GNU Linux-libre 4.9.337-gnu1
[releases.git] / drivers / target / target_core_user.c
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 area 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, struct tcmu_cmd *cmd,
310                              bool bidi)
311 {
312         struct se_cmd *se_cmd = cmd->se_cmd;
313         int i, block;
314         int block_remaining = 0;
315         void *from, *to;
316         size_t copy_bytes, from_offset;
317         struct scatterlist *sg, *data_sg;
318         unsigned int data_nents;
319         DECLARE_BITMAP(bitmap, DATA_BLOCK_BITS);
320
321         bitmap_copy(bitmap, cmd->data_bitmap, DATA_BLOCK_BITS);
322
323         if (!bidi) {
324                 data_sg = se_cmd->t_data_sg;
325                 data_nents = se_cmd->t_data_nents;
326         } else {
327                 uint32_t count;
328
329                 /*
330                  * For bidi case, the first count blocks are for Data-Out
331                  * buffer blocks, and before gathering the Data-In buffer
332                  * the Data-Out buffer blocks should be discarded.
333                  */
334                 count = DIV_ROUND_UP(se_cmd->data_length, DATA_BLOCK_SIZE);
335                 while (count--) {
336                         block = find_first_bit(bitmap, DATA_BLOCK_BITS);
337                         clear_bit(block, bitmap);
338                 }
339
340                 data_sg = se_cmd->t_bidi_data_sg;
341                 data_nents = se_cmd->t_bidi_data_nents;
342         }
343
344         for_each_sg(data_sg, sg, data_nents, i) {
345                 int sg_remaining = sg->length;
346                 to = kmap_atomic(sg_page(sg)) + sg->offset;
347                 while (sg_remaining > 0) {
348                         if (block_remaining == 0) {
349                                 block = find_first_bit(bitmap,
350                                                 DATA_BLOCK_BITS);
351                                 block_remaining = DATA_BLOCK_SIZE;
352                                 clear_bit(block, bitmap);
353                         }
354                         copy_bytes = min_t(size_t, sg_remaining,
355                                         block_remaining);
356                         from_offset = get_block_offset(udev, block,
357                                         block_remaining);
358                         from = (void *) udev->mb_addr + from_offset;
359                         tcmu_flush_dcache_range(from, copy_bytes);
360                         memcpy(to + sg->length - sg_remaining, from,
361                                         copy_bytes);
362
363                         sg_remaining -= copy_bytes;
364                         block_remaining -= copy_bytes;
365                 }
366                 kunmap_atomic(to - sg->offset);
367         }
368 }
369
370 static inline size_t spc_bitmap_free(unsigned long *bitmap)
371 {
372         return DATA_BLOCK_SIZE * (DATA_BLOCK_BITS -
373                         bitmap_weight(bitmap, DATA_BLOCK_BITS));
374 }
375
376 /*
377  * We can't queue a command until we have space available on the cmd ring *and*
378  * space available on the data area.
379  *
380  * Called with ring lock held.
381  */
382 static bool is_ring_space_avail(struct tcmu_dev *udev, size_t cmd_size, size_t data_needed)
383 {
384         struct tcmu_mailbox *mb = udev->mb_addr;
385         size_t space, cmd_needed;
386         u32 cmd_head;
387
388         tcmu_flush_dcache_range(mb, sizeof(*mb));
389
390         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
391
392         /*
393          * If cmd end-of-ring space is too small then we need space for a NOP plus
394          * original cmd - cmds are internally contiguous.
395          */
396         if (head_to_end(cmd_head, udev->cmdr_size) >= cmd_size)
397                 cmd_needed = cmd_size;
398         else
399                 cmd_needed = cmd_size + head_to_end(cmd_head, udev->cmdr_size);
400
401         space = spc_free(cmd_head, udev->cmdr_last_cleaned, udev->cmdr_size);
402         if (space < cmd_needed) {
403                 pr_debug("no cmd space: %u %u %u\n", cmd_head,
404                        udev->cmdr_last_cleaned, udev->cmdr_size);
405                 return false;
406         }
407
408         space = spc_bitmap_free(udev->data_bitmap);
409         if (space < data_needed) {
410                 pr_debug("no data space: only %zu available, but ask for %zu\n",
411                                 space, data_needed);
412                 return false;
413         }
414
415         return true;
416 }
417
418 static inline size_t tcmu_cmd_get_data_length(struct tcmu_cmd *tcmu_cmd)
419 {
420         struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
421         size_t data_length = round_up(se_cmd->data_length, DATA_BLOCK_SIZE);
422
423         if (se_cmd->se_cmd_flags & SCF_BIDI) {
424                 BUG_ON(!(se_cmd->t_bidi_data_sg && se_cmd->t_bidi_data_nents));
425                 data_length += round_up(se_cmd->t_bidi_data_sg->length,
426                                 DATA_BLOCK_SIZE);
427         }
428
429         return data_length;
430 }
431
432 static inline uint32_t tcmu_cmd_get_block_cnt(struct tcmu_cmd *tcmu_cmd)
433 {
434         size_t data_length = tcmu_cmd_get_data_length(tcmu_cmd);
435
436         return data_length / DATA_BLOCK_SIZE;
437 }
438
439 static sense_reason_t
440 tcmu_queue_cmd_ring(struct tcmu_cmd *tcmu_cmd)
441 {
442         struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
443         struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
444         size_t base_command_size, command_size;
445         struct tcmu_mailbox *mb;
446         struct tcmu_cmd_entry *entry;
447         struct iovec *iov;
448         int iov_cnt;
449         uint32_t cmd_head;
450         uint64_t cdb_off;
451         bool copy_to_data_area;
452         size_t data_length = tcmu_cmd_get_data_length(tcmu_cmd);
453         DECLARE_BITMAP(old_bitmap, DATA_BLOCK_BITS);
454
455         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags))
456                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
457
458         /*
459          * Must be a certain minimum size for response sense info, but
460          * also may be larger if the iov array is large.
461          *
462          * We prepare way too many iovs for potential uses here, because it's
463          * expensive to tell how many regions are freed in the bitmap
464         */
465         base_command_size = max(offsetof(struct tcmu_cmd_entry,
466                                 req.iov[tcmu_cmd_get_block_cnt(tcmu_cmd)]),
467                                 sizeof(struct tcmu_cmd_entry));
468         command_size = base_command_size
469                 + round_up(scsi_command_size(se_cmd->t_task_cdb), TCMU_OP_ALIGN_SIZE);
470
471         WARN_ON(command_size & (TCMU_OP_ALIGN_SIZE-1));
472
473         spin_lock_irq(&udev->cmdr_lock);
474
475         mb = udev->mb_addr;
476         cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
477         if ((command_size > (udev->cmdr_size / 2)) ||
478             data_length > udev->data_size) {
479                 pr_warn("TCMU: Request of size %zu/%zu is too big for %u/%zu "
480                         "cmd ring/data area\n", command_size, data_length,
481                         udev->cmdr_size, udev->data_size);
482                 spin_unlock_irq(&udev->cmdr_lock);
483                 return TCM_INVALID_CDB_FIELD;
484         }
485
486         while (!is_ring_space_avail(udev, command_size, data_length)) {
487                 int ret;
488                 DEFINE_WAIT(__wait);
489
490                 prepare_to_wait(&udev->wait_cmdr, &__wait, TASK_INTERRUPTIBLE);
491
492                 pr_debug("sleeping for ring space\n");
493                 spin_unlock_irq(&udev->cmdr_lock);
494                 ret = schedule_timeout(msecs_to_jiffies(TCMU_TIME_OUT));
495                 finish_wait(&udev->wait_cmdr, &__wait);
496                 if (!ret) {
497                         pr_warn("tcmu: command timed out\n");
498                         return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
499                 }
500
501                 spin_lock_irq(&udev->cmdr_lock);
502
503                 /* We dropped cmdr_lock, cmd_head is stale */
504                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
505         }
506
507         /* Insert a PAD if end-of-ring space is too small */
508         if (head_to_end(cmd_head, udev->cmdr_size) < command_size) {
509                 size_t pad_size = head_to_end(cmd_head, udev->cmdr_size);
510
511                 entry = (void *) mb + CMDR_OFF + cmd_head;
512                 tcmu_flush_dcache_range(entry, sizeof(*entry));
513                 tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_PAD);
514                 tcmu_hdr_set_len(&entry->hdr.len_op, pad_size);
515                 entry->hdr.cmd_id = 0; /* not used for PAD */
516                 entry->hdr.kflags = 0;
517                 entry->hdr.uflags = 0;
518
519                 UPDATE_HEAD(mb->cmd_head, pad_size, udev->cmdr_size);
520
521                 cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
522                 WARN_ON(cmd_head != 0);
523         }
524
525         entry = (void *) mb + CMDR_OFF + cmd_head;
526         tcmu_flush_dcache_range(entry, sizeof(*entry));
527         tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_CMD);
528         tcmu_hdr_set_len(&entry->hdr.len_op, command_size);
529         entry->hdr.cmd_id = tcmu_cmd->cmd_id;
530         entry->hdr.kflags = 0;
531         entry->hdr.uflags = 0;
532
533         bitmap_copy(old_bitmap, udev->data_bitmap, DATA_BLOCK_BITS);
534
535         /* Handle allocating space from the data area */
536         iov = &entry->req.iov[0];
537         iov_cnt = 0;
538         copy_to_data_area = (se_cmd->data_direction == DMA_TO_DEVICE
539                 || se_cmd->se_cmd_flags & SCF_BIDI);
540         alloc_and_scatter_data_area(udev, se_cmd->t_data_sg,
541                 se_cmd->t_data_nents, &iov, &iov_cnt, copy_to_data_area);
542         entry->req.iov_cnt = iov_cnt;
543         entry->req.iov_dif_cnt = 0;
544
545         /* Handle BIDI commands */
546         if (se_cmd->se_cmd_flags & SCF_BIDI) {
547                 iov_cnt = 0;
548                 iov++;
549                 alloc_and_scatter_data_area(udev, se_cmd->t_bidi_data_sg,
550                                 se_cmd->t_bidi_data_nents, &iov, &iov_cnt,
551                                 false);
552                 entry->req.iov_bidi_cnt = iov_cnt;
553         }
554         /* cmd's data_bitmap is what changed in process */
555         bitmap_xor(tcmu_cmd->data_bitmap, old_bitmap, udev->data_bitmap,
556                         DATA_BLOCK_BITS);
557
558         /* All offsets relative to mb_addr, not start of entry! */
559         cdb_off = CMDR_OFF + cmd_head + base_command_size;
560         memcpy((void *) mb + cdb_off, se_cmd->t_task_cdb, scsi_command_size(se_cmd->t_task_cdb));
561         entry->req.cdb_off = cdb_off;
562         tcmu_flush_dcache_range(entry, sizeof(*entry));
563
564         UPDATE_HEAD(mb->cmd_head, command_size, udev->cmdr_size);
565         tcmu_flush_dcache_range(mb, sizeof(*mb));
566
567         spin_unlock_irq(&udev->cmdr_lock);
568
569         /* TODO: only if FLUSH and FUA? */
570         uio_event_notify(&udev->uio_info);
571
572         mod_timer(&udev->timeout,
573                 round_jiffies_up(jiffies + msecs_to_jiffies(TCMU_TIME_OUT)));
574
575         return TCM_NO_SENSE;
576 }
577
578 static sense_reason_t
579 tcmu_queue_cmd(struct se_cmd *se_cmd)
580 {
581         struct se_device *se_dev = se_cmd->se_dev;
582         struct tcmu_dev *udev = TCMU_DEV(se_dev);
583         struct tcmu_cmd *tcmu_cmd;
584         int ret;
585
586         tcmu_cmd = tcmu_alloc_cmd(se_cmd);
587         if (!tcmu_cmd)
588                 return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
589
590         ret = tcmu_queue_cmd_ring(tcmu_cmd);
591         if (ret != TCM_NO_SENSE) {
592                 pr_err("TCMU: Could not queue command\n");
593                 spin_lock_irq(&udev->commands_lock);
594                 idr_remove(&udev->commands, tcmu_cmd->cmd_id);
595                 spin_unlock_irq(&udev->commands_lock);
596
597                 kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
598         }
599
600         return ret;
601 }
602
603 static void tcmu_handle_completion(struct tcmu_cmd *cmd, struct tcmu_cmd_entry *entry)
604 {
605         struct se_cmd *se_cmd = cmd->se_cmd;
606         struct tcmu_dev *udev = cmd->tcmu_dev;
607
608         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags)) {
609                 /*
610                  * cmd has been completed already from timeout, just reclaim
611                  * data area space and free cmd
612                  */
613                 free_data_area(udev, cmd);
614
615                 kmem_cache_free(tcmu_cmd_cache, cmd);
616                 return;
617         }
618
619         if (entry->hdr.uflags & TCMU_UFLAG_UNKNOWN_OP) {
620                 free_data_area(udev, cmd);
621                 pr_warn("TCMU: Userspace set UNKNOWN_OP flag on se_cmd %p\n",
622                         cmd->se_cmd);
623                 entry->rsp.scsi_status = SAM_STAT_CHECK_CONDITION;
624         } else if (entry->rsp.scsi_status == SAM_STAT_CHECK_CONDITION) {
625                 memcpy(se_cmd->sense_buffer, entry->rsp.sense_buffer,
626                                se_cmd->scsi_sense_length);
627                 free_data_area(udev, cmd);
628         } else if (se_cmd->se_cmd_flags & SCF_BIDI) {
629                 /* Get Data-In buffer before clean up */
630                 gather_data_area(udev, cmd, true);
631                 free_data_area(udev, cmd);
632         } else if (se_cmd->data_direction == DMA_FROM_DEVICE) {
633                 gather_data_area(udev, cmd, false);
634                 free_data_area(udev, cmd);
635         } else if (se_cmd->data_direction == DMA_TO_DEVICE) {
636                 free_data_area(udev, cmd);
637         } else if (se_cmd->data_direction != DMA_NONE) {
638                 pr_warn("TCMU: data direction was %d!\n",
639                         se_cmd->data_direction);
640         }
641
642         target_complete_cmd(cmd->se_cmd, entry->rsp.scsi_status);
643         cmd->se_cmd = NULL;
644
645         kmem_cache_free(tcmu_cmd_cache, cmd);
646 }
647
648 static unsigned int tcmu_handle_completions(struct tcmu_dev *udev)
649 {
650         struct tcmu_mailbox *mb;
651         unsigned long flags;
652         int handled = 0;
653
654         if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags)) {
655                 pr_err("ring broken, not handling completions\n");
656                 return 0;
657         }
658
659         spin_lock_irqsave(&udev->cmdr_lock, flags);
660
661         mb = udev->mb_addr;
662         tcmu_flush_dcache_range(mb, sizeof(*mb));
663
664         while (udev->cmdr_last_cleaned != ACCESS_ONCE(mb->cmd_tail)) {
665
666                 struct tcmu_cmd_entry *entry = (void *) mb + CMDR_OFF + udev->cmdr_last_cleaned;
667                 struct tcmu_cmd *cmd;
668
669                 /*
670                  * Flush max. up to end of cmd ring since current entry might
671                  * be a padding that is shorter than sizeof(*entry)
672                  */
673                 size_t ring_left = head_to_end(udev->cmdr_last_cleaned,
674                                                udev->cmdr_size);
675                 tcmu_flush_dcache_range(entry, ring_left < sizeof(*entry) ?
676                                         ring_left : sizeof(*entry));
677
678                 if (tcmu_hdr_get_op(entry->hdr.len_op) == TCMU_OP_PAD) {
679                         UPDATE_HEAD(udev->cmdr_last_cleaned,
680                                     tcmu_hdr_get_len(entry->hdr.len_op),
681                                     udev->cmdr_size);
682                         continue;
683                 }
684                 WARN_ON(tcmu_hdr_get_op(entry->hdr.len_op) != TCMU_OP_CMD);
685
686                 spin_lock(&udev->commands_lock);
687                 cmd = idr_find(&udev->commands, entry->hdr.cmd_id);
688                 if (cmd)
689                         idr_remove(&udev->commands, cmd->cmd_id);
690                 spin_unlock(&udev->commands_lock);
691
692                 if (!cmd) {
693                         pr_err("cmd_id not found, ring is broken\n");
694                         set_bit(TCMU_DEV_BIT_BROKEN, &udev->flags);
695                         break;
696                 }
697
698                 tcmu_handle_completion(cmd, entry);
699
700                 UPDATE_HEAD(udev->cmdr_last_cleaned,
701                             tcmu_hdr_get_len(entry->hdr.len_op),
702                             udev->cmdr_size);
703
704                 handled++;
705         }
706
707         if (mb->cmd_tail == mb->cmd_head)
708                 del_timer(&udev->timeout); /* no more pending cmds */
709
710         spin_unlock_irqrestore(&udev->cmdr_lock, flags);
711
712         wake_up(&udev->wait_cmdr);
713
714         return handled;
715 }
716
717 static int tcmu_check_expired_cmd(int id, void *p, void *data)
718 {
719         struct tcmu_cmd *cmd = p;
720
721         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
722                 return 0;
723
724         if (!time_after(jiffies, cmd->deadline))
725                 return 0;
726
727         set_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags);
728         target_complete_cmd(cmd->se_cmd, SAM_STAT_CHECK_CONDITION);
729         cmd->se_cmd = NULL;
730
731         return 0;
732 }
733
734 static void tcmu_device_timedout(unsigned long data)
735 {
736         struct tcmu_dev *udev = (struct tcmu_dev *)data;
737         unsigned long flags;
738         int handled;
739
740         handled = tcmu_handle_completions(udev);
741
742         pr_warn("%d completions handled from timeout\n", handled);
743
744         spin_lock_irqsave(&udev->commands_lock, flags);
745         idr_for_each(&udev->commands, tcmu_check_expired_cmd, NULL);
746         spin_unlock_irqrestore(&udev->commands_lock, flags);
747
748         /*
749          * We don't need to wakeup threads on wait_cmdr since they have their
750          * own timeout.
751          */
752 }
753
754 static int tcmu_attach_hba(struct se_hba *hba, u32 host_id)
755 {
756         struct tcmu_hba *tcmu_hba;
757
758         tcmu_hba = kzalloc(sizeof(struct tcmu_hba), GFP_KERNEL);
759         if (!tcmu_hba)
760                 return -ENOMEM;
761
762         tcmu_hba->host_id = host_id;
763         hba->hba_ptr = tcmu_hba;
764
765         return 0;
766 }
767
768 static void tcmu_detach_hba(struct se_hba *hba)
769 {
770         kfree(hba->hba_ptr);
771         hba->hba_ptr = NULL;
772 }
773
774 static struct se_device *tcmu_alloc_device(struct se_hba *hba, const char *name)
775 {
776         struct tcmu_dev *udev;
777
778         udev = kzalloc(sizeof(struct tcmu_dev), GFP_KERNEL);
779         if (!udev)
780                 return NULL;
781
782         udev->name = kstrdup(name, GFP_KERNEL);
783         if (!udev->name) {
784                 kfree(udev);
785                 return NULL;
786         }
787
788         udev->hba = hba;
789
790         init_waitqueue_head(&udev->wait_cmdr);
791         spin_lock_init(&udev->cmdr_lock);
792
793         idr_init(&udev->commands);
794         spin_lock_init(&udev->commands_lock);
795
796         setup_timer(&udev->timeout, tcmu_device_timedout,
797                 (unsigned long)udev);
798
799         return &udev->se_dev;
800 }
801
802 static int tcmu_irqcontrol(struct uio_info *info, s32 irq_on)
803 {
804         struct tcmu_dev *tcmu_dev = container_of(info, struct tcmu_dev, uio_info);
805
806         tcmu_handle_completions(tcmu_dev);
807
808         return 0;
809 }
810
811 /*
812  * mmap code from uio.c. Copied here because we want to hook mmap()
813  * and this stuff must come along.
814  */
815 static int tcmu_find_mem_index(struct vm_area_struct *vma)
816 {
817         struct tcmu_dev *udev = vma->vm_private_data;
818         struct uio_info *info = &udev->uio_info;
819
820         if (vma->vm_pgoff < MAX_UIO_MAPS) {
821                 if (info->mem[vma->vm_pgoff].size == 0)
822                         return -1;
823                 return (int)vma->vm_pgoff;
824         }
825         return -1;
826 }
827
828 static int tcmu_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
829 {
830         struct tcmu_dev *udev = vma->vm_private_data;
831         struct uio_info *info = &udev->uio_info;
832         struct page *page;
833         unsigned long offset;
834         void *addr;
835
836         int mi = tcmu_find_mem_index(vma);
837         if (mi < 0)
838                 return VM_FAULT_SIGBUS;
839
840         /*
841          * We need to subtract mi because userspace uses offset = N*PAGE_SIZE
842          * to use mem[N].
843          */
844         offset = (vmf->pgoff - mi) << PAGE_SHIFT;
845
846         addr = (void *)(unsigned long)info->mem[mi].addr + offset;
847         if (info->mem[mi].memtype == UIO_MEM_LOGICAL)
848                 page = virt_to_page(addr);
849         else
850                 page = vmalloc_to_page(addr);
851         get_page(page);
852         vmf->page = page;
853         return 0;
854 }
855
856 static const struct vm_operations_struct tcmu_vm_ops = {
857         .fault = tcmu_vma_fault,
858 };
859
860 static int tcmu_mmap(struct uio_info *info, struct vm_area_struct *vma)
861 {
862         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
863
864         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
865         vma->vm_ops = &tcmu_vm_ops;
866
867         vma->vm_private_data = udev;
868
869         /* Ensure the mmap is exactly the right size */
870         if (vma_pages(vma) != (TCMU_RING_SIZE >> PAGE_SHIFT))
871                 return -EINVAL;
872
873         return 0;
874 }
875
876 static int tcmu_open(struct uio_info *info, struct inode *inode)
877 {
878         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
879
880         /* O_EXCL not supported for char devs, so fake it? */
881         if (test_and_set_bit(TCMU_DEV_BIT_OPEN, &udev->flags))
882                 return -EBUSY;
883
884         pr_debug("open\n");
885
886         return 0;
887 }
888
889 static int tcmu_release(struct uio_info *info, struct inode *inode)
890 {
891         struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
892
893         clear_bit(TCMU_DEV_BIT_OPEN, &udev->flags);
894
895         pr_debug("close\n");
896
897         return 0;
898 }
899
900 static int tcmu_netlink_event(enum tcmu_genl_cmd cmd, const char *name, int minor)
901 {
902         struct sk_buff *skb;
903         void *msg_header;
904         int ret = -ENOMEM;
905
906         skb = genlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
907         if (!skb)
908                 return ret;
909
910         msg_header = genlmsg_put(skb, 0, 0, &tcmu_genl_family, 0, cmd);
911         if (!msg_header)
912                 goto free_skb;
913
914         ret = nla_put_string(skb, TCMU_ATTR_DEVICE, name);
915         if (ret < 0)
916                 goto free_skb;
917
918         ret = nla_put_u32(skb, TCMU_ATTR_MINOR, minor);
919         if (ret < 0)
920                 goto free_skb;
921
922         genlmsg_end(skb, msg_header);
923
924         ret = genlmsg_multicast_allns(&tcmu_genl_family, skb, 0,
925                                 TCMU_MCGRP_CONFIG, GFP_KERNEL);
926
927         /* We don't care if no one is listening */
928         if (ret == -ESRCH)
929                 ret = 0;
930
931         return ret;
932 free_skb:
933         nlmsg_free(skb);
934         return ret;
935 }
936
937 static int tcmu_configure_device(struct se_device *dev)
938 {
939         struct tcmu_dev *udev = TCMU_DEV(dev);
940         struct tcmu_hba *hba = udev->hba->hba_ptr;
941         struct uio_info *info;
942         struct tcmu_mailbox *mb;
943         size_t size;
944         size_t used;
945         int ret = 0;
946         char *str;
947
948         info = &udev->uio_info;
949
950         size = snprintf(NULL, 0, "tcm-user/%u/%s/%s", hba->host_id, udev->name,
951                         udev->dev_config);
952         size += 1; /* for \0 */
953         str = kmalloc(size, GFP_KERNEL);
954         if (!str)
955                 return -ENOMEM;
956
957         used = snprintf(str, size, "tcm-user/%u/%s", hba->host_id, udev->name);
958
959         if (udev->dev_config[0])
960                 snprintf(str + used, size - used, "/%s", udev->dev_config);
961
962         info->name = str;
963
964         udev->mb_addr = vzalloc(TCMU_RING_SIZE);
965         if (!udev->mb_addr) {
966                 ret = -ENOMEM;
967                 goto err_vzalloc;
968         }
969
970         /* mailbox fits in first part of CMDR space */
971         udev->cmdr_size = CMDR_SIZE - CMDR_OFF;
972         udev->data_off = CMDR_SIZE;
973         udev->data_size = TCMU_RING_SIZE - CMDR_SIZE;
974
975         mb = udev->mb_addr;
976         mb->version = TCMU_MAILBOX_VERSION;
977         mb->flags = TCMU_MAILBOX_FLAG_CAP_OOOC;
978         mb->cmdr_off = CMDR_OFF;
979         mb->cmdr_size = udev->cmdr_size;
980
981         WARN_ON(!PAGE_ALIGNED(udev->data_off));
982         WARN_ON(udev->data_size % PAGE_SIZE);
983         WARN_ON(udev->data_size % DATA_BLOCK_SIZE);
984
985         info->version = __stringify(TCMU_MAILBOX_VERSION);
986
987         info->mem[0].name = "tcm-user command & data buffer";
988         info->mem[0].addr = (phys_addr_t)(uintptr_t)udev->mb_addr;
989         info->mem[0].size = TCMU_RING_SIZE;
990         info->mem[0].memtype = UIO_MEM_VIRTUAL;
991
992         info->irqcontrol = tcmu_irqcontrol;
993         info->irq = UIO_IRQ_CUSTOM;
994
995         info->mmap = tcmu_mmap;
996         info->open = tcmu_open;
997         info->release = tcmu_release;
998
999         ret = uio_register_device(tcmu_root_device, info);
1000         if (ret)
1001                 goto err_register;
1002
1003         /* User can set hw_block_size before enable the device */
1004         if (dev->dev_attrib.hw_block_size == 0)
1005                 dev->dev_attrib.hw_block_size = 512;
1006         /* Other attributes can be configured in userspace */
1007         dev->dev_attrib.hw_max_sectors = 128;
1008         dev->dev_attrib.hw_queue_depth = 128;
1009
1010         ret = tcmu_netlink_event(TCMU_CMD_ADDED_DEVICE, udev->uio_info.name,
1011                                  udev->uio_info.uio_dev->minor);
1012         if (ret)
1013                 goto err_netlink;
1014
1015         return 0;
1016
1017 err_netlink:
1018         uio_unregister_device(&udev->uio_info);
1019 err_register:
1020         vfree(udev->mb_addr);
1021 err_vzalloc:
1022         kfree(info->name);
1023
1024         return ret;
1025 }
1026
1027 static int tcmu_check_and_free_pending_cmd(struct tcmu_cmd *cmd)
1028 {
1029         if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags)) {
1030                 kmem_cache_free(tcmu_cmd_cache, cmd);
1031                 return 0;
1032         }
1033         return -EINVAL;
1034 }
1035
1036 static void tcmu_dev_call_rcu(struct rcu_head *p)
1037 {
1038         struct se_device *dev = container_of(p, struct se_device, rcu_head);
1039         struct tcmu_dev *udev = TCMU_DEV(dev);
1040
1041         kfree(udev);
1042 }
1043
1044 static void tcmu_free_device(struct se_device *dev)
1045 {
1046         struct tcmu_dev *udev = TCMU_DEV(dev);
1047         struct tcmu_cmd *cmd;
1048         bool all_expired = true;
1049         int i;
1050
1051         del_timer_sync(&udev->timeout);
1052
1053         vfree(udev->mb_addr);
1054
1055         /* Upper layer should drain all requests before calling this */
1056         spin_lock_irq(&udev->commands_lock);
1057         idr_for_each_entry(&udev->commands, cmd, i) {
1058                 if (tcmu_check_and_free_pending_cmd(cmd) != 0)
1059                         all_expired = false;
1060         }
1061         idr_destroy(&udev->commands);
1062         spin_unlock_irq(&udev->commands_lock);
1063         WARN_ON(!all_expired);
1064
1065         /* Device was configured */
1066         if (udev->uio_info.uio_dev) {
1067                 tcmu_netlink_event(TCMU_CMD_REMOVED_DEVICE, udev->uio_info.name,
1068                                    udev->uio_info.uio_dev->minor);
1069
1070                 uio_unregister_device(&udev->uio_info);
1071                 kfree(udev->uio_info.name);
1072                 kfree(udev->name);
1073         }
1074         call_rcu(&dev->rcu_head, tcmu_dev_call_rcu);
1075 }
1076
1077 enum {
1078         Opt_dev_config, Opt_dev_size, Opt_hw_block_size, Opt_err,
1079 };
1080
1081 static match_table_t tokens = {
1082         {Opt_dev_config, "dev_config=%s"},
1083         {Opt_dev_size, "dev_size=%u"},
1084         {Opt_hw_block_size, "hw_block_size=%u"},
1085         {Opt_err, NULL}
1086 };
1087
1088 static ssize_t tcmu_set_configfs_dev_params(struct se_device *dev,
1089                 const char *page, ssize_t count)
1090 {
1091         struct tcmu_dev *udev = TCMU_DEV(dev);
1092         char *orig, *ptr, *opts, *arg_p;
1093         substring_t args[MAX_OPT_ARGS];
1094         int ret = 0, token;
1095         unsigned long tmp_ul;
1096
1097         opts = kstrdup(page, GFP_KERNEL);
1098         if (!opts)
1099                 return -ENOMEM;
1100
1101         orig = opts;
1102
1103         while ((ptr = strsep(&opts, ",\n")) != NULL) {
1104                 if (!*ptr)
1105                         continue;
1106
1107                 token = match_token(ptr, tokens, args);
1108                 switch (token) {
1109                 case Opt_dev_config:
1110                         if (match_strlcpy(udev->dev_config, &args[0],
1111                                           TCMU_CONFIG_LEN) == 0) {
1112                                 ret = -EINVAL;
1113                                 break;
1114                         }
1115                         pr_debug("TCMU: Referencing Path: %s\n", udev->dev_config);
1116                         break;
1117                 case Opt_dev_size:
1118                         arg_p = match_strdup(&args[0]);
1119                         if (!arg_p) {
1120                                 ret = -ENOMEM;
1121                                 break;
1122                         }
1123                         ret = kstrtoul(arg_p, 0, (unsigned long *) &udev->dev_size);
1124                         kfree(arg_p);
1125                         if (ret < 0)
1126                                 pr_err("kstrtoul() failed for dev_size=\n");
1127                         break;
1128                 case Opt_hw_block_size:
1129                         arg_p = match_strdup(&args[0]);
1130                         if (!arg_p) {
1131                                 ret = -ENOMEM;
1132                                 break;
1133                         }
1134                         ret = kstrtoul(arg_p, 0, &tmp_ul);
1135                         kfree(arg_p);
1136                         if (ret < 0) {
1137                                 pr_err("kstrtoul() failed for hw_block_size=\n");
1138                                 break;
1139                         }
1140                         if (!tmp_ul) {
1141                                 pr_err("hw_block_size must be nonzero\n");
1142                                 break;
1143                         }
1144                         dev->dev_attrib.hw_block_size = tmp_ul;
1145                         break;
1146                 default:
1147                         break;
1148                 }
1149         }
1150
1151         kfree(orig);
1152         return (!ret) ? count : ret;
1153 }
1154
1155 static ssize_t tcmu_show_configfs_dev_params(struct se_device *dev, char *b)
1156 {
1157         struct tcmu_dev *udev = TCMU_DEV(dev);
1158         ssize_t bl = 0;
1159
1160         bl = sprintf(b + bl, "Config: %s ",
1161                      udev->dev_config[0] ? udev->dev_config : "NULL");
1162         bl += sprintf(b + bl, "Size: %zu\n", udev->dev_size);
1163
1164         return bl;
1165 }
1166
1167 static sector_t tcmu_get_blocks(struct se_device *dev)
1168 {
1169         struct tcmu_dev *udev = TCMU_DEV(dev);
1170
1171         return div_u64(udev->dev_size - dev->dev_attrib.block_size,
1172                        dev->dev_attrib.block_size);
1173 }
1174
1175 static sense_reason_t
1176 tcmu_parse_cdb(struct se_cmd *cmd)
1177 {
1178         return passthrough_parse_cdb(cmd, tcmu_queue_cmd);
1179 }
1180
1181 static const struct target_backend_ops tcmu_ops = {
1182         .name                   = "user",
1183         .owner                  = THIS_MODULE,
1184         .transport_flags        = TRANSPORT_FLAG_PASSTHROUGH,
1185         .attach_hba             = tcmu_attach_hba,
1186         .detach_hba             = tcmu_detach_hba,
1187         .alloc_device           = tcmu_alloc_device,
1188         .configure_device       = tcmu_configure_device,
1189         .free_device            = tcmu_free_device,
1190         .parse_cdb              = tcmu_parse_cdb,
1191         .set_configfs_dev_params = tcmu_set_configfs_dev_params,
1192         .show_configfs_dev_params = tcmu_show_configfs_dev_params,
1193         .get_device_type        = sbc_get_device_type,
1194         .get_blocks             = tcmu_get_blocks,
1195         .tb_dev_attrib_attrs    = passthrough_attrib_attrs,
1196 };
1197
1198 static int __init tcmu_module_init(void)
1199 {
1200         int ret;
1201
1202         BUILD_BUG_ON((sizeof(struct tcmu_cmd_entry) % TCMU_OP_ALIGN_SIZE) != 0);
1203
1204         tcmu_cmd_cache = kmem_cache_create("tcmu_cmd_cache",
1205                                 sizeof(struct tcmu_cmd),
1206                                 __alignof__(struct tcmu_cmd),
1207                                 0, NULL);
1208         if (!tcmu_cmd_cache)
1209                 return -ENOMEM;
1210
1211         tcmu_root_device = root_device_register("tcm_user");
1212         if (IS_ERR(tcmu_root_device)) {
1213                 ret = PTR_ERR(tcmu_root_device);
1214                 goto out_free_cache;
1215         }
1216
1217         ret = genl_register_family(&tcmu_genl_family);
1218         if (ret < 0) {
1219                 goto out_unreg_device;
1220         }
1221
1222         ret = transport_backend_register(&tcmu_ops);
1223         if (ret)
1224                 goto out_unreg_genl;
1225
1226         return 0;
1227
1228 out_unreg_genl:
1229         genl_unregister_family(&tcmu_genl_family);
1230 out_unreg_device:
1231         root_device_unregister(tcmu_root_device);
1232 out_free_cache:
1233         kmem_cache_destroy(tcmu_cmd_cache);
1234
1235         return ret;
1236 }
1237
1238 static void __exit tcmu_module_exit(void)
1239 {
1240         target_backend_unregister(&tcmu_ops);
1241         genl_unregister_family(&tcmu_genl_family);
1242         root_device_unregister(tcmu_root_device);
1243         kmem_cache_destroy(tcmu_cmd_cache);
1244 }
1245
1246 MODULE_DESCRIPTION("TCM USER subsystem plugin");
1247 MODULE_AUTHOR("Shaohua Li <shli@kernel.org>");
1248 MODULE_AUTHOR("Andy Grover <agrover@redhat.com>");
1249 MODULE_LICENSE("GPL");
1250
1251 module_init(tcmu_module_init);
1252 module_exit(tcmu_module_exit);