
765 }
766 
767 /*
768  * mmap code from uio.c. Copied here because we want to hook mmap()
769  * and this stuff must come along.
770  */
771 static int tcmu_find_mem_index(struct vm_area_struct *vma)
772 {
773 	struct tcmu_dev *udev = vma->vm_private_data;
774 	struct uio_info *info = &udev->uio_info;
775 
776 	if (vma->vm_pgoff < MAX_UIO_MAPS) {
777 		if (info->mem[vma->vm_pgoff].size == 0)
778 			return -1;
779 		return (int)vma->vm_pgoff;
780 	}
781 	return -1;
782 }
783 
784 static int tcmu_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
785 {
786 	struct tcmu_dev *udev = vma->vm_private_data;
787 	struct uio_info *info = &udev->uio_info;
788 	struct page *page;
789 	unsigned long offset;
790 	void *addr;
791 
792 	int mi = tcmu_find_mem_index(vma);
793 	if (mi < 0)
794 		return VM_FAULT_SIGBUS;
795 
796 	/*
797 	 * We need to subtract mi because userspace uses offset = N*PAGE_SIZE
798 	 * to use mem[N].
799 	 */
800 	offset = (vmf->pgoff - mi) << PAGE_SHIFT;
801 
802 	addr = (void *)(unsigned long)info->mem[mi].addr + offset;
803 	if (info->mem[mi].memtype == UIO_MEM_LOGICAL)
804 		page = virt_to_page(addr);
805 	else
806 		page = vmalloc_to_page(addr);
807 	get_page(page);
808 	vmf->page = page;
809 	return 0;
810 }
811 
812 static const struct vm_operations_struct tcmu_vm_ops = {
813 	.fault = tcmu_vma_fault,
814 };
815 
816 static int tcmu_mmap(struct uio_info *info, struct vm_area_struct *vma)
817 {
818 	struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
819 
820 	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
821 	vma->vm_ops = &tcmu_vm_ops;
822 
823 	vma->vm_private_data = udev;
824 
825 	/* Ensure the mmap is exactly the right size */
826 	if (vma_pages(vma) != (TCMU_RING_SIZE >> PAGE_SHIFT))
827 		return -EINVAL;
828 
829 	return 0;
830 }
831 
832 static int tcmu_open(struct uio_info *info, struct inode *inode)
833 {
834 	struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
835 
836 	/* O_EXCL not supported for char devs, so fake it? */
837 	if (test_and_set_bit(TCMU_DEV_BIT_OPEN, &udev->flags))
838 		return -EBUSY;
839 
840 	pr_debug("open\n");
841 
842 	return 0;
843 }
844 
845 static int tcmu_release(struct uio_info *info, struct inode *inode)
846 {
847 	struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
848 
849 	clear_bit(TCMU_DEV_BIT_OPEN, &udev->flags);
850 
851 	pr_debug("close\n");
852 
853 	return 0;
854 }
855 
856 static int tcmu_netlink_event(enum tcmu_genl_cmd cmd, const char *name, int minor)
857 {
858 	struct sk_buff *skb;
859 	void *msg_header;
860 	int ret = -ENOMEM;
861 
862 	skb = genlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
863 	if (!skb)
864 		return ret;
865 
866 	msg_header = genlmsg_put(skb, 0, 0, &tcmu_genl_family, 0, cmd);
867 	if (!msg_header)
868 		goto free_skb;
869 
870 	ret = nla_put_string(skb, TCMU_ATTR_DEVICE, name);
871 	if (ret < 0)
872 		goto free_skb;
873 
874 	ret = nla_put_u32(skb, TCMU_ATTR_MINOR, minor);
875 	if (ret < 0)
876 		goto free_skb;
877 
878 	genlmsg_end(skb, msg_header);
879 
880 	ret = genlmsg_multicast_allns(&tcmu_genl_family, skb, 0,
881 				TCMU_MCGRP_CONFIG, GFP_KERNEL);
882 
883 	/* We don't care if no one is listening */
884 	if (ret == -ESRCH)
885 		ret = 0;
886 
887 	return ret;
888 free_skb:
889 	nlmsg_free(skb);
890 	return ret;
891 }
892 
893 static int tcmu_configure_device(struct se_device *dev)
894 {
895 	struct tcmu_dev *udev = TCMU_DEV(dev);
896 	struct tcmu_hba *hba = udev->hba->hba_ptr;
897 	struct uio_info *info;
898 	struct tcmu_mailbox *mb;
899 	size_t size;
900 	size_t used;
901 	int ret = 0;
902 	char *str;
903 
904 	info = &udev->uio_info;
905 
906 	size = snprintf(NULL, 0, "tcm-user/%u/%s/%s", hba->host_id, udev->name,
907 			udev->dev_config);
908 	size += 1; /* for \0 */
909 	str = kmalloc(size, GFP_KERNEL);
910 	if (!str)
911 		return -ENOMEM;
912 
913 	used = snprintf(str, size, "tcm-user/%u/%s", hba->host_id, udev->name);
914 
915 	if (udev->dev_config[0])
916 		snprintf(str + used, size - used, "/%s", udev->dev_config);
917 
918 	info->name = str;
919 
920 	udev->mb_addr = vzalloc(TCMU_RING_SIZE);
921 	if (!udev->mb_addr) {
922 		ret = -ENOMEM;
923 		goto err_vzalloc;
924 	}
925 
926 	/* mailbox fits in first part of CMDR space */
927 	udev->cmdr_size = CMDR_SIZE - CMDR_OFF;
928 	udev->data_off = CMDR_SIZE;
929 	udev->data_size = TCMU_RING_SIZE - CMDR_SIZE;
930 
931 	mb = udev->mb_addr;
932 	mb->version = TCMU_MAILBOX_VERSION;
933 	mb->flags = TCMU_MAILBOX_FLAG_CAP_OOOC;
934 	mb->cmdr_off = CMDR_OFF;
935 	mb->cmdr_size = udev->cmdr_size;
936 
937 	WARN_ON(!PAGE_ALIGNED(udev->data_off));
938 	WARN_ON(udev->data_size % PAGE_SIZE);
939 	WARN_ON(udev->data_size % DATA_BLOCK_SIZE);
940 
941 	info->version = __stringify(TCMU_MAILBOX_VERSION);
942 
943 	info->mem[0].name = "tcm-user command & data buffer";
944 	info->mem[0].addr = (phys_addr_t)(uintptr_t)udev->mb_addr;
945 	info->mem[0].size = TCMU_RING_SIZE;
946 	info->mem[0].memtype = UIO_MEM_VIRTUAL;
947 
948 	info->irqcontrol = tcmu_irqcontrol;
949 	info->irq = UIO_IRQ_CUSTOM;
950 
951 	info->mmap = tcmu_mmap;
952 	info->open = tcmu_open;
953 	info->release = tcmu_release;
954 
955 	ret = uio_register_device(tcmu_root_device, info);
956 	if (ret)
957 		goto err_register;
958 
959 	/* User can set hw_block_size before enable the device */
960 	if (dev->dev_attrib.hw_block_size == 0)
961 		dev->dev_attrib.hw_block_size = 512;
962 	/* Other attributes can be configured in userspace */
963 	dev->dev_attrib.hw_max_sectors = 128;
964 	dev->dev_attrib.hw_queue_depth = 128;
965 
966 	ret = tcmu_netlink_event(TCMU_CMD_ADDED_DEVICE, udev->uio_info.name,
967 				 udev->uio_info.uio_dev->minor);
968 	if (ret)
969 		goto err_netlink;
970 
971 	return 0;
972 
973 err_netlink:
974 	uio_unregister_device(&udev->uio_info);
975 err_register:
976 	vfree(udev->mb_addr);
977 err_vzalloc:
978 	kfree(info->name);
979 
980 	return ret;
981 }
982 
983 static int tcmu_check_and_free_pending_cmd(struct tcmu_cmd *cmd)
984 {
985 	if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags)) {
986 		kmem_cache_free(tcmu_cmd_cache, cmd);
987 		return 0;
988 	}
989 	return -EINVAL;
990 }
991 
992 static void tcmu_dev_call_rcu(struct rcu_head *p)
993 {
994 	struct se_device *dev = container_of(p, struct se_device, rcu_head);
995 	struct tcmu_dev *udev = TCMU_DEV(dev);
996 
997 	kfree(udev);
998 }
999 
1000 static void tcmu_free_device(struct se_device *dev)
1001 {
1002 	struct tcmu_dev *udev = TCMU_DEV(dev);
1003 	struct tcmu_cmd *cmd;
1004 	bool all_expired = true;
1005 	int i;
1006 
1007 	del_timer_sync(&udev->timeout);
1008 
1009 	vfree(udev->mb_addr);
1010 
1011 	/* Upper layer should drain all requests before calling this */
1012 	spin_lock_irq(&udev->commands_lock);
1013 	idr_for_each_entry(&udev->commands, cmd, i) {
1014 		if (tcmu_check_and_free_pending_cmd(cmd) != 0)
1015 			all_expired = false;
1016 	}
1017 	idr_destroy(&udev->commands);
1018 	spin_unlock_irq(&udev->commands_lock);
1019 	WARN_ON(!all_expired);
1020 
1021 	/* Device was configured */
1022 	if (udev->uio_info.uio_dev) {
1023 		tcmu_netlink_event(TCMU_CMD_REMOVED_DEVICE, udev->uio_info.name,
1024 				   udev->uio_info.uio_dev->minor);
1025 
1026 		uio_unregister_device(&udev->uio_info);
1027 		kfree(udev->uio_info.name);
1028 		kfree(udev->name);
1029 	}
1030 	call_rcu(&dev->rcu_head, tcmu_dev_call_rcu);
1031 }
1032 
1033 enum {
1034 	Opt_dev_config, Opt_dev_size, Opt_hw_block_size, Opt_err,
1035 };
1036 
1037 static match_table_t tokens = {
1038 	{Opt_dev_config, "dev_config=%s"},
1039 	{Opt_dev_size, "dev_size=%u"},
1040 	{Opt_hw_block_size, "hw_block_size=%u"},
1041 	{Opt_err, NULL}
1042 };
1043 
1044 static ssize_t tcmu_set_configfs_dev_params(struct se_device *dev,
1045 		const char *page, ssize_t count)
1046 {
1047 	struct tcmu_dev *udev = TCMU_DEV(dev);
1048 	char *orig, *ptr, *opts, *arg_p;
1049 	substring_t args[MAX_OPT_ARGS];
1050 	int ret = 0, token;
1051 	unsigned long tmp_ul;
1052 
1053 	opts = kstrdup(page, GFP_KERNEL);
1054 	if (!opts)
1055 		return -ENOMEM;
1056 
1057 	orig = opts;
1058 
1059 	while ((ptr = strsep(&opts, ",\n")) != NULL) {
1060 		if (!*ptr)
1061 			continue;
1062 
1063 		token = match_token(ptr, tokens, args);
1064 		switch (token) {
1065 		case Opt_dev_config:
1066 			if (match_strlcpy(udev->dev_config, &args[0],
1067 					  TCMU_CONFIG_LEN) == 0) {
1068 				ret = -EINVAL;
1069 				break;
1070 			}
1071 			pr_debug("TCMU: Referencing Path: %s\n", udev->dev_config);
1072 			break;
1073 		case Opt_dev_size:
1074 			arg_p = match_strdup(&args[0]);
1075 			if (!arg_p) {
1076 				ret = -ENOMEM;
1077 				break;
1078 			}
1079 			ret = kstrtoul(arg_p, 0, (unsigned long *) &udev->dev_size);
1080 			kfree(arg_p);
1081 			if (ret < 0)
1082 				pr_err("kstrtoul() failed for dev_size=\n");
1083 			break;
1084 		case Opt_hw_block_size:
1085 			arg_p = match_strdup(&args[0]);
1086 			if (!arg_p) {
1087 				ret = -ENOMEM;
1088 				break;
1089 			}
1090 			ret = kstrtoul(arg_p, 0, &tmp_ul);
1091 			kfree(arg_p);
1092 			if (ret < 0) {
1093 				pr_err("kstrtoul() failed for hw_block_size=\n");
1094 				break;
1095 			}
1096 			if (!tmp_ul) {
1097 				pr_err("hw_block_size must be nonzero\n");
1098 				break;
1099 			}
1100 			dev->dev_attrib.hw_block_size = tmp_ul;
1101 			break;
1102 		default:
1103 			break;
1104 		}
1105 	}
1106 
1107 	kfree(orig);
1108 	return (!ret) ? count : ret;
1109 }
1110 
1111 static ssize_t tcmu_show_configfs_dev_params(struct se_device *dev, char *b)
1112 {
1113 	struct tcmu_dev *udev = TCMU_DEV(dev);
1114 	ssize_t bl = 0;
1115 
1116 	bl = sprintf(b + bl, "Config: %s ",
1117 		     udev->dev_config[0] ? udev->dev_config : "NULL");
1118 	bl += sprintf(b + bl, "Size: %zu\n", udev->dev_size);
1119 
1120 	return bl;
1121 }
1122 
1123 static sector_t tcmu_get_blocks(struct se_device *dev)
1124 {
1125 	struct tcmu_dev *udev = TCMU_DEV(dev);
1126 
1127 	return div_u64(udev->dev_size - dev->dev_attrib.block_size,
1128 		       dev->dev_attrib.block_size);
1129 }
1130 
1131 static sense_reason_t
1132 tcmu_pass_op(struct se_cmd *se_cmd)
1133 {
1134 	int ret = tcmu_queue_cmd(se_cmd);
1135 
1136 	if (ret != 0)
1137 		return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
1138 	else
1139 		return TCM_NO_SENSE;
1140 }
1141 
1142 static sense_reason_t
1143 tcmu_parse_cdb(struct se_cmd *cmd)
1144 {
1145 	return passthrough_parse_cdb(cmd, tcmu_pass_op);
1146 }
1147 
1148 static const struct target_backend_ops tcmu_ops = {
1149 	.name			= "user",
1150 	.owner			= THIS_MODULE,
1151 	.transport_flags	= TRANSPORT_FLAG_PASSTHROUGH,
1152 	.attach_hba		= tcmu_attach_hba,
1153 	.detach_hba		= tcmu_detach_hba,
1154 	.alloc_device		= tcmu_alloc_device,
1155 	.configure_device	= tcmu_configure_device,
1156 	.free_device		= tcmu_free_device,
1157 	.parse_cdb		= tcmu_parse_cdb,
1158 	.set_configfs_dev_params = tcmu_set_configfs_dev_params,
1159 	.show_configfs_dev_params = tcmu_show_configfs_dev_params,
1160 	.get_device_type	= sbc_get_device_type,
1161 	.get_blocks		= tcmu_get_blocks,
1162 	.tb_dev_attrib_attrs	= passthrough_attrib_attrs,
1163 };
1164 
1165 static int __init tcmu_module_init(void)
1166 {
1167 	int ret;
1168 
1169 	BUILD_BUG_ON((sizeof(struct tcmu_cmd_entry) % TCMU_OP_ALIGN_SIZE) != 0);
1170 
1171 	tcmu_cmd_cache = kmem_cache_create("tcmu_cmd_cache",
1172 				sizeof(struct tcmu_cmd),
1173 				__alignof__(struct tcmu_cmd),
1174 				0, NULL);
1175 	if (!tcmu_cmd_cache)
1176 		return -ENOMEM;
1177 
1178 	tcmu_root_device = root_device_register("tcm_user");
1179 	if (IS_ERR(tcmu_root_device)) {
1180 		ret = PTR_ERR(tcmu_root_device);
1181 		goto out_free_cache;
1182 	}
1183 
1184 	ret = genl_register_family(&tcmu_genl_family);
1185 	if (ret < 0) {
1186 		goto out_unreg_device;
1187 	}
1188 
1189 	ret = transport_backend_register(&tcmu_ops);
1190 	if (ret)
1191 		goto out_unreg_genl;
1192 
1193 	return 0;
1194 
1195 out_unreg_genl:
1196 	genl_unregister_family(&tcmu_genl_family);
1197 out_unreg_device:
1198 	root_device_unregister(tcmu_root_device);
1199 out_free_cache:
1200 	kmem_cache_destroy(tcmu_cmd_cache);
1201 
1202 	return ret;
1203 }
1204 
1205 static void __exit tcmu_module_exit(void)
1206 {
1207 	target_backend_unregister(&tcmu_ops);
1208 	genl_unregister_family(&tcmu_genl_family);
1209 	root_device_unregister(tcmu_root_device);
1210 	kmem_cache_destroy(tcmu_cmd_cache);
1211 }
1212 
1213 MODULE_DESCRIPTION("TCM USER subsystem plugin");
1214 MODULE_AUTHOR("Shaohua Li <shli@kernel.org>");
1215 MODULE_AUTHOR("Andy Grover <agrover@redhat.com>");
1216 MODULE_LICENSE("GPL");
1217 
1218 module_init(tcmu_module_init);
1219 module_exit(tcmu_module_exit);
1220