]> git.karo-electronics.de Git - karo-tx-linux.git/blob - drivers/scsi/sd.c
xfs: fix spurious spin_is_locked() assert failures on non-smp kernels
[karo-tx-linux.git] / drivers / scsi / sd.c
1 /*
2  *      sd.c Copyright (C) 1992 Drew Eckhardt
3  *           Copyright (C) 1993, 1994, 1995, 1999 Eric Youngdale
4  *
5  *      Linux scsi disk driver
6  *              Initial versions: Drew Eckhardt
7  *              Subsequent revisions: Eric Youngdale
8  *      Modification history:
9  *       - Drew Eckhardt <drew@colorado.edu> original
10  *       - Eric Youngdale <eric@andante.org> add scatter-gather, multiple 
11  *         outstanding request, and other enhancements.
12  *         Support loadable low-level scsi drivers.
13  *       - Jirka Hanika <geo@ff.cuni.cz> support more scsi disks using 
14  *         eight major numbers.
15  *       - Richard Gooch <rgooch@atnf.csiro.au> support devfs.
16  *       - Torben Mathiasen <tmm@image.dk> Resource allocation fixes in 
17  *         sd_init and cleanups.
18  *       - Alex Davis <letmein@erols.com> Fix problem where partition info
19  *         not being read in sd_open. Fix problem where removable media 
20  *         could be ejected after sd_open.
21  *       - Douglas Gilbert <dgilbert@interlog.com> cleanup for lk 2.5.x
22  *       - Badari Pulavarty <pbadari@us.ibm.com>, Matthew Wilcox 
23  *         <willy@debian.org>, Kurt Garloff <garloff@suse.de>: 
24  *         Support 32k/1M disks.
25  *
26  *      Logging policy (needs CONFIG_SCSI_LOGGING defined):
27  *       - setting up transfer: SCSI_LOG_HLQUEUE levels 1 and 2
28  *       - end of transfer (bh + scsi_lib): SCSI_LOG_HLCOMPLETE level 1
29  *       - entering sd_ioctl: SCSI_LOG_IOCTL level 1
30  *       - entering other commands: SCSI_LOG_HLQUEUE level 3
31  *      Note: when the logging level is set by the user, it must be greater
32  *      than the level indicated above to trigger output.       
33  */
34
35 #include <linux/module.h>
36 #include <linux/fs.h>
37 #include <linux/kernel.h>
38 #include <linux/mm.h>
39 #include <linux/bio.h>
40 #include <linux/genhd.h>
41 #include <linux/hdreg.h>
42 #include <linux/errno.h>
43 #include <linux/idr.h>
44 #include <linux/interrupt.h>
45 #include <linux/init.h>
46 #include <linux/blkdev.h>
47 #include <linux/blkpg.h>
48 #include <linux/delay.h>
49 #include <linux/mutex.h>
50 #include <linux/string_helpers.h>
51 #include <linux/async.h>
52 #include <linux/slab.h>
53 #include <linux/pm_runtime.h>
54 #include <linux/pr.h>
55 #include <linux/t10-pi.h>
56 #include <linux/uaccess.h>
57 #include <asm/unaligned.h>
58
59 #include <scsi/scsi.h>
60 #include <scsi/scsi_cmnd.h>
61 #include <scsi/scsi_dbg.h>
62 #include <scsi/scsi_device.h>
63 #include <scsi/scsi_driver.h>
64 #include <scsi/scsi_eh.h>
65 #include <scsi/scsi_host.h>
66 #include <scsi/scsi_ioctl.h>
67 #include <scsi/scsicam.h>
68
69 #include "sd.h"
70 #include "scsi_priv.h"
71 #include "scsi_logging.h"
72
73 MODULE_AUTHOR("Eric Youngdale");
74 MODULE_DESCRIPTION("SCSI disk (sd) driver");
75 MODULE_LICENSE("GPL");
76
77 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK0_MAJOR);
78 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK1_MAJOR);
79 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK2_MAJOR);
80 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK3_MAJOR);
81 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK4_MAJOR);
82 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK5_MAJOR);
83 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK6_MAJOR);
84 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK7_MAJOR);
85 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK8_MAJOR);
86 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK9_MAJOR);
87 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK10_MAJOR);
88 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK11_MAJOR);
89 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK12_MAJOR);
90 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK13_MAJOR);
91 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK14_MAJOR);
92 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK15_MAJOR);
93 MODULE_ALIAS_SCSI_DEVICE(TYPE_DISK);
94 MODULE_ALIAS_SCSI_DEVICE(TYPE_MOD);
95 MODULE_ALIAS_SCSI_DEVICE(TYPE_RBC);
96 MODULE_ALIAS_SCSI_DEVICE(TYPE_ZBC);
97
98 #if !defined(CONFIG_DEBUG_BLOCK_EXT_DEVT)
99 #define SD_MINORS       16
100 #else
101 #define SD_MINORS       0
102 #endif
103
104 static void sd_config_discard(struct scsi_disk *, unsigned int);
105 static void sd_config_write_same(struct scsi_disk *);
106 static int  sd_revalidate_disk(struct gendisk *);
107 static void sd_unlock_native_capacity(struct gendisk *disk);
108 static int  sd_probe(struct device *);
109 static int  sd_remove(struct device *);
110 static void sd_shutdown(struct device *);
111 static int sd_suspend_system(struct device *);
112 static int sd_suspend_runtime(struct device *);
113 static int sd_resume(struct device *);
114 static void sd_rescan(struct device *);
115 static int sd_init_command(struct scsi_cmnd *SCpnt);
116 static void sd_uninit_command(struct scsi_cmnd *SCpnt);
117 static int sd_done(struct scsi_cmnd *);
118 static void sd_eh_reset(struct scsi_cmnd *);
119 static int sd_eh_action(struct scsi_cmnd *, int);
120 static void sd_read_capacity(struct scsi_disk *sdkp, unsigned char *buffer);
121 static void scsi_disk_release(struct device *cdev);
122 static void sd_print_sense_hdr(struct scsi_disk *, struct scsi_sense_hdr *);
123 static void sd_print_result(const struct scsi_disk *, const char *, int);
124
125 static DEFINE_SPINLOCK(sd_index_lock);
126 static DEFINE_IDA(sd_index_ida);
127
128 /* This semaphore is used to mediate the 0->1 reference get in the
129  * face of object destruction (i.e. we can't allow a get on an
130  * object after last put) */
131 static DEFINE_MUTEX(sd_ref_mutex);
132
133 static struct kmem_cache *sd_cdb_cache;
134 static mempool_t *sd_cdb_pool;
135
136 static const char *sd_cache_types[] = {
137         "write through", "none", "write back",
138         "write back, no read (daft)"
139 };
140
141 static void sd_set_flush_flag(struct scsi_disk *sdkp)
142 {
143         bool wc = false, fua = false;
144
145         if (sdkp->WCE) {
146                 wc = true;
147                 if (sdkp->DPOFUA)
148                         fua = true;
149         }
150
151         blk_queue_write_cache(sdkp->disk->queue, wc, fua);
152 }
153
154 static ssize_t
155 cache_type_store(struct device *dev, struct device_attribute *attr,
156                  const char *buf, size_t count)
157 {
158         int i, ct = -1, rcd, wce, sp;
159         struct scsi_disk *sdkp = to_scsi_disk(dev);
160         struct scsi_device *sdp = sdkp->device;
161         char buffer[64];
162         char *buffer_data;
163         struct scsi_mode_data data;
164         struct scsi_sense_hdr sshdr;
165         static const char temp[] = "temporary ";
166         int len;
167
168         if (sdp->type != TYPE_DISK && sdp->type != TYPE_ZBC)
169                 /* no cache control on RBC devices; theoretically they
170                  * can do it, but there's probably so many exceptions
171                  * it's not worth the risk */
172                 return -EINVAL;
173
174         if (strncmp(buf, temp, sizeof(temp) - 1) == 0) {
175                 buf += sizeof(temp) - 1;
176                 sdkp->cache_override = 1;
177         } else {
178                 sdkp->cache_override = 0;
179         }
180
181         for (i = 0; i < ARRAY_SIZE(sd_cache_types); i++) {
182                 len = strlen(sd_cache_types[i]);
183                 if (strncmp(sd_cache_types[i], buf, len) == 0 &&
184                     buf[len] == '\n') {
185                         ct = i;
186                         break;
187                 }
188         }
189         if (ct < 0)
190                 return -EINVAL;
191         rcd = ct & 0x01 ? 1 : 0;
192         wce = (ct & 0x02) && !sdkp->write_prot ? 1 : 0;
193
194         if (sdkp->cache_override) {
195                 sdkp->WCE = wce;
196                 sdkp->RCD = rcd;
197                 sd_set_flush_flag(sdkp);
198                 return count;
199         }
200
201         if (scsi_mode_sense(sdp, 0x08, 8, buffer, sizeof(buffer), SD_TIMEOUT,
202                             SD_MAX_RETRIES, &data, NULL))
203                 return -EINVAL;
204         len = min_t(size_t, sizeof(buffer), data.length - data.header_length -
205                   data.block_descriptor_length);
206         buffer_data = buffer + data.header_length +
207                 data.block_descriptor_length;
208         buffer_data[2] &= ~0x05;
209         buffer_data[2] |= wce << 2 | rcd;
210         sp = buffer_data[0] & 0x80 ? 1 : 0;
211         buffer_data[0] &= ~0x80;
212
213         if (scsi_mode_select(sdp, 1, sp, 8, buffer_data, len, SD_TIMEOUT,
214                              SD_MAX_RETRIES, &data, &sshdr)) {
215                 if (scsi_sense_valid(&sshdr))
216                         sd_print_sense_hdr(sdkp, &sshdr);
217                 return -EINVAL;
218         }
219         revalidate_disk(sdkp->disk);
220         return count;
221 }
222
223 static ssize_t
224 manage_start_stop_show(struct device *dev, struct device_attribute *attr,
225                        char *buf)
226 {
227         struct scsi_disk *sdkp = to_scsi_disk(dev);
228         struct scsi_device *sdp = sdkp->device;
229
230         return snprintf(buf, 20, "%u\n", sdp->manage_start_stop);
231 }
232
233 static ssize_t
234 manage_start_stop_store(struct device *dev, struct device_attribute *attr,
235                         const char *buf, size_t count)
236 {
237         struct scsi_disk *sdkp = to_scsi_disk(dev);
238         struct scsi_device *sdp = sdkp->device;
239
240         if (!capable(CAP_SYS_ADMIN))
241                 return -EACCES;
242
243         sdp->manage_start_stop = simple_strtoul(buf, NULL, 10);
244
245         return count;
246 }
247 static DEVICE_ATTR_RW(manage_start_stop);
248
249 static ssize_t
250 allow_restart_show(struct device *dev, struct device_attribute *attr, char *buf)
251 {
252         struct scsi_disk *sdkp = to_scsi_disk(dev);
253
254         return snprintf(buf, 40, "%d\n", sdkp->device->allow_restart);
255 }
256
257 static ssize_t
258 allow_restart_store(struct device *dev, struct device_attribute *attr,
259                     const char *buf, size_t count)
260 {
261         struct scsi_disk *sdkp = to_scsi_disk(dev);
262         struct scsi_device *sdp = sdkp->device;
263
264         if (!capable(CAP_SYS_ADMIN))
265                 return -EACCES;
266
267         if (sdp->type != TYPE_DISK && sdp->type != TYPE_ZBC)
268                 return -EINVAL;
269
270         sdp->allow_restart = simple_strtoul(buf, NULL, 10);
271
272         return count;
273 }
274 static DEVICE_ATTR_RW(allow_restart);
275
276 static ssize_t
277 cache_type_show(struct device *dev, struct device_attribute *attr, char *buf)
278 {
279         struct scsi_disk *sdkp = to_scsi_disk(dev);
280         int ct = sdkp->RCD + 2*sdkp->WCE;
281
282         return snprintf(buf, 40, "%s\n", sd_cache_types[ct]);
283 }
284 static DEVICE_ATTR_RW(cache_type);
285
286 static ssize_t
287 FUA_show(struct device *dev, struct device_attribute *attr, char *buf)
288 {
289         struct scsi_disk *sdkp = to_scsi_disk(dev);
290
291         return snprintf(buf, 20, "%u\n", sdkp->DPOFUA);
292 }
293 static DEVICE_ATTR_RO(FUA);
294
295 static ssize_t
296 protection_type_show(struct device *dev, struct device_attribute *attr,
297                      char *buf)
298 {
299         struct scsi_disk *sdkp = to_scsi_disk(dev);
300
301         return snprintf(buf, 20, "%u\n", sdkp->protection_type);
302 }
303
304 static ssize_t
305 protection_type_store(struct device *dev, struct device_attribute *attr,
306                       const char *buf, size_t count)
307 {
308         struct scsi_disk *sdkp = to_scsi_disk(dev);
309         unsigned int val;
310         int err;
311
312         if (!capable(CAP_SYS_ADMIN))
313                 return -EACCES;
314
315         err = kstrtouint(buf, 10, &val);
316
317         if (err)
318                 return err;
319
320         if (val >= 0 && val <= T10_PI_TYPE3_PROTECTION)
321                 sdkp->protection_type = val;
322
323         return count;
324 }
325 static DEVICE_ATTR_RW(protection_type);
326
327 static ssize_t
328 protection_mode_show(struct device *dev, struct device_attribute *attr,
329                      char *buf)
330 {
331         struct scsi_disk *sdkp = to_scsi_disk(dev);
332         struct scsi_device *sdp = sdkp->device;
333         unsigned int dif, dix;
334
335         dif = scsi_host_dif_capable(sdp->host, sdkp->protection_type);
336         dix = scsi_host_dix_capable(sdp->host, sdkp->protection_type);
337
338         if (!dix && scsi_host_dix_capable(sdp->host, T10_PI_TYPE0_PROTECTION)) {
339                 dif = 0;
340                 dix = 1;
341         }
342
343         if (!dif && !dix)
344                 return snprintf(buf, 20, "none\n");
345
346         return snprintf(buf, 20, "%s%u\n", dix ? "dix" : "dif", dif);
347 }
348 static DEVICE_ATTR_RO(protection_mode);
349
350 static ssize_t
351 app_tag_own_show(struct device *dev, struct device_attribute *attr, char *buf)
352 {
353         struct scsi_disk *sdkp = to_scsi_disk(dev);
354
355         return snprintf(buf, 20, "%u\n", sdkp->ATO);
356 }
357 static DEVICE_ATTR_RO(app_tag_own);
358
359 static ssize_t
360 thin_provisioning_show(struct device *dev, struct device_attribute *attr,
361                        char *buf)
362 {
363         struct scsi_disk *sdkp = to_scsi_disk(dev);
364
365         return snprintf(buf, 20, "%u\n", sdkp->lbpme);
366 }
367 static DEVICE_ATTR_RO(thin_provisioning);
368
369 static const char *lbp_mode[] = {
370         [SD_LBP_FULL]           = "full",
371         [SD_LBP_UNMAP]          = "unmap",
372         [SD_LBP_WS16]           = "writesame_16",
373         [SD_LBP_WS10]           = "writesame_10",
374         [SD_LBP_ZERO]           = "writesame_zero",
375         [SD_LBP_DISABLE]        = "disabled",
376 };
377
378 static ssize_t
379 provisioning_mode_show(struct device *dev, struct device_attribute *attr,
380                        char *buf)
381 {
382         struct scsi_disk *sdkp = to_scsi_disk(dev);
383
384         return snprintf(buf, 20, "%s\n", lbp_mode[sdkp->provisioning_mode]);
385 }
386
387 static ssize_t
388 provisioning_mode_store(struct device *dev, struct device_attribute *attr,
389                         const char *buf, size_t count)
390 {
391         struct scsi_disk *sdkp = to_scsi_disk(dev);
392         struct scsi_device *sdp = sdkp->device;
393
394         if (!capable(CAP_SYS_ADMIN))
395                 return -EACCES;
396
397         if (sd_is_zoned(sdkp)) {
398                 sd_config_discard(sdkp, SD_LBP_DISABLE);
399                 return count;
400         }
401
402         if (sdp->type != TYPE_DISK)
403                 return -EINVAL;
404
405         if (!strncmp(buf, lbp_mode[SD_LBP_UNMAP], 20))
406                 sd_config_discard(sdkp, SD_LBP_UNMAP);
407         else if (!strncmp(buf, lbp_mode[SD_LBP_WS16], 20))
408                 sd_config_discard(sdkp, SD_LBP_WS16);
409         else if (!strncmp(buf, lbp_mode[SD_LBP_WS10], 20))
410                 sd_config_discard(sdkp, SD_LBP_WS10);
411         else if (!strncmp(buf, lbp_mode[SD_LBP_ZERO], 20))
412                 sd_config_discard(sdkp, SD_LBP_ZERO);
413         else if (!strncmp(buf, lbp_mode[SD_LBP_DISABLE], 20))
414                 sd_config_discard(sdkp, SD_LBP_DISABLE);
415         else
416                 return -EINVAL;
417
418         return count;
419 }
420 static DEVICE_ATTR_RW(provisioning_mode);
421
422 static const char *zeroing_mode[] = {
423         [SD_ZERO_WRITE]         = "write",
424         [SD_ZERO_WS]            = "writesame",
425         [SD_ZERO_WS16_UNMAP]    = "writesame_16_unmap",
426         [SD_ZERO_WS10_UNMAP]    = "writesame_10_unmap",
427 };
428
429 static ssize_t
430 zeroing_mode_show(struct device *dev, struct device_attribute *attr,
431                   char *buf)
432 {
433         struct scsi_disk *sdkp = to_scsi_disk(dev);
434
435         return snprintf(buf, 20, "%s\n", zeroing_mode[sdkp->zeroing_mode]);
436 }
437
438 static ssize_t
439 zeroing_mode_store(struct device *dev, struct device_attribute *attr,
440                    const char *buf, size_t count)
441 {
442         struct scsi_disk *sdkp = to_scsi_disk(dev);
443
444         if (!capable(CAP_SYS_ADMIN))
445                 return -EACCES;
446
447         if (!strncmp(buf, zeroing_mode[SD_ZERO_WRITE], 20))
448                 sdkp->zeroing_mode = SD_ZERO_WRITE;
449         else if (!strncmp(buf, zeroing_mode[SD_ZERO_WS], 20))
450                 sdkp->zeroing_mode = SD_ZERO_WS;
451         else if (!strncmp(buf, zeroing_mode[SD_ZERO_WS16_UNMAP], 20))
452                 sdkp->zeroing_mode = SD_ZERO_WS16_UNMAP;
453         else if (!strncmp(buf, zeroing_mode[SD_ZERO_WS10_UNMAP], 20))
454                 sdkp->zeroing_mode = SD_ZERO_WS10_UNMAP;
455         else
456                 return -EINVAL;
457
458         return count;
459 }
460 static DEVICE_ATTR_RW(zeroing_mode);
461
462 static ssize_t
463 max_medium_access_timeouts_show(struct device *dev,
464                                 struct device_attribute *attr, char *buf)
465 {
466         struct scsi_disk *sdkp = to_scsi_disk(dev);
467
468         return snprintf(buf, 20, "%u\n", sdkp->max_medium_access_timeouts);
469 }
470
471 static ssize_t
472 max_medium_access_timeouts_store(struct device *dev,
473                                  struct device_attribute *attr, const char *buf,
474                                  size_t count)
475 {
476         struct scsi_disk *sdkp = to_scsi_disk(dev);
477         int err;
478
479         if (!capable(CAP_SYS_ADMIN))
480                 return -EACCES;
481
482         err = kstrtouint(buf, 10, &sdkp->max_medium_access_timeouts);
483
484         return err ? err : count;
485 }
486 static DEVICE_ATTR_RW(max_medium_access_timeouts);
487
488 static ssize_t
489 max_write_same_blocks_show(struct device *dev, struct device_attribute *attr,
490                            char *buf)
491 {
492         struct scsi_disk *sdkp = to_scsi_disk(dev);
493
494         return snprintf(buf, 20, "%u\n", sdkp->max_ws_blocks);
495 }
496
497 static ssize_t
498 max_write_same_blocks_store(struct device *dev, struct device_attribute *attr,
499                             const char *buf, size_t count)
500 {
501         struct scsi_disk *sdkp = to_scsi_disk(dev);
502         struct scsi_device *sdp = sdkp->device;
503         unsigned long max;
504         int err;
505
506         if (!capable(CAP_SYS_ADMIN))
507                 return -EACCES;
508
509         if (sdp->type != TYPE_DISK && sdp->type != TYPE_ZBC)
510                 return -EINVAL;
511
512         err = kstrtoul(buf, 10, &max);
513
514         if (err)
515                 return err;
516
517         if (max == 0)
518                 sdp->no_write_same = 1;
519         else if (max <= SD_MAX_WS16_BLOCKS) {
520                 sdp->no_write_same = 0;
521                 sdkp->max_ws_blocks = max;
522         }
523
524         sd_config_write_same(sdkp);
525
526         return count;
527 }
528 static DEVICE_ATTR_RW(max_write_same_blocks);
529
530 static struct attribute *sd_disk_attrs[] = {
531         &dev_attr_cache_type.attr,
532         &dev_attr_FUA.attr,
533         &dev_attr_allow_restart.attr,
534         &dev_attr_manage_start_stop.attr,
535         &dev_attr_protection_type.attr,
536         &dev_attr_protection_mode.attr,
537         &dev_attr_app_tag_own.attr,
538         &dev_attr_thin_provisioning.attr,
539         &dev_attr_provisioning_mode.attr,
540         &dev_attr_zeroing_mode.attr,
541         &dev_attr_max_write_same_blocks.attr,
542         &dev_attr_max_medium_access_timeouts.attr,
543         NULL,
544 };
545 ATTRIBUTE_GROUPS(sd_disk);
546
547 static struct class sd_disk_class = {
548         .name           = "scsi_disk",
549         .owner          = THIS_MODULE,
550         .dev_release    = scsi_disk_release,
551         .dev_groups     = sd_disk_groups,
552 };
553
554 static const struct dev_pm_ops sd_pm_ops = {
555         .suspend                = sd_suspend_system,
556         .resume                 = sd_resume,
557         .poweroff               = sd_suspend_system,
558         .restore                = sd_resume,
559         .runtime_suspend        = sd_suspend_runtime,
560         .runtime_resume         = sd_resume,
561 };
562
563 static struct scsi_driver sd_template = {
564         .gendrv = {
565                 .name           = "sd",
566                 .owner          = THIS_MODULE,
567                 .probe          = sd_probe,
568                 .remove         = sd_remove,
569                 .shutdown       = sd_shutdown,
570                 .pm             = &sd_pm_ops,
571         },
572         .rescan                 = sd_rescan,
573         .init_command           = sd_init_command,
574         .uninit_command         = sd_uninit_command,
575         .done                   = sd_done,
576         .eh_action              = sd_eh_action,
577         .eh_reset               = sd_eh_reset,
578 };
579
580 /*
581  * Dummy kobj_map->probe function.
582  * The default ->probe function will call modprobe, which is
583  * pointless as this module is already loaded.
584  */
585 static struct kobject *sd_default_probe(dev_t devt, int *partno, void *data)
586 {
587         return NULL;
588 }
589
590 /*
591  * Device no to disk mapping:
592  * 
593  *       major         disc2     disc  p1
594  *   |............|.............|....|....| <- dev_t
595  *    31        20 19          8 7  4 3  0
596  * 
597  * Inside a major, we have 16k disks, however mapped non-
598  * contiguously. The first 16 disks are for major0, the next
599  * ones with major1, ... Disk 256 is for major0 again, disk 272 
600  * for major1, ... 
601  * As we stay compatible with our numbering scheme, we can reuse 
602  * the well-know SCSI majors 8, 65--71, 136--143.
603  */
604 static int sd_major(int major_idx)
605 {
606         switch (major_idx) {
607         case 0:
608                 return SCSI_DISK0_MAJOR;
609         case 1 ... 7:
610                 return SCSI_DISK1_MAJOR + major_idx - 1;
611         case 8 ... 15:
612                 return SCSI_DISK8_MAJOR + major_idx - 8;
613         default:
614                 BUG();
615                 return 0;       /* shut up gcc */
616         }
617 }
618
619 static struct scsi_disk *scsi_disk_get(struct gendisk *disk)
620 {
621         struct scsi_disk *sdkp = NULL;
622
623         mutex_lock(&sd_ref_mutex);
624
625         if (disk->private_data) {
626                 sdkp = scsi_disk(disk);
627                 if (scsi_device_get(sdkp->device) == 0)
628                         get_device(&sdkp->dev);
629                 else
630                         sdkp = NULL;
631         }
632         mutex_unlock(&sd_ref_mutex);
633         return sdkp;
634 }
635
636 static void scsi_disk_put(struct scsi_disk *sdkp)
637 {
638         struct scsi_device *sdev = sdkp->device;
639
640         mutex_lock(&sd_ref_mutex);
641         put_device(&sdkp->dev);
642         scsi_device_put(sdev);
643         mutex_unlock(&sd_ref_mutex);
644 }
645
646 static unsigned char sd_setup_protect_cmnd(struct scsi_cmnd *scmd,
647                                            unsigned int dix, unsigned int dif)
648 {
649         struct bio *bio = scmd->request->bio;
650         unsigned int prot_op = sd_prot_op(rq_data_dir(scmd->request), dix, dif);
651         unsigned int protect = 0;
652
653         if (dix) {                              /* DIX Type 0, 1, 2, 3 */
654                 if (bio_integrity_flagged(bio, BIP_IP_CHECKSUM))
655                         scmd->prot_flags |= SCSI_PROT_IP_CHECKSUM;
656
657                 if (bio_integrity_flagged(bio, BIP_CTRL_NOCHECK) == false)
658                         scmd->prot_flags |= SCSI_PROT_GUARD_CHECK;
659         }
660
661         if (dif != T10_PI_TYPE3_PROTECTION) {   /* DIX/DIF Type 0, 1, 2 */
662                 scmd->prot_flags |= SCSI_PROT_REF_INCREMENT;
663
664                 if (bio_integrity_flagged(bio, BIP_CTRL_NOCHECK) == false)
665                         scmd->prot_flags |= SCSI_PROT_REF_CHECK;
666         }
667
668         if (dif) {                              /* DIX/DIF Type 1, 2, 3 */
669                 scmd->prot_flags |= SCSI_PROT_TRANSFER_PI;
670
671                 if (bio_integrity_flagged(bio, BIP_DISK_NOCHECK))
672                         protect = 3 << 5;       /* Disable target PI checking */
673                 else
674                         protect = 1 << 5;       /* Enable target PI checking */
675         }
676
677         scsi_set_prot_op(scmd, prot_op);
678         scsi_set_prot_type(scmd, dif);
679         scmd->prot_flags &= sd_prot_flag_mask(prot_op);
680
681         return protect;
682 }
683
684 static void sd_config_discard(struct scsi_disk *sdkp, unsigned int mode)
685 {
686         struct request_queue *q = sdkp->disk->queue;
687         unsigned int logical_block_size = sdkp->device->sector_size;
688         unsigned int max_blocks = 0;
689
690         q->limits.discard_alignment =
691                 sdkp->unmap_alignment * logical_block_size;
692         q->limits.discard_granularity =
693                 max(sdkp->physical_block_size,
694                     sdkp->unmap_granularity * logical_block_size);
695         sdkp->provisioning_mode = mode;
696
697         switch (mode) {
698
699         case SD_LBP_DISABLE:
700                 blk_queue_max_discard_sectors(q, 0);
701                 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, q);
702                 return;
703
704         case SD_LBP_UNMAP:
705                 max_blocks = min_not_zero(sdkp->max_unmap_blocks,
706                                           (u32)SD_MAX_WS16_BLOCKS);
707                 break;
708
709         case SD_LBP_WS16:
710                 max_blocks = min_not_zero(sdkp->max_ws_blocks,
711                                           (u32)SD_MAX_WS16_BLOCKS);
712                 break;
713
714         case SD_LBP_WS10:
715                 max_blocks = min_not_zero(sdkp->max_ws_blocks,
716                                           (u32)SD_MAX_WS10_BLOCKS);
717                 break;
718
719         case SD_LBP_ZERO:
720                 max_blocks = min_not_zero(sdkp->max_ws_blocks,
721                                           (u32)SD_MAX_WS10_BLOCKS);
722                 break;
723         }
724
725         blk_queue_max_discard_sectors(q, max_blocks * (logical_block_size >> 9));
726         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
727 }
728
729 static int sd_setup_unmap_cmnd(struct scsi_cmnd *cmd)
730 {
731         struct scsi_device *sdp = cmd->device;
732         struct request *rq = cmd->request;
733         u64 sector = blk_rq_pos(rq) >> (ilog2(sdp->sector_size) - 9);
734         u32 nr_sectors = blk_rq_sectors(rq) >> (ilog2(sdp->sector_size) - 9);
735         unsigned int data_len = 24;
736         char *buf;
737
738         rq->special_vec.bv_page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
739         if (!rq->special_vec.bv_page)
740                 return BLKPREP_DEFER;
741         rq->special_vec.bv_offset = 0;
742         rq->special_vec.bv_len = data_len;
743         rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
744
745         cmd->cmd_len = 10;
746         cmd->cmnd[0] = UNMAP;
747         cmd->cmnd[8] = 24;
748
749         buf = page_address(rq->special_vec.bv_page);
750         put_unaligned_be16(6 + 16, &buf[0]);
751         put_unaligned_be16(16, &buf[2]);
752         put_unaligned_be64(sector, &buf[8]);
753         put_unaligned_be32(nr_sectors, &buf[16]);
754
755         cmd->allowed = SD_MAX_RETRIES;
756         cmd->transfersize = data_len;
757         rq->timeout = SD_TIMEOUT;
758         scsi_req(rq)->resid_len = data_len;
759
760         return scsi_init_io(cmd);
761 }
762
763 static int sd_setup_write_same16_cmnd(struct scsi_cmnd *cmd, bool unmap)
764 {
765         struct scsi_device *sdp = cmd->device;
766         struct request *rq = cmd->request;
767         u64 sector = blk_rq_pos(rq) >> (ilog2(sdp->sector_size) - 9);
768         u32 nr_sectors = blk_rq_sectors(rq) >> (ilog2(sdp->sector_size) - 9);
769         u32 data_len = sdp->sector_size;
770
771         rq->special_vec.bv_page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
772         if (!rq->special_vec.bv_page)
773                 return BLKPREP_DEFER;
774         rq->special_vec.bv_offset = 0;
775         rq->special_vec.bv_len = data_len;
776         rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
777
778         cmd->cmd_len = 16;
779         cmd->cmnd[0] = WRITE_SAME_16;
780         if (unmap)
781                 cmd->cmnd[1] = 0x8; /* UNMAP */
782         put_unaligned_be64(sector, &cmd->cmnd[2]);
783         put_unaligned_be32(nr_sectors, &cmd->cmnd[10]);
784
785         cmd->allowed = SD_MAX_RETRIES;
786         cmd->transfersize = data_len;
787         rq->timeout = unmap ? SD_TIMEOUT : SD_WRITE_SAME_TIMEOUT;
788         scsi_req(rq)->resid_len = data_len;
789
790         return scsi_init_io(cmd);
791 }
792
793 static int sd_setup_write_same10_cmnd(struct scsi_cmnd *cmd, bool unmap)
794 {
795         struct scsi_device *sdp = cmd->device;
796         struct request *rq = cmd->request;
797         u64 sector = blk_rq_pos(rq) >> (ilog2(sdp->sector_size) - 9);
798         u32 nr_sectors = blk_rq_sectors(rq) >> (ilog2(sdp->sector_size) - 9);
799         u32 data_len = sdp->sector_size;
800
801         rq->special_vec.bv_page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
802         if (!rq->special_vec.bv_page)
803                 return BLKPREP_DEFER;
804         rq->special_vec.bv_offset = 0;
805         rq->special_vec.bv_len = data_len;
806         rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
807
808         cmd->cmd_len = 10;
809         cmd->cmnd[0] = WRITE_SAME;
810         if (unmap)
811                 cmd->cmnd[1] = 0x8; /* UNMAP */
812         put_unaligned_be32(sector, &cmd->cmnd[2]);
813         put_unaligned_be16(nr_sectors, &cmd->cmnd[7]);
814
815         cmd->allowed = SD_MAX_RETRIES;
816         cmd->transfersize = data_len;
817         rq->timeout = unmap ? SD_TIMEOUT : SD_WRITE_SAME_TIMEOUT;
818         scsi_req(rq)->resid_len = data_len;
819
820         return scsi_init_io(cmd);
821 }
822
823 static int sd_setup_write_zeroes_cmnd(struct scsi_cmnd *cmd)
824 {
825         struct request *rq = cmd->request;
826         struct scsi_device *sdp = cmd->device;
827         struct scsi_disk *sdkp = scsi_disk(rq->rq_disk);
828         u64 sector = blk_rq_pos(rq) >> (ilog2(sdp->sector_size) - 9);
829         u32 nr_sectors = blk_rq_sectors(rq) >> (ilog2(sdp->sector_size) - 9);
830
831         if (!(rq->cmd_flags & REQ_NOUNMAP)) {
832                 switch (sdkp->zeroing_mode) {
833                 case SD_ZERO_WS16_UNMAP:
834                         return sd_setup_write_same16_cmnd(cmd, true);
835                 case SD_ZERO_WS10_UNMAP:
836                         return sd_setup_write_same10_cmnd(cmd, true);
837                 }
838         }
839
840         if (sdp->no_write_same)
841                 return BLKPREP_INVALID;
842         if (sdkp->ws16 || sector > 0xffffffff || nr_sectors > 0xffff)
843                 return sd_setup_write_same16_cmnd(cmd, false);
844         return sd_setup_write_same10_cmnd(cmd, false);
845 }
846
847 static void sd_config_write_same(struct scsi_disk *sdkp)
848 {
849         struct request_queue *q = sdkp->disk->queue;
850         unsigned int logical_block_size = sdkp->device->sector_size;
851
852         if (sdkp->device->no_write_same) {
853                 sdkp->max_ws_blocks = 0;
854                 goto out;
855         }
856
857         /* Some devices can not handle block counts above 0xffff despite
858          * supporting WRITE SAME(16). Consequently we default to 64k
859          * blocks per I/O unless the device explicitly advertises a
860          * bigger limit.
861          */
862         if (sdkp->max_ws_blocks > SD_MAX_WS10_BLOCKS)
863                 sdkp->max_ws_blocks = min_not_zero(sdkp->max_ws_blocks,
864                                                    (u32)SD_MAX_WS16_BLOCKS);
865         else if (sdkp->ws16 || sdkp->ws10 || sdkp->device->no_report_opcodes)
866                 sdkp->max_ws_blocks = min_not_zero(sdkp->max_ws_blocks,
867                                                    (u32)SD_MAX_WS10_BLOCKS);
868         else {
869                 sdkp->device->no_write_same = 1;
870                 sdkp->max_ws_blocks = 0;
871         }
872
873         if (sdkp->lbprz && sdkp->lbpws)
874                 sdkp->zeroing_mode = SD_ZERO_WS16_UNMAP;
875         else if (sdkp->lbprz && sdkp->lbpws10)
876                 sdkp->zeroing_mode = SD_ZERO_WS10_UNMAP;
877         else if (sdkp->max_ws_blocks)
878                 sdkp->zeroing_mode = SD_ZERO_WS;
879         else
880                 sdkp->zeroing_mode = SD_ZERO_WRITE;
881
882 out:
883         blk_queue_max_write_same_sectors(q, sdkp->max_ws_blocks *
884                                          (logical_block_size >> 9));
885         blk_queue_max_write_zeroes_sectors(q, sdkp->max_ws_blocks *
886                                          (logical_block_size >> 9));
887 }
888
889 /**
890  * sd_setup_write_same_cmnd - write the same data to multiple blocks
891  * @cmd: command to prepare
892  *
893  * Will set up either WRITE SAME(10) or WRITE SAME(16) depending on
894  * the preference indicated by the target device.
895  **/
896 static int sd_setup_write_same_cmnd(struct scsi_cmnd *cmd)
897 {
898         struct request *rq = cmd->request;
899         struct scsi_device *sdp = cmd->device;
900         struct scsi_disk *sdkp = scsi_disk(rq->rq_disk);
901         struct bio *bio = rq->bio;
902         sector_t sector = blk_rq_pos(rq);
903         unsigned int nr_sectors = blk_rq_sectors(rq);
904         unsigned int nr_bytes = blk_rq_bytes(rq);
905         int ret;
906
907         if (sdkp->device->no_write_same)
908                 return BLKPREP_INVALID;
909
910         BUG_ON(bio_offset(bio) || bio_iovec(bio).bv_len != sdp->sector_size);
911
912         if (sd_is_zoned(sdkp)) {
913                 ret = sd_zbc_write_lock_zone(cmd);
914                 if (ret != BLKPREP_OK)
915                         return ret;
916         }
917
918         sector >>= ilog2(sdp->sector_size) - 9;
919         nr_sectors >>= ilog2(sdp->sector_size) - 9;
920
921         rq->timeout = SD_WRITE_SAME_TIMEOUT;
922
923         if (sdkp->ws16 || sector > 0xffffffff || nr_sectors > 0xffff) {
924                 cmd->cmd_len = 16;
925                 cmd->cmnd[0] = WRITE_SAME_16;
926                 put_unaligned_be64(sector, &cmd->cmnd[2]);
927                 put_unaligned_be32(nr_sectors, &cmd->cmnd[10]);
928         } else {
929                 cmd->cmd_len = 10;
930                 cmd->cmnd[0] = WRITE_SAME;
931                 put_unaligned_be32(sector, &cmd->cmnd[2]);
932                 put_unaligned_be16(nr_sectors, &cmd->cmnd[7]);
933         }
934
935         cmd->transfersize = sdp->sector_size;
936         cmd->allowed = SD_MAX_RETRIES;
937
938         /*
939          * For WRITE SAME the data transferred via the DATA OUT buffer is
940          * different from the amount of data actually written to the target.
941          *
942          * We set up __data_len to the amount of data transferred via the
943          * DATA OUT buffer so that blk_rq_map_sg sets up the proper S/G list
944          * to transfer a single sector of data first, but then reset it to
945          * the amount of data to be written right after so that the I/O path
946          * knows how much to actually write.
947          */
948         rq->__data_len = sdp->sector_size;
949         ret = scsi_init_io(cmd);
950         rq->__data_len = nr_bytes;
951         return ret;
952 }
953
954 static int sd_setup_flush_cmnd(struct scsi_cmnd *cmd)
955 {
956         struct request *rq = cmd->request;
957
958         /* flush requests don't perform I/O, zero the S/G table */
959         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
960
961         cmd->cmnd[0] = SYNCHRONIZE_CACHE;
962         cmd->cmd_len = 10;
963         cmd->transfersize = 0;
964         cmd->allowed = SD_MAX_RETRIES;
965
966         rq->timeout = rq->q->rq_timeout * SD_FLUSH_TIMEOUT_MULTIPLIER;
967         return BLKPREP_OK;
968 }
969
970 static int sd_setup_read_write_cmnd(struct scsi_cmnd *SCpnt)
971 {
972         struct request *rq = SCpnt->request;
973         struct scsi_device *sdp = SCpnt->device;
974         struct gendisk *disk = rq->rq_disk;
975         struct scsi_disk *sdkp = scsi_disk(disk);
976         sector_t block = blk_rq_pos(rq);
977         sector_t threshold;
978         unsigned int this_count = blk_rq_sectors(rq);
979         unsigned int dif, dix;
980         bool zoned_write = sd_is_zoned(sdkp) && rq_data_dir(rq) == WRITE;
981         int ret;
982         unsigned char protect;
983
984         if (zoned_write) {
985                 ret = sd_zbc_write_lock_zone(SCpnt);
986                 if (ret != BLKPREP_OK)
987                         return ret;
988         }
989
990         ret = scsi_init_io(SCpnt);
991         if (ret != BLKPREP_OK)
992                 goto out;
993         SCpnt = rq->special;
994
995         /* from here on until we're complete, any goto out
996          * is used for a killable error condition */
997         ret = BLKPREP_KILL;
998
999         SCSI_LOG_HLQUEUE(1,
1000                 scmd_printk(KERN_INFO, SCpnt,
1001                         "%s: block=%llu, count=%d\n",
1002                         __func__, (unsigned long long)block, this_count));
1003
1004         if (!sdp || !scsi_device_online(sdp) ||
1005             block + blk_rq_sectors(rq) > get_capacity(disk)) {
1006                 SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt,
1007                                                 "Finishing %u sectors\n",
1008                                                 blk_rq_sectors(rq)));
1009                 SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt,
1010                                                 "Retry with 0x%p\n", SCpnt));
1011                 goto out;
1012         }
1013
1014         if (sdp->changed) {
1015                 /*
1016                  * quietly refuse to do anything to a changed disc until 
1017                  * the changed bit has been reset
1018                  */
1019                 /* printk("SCSI disk has been changed or is not present. Prohibiting further I/O.\n"); */
1020                 goto out;
1021         }
1022
1023         /*
1024          * Some SD card readers can't handle multi-sector accesses which touch
1025          * the last one or two hardware sectors.  Split accesses as needed.
1026          */
1027         threshold = get_capacity(disk) - SD_LAST_BUGGY_SECTORS *
1028                 (sdp->sector_size / 512);
1029
1030         if (unlikely(sdp->last_sector_bug && block + this_count > threshold)) {
1031                 if (block < threshold) {
1032                         /* Access up to the threshold but not beyond */
1033                         this_count = threshold - block;
1034                 } else {
1035                         /* Access only a single hardware sector */
1036                         this_count = sdp->sector_size / 512;
1037                 }
1038         }
1039
1040         SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt, "block=%llu\n",
1041                                         (unsigned long long)block));
1042
1043         /*
1044          * If we have a 1K hardware sectorsize, prevent access to single
1045          * 512 byte sectors.  In theory we could handle this - in fact
1046          * the scsi cdrom driver must be able to handle this because
1047          * we typically use 1K blocksizes, and cdroms typically have
1048          * 2K hardware sectorsizes.  Of course, things are simpler
1049          * with the cdrom, since it is read-only.  For performance
1050          * reasons, the filesystems should be able to handle this
1051          * and not force the scsi disk driver to use bounce buffers
1052          * for this.
1053          */
1054         if (sdp->sector_size == 1024) {
1055                 if ((block & 1) || (blk_rq_sectors(rq) & 1)) {
1056                         scmd_printk(KERN_ERR, SCpnt,
1057                                     "Bad block number requested\n");
1058                         goto out;
1059                 } else {
1060                         block = block >> 1;
1061                         this_count = this_count >> 1;
1062                 }
1063         }
1064         if (sdp->sector_size == 2048) {
1065                 if ((block & 3) || (blk_rq_sectors(rq) & 3)) {
1066                         scmd_printk(KERN_ERR, SCpnt,
1067                                     "Bad block number requested\n");
1068                         goto out;
1069                 } else {
1070                         block = block >> 2;
1071                         this_count = this_count >> 2;
1072                 }
1073         }
1074         if (sdp->sector_size == 4096) {
1075                 if ((block & 7) || (blk_rq_sectors(rq) & 7)) {
1076                         scmd_printk(KERN_ERR, SCpnt,
1077                                     "Bad block number requested\n");
1078                         goto out;
1079                 } else {
1080                         block = block >> 3;
1081                         this_count = this_count >> 3;
1082                 }
1083         }
1084         if (rq_data_dir(rq) == WRITE) {
1085                 SCpnt->cmnd[0] = WRITE_6;
1086
1087                 if (blk_integrity_rq(rq))
1088                         sd_dif_prepare(SCpnt);
1089
1090         } else if (rq_data_dir(rq) == READ) {
1091                 SCpnt->cmnd[0] = READ_6;
1092         } else {
1093                 scmd_printk(KERN_ERR, SCpnt, "Unknown command %d\n", req_op(rq));
1094                 goto out;
1095         }
1096
1097         SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt,
1098                                         "%s %d/%u 512 byte blocks.\n",
1099                                         (rq_data_dir(rq) == WRITE) ?
1100                                         "writing" : "reading", this_count,
1101                                         blk_rq_sectors(rq)));
1102
1103         dix = scsi_prot_sg_count(SCpnt);
1104         dif = scsi_host_dif_capable(SCpnt->device->host, sdkp->protection_type);
1105
1106         if (dif || dix)
1107                 protect = sd_setup_protect_cmnd(SCpnt, dix, dif);
1108         else
1109                 protect = 0;
1110
1111         if (protect && sdkp->protection_type == T10_PI_TYPE2_PROTECTION) {
1112                 SCpnt->cmnd = mempool_alloc(sd_cdb_pool, GFP_ATOMIC);
1113
1114                 if (unlikely(SCpnt->cmnd == NULL)) {
1115                         ret = BLKPREP_DEFER;
1116                         goto out;
1117                 }
1118
1119                 SCpnt->cmd_len = SD_EXT_CDB_SIZE;
1120                 memset(SCpnt->cmnd, 0, SCpnt->cmd_len);
1121                 SCpnt->cmnd[0] = VARIABLE_LENGTH_CMD;
1122                 SCpnt->cmnd[7] = 0x18;
1123                 SCpnt->cmnd[9] = (rq_data_dir(rq) == READ) ? READ_32 : WRITE_32;
1124                 SCpnt->cmnd[10] = protect | ((rq->cmd_flags & REQ_FUA) ? 0x8 : 0);
1125
1126                 /* LBA */
1127                 SCpnt->cmnd[12] = sizeof(block) > 4 ? (unsigned char) (block >> 56) & 0xff : 0;
1128                 SCpnt->cmnd[13] = sizeof(block) > 4 ? (unsigned char) (block >> 48) & 0xff : 0;
1129                 SCpnt->cmnd[14] = sizeof(block) > 4 ? (unsigned char) (block >> 40) & 0xff : 0;
1130                 SCpnt->cmnd[15] = sizeof(block) > 4 ? (unsigned char) (block >> 32) & 0xff : 0;
1131                 SCpnt->cmnd[16] = (unsigned char) (block >> 24) & 0xff;
1132                 SCpnt->cmnd[17] = (unsigned char) (block >> 16) & 0xff;
1133                 SCpnt->cmnd[18] = (unsigned char) (block >> 8) & 0xff;
1134                 SCpnt->cmnd[19] = (unsigned char) block & 0xff;
1135
1136                 /* Expected Indirect LBA */
1137                 SCpnt->cmnd[20] = (unsigned char) (block >> 24) & 0xff;
1138                 SCpnt->cmnd[21] = (unsigned char) (block >> 16) & 0xff;
1139                 SCpnt->cmnd[22] = (unsigned char) (block >> 8) & 0xff;
1140                 SCpnt->cmnd[23] = (unsigned char) block & 0xff;
1141
1142                 /* Transfer length */
1143                 SCpnt->cmnd[28] = (unsigned char) (this_count >> 24) & 0xff;
1144                 SCpnt->cmnd[29] = (unsigned char) (this_count >> 16) & 0xff;
1145                 SCpnt->cmnd[30] = (unsigned char) (this_count >> 8) & 0xff;
1146                 SCpnt->cmnd[31] = (unsigned char) this_count & 0xff;
1147         } else if (sdp->use_16_for_rw || (this_count > 0xffff)) {
1148                 SCpnt->cmnd[0] += READ_16 - READ_6;
1149                 SCpnt->cmnd[1] = protect | ((rq->cmd_flags & REQ_FUA) ? 0x8 : 0);
1150                 SCpnt->cmnd[2] = sizeof(block) > 4 ? (unsigned char) (block >> 56) & 0xff : 0;
1151                 SCpnt->cmnd[3] = sizeof(block) > 4 ? (unsigned char) (block >> 48) & 0xff : 0;
1152                 SCpnt->cmnd[4] = sizeof(block) > 4 ? (unsigned char) (block >> 40) & 0xff : 0;
1153                 SCpnt->cmnd[5] = sizeof(block) > 4 ? (unsigned char) (block >> 32) & 0xff : 0;
1154                 SCpnt->cmnd[6] = (unsigned char) (block >> 24) & 0xff;
1155                 SCpnt->cmnd[7] = (unsigned char) (block >> 16) & 0xff;
1156                 SCpnt->cmnd[8] = (unsigned char) (block >> 8) & 0xff;
1157                 SCpnt->cmnd[9] = (unsigned char) block & 0xff;
1158                 SCpnt->cmnd[10] = (unsigned char) (this_count >> 24) & 0xff;
1159                 SCpnt->cmnd[11] = (unsigned char) (this_count >> 16) & 0xff;
1160                 SCpnt->cmnd[12] = (unsigned char) (this_count >> 8) & 0xff;
1161                 SCpnt->cmnd[13] = (unsigned char) this_count & 0xff;
1162                 SCpnt->cmnd[14] = SCpnt->cmnd[15] = 0;
1163         } else if ((this_count > 0xff) || (block > 0x1fffff) ||
1164                    scsi_device_protection(SCpnt->device) ||
1165                    SCpnt->device->use_10_for_rw) {
1166                 SCpnt->cmnd[0] += READ_10 - READ_6;
1167                 SCpnt->cmnd[1] = protect | ((rq->cmd_flags & REQ_FUA) ? 0x8 : 0);
1168                 SCpnt->cmnd[2] = (unsigned char) (block >> 24) & 0xff;
1169                 SCpnt->cmnd[3] = (unsigned char) (block >> 16) & 0xff;
1170                 SCpnt->cmnd[4] = (unsigned char) (block >> 8) & 0xff;
1171                 SCpnt->cmnd[5] = (unsigned char) block & 0xff;
1172                 SCpnt->cmnd[6] = SCpnt->cmnd[9] = 0;
1173                 SCpnt->cmnd[7] = (unsigned char) (this_count >> 8) & 0xff;
1174                 SCpnt->cmnd[8] = (unsigned char) this_count & 0xff;
1175         } else {
1176                 if (unlikely(rq->cmd_flags & REQ_FUA)) {
1177                         /*
1178                          * This happens only if this drive failed
1179                          * 10byte rw command with ILLEGAL_REQUEST
1180                          * during operation and thus turned off
1181                          * use_10_for_rw.
1182                          */
1183                         scmd_printk(KERN_ERR, SCpnt,
1184                                     "FUA write on READ/WRITE(6) drive\n");
1185                         goto out;
1186                 }
1187
1188                 SCpnt->cmnd[1] |= (unsigned char) ((block >> 16) & 0x1f);
1189                 SCpnt->cmnd[2] = (unsigned char) ((block >> 8) & 0xff);
1190                 SCpnt->cmnd[3] = (unsigned char) block & 0xff;
1191                 SCpnt->cmnd[4] = (unsigned char) this_count;
1192                 SCpnt->cmnd[5] = 0;
1193         }
1194         SCpnt->sdb.length = this_count * sdp->sector_size;
1195
1196         /*
1197          * We shouldn't disconnect in the middle of a sector, so with a dumb
1198          * host adapter, it's safe to assume that we can at least transfer
1199          * this many bytes between each connect / disconnect.
1200          */
1201         SCpnt->transfersize = sdp->sector_size;
1202         SCpnt->underflow = this_count << 9;
1203         SCpnt->allowed = SD_MAX_RETRIES;
1204
1205         /*
1206          * This indicates that the command is ready from our end to be
1207          * queued.
1208          */
1209         ret = BLKPREP_OK;
1210  out:
1211         if (zoned_write && ret != BLKPREP_OK)
1212                 sd_zbc_write_unlock_zone(SCpnt);
1213
1214         return ret;
1215 }
1216
1217 static int sd_init_command(struct scsi_cmnd *cmd)
1218 {
1219         struct request *rq = cmd->request;
1220
1221         switch (req_op(rq)) {
1222         case REQ_OP_DISCARD:
1223                 switch (scsi_disk(rq->rq_disk)->provisioning_mode) {
1224                 case SD_LBP_UNMAP:
1225                         return sd_setup_unmap_cmnd(cmd);
1226                 case SD_LBP_WS16:
1227                         return sd_setup_write_same16_cmnd(cmd, true);
1228                 case SD_LBP_WS10:
1229                         return sd_setup_write_same10_cmnd(cmd, true);
1230                 case SD_LBP_ZERO:
1231                         return sd_setup_write_same10_cmnd(cmd, false);
1232                 default:
1233                         return BLKPREP_INVALID;
1234                 }
1235         case REQ_OP_WRITE_ZEROES:
1236                 return sd_setup_write_zeroes_cmnd(cmd);
1237         case REQ_OP_WRITE_SAME:
1238                 return sd_setup_write_same_cmnd(cmd);
1239         case REQ_OP_FLUSH:
1240                 return sd_setup_flush_cmnd(cmd);
1241         case REQ_OP_READ:
1242         case REQ_OP_WRITE:
1243                 return sd_setup_read_write_cmnd(cmd);
1244         case REQ_OP_ZONE_REPORT:
1245                 return sd_zbc_setup_report_cmnd(cmd);
1246         case REQ_OP_ZONE_RESET:
1247                 return sd_zbc_setup_reset_cmnd(cmd);
1248         default:
1249                 BUG();
1250         }
1251 }
1252
1253 static void sd_uninit_command(struct scsi_cmnd *SCpnt)
1254 {
1255         struct request *rq = SCpnt->request;
1256
1257         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
1258                 __free_page(rq->special_vec.bv_page);
1259
1260         if (SCpnt->cmnd != scsi_req(rq)->cmd) {
1261                 mempool_free(SCpnt->cmnd, sd_cdb_pool);
1262                 SCpnt->cmnd = NULL;
1263                 SCpnt->cmd_len = 0;
1264         }
1265 }
1266
1267 /**
1268  *      sd_open - open a scsi disk device
1269  *      @bdev: Block device of the scsi disk to open
1270  *      @mode: FMODE_* mask
1271  *
1272  *      Returns 0 if successful. Returns a negated errno value in case 
1273  *      of error.
1274  *
1275  *      Note: This can be called from a user context (e.g. fsck(1) )
1276  *      or from within the kernel (e.g. as a result of a mount(1) ).
1277  *      In the latter case @inode and @filp carry an abridged amount
1278  *      of information as noted above.
1279  *
1280  *      Locking: called with bdev->bd_mutex held.
1281  **/
1282 static int sd_open(struct block_device *bdev, fmode_t mode)
1283 {
1284         struct scsi_disk *sdkp = scsi_disk_get(bdev->bd_disk);
1285         struct scsi_device *sdev;
1286         int retval;
1287
1288         if (!sdkp)
1289                 return -ENXIO;
1290
1291         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp, "sd_open\n"));
1292
1293         sdev = sdkp->device;
1294
1295         /*
1296          * If the device is in error recovery, wait until it is done.
1297          * If the device is offline, then disallow any access to it.
1298          */
1299         retval = -ENXIO;
1300         if (!scsi_block_when_processing_errors(sdev))
1301                 goto error_out;
1302
1303         if (sdev->removable || sdkp->write_prot)
1304                 check_disk_change(bdev);
1305
1306         /*
1307          * If the drive is empty, just let the open fail.
1308          */
1309         retval = -ENOMEDIUM;
1310         if (sdev->removable && !sdkp->media_present && !(mode & FMODE_NDELAY))
1311                 goto error_out;
1312
1313         /*
1314          * If the device has the write protect tab set, have the open fail
1315          * if the user expects to be able to write to the thing.
1316          */
1317         retval = -EROFS;
1318         if (sdkp->write_prot && (mode & FMODE_WRITE))
1319                 goto error_out;
1320
1321         /*
1322          * It is possible that the disk changing stuff resulted in
1323          * the device being taken offline.  If this is the case,
1324          * report this to the user, and don't pretend that the
1325          * open actually succeeded.
1326          */
1327         retval = -ENXIO;
1328         if (!scsi_device_online(sdev))
1329                 goto error_out;
1330
1331         if ((atomic_inc_return(&sdkp->openers) == 1) && sdev->removable) {
1332                 if (scsi_block_when_processing_errors(sdev))
1333                         scsi_set_medium_removal(sdev, SCSI_REMOVAL_PREVENT);
1334         }
1335
1336         return 0;
1337
1338 error_out:
1339         scsi_disk_put(sdkp);
1340         return retval;  
1341 }
1342
1343 /**
1344  *      sd_release - invoked when the (last) close(2) is called on this
1345  *      scsi disk.
1346  *      @disk: disk to release
1347  *      @mode: FMODE_* mask
1348  *
1349  *      Returns 0. 
1350  *
1351  *      Note: may block (uninterruptible) if error recovery is underway
1352  *      on this disk.
1353  *
1354  *      Locking: called with bdev->bd_mutex held.
1355  **/
1356 static void sd_release(struct gendisk *disk, fmode_t mode)
1357 {
1358         struct scsi_disk *sdkp = scsi_disk(disk);
1359         struct scsi_device *sdev = sdkp->device;
1360
1361         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp, "sd_release\n"));
1362
1363         if (atomic_dec_return(&sdkp->openers) == 0 && sdev->removable) {
1364                 if (scsi_block_when_processing_errors(sdev))
1365                         scsi_set_medium_removal(sdev, SCSI_REMOVAL_ALLOW);
1366         }
1367
1368         /*
1369          * XXX and what if there are packets in flight and this close()
1370          * XXX is followed by a "rmmod sd_mod"?
1371          */
1372
1373         scsi_disk_put(sdkp);
1374 }
1375
1376 static int sd_getgeo(struct block_device *bdev, struct hd_geometry *geo)
1377 {
1378         struct scsi_disk *sdkp = scsi_disk(bdev->bd_disk);
1379         struct scsi_device *sdp = sdkp->device;
1380         struct Scsi_Host *host = sdp->host;
1381         sector_t capacity = logical_to_sectors(sdp, sdkp->capacity);
1382         int diskinfo[4];
1383
1384         /* default to most commonly used values */
1385         diskinfo[0] = 0x40;     /* 1 << 6 */
1386         diskinfo[1] = 0x20;     /* 1 << 5 */
1387         diskinfo[2] = capacity >> 11;
1388
1389         /* override with calculated, extended default, or driver values */
1390         if (host->hostt->bios_param)
1391                 host->hostt->bios_param(sdp, bdev, capacity, diskinfo);
1392         else
1393                 scsicam_bios_param(bdev, capacity, diskinfo);
1394
1395         geo->heads = diskinfo[0];
1396         geo->sectors = diskinfo[1];
1397         geo->cylinders = diskinfo[2];
1398         return 0;
1399 }
1400
1401 /**
1402  *      sd_ioctl - process an ioctl
1403  *      @bdev: target block device
1404  *      @mode: FMODE_* mask
1405  *      @cmd: ioctl command number
1406  *      @arg: this is third argument given to ioctl(2) system call.
1407  *      Often contains a pointer.
1408  *
1409  *      Returns 0 if successful (some ioctls return positive numbers on
1410  *      success as well). Returns a negated errno value in case of error.
1411  *
1412  *      Note: most ioctls are forward onto the block subsystem or further
1413  *      down in the scsi subsystem.
1414  **/
1415 static int sd_ioctl(struct block_device *bdev, fmode_t mode,
1416                     unsigned int cmd, unsigned long arg)
1417 {
1418         struct gendisk *disk = bdev->bd_disk;
1419         struct scsi_disk *sdkp = scsi_disk(disk);
1420         struct scsi_device *sdp = sdkp->device;
1421         void __user *p = (void __user *)arg;
1422         int error;
1423     
1424         SCSI_LOG_IOCTL(1, sd_printk(KERN_INFO, sdkp, "sd_ioctl: disk=%s, "
1425                                     "cmd=0x%x\n", disk->disk_name, cmd));
1426
1427         error = scsi_verify_blk_ioctl(bdev, cmd);
1428         if (error < 0)
1429                 return error;
1430
1431         /*
1432          * If we are in the middle of error recovery, don't let anyone
1433          * else try and use this device.  Also, if error recovery fails, it
1434          * may try and take the device offline, in which case all further
1435          * access to the device is prohibited.
1436          */
1437         error = scsi_ioctl_block_when_processing_errors(sdp, cmd,
1438                         (mode & FMODE_NDELAY) != 0);
1439         if (error)
1440                 goto out;
1441
1442         /*
1443          * Send SCSI addressing ioctls directly to mid level, send other
1444          * ioctls to block level and then onto mid level if they can't be
1445          * resolved.
1446          */
1447         switch (cmd) {
1448                 case SCSI_IOCTL_GET_IDLUN:
1449                 case SCSI_IOCTL_GET_BUS_NUMBER:
1450                         error = scsi_ioctl(sdp, cmd, p);
1451                         break;
1452                 default:
1453                         error = scsi_cmd_blk_ioctl(bdev, mode, cmd, p);
1454                         if (error != -ENOTTY)
1455                                 break;
1456                         error = scsi_ioctl(sdp, cmd, p);
1457                         break;
1458         }
1459 out:
1460         return error;
1461 }
1462
1463 static void set_media_not_present(struct scsi_disk *sdkp)
1464 {
1465         if (sdkp->media_present)
1466                 sdkp->device->changed = 1;
1467
1468         if (sdkp->device->removable) {
1469                 sdkp->media_present = 0;
1470                 sdkp->capacity = 0;
1471         }
1472 }
1473
1474 static int media_not_present(struct scsi_disk *sdkp,
1475                              struct scsi_sense_hdr *sshdr)
1476 {
1477         if (!scsi_sense_valid(sshdr))
1478                 return 0;
1479
1480         /* not invoked for commands that could return deferred errors */
1481         switch (sshdr->sense_key) {
1482         case UNIT_ATTENTION:
1483         case NOT_READY:
1484                 /* medium not present */
1485                 if (sshdr->asc == 0x3A) {
1486                         set_media_not_present(sdkp);
1487                         return 1;
1488                 }
1489         }
1490         return 0;
1491 }
1492
1493 /**
1494  *      sd_check_events - check media events
1495  *      @disk: kernel device descriptor
1496  *      @clearing: disk events currently being cleared
1497  *
1498  *      Returns mask of DISK_EVENT_*.
1499  *
1500  *      Note: this function is invoked from the block subsystem.
1501  **/
1502 static unsigned int sd_check_events(struct gendisk *disk, unsigned int clearing)
1503 {
1504         struct scsi_disk *sdkp = scsi_disk_get(disk);
1505         struct scsi_device *sdp;
1506         int retval;
1507
1508         if (!sdkp)
1509                 return 0;
1510
1511         sdp = sdkp->device;
1512         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp, "sd_check_events\n"));
1513
1514         /*
1515          * If the device is offline, don't send any commands - just pretend as
1516          * if the command failed.  If the device ever comes back online, we
1517          * can deal with it then.  It is only because of unrecoverable errors
1518          * that we would ever take a device offline in the first place.
1519          */
1520         if (!scsi_device_online(sdp)) {
1521                 set_media_not_present(sdkp);
1522                 goto out;
1523         }
1524
1525         /*
1526          * Using TEST_UNIT_READY enables differentiation between drive with
1527          * no cartridge loaded - NOT READY, drive with changed cartridge -
1528          * UNIT ATTENTION, or with same cartridge - GOOD STATUS.
1529          *
1530          * Drives that auto spin down. eg iomega jaz 1G, will be started
1531          * by sd_spinup_disk() from sd_revalidate_disk(), which happens whenever
1532          * sd_revalidate() is called.
1533          */
1534         if (scsi_block_when_processing_errors(sdp)) {
1535                 struct scsi_sense_hdr sshdr = { 0, };
1536
1537                 retval = scsi_test_unit_ready(sdp, SD_TIMEOUT, SD_MAX_RETRIES,
1538                                               &sshdr);
1539
1540                 /* failed to execute TUR, assume media not present */
1541                 if (host_byte(retval)) {
1542                         set_media_not_present(sdkp);
1543                         goto out;
1544                 }
1545
1546                 if (media_not_present(sdkp, &sshdr))
1547                         goto out;
1548         }
1549
1550         /*
1551          * For removable scsi disk we have to recognise the presence
1552          * of a disk in the drive.
1553          */
1554         if (!sdkp->media_present)
1555                 sdp->changed = 1;
1556         sdkp->media_present = 1;
1557 out:
1558         /*
1559          * sdp->changed is set under the following conditions:
1560          *
1561          *      Medium present state has changed in either direction.
1562          *      Device has indicated UNIT_ATTENTION.
1563          */
1564         retval = sdp->changed ? DISK_EVENT_MEDIA_CHANGE : 0;
1565         sdp->changed = 0;
1566         scsi_disk_put(sdkp);
1567         return retval;
1568 }
1569
1570 static int sd_sync_cache(struct scsi_disk *sdkp)
1571 {
1572         int retries, res;
1573         struct scsi_device *sdp = sdkp->device;
1574         const int timeout = sdp->request_queue->rq_timeout
1575                 * SD_FLUSH_TIMEOUT_MULTIPLIER;
1576         struct scsi_sense_hdr sshdr;
1577
1578         if (!scsi_device_online(sdp))
1579                 return -ENODEV;
1580
1581         for (retries = 3; retries > 0; --retries) {
1582                 unsigned char cmd[10] = { 0 };
1583
1584                 cmd[0] = SYNCHRONIZE_CACHE;
1585                 /*
1586                  * Leave the rest of the command zero to indicate
1587                  * flush everything.
1588                  */
1589                 res = scsi_execute(sdp, cmd, DMA_NONE, NULL, 0, NULL, &sshdr,
1590                                 timeout, SD_MAX_RETRIES, 0, RQF_PM, NULL);
1591                 if (res == 0)
1592                         break;
1593         }
1594
1595         if (res) {
1596                 sd_print_result(sdkp, "Synchronize Cache(10) failed", res);
1597
1598                 if (driver_byte(res) & DRIVER_SENSE)
1599                         sd_print_sense_hdr(sdkp, &sshdr);
1600                 /* we need to evaluate the error return  */
1601                 if (scsi_sense_valid(&sshdr) &&
1602                         (sshdr.asc == 0x3a ||   /* medium not present */
1603                          sshdr.asc == 0x20))    /* invalid command */
1604                                 /* this is no error here */
1605                                 return 0;
1606
1607                 switch (host_byte(res)) {
1608                 /* ignore errors due to racing a disconnection */
1609                 case DID_BAD_TARGET:
1610                 case DID_NO_CONNECT:
1611                         return 0;
1612                 /* signal the upper layer it might try again */
1613                 case DID_BUS_BUSY:
1614                 case DID_IMM_RETRY:
1615                 case DID_REQUEUE:
1616                 case DID_SOFT_ERROR:
1617                         return -EBUSY;
1618                 default:
1619                         return -EIO;
1620                 }
1621         }
1622         return 0;
1623 }
1624
1625 static void sd_rescan(struct device *dev)
1626 {
1627         struct scsi_disk *sdkp = dev_get_drvdata(dev);
1628
1629         revalidate_disk(sdkp->disk);
1630 }
1631
1632
1633 #ifdef CONFIG_COMPAT
1634 /* 
1635  * This gets directly called from VFS. When the ioctl 
1636  * is not recognized we go back to the other translation paths. 
1637  */
1638 static int sd_compat_ioctl(struct block_device *bdev, fmode_t mode,
1639                            unsigned int cmd, unsigned long arg)
1640 {
1641         struct scsi_device *sdev = scsi_disk(bdev->bd_disk)->device;
1642         int error;
1643
1644         error = scsi_ioctl_block_when_processing_errors(sdev, cmd,
1645                         (mode & FMODE_NDELAY) != 0);
1646         if (error)
1647                 return error;
1648                
1649         /* 
1650          * Let the static ioctl translation table take care of it.
1651          */
1652         if (!sdev->host->hostt->compat_ioctl)
1653                 return -ENOIOCTLCMD; 
1654         return sdev->host->hostt->compat_ioctl(sdev, cmd, (void __user *)arg);
1655 }
1656 #endif
1657
1658 static char sd_pr_type(enum pr_type type)
1659 {
1660         switch (type) {
1661         case PR_WRITE_EXCLUSIVE:
1662                 return 0x01;
1663         case PR_EXCLUSIVE_ACCESS:
1664                 return 0x03;
1665         case PR_WRITE_EXCLUSIVE_REG_ONLY:
1666                 return 0x05;
1667         case PR_EXCLUSIVE_ACCESS_REG_ONLY:
1668                 return 0x06;
1669         case PR_WRITE_EXCLUSIVE_ALL_REGS:
1670                 return 0x07;
1671         case PR_EXCLUSIVE_ACCESS_ALL_REGS:
1672                 return 0x08;
1673         default:
1674                 return 0;
1675         }
1676 };
1677
1678 static int sd_pr_command(struct block_device *bdev, u8 sa,
1679                 u64 key, u64 sa_key, u8 type, u8 flags)
1680 {
1681         struct scsi_device *sdev = scsi_disk(bdev->bd_disk)->device;
1682         struct scsi_sense_hdr sshdr;
1683         int result;
1684         u8 cmd[16] = { 0, };
1685         u8 data[24] = { 0, };
1686
1687         cmd[0] = PERSISTENT_RESERVE_OUT;
1688         cmd[1] = sa;
1689         cmd[2] = type;
1690         put_unaligned_be32(sizeof(data), &cmd[5]);
1691
1692         put_unaligned_be64(key, &data[0]);
1693         put_unaligned_be64(sa_key, &data[8]);
1694         data[20] = flags;
1695
1696         result = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, &data, sizeof(data),
1697                         &sshdr, SD_TIMEOUT, SD_MAX_RETRIES, NULL);
1698
1699         if ((driver_byte(result) & DRIVER_SENSE) &&
1700             (scsi_sense_valid(&sshdr))) {
1701                 sdev_printk(KERN_INFO, sdev, "PR command failed: %d\n", result);
1702                 scsi_print_sense_hdr(sdev, NULL, &sshdr);
1703         }
1704
1705         return result;
1706 }
1707
1708 static int sd_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
1709                 u32 flags)
1710 {
1711         if (flags & ~PR_FL_IGNORE_KEY)
1712                 return -EOPNOTSUPP;
1713         return sd_pr_command(bdev, (flags & PR_FL_IGNORE_KEY) ? 0x06 : 0x00,
1714                         old_key, new_key, 0,
1715                         (1 << 0) /* APTPL */);
1716 }
1717
1718 static int sd_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
1719                 u32 flags)
1720 {
1721         if (flags)
1722                 return -EOPNOTSUPP;
1723         return sd_pr_command(bdev, 0x01, key, 0, sd_pr_type(type), 0);
1724 }
1725
1726 static int sd_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
1727 {
1728         return sd_pr_command(bdev, 0x02, key, 0, sd_pr_type(type), 0);
1729 }
1730
1731 static int sd_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
1732                 enum pr_type type, bool abort)
1733 {
1734         return sd_pr_command(bdev, abort ? 0x05 : 0x04, old_key, new_key,
1735                              sd_pr_type(type), 0);
1736 }
1737
1738 static int sd_pr_clear(struct block_device *bdev, u64 key)
1739 {
1740         return sd_pr_command(bdev, 0x03, key, 0, 0, 0);
1741 }
1742
1743 static const struct pr_ops sd_pr_ops = {
1744         .pr_register    = sd_pr_register,
1745         .pr_reserve     = sd_pr_reserve,
1746         .pr_release     = sd_pr_release,
1747         .pr_preempt     = sd_pr_preempt,
1748         .pr_clear       = sd_pr_clear,
1749 };
1750
1751 static const struct block_device_operations sd_fops = {
1752         .owner                  = THIS_MODULE,
1753         .open                   = sd_open,
1754         .release                = sd_release,
1755         .ioctl                  = sd_ioctl,
1756         .getgeo                 = sd_getgeo,
1757 #ifdef CONFIG_COMPAT
1758         .compat_ioctl           = sd_compat_ioctl,
1759 #endif
1760         .check_events           = sd_check_events,
1761         .revalidate_disk        = sd_revalidate_disk,
1762         .unlock_native_capacity = sd_unlock_native_capacity,
1763         .pr_ops                 = &sd_pr_ops,
1764 };
1765
1766 /**
1767  *      sd_eh_reset - reset error handling callback
1768  *      @scmd:          sd-issued command that has failed
1769  *
1770  *      This function is called by the SCSI midlayer before starting
1771  *      SCSI EH. When counting medium access failures we have to be
1772  *      careful to register it only only once per device and SCSI EH run;
1773  *      there might be several timed out commands which will cause the
1774  *      'max_medium_access_timeouts' counter to trigger after the first
1775  *      SCSI EH run already and set the device to offline.
1776  *      So this function resets the internal counter before starting SCSI EH.
1777  **/
1778 static void sd_eh_reset(struct scsi_cmnd *scmd)
1779 {
1780         struct scsi_disk *sdkp = scsi_disk(scmd->request->rq_disk);
1781
1782         /* New SCSI EH run, reset gate variable */
1783         sdkp->ignore_medium_access_errors = false;
1784 }
1785
1786 /**
1787  *      sd_eh_action - error handling callback
1788  *      @scmd:          sd-issued command that has failed
1789  *      @eh_disp:       The recovery disposition suggested by the midlayer
1790  *
1791  *      This function is called by the SCSI midlayer upon completion of an
1792  *      error test command (currently TEST UNIT READY). The result of sending
1793  *      the eh command is passed in eh_disp.  We're looking for devices that
1794  *      fail medium access commands but are OK with non access commands like
1795  *      test unit ready (so wrongly see the device as having a successful
1796  *      recovery)
1797  **/
1798 static int sd_eh_action(struct scsi_cmnd *scmd, int eh_disp)
1799 {
1800         struct scsi_disk *sdkp = scsi_disk(scmd->request->rq_disk);
1801
1802         if (!scsi_device_online(scmd->device) ||
1803             !scsi_medium_access_command(scmd) ||
1804             host_byte(scmd->result) != DID_TIME_OUT ||
1805             eh_disp != SUCCESS)
1806                 return eh_disp;
1807
1808         /*
1809          * The device has timed out executing a medium access command.
1810          * However, the TEST UNIT READY command sent during error
1811          * handling completed successfully. Either the device is in the
1812          * process of recovering or has it suffered an internal failure
1813          * that prevents access to the storage medium.
1814          */
1815         if (!sdkp->ignore_medium_access_errors) {
1816                 sdkp->medium_access_timed_out++;
1817                 sdkp->ignore_medium_access_errors = true;
1818         }
1819
1820         /*
1821          * If the device keeps failing read/write commands but TEST UNIT
1822          * READY always completes successfully we assume that medium
1823          * access is no longer possible and take the device offline.
1824          */
1825         if (sdkp->medium_access_timed_out >= sdkp->max_medium_access_timeouts) {
1826                 scmd_printk(KERN_ERR, scmd,
1827                             "Medium access timeout failure. Offlining disk!\n");
1828                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1829
1830                 return SUCCESS;
1831         }
1832
1833         return eh_disp;
1834 }
1835
1836 static unsigned int sd_completed_bytes(struct scsi_cmnd *scmd)
1837 {
1838         struct request *req = scmd->request;
1839         struct scsi_device *sdev = scmd->device;
1840         unsigned int transferred, good_bytes;
1841         u64 start_lba, end_lba, bad_lba;
1842
1843         /*
1844          * Some commands have a payload smaller than the device logical
1845          * block size (e.g. INQUIRY on a 4K disk).
1846          */
1847         if (scsi_bufflen(scmd) <= sdev->sector_size)
1848                 return 0;
1849
1850         /* Check if we have a 'bad_lba' information */
1851         if (!scsi_get_sense_info_fld(scmd->sense_buffer,
1852                                      SCSI_SENSE_BUFFERSIZE,
1853                                      &bad_lba))
1854                 return 0;
1855
1856         /*
1857          * If the bad lba was reported incorrectly, we have no idea where
1858          * the error is.
1859          */
1860         start_lba = sectors_to_logical(sdev, blk_rq_pos(req));
1861         end_lba = start_lba + bytes_to_logical(sdev, scsi_bufflen(scmd));
1862         if (bad_lba < start_lba || bad_lba >= end_lba)
1863                 return 0;
1864
1865         /*
1866          * resid is optional but mostly filled in.  When it's unused,
1867          * its value is zero, so we assume the whole buffer transferred
1868          */
1869         transferred = scsi_bufflen(scmd) - scsi_get_resid(scmd);
1870
1871         /* This computation should always be done in terms of the
1872          * resolution of the device's medium.
1873          */
1874         good_bytes = logical_to_bytes(sdev, bad_lba - start_lba);
1875
1876         return min(good_bytes, transferred);
1877 }
1878
1879 /**
1880  *      sd_done - bottom half handler: called when the lower level
1881  *      driver has completed (successfully or otherwise) a scsi command.
1882  *      @SCpnt: mid-level's per command structure.
1883  *
1884  *      Note: potentially run from within an ISR. Must not block.
1885  **/
1886 static int sd_done(struct scsi_cmnd *SCpnt)
1887 {
1888         int result = SCpnt->result;
1889         unsigned int good_bytes = result ? 0 : scsi_bufflen(SCpnt);
1890         unsigned int sector_size = SCpnt->device->sector_size;
1891         unsigned int resid;
1892         struct scsi_sense_hdr sshdr;
1893         struct scsi_disk *sdkp = scsi_disk(SCpnt->request->rq_disk);
1894         struct request *req = SCpnt->request;
1895         int sense_valid = 0;
1896         int sense_deferred = 0;
1897
1898         switch (req_op(req)) {
1899         case REQ_OP_DISCARD:
1900         case REQ_OP_WRITE_ZEROES:
1901         case REQ_OP_WRITE_SAME:
1902         case REQ_OP_ZONE_RESET:
1903                 if (!result) {
1904                         good_bytes = blk_rq_bytes(req);
1905                         scsi_set_resid(SCpnt, 0);
1906                 } else {
1907                         good_bytes = 0;
1908                         scsi_set_resid(SCpnt, blk_rq_bytes(req));
1909                 }
1910                 break;
1911         case REQ_OP_ZONE_REPORT:
1912                 if (!result) {
1913                         good_bytes = scsi_bufflen(SCpnt)
1914                                 - scsi_get_resid(SCpnt);
1915                         scsi_set_resid(SCpnt, 0);
1916                 } else {
1917                         good_bytes = 0;
1918                         scsi_set_resid(SCpnt, blk_rq_bytes(req));
1919                 }
1920                 break;
1921         default:
1922                 /*
1923                  * In case of bogus fw or device, we could end up having
1924                  * an unaligned partial completion. Check this here and force
1925                  * alignment.
1926                  */
1927                 resid = scsi_get_resid(SCpnt);
1928                 if (resid & (sector_size - 1)) {
1929                         sd_printk(KERN_INFO, sdkp,
1930                                 "Unaligned partial completion (resid=%u, sector_sz=%u)\n",
1931                                 resid, sector_size);
1932                         resid = min(scsi_bufflen(SCpnt),
1933                                     round_up(resid, sector_size));
1934                         scsi_set_resid(SCpnt, resid);
1935                 }
1936         }
1937
1938         if (result) {
1939                 sense_valid = scsi_command_normalize_sense(SCpnt, &sshdr);
1940                 if (sense_valid)
1941                         sense_deferred = scsi_sense_is_deferred(&sshdr);
1942         }
1943         sdkp->medium_access_timed_out = 0;
1944
1945         if (driver_byte(result) != DRIVER_SENSE &&
1946             (!sense_valid || sense_deferred))
1947                 goto out;
1948
1949         switch (sshdr.sense_key) {
1950         case HARDWARE_ERROR:
1951         case MEDIUM_ERROR:
1952                 good_bytes = sd_completed_bytes(SCpnt);
1953                 break;
1954         case RECOVERED_ERROR:
1955                 good_bytes = scsi_bufflen(SCpnt);
1956                 break;
1957         case NO_SENSE:
1958                 /* This indicates a false check condition, so ignore it.  An
1959                  * unknown amount of data was transferred so treat it as an
1960                  * error.
1961                  */
1962                 SCpnt->result = 0;
1963                 memset(SCpnt->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
1964                 break;
1965         case ABORTED_COMMAND:
1966                 if (sshdr.asc == 0x10)  /* DIF: Target detected corruption */
1967                         good_bytes = sd_completed_bytes(SCpnt);
1968                 break;
1969         case ILLEGAL_REQUEST:
1970                 switch (sshdr.asc) {
1971                 case 0x10:      /* DIX: Host detected corruption */
1972                         good_bytes = sd_completed_bytes(SCpnt);
1973                         break;
1974                 case 0x20:      /* INVALID COMMAND OPCODE */
1975                 case 0x24:      /* INVALID FIELD IN CDB */
1976                         switch (SCpnt->cmnd[0]) {
1977                         case UNMAP:
1978                                 sd_config_discard(sdkp, SD_LBP_DISABLE);
1979                                 break;
1980                         case WRITE_SAME_16:
1981                         case WRITE_SAME:
1982                                 if (SCpnt->cmnd[1] & 8) { /* UNMAP */
1983                                         sd_config_discard(sdkp, SD_LBP_DISABLE);
1984                                 } else {
1985                                         sdkp->device->no_write_same = 1;
1986                                         sd_config_write_same(sdkp);
1987                                         req->__data_len = blk_rq_bytes(req);
1988                                         req->rq_flags |= RQF_QUIET;
1989                                 }
1990                                 break;
1991                         }
1992                 }
1993                 break;
1994         default:
1995                 break;
1996         }
1997
1998  out:
1999         if (sd_is_zoned(sdkp))
2000                 sd_zbc_complete(SCpnt, good_bytes, &sshdr);
2001
2002         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, SCpnt,
2003                                            "sd_done: completed %d of %d bytes\n",
2004                                            good_bytes, scsi_bufflen(SCpnt)));
2005
2006         if (rq_data_dir(SCpnt->request) == READ && scsi_prot_sg_count(SCpnt))
2007                 sd_dif_complete(SCpnt, good_bytes);
2008
2009         return good_bytes;
2010 }
2011
2012 /*
2013  * spinup disk - called only in sd_revalidate_disk()
2014  */
2015 static void
2016 sd_spinup_disk(struct scsi_disk *sdkp)
2017 {
2018         unsigned char cmd[10];
2019         unsigned long spintime_expire = 0;
2020         int retries, spintime;
2021         unsigned int the_result;
2022         struct scsi_sense_hdr sshdr;
2023         int sense_valid = 0;
2024
2025         spintime = 0;
2026
2027         /* Spin up drives, as required.  Only do this at boot time */
2028         /* Spinup needs to be done for module loads too. */
2029         do {
2030                 retries = 0;
2031
2032                 do {
2033                         cmd[0] = TEST_UNIT_READY;
2034                         memset((void *) &cmd[1], 0, 9);
2035
2036                         the_result = scsi_execute_req(sdkp->device, cmd,
2037                                                       DMA_NONE, NULL, 0,
2038                                                       &sshdr, SD_TIMEOUT,
2039                                                       SD_MAX_RETRIES, NULL);
2040
2041                         /*
2042                          * If the drive has indicated to us that it
2043                          * doesn't have any media in it, don't bother
2044                          * with any more polling.
2045                          */
2046                         if (media_not_present(sdkp, &sshdr))
2047                                 return;
2048
2049                         if (the_result)
2050                                 sense_valid = scsi_sense_valid(&sshdr);
2051                         retries++;
2052                 } while (retries < 3 && 
2053                          (!scsi_status_is_good(the_result) ||
2054                           ((driver_byte(the_result) & DRIVER_SENSE) &&
2055                           sense_valid && sshdr.sense_key == UNIT_ATTENTION)));
2056
2057                 if ((driver_byte(the_result) & DRIVER_SENSE) == 0) {
2058                         /* no sense, TUR either succeeded or failed
2059                          * with a status error */
2060                         if(!spintime && !scsi_status_is_good(the_result)) {
2061                                 sd_print_result(sdkp, "Test Unit Ready failed",
2062                                                 the_result);
2063                         }
2064                         break;
2065                 }
2066
2067                 /*
2068                  * The device does not want the automatic start to be issued.
2069                  */
2070                 if (sdkp->device->no_start_on_add)
2071                         break;
2072
2073                 if (sense_valid && sshdr.sense_key == NOT_READY) {
2074                         if (sshdr.asc == 4 && sshdr.ascq == 3)
2075                                 break;  /* manual intervention required */
2076                         if (sshdr.asc == 4 && sshdr.ascq == 0xb)
2077                                 break;  /* standby */
2078                         if (sshdr.asc == 4 && sshdr.ascq == 0xc)
2079                                 break;  /* unavailable */
2080                         /*
2081                          * Issue command to spin up drive when not ready
2082                          */
2083                         if (!spintime) {
2084                                 sd_printk(KERN_NOTICE, sdkp, "Spinning up disk...");
2085                                 cmd[0] = START_STOP;
2086                                 cmd[1] = 1;     /* Return immediately */
2087                                 memset((void *) &cmd[2], 0, 8);
2088                                 cmd[4] = 1;     /* Start spin cycle */
2089                                 if (sdkp->device->start_stop_pwr_cond)
2090                                         cmd[4] |= 1 << 4;
2091                                 scsi_execute_req(sdkp->device, cmd, DMA_NONE,
2092                                                  NULL, 0, &sshdr,
2093                                                  SD_TIMEOUT, SD_MAX_RETRIES,
2094                                                  NULL);
2095                                 spintime_expire = jiffies + 100 * HZ;
2096                                 spintime = 1;
2097                         }
2098                         /* Wait 1 second for next try */
2099                         msleep(1000);
2100                         printk(".");
2101
2102                 /*
2103                  * Wait for USB flash devices with slow firmware.
2104                  * Yes, this sense key/ASC combination shouldn't
2105                  * occur here.  It's characteristic of these devices.
2106                  */
2107                 } else if (sense_valid &&
2108                                 sshdr.sense_key == UNIT_ATTENTION &&
2109                                 sshdr.asc == 0x28) {
2110                         if (!spintime) {
2111                                 spintime_expire = jiffies + 5 * HZ;
2112                                 spintime = 1;
2113                         }
2114                         /* Wait 1 second for next try */
2115                         msleep(1000);
2116                 } else {
2117                         /* we don't understand the sense code, so it's
2118                          * probably pointless to loop */
2119                         if(!spintime) {
2120                                 sd_printk(KERN_NOTICE, sdkp, "Unit Not Ready\n");
2121                                 sd_print_sense_hdr(sdkp, &sshdr);
2122                         }
2123                         break;
2124                 }
2125                                 
2126         } while (spintime && time_before_eq(jiffies, spintime_expire));
2127
2128         if (spintime) {
2129                 if (scsi_status_is_good(the_result))
2130                         printk("ready\n");
2131                 else
2132                         printk("not responding...\n");
2133         }
2134 }
2135
2136 /*
2137  * Determine whether disk supports Data Integrity Field.
2138  */
2139 static int sd_read_protection_type(struct scsi_disk *sdkp, unsigned char *buffer)
2140 {
2141         struct scsi_device *sdp = sdkp->device;
2142         u8 type;
2143         int ret = 0;
2144
2145         if (scsi_device_protection(sdp) == 0 || (buffer[12] & 1) == 0)
2146                 return ret;
2147
2148         type = ((buffer[12] >> 1) & 7) + 1; /* P_TYPE 0 = Type 1 */
2149
2150         if (type > T10_PI_TYPE3_PROTECTION)
2151                 ret = -ENODEV;
2152         else if (scsi_host_dif_capable(sdp->host, type))
2153                 ret = 1;
2154
2155         if (sdkp->first_scan || type != sdkp->protection_type)
2156                 switch (ret) {
2157                 case -ENODEV:
2158                         sd_printk(KERN_ERR, sdkp, "formatted with unsupported" \
2159                                   " protection type %u. Disabling disk!\n",
2160                                   type);
2161                         break;
2162                 case 1:
2163                         sd_printk(KERN_NOTICE, sdkp,
2164                                   "Enabling DIF Type %u protection\n", type);
2165                         break;
2166                 case 0:
2167                         sd_printk(KERN_NOTICE, sdkp,
2168                                   "Disabling DIF Type %u protection\n", type);
2169                         break;
2170                 }
2171
2172         sdkp->protection_type = type;
2173
2174         return ret;
2175 }
2176
2177 static void read_capacity_error(struct scsi_disk *sdkp, struct scsi_device *sdp,
2178                         struct scsi_sense_hdr *sshdr, int sense_valid,
2179                         int the_result)
2180 {
2181         if (driver_byte(the_result) & DRIVER_SENSE)
2182                 sd_print_sense_hdr(sdkp, sshdr);
2183         else
2184                 sd_printk(KERN_NOTICE, sdkp, "Sense not available.\n");
2185
2186         /*
2187          * Set dirty bit for removable devices if not ready -
2188          * sometimes drives will not report this properly.
2189          */
2190         if (sdp->removable &&
2191             sense_valid && sshdr->sense_key == NOT_READY)
2192                 set_media_not_present(sdkp);
2193
2194         /*
2195          * We used to set media_present to 0 here to indicate no media
2196          * in the drive, but some drives fail read capacity even with
2197          * media present, so we can't do that.
2198          */
2199         sdkp->capacity = 0; /* unknown mapped to zero - as usual */
2200 }
2201
2202 #define RC16_LEN 32
2203 #if RC16_LEN > SD_BUF_SIZE
2204 #error RC16_LEN must not be more than SD_BUF_SIZE
2205 #endif
2206
2207 #define READ_CAPACITY_RETRIES_ON_RESET  10
2208
2209 /*
2210  * Ensure that we don't overflow sector_t when CONFIG_LBDAF is not set
2211  * and the reported logical block size is bigger than 512 bytes. Note
2212  * that last_sector is a u64 and therefore logical_to_sectors() is not
2213  * applicable.
2214  */
2215 static bool sd_addressable_capacity(u64 lba, unsigned int sector_size)
2216 {
2217         u64 last_sector = (lba + 1ULL) << (ilog2(sector_size) - 9);
2218
2219         if (sizeof(sector_t) == 4 && last_sector > U32_MAX)
2220                 return false;
2221
2222         return true;
2223 }
2224
2225 static int read_capacity_16(struct scsi_disk *sdkp, struct scsi_device *sdp,
2226                                                 unsigned char *buffer)
2227 {
2228         unsigned char cmd[16];
2229         struct scsi_sense_hdr sshdr;
2230         int sense_valid = 0;
2231         int the_result;
2232         int retries = 3, reset_retries = READ_CAPACITY_RETRIES_ON_RESET;
2233         unsigned int alignment;
2234         unsigned long long lba;
2235         unsigned sector_size;
2236
2237         if (sdp->no_read_capacity_16)
2238                 return -EINVAL;
2239
2240         do {
2241                 memset(cmd, 0, 16);
2242                 cmd[0] = SERVICE_ACTION_IN_16;
2243                 cmd[1] = SAI_READ_CAPACITY_16;
2244                 cmd[13] = RC16_LEN;
2245                 memset(buffer, 0, RC16_LEN);
2246
2247                 the_result = scsi_execute_req(sdp, cmd, DMA_FROM_DEVICE,
2248                                         buffer, RC16_LEN, &sshdr,
2249                                         SD_TIMEOUT, SD_MAX_RETRIES, NULL);
2250
2251                 if (media_not_present(sdkp, &sshdr))
2252                         return -ENODEV;
2253
2254                 if (the_result) {
2255                         sense_valid = scsi_sense_valid(&sshdr);
2256                         if (sense_valid &&
2257                             sshdr.sense_key == ILLEGAL_REQUEST &&
2258                             (sshdr.asc == 0x20 || sshdr.asc == 0x24) &&
2259                             sshdr.ascq == 0x00)
2260                                 /* Invalid Command Operation Code or
2261                                  * Invalid Field in CDB, just retry
2262                                  * silently with RC10 */
2263                                 return -EINVAL;
2264                         if (sense_valid &&
2265                             sshdr.sense_key == UNIT_ATTENTION &&
2266                             sshdr.asc == 0x29 && sshdr.ascq == 0x00)
2267                                 /* Device reset might occur several times,
2268                                  * give it one more chance */
2269                                 if (--reset_retries > 0)
2270                                         continue;
2271                 }
2272                 retries--;
2273
2274         } while (the_result && retries);
2275
2276         if (the_result) {
2277                 sd_print_result(sdkp, "Read Capacity(16) failed", the_result);
2278                 read_capacity_error(sdkp, sdp, &sshdr, sense_valid, the_result);
2279                 return -EINVAL;
2280         }
2281
2282         sector_size = get_unaligned_be32(&buffer[8]);
2283         lba = get_unaligned_be64(&buffer[0]);
2284
2285         if (sd_read_protection_type(sdkp, buffer) < 0) {
2286                 sdkp->capacity = 0;
2287                 return -ENODEV;
2288         }
2289
2290         if (!sd_addressable_capacity(lba, sector_size)) {
2291                 sd_printk(KERN_ERR, sdkp, "Too big for this kernel. Use a "
2292                         "kernel compiled with support for large block "
2293                         "devices.\n");
2294                 sdkp->capacity = 0;
2295                 return -EOVERFLOW;
2296         }
2297
2298         /* Logical blocks per physical block exponent */
2299         sdkp->physical_block_size = (1 << (buffer[13] & 0xf)) * sector_size;
2300
2301         /* RC basis */
2302         sdkp->rc_basis = (buffer[12] >> 4) & 0x3;
2303
2304         /* Lowest aligned logical block */
2305         alignment = ((buffer[14] & 0x3f) << 8 | buffer[15]) * sector_size;
2306         blk_queue_alignment_offset(sdp->request_queue, alignment);
2307         if (alignment && sdkp->first_scan)
2308                 sd_printk(KERN_NOTICE, sdkp,
2309                           "physical block alignment offset: %u\n", alignment);
2310
2311         if (buffer[14] & 0x80) { /* LBPME */
2312                 sdkp->lbpme = 1;
2313
2314                 if (buffer[14] & 0x40) /* LBPRZ */
2315                         sdkp->lbprz = 1;
2316
2317                 sd_config_discard(sdkp, SD_LBP_WS16);
2318         }
2319
2320         sdkp->capacity = lba + 1;
2321         return sector_size;
2322 }
2323
2324 static int read_capacity_10(struct scsi_disk *sdkp, struct scsi_device *sdp,
2325                                                 unsigned char *buffer)
2326 {
2327         unsigned char cmd[16];
2328         struct scsi_sense_hdr sshdr;
2329         int sense_valid = 0;
2330         int the_result;
2331         int retries = 3, reset_retries = READ_CAPACITY_RETRIES_ON_RESET;
2332         sector_t lba;
2333         unsigned sector_size;
2334
2335         do {
2336                 cmd[0] = READ_CAPACITY;
2337                 memset(&cmd[1], 0, 9);
2338                 memset(buffer, 0, 8);
2339
2340                 the_result = scsi_execute_req(sdp, cmd, DMA_FROM_DEVICE,
2341                                         buffer, 8, &sshdr,
2342                                         SD_TIMEOUT, SD_MAX_RETRIES, NULL);
2343
2344                 if (media_not_present(sdkp, &sshdr))
2345                         return -ENODEV;
2346
2347                 if (the_result) {
2348                         sense_valid = scsi_sense_valid(&sshdr);
2349                         if (sense_valid &&
2350                             sshdr.sense_key == UNIT_ATTENTION &&
2351                             sshdr.asc == 0x29 && sshdr.ascq == 0x00)
2352                                 /* Device reset might occur several times,
2353                                  * give it one more chance */
2354                                 if (--reset_retries > 0)
2355                                         continue;
2356                 }
2357                 retries--;
2358
2359         } while (the_result && retries);
2360
2361         if (the_result) {
2362                 sd_print_result(sdkp, "Read Capacity(10) failed", the_result);
2363                 read_capacity_error(sdkp, sdp, &sshdr, sense_valid, the_result);
2364                 return -EINVAL;
2365         }
2366
2367         sector_size = get_unaligned_be32(&buffer[4]);
2368         lba = get_unaligned_be32(&buffer[0]);
2369
2370         if (sdp->no_read_capacity_16 && (lba == 0xffffffff)) {
2371                 /* Some buggy (usb cardreader) devices return an lba of
2372                    0xffffffff when the want to report a size of 0 (with
2373                    which they really mean no media is present) */
2374                 sdkp->capacity = 0;
2375                 sdkp->physical_block_size = sector_size;
2376                 return sector_size;
2377         }
2378
2379         if (!sd_addressable_capacity(lba, sector_size)) {
2380                 sd_printk(KERN_ERR, sdkp, "Too big for this kernel. Use a "
2381                         "kernel compiled with support for large block "
2382                         "devices.\n");
2383                 sdkp->capacity = 0;
2384                 return -EOVERFLOW;
2385         }
2386
2387         sdkp->capacity = lba + 1;
2388         sdkp->physical_block_size = sector_size;
2389         return sector_size;
2390 }
2391
2392 static int sd_try_rc16_first(struct scsi_device *sdp)
2393 {
2394         if (sdp->host->max_cmd_len < 16)
2395                 return 0;
2396         if (sdp->try_rc_10_first)
2397                 return 0;
2398         if (sdp->scsi_level > SCSI_SPC_2)
2399                 return 1;
2400         if (scsi_device_protection(sdp))
2401                 return 1;
2402         return 0;
2403 }
2404
2405 /*
2406  * read disk capacity
2407  */
2408 static void
2409 sd_read_capacity(struct scsi_disk *sdkp, unsigned char *buffer)
2410 {
2411         int sector_size;
2412         struct scsi_device *sdp = sdkp->device;
2413
2414         if (sd_try_rc16_first(sdp)) {
2415                 sector_size = read_capacity_16(sdkp, sdp, buffer);
2416                 if (sector_size == -EOVERFLOW)
2417                         goto got_data;
2418                 if (sector_size == -ENODEV)
2419                         return;
2420                 if (sector_size < 0)
2421                         sector_size = read_capacity_10(sdkp, sdp, buffer);
2422                 if (sector_size < 0)
2423                         return;
2424         } else {
2425                 sector_size = read_capacity_10(sdkp, sdp, buffer);
2426                 if (sector_size == -EOVERFLOW)
2427                         goto got_data;
2428                 if (sector_size < 0)
2429                         return;
2430                 if ((sizeof(sdkp->capacity) > 4) &&
2431                     (sdkp->capacity > 0xffffffffULL)) {
2432                         int old_sector_size = sector_size;
2433                         sd_printk(KERN_NOTICE, sdkp, "Very big device. "
2434                                         "Trying to use READ CAPACITY(16).\n");
2435                         sector_size = read_capacity_16(sdkp, sdp, buffer);
2436                         if (sector_size < 0) {
2437                                 sd_printk(KERN_NOTICE, sdkp,
2438                                         "Using 0xffffffff as device size\n");
2439                                 sdkp->capacity = 1 + (sector_t) 0xffffffff;
2440                                 sector_size = old_sector_size;
2441                                 goto got_data;
2442                         }
2443                 }
2444         }
2445
2446         /* Some devices are known to return the total number of blocks,
2447          * not the highest block number.  Some devices have versions
2448          * which do this and others which do not.  Some devices we might
2449          * suspect of doing this but we don't know for certain.
2450          *
2451          * If we know the reported capacity is wrong, decrement it.  If
2452          * we can only guess, then assume the number of blocks is even
2453          * (usually true but not always) and err on the side of lowering
2454          * the capacity.
2455          */
2456         if (sdp->fix_capacity ||
2457             (sdp->guess_capacity && (sdkp->capacity & 0x01))) {
2458                 sd_printk(KERN_INFO, sdkp, "Adjusting the sector count "
2459                                 "from its reported value: %llu\n",
2460                                 (unsigned long long) sdkp->capacity);
2461                 --sdkp->capacity;
2462         }
2463
2464 got_data:
2465         if (sector_size == 0) {
2466                 sector_size = 512;
2467                 sd_printk(KERN_NOTICE, sdkp, "Sector size 0 reported, "
2468                           "assuming 512.\n");
2469         }
2470
2471         if (sector_size != 512 &&
2472             sector_size != 1024 &&
2473             sector_size != 2048 &&
2474             sector_size != 4096) {
2475                 sd_printk(KERN_NOTICE, sdkp, "Unsupported sector size %d.\n",
2476                           sector_size);
2477                 /*
2478                  * The user might want to re-format the drive with
2479                  * a supported sectorsize.  Once this happens, it
2480                  * would be relatively trivial to set the thing up.
2481                  * For this reason, we leave the thing in the table.
2482                  */
2483                 sdkp->capacity = 0;
2484                 /*
2485                  * set a bogus sector size so the normal read/write
2486                  * logic in the block layer will eventually refuse any
2487                  * request on this device without tripping over power
2488                  * of two sector size assumptions
2489                  */
2490                 sector_size = 512;
2491         }
2492         blk_queue_logical_block_size(sdp->request_queue, sector_size);
2493         blk_queue_physical_block_size(sdp->request_queue,
2494                                       sdkp->physical_block_size);
2495         sdkp->device->sector_size = sector_size;
2496
2497         if (sdkp->capacity > 0xffffffff)
2498                 sdp->use_16_for_rw = 1;
2499
2500 }
2501
2502 /*
2503  * Print disk capacity
2504  */
2505 static void
2506 sd_print_capacity(struct scsi_disk *sdkp,
2507                   sector_t old_capacity)
2508 {
2509         int sector_size = sdkp->device->sector_size;
2510         char cap_str_2[10], cap_str_10[10];
2511
2512         string_get_size(sdkp->capacity, sector_size,
2513                         STRING_UNITS_2, cap_str_2, sizeof(cap_str_2));
2514         string_get_size(sdkp->capacity, sector_size,
2515                         STRING_UNITS_10, cap_str_10,
2516                         sizeof(cap_str_10));
2517
2518         if (sdkp->first_scan || old_capacity != sdkp->capacity) {
2519                 sd_printk(KERN_NOTICE, sdkp,
2520                           "%llu %d-byte logical blocks: (%s/%s)\n",
2521                           (unsigned long long)sdkp->capacity,
2522                           sector_size, cap_str_10, cap_str_2);
2523
2524                 if (sdkp->physical_block_size != sector_size)
2525                         sd_printk(KERN_NOTICE, sdkp,
2526                                   "%u-byte physical blocks\n",
2527                                   sdkp->physical_block_size);
2528
2529                 sd_zbc_print_zones(sdkp);
2530         }
2531 }
2532
2533 /* called with buffer of length 512 */
2534 static inline int
2535 sd_do_mode_sense(struct scsi_device *sdp, int dbd, int modepage,
2536                  unsigned char *buffer, int len, struct scsi_mode_data *data,
2537                  struct scsi_sense_hdr *sshdr)
2538 {
2539         return scsi_mode_sense(sdp, dbd, modepage, buffer, len,
2540                                SD_TIMEOUT, SD_MAX_RETRIES, data,
2541                                sshdr);
2542 }
2543
2544 /*
2545  * read write protect setting, if possible - called only in sd_revalidate_disk()
2546  * called with buffer of length SD_BUF_SIZE
2547  */
2548 static void
2549 sd_read_write_protect_flag(struct scsi_disk *sdkp, unsigned char *buffer)
2550 {
2551         int res;
2552         struct scsi_device *sdp = sdkp->device;
2553         struct scsi_mode_data data;
2554         int old_wp = sdkp->write_prot;
2555
2556         set_disk_ro(sdkp->disk, 0);
2557         if (sdp->skip_ms_page_3f) {
2558                 sd_first_printk(KERN_NOTICE, sdkp, "Assuming Write Enabled\n");
2559                 return;
2560         }
2561
2562         if (sdp->use_192_bytes_for_3f) {
2563                 res = sd_do_mode_sense(sdp, 0, 0x3F, buffer, 192, &data, NULL);
2564         } else {
2565                 /*
2566                  * First attempt: ask for all pages (0x3F), but only 4 bytes.
2567                  * We have to start carefully: some devices hang if we ask
2568                  * for more than is available.
2569                  */
2570                 res = sd_do_mode_sense(sdp, 0, 0x3F, buffer, 4, &data, NULL);
2571
2572                 /*
2573                  * Second attempt: ask for page 0 When only page 0 is
2574                  * implemented, a request for page 3F may return Sense Key
2575                  * 5: Illegal Request, Sense Code 24: Invalid field in
2576                  * CDB.
2577                  */
2578                 if (!scsi_status_is_good(res))
2579                         res = sd_do_mode_sense(sdp, 0, 0, buffer, 4, &data, NULL);
2580
2581                 /*
2582                  * Third attempt: ask 255 bytes, as we did earlier.
2583                  */
2584                 if (!scsi_status_is_good(res))
2585                         res = sd_do_mode_sense(sdp, 0, 0x3F, buffer, 255,
2586                                                &data, NULL);
2587         }
2588
2589         if (!scsi_status_is_good(res)) {
2590                 sd_first_printk(KERN_WARNING, sdkp,
2591                           "Test WP failed, assume Write Enabled\n");
2592         } else {
2593                 sdkp->write_prot = ((data.device_specific & 0x80) != 0);
2594                 set_disk_ro(sdkp->disk, sdkp->write_prot);
2595                 if (sdkp->first_scan || old_wp != sdkp->write_prot) {
2596                         sd_printk(KERN_NOTICE, sdkp, "Write Protect is %s\n",
2597                                   sdkp->write_prot ? "on" : "off");
2598                         sd_printk(KERN_DEBUG, sdkp, "Mode Sense: %4ph\n", buffer);
2599                 }
2600         }
2601 }
2602
2603 /*
2604  * sd_read_cache_type - called only from sd_revalidate_disk()
2605  * called with buffer of length SD_BUF_SIZE
2606  */
2607 static void
2608 sd_read_cache_type(struct scsi_disk *sdkp, unsigned char *buffer)
2609 {
2610         int len = 0, res;
2611         struct scsi_device *sdp = sdkp->device;
2612
2613         int dbd;
2614         int modepage;
2615         int first_len;
2616         struct scsi_mode_data data;
2617         struct scsi_sense_hdr sshdr;
2618         int old_wce = sdkp->WCE;
2619         int old_rcd = sdkp->RCD;
2620         int old_dpofua = sdkp->DPOFUA;
2621
2622
2623         if (sdkp->cache_override)
2624                 return;
2625
2626         first_len = 4;
2627         if (sdp->skip_ms_page_8) {
2628                 if (sdp->type == TYPE_RBC)
2629                         goto defaults;
2630                 else {
2631                         if (sdp->skip_ms_page_3f)
2632                                 goto defaults;
2633                         modepage = 0x3F;
2634                         if (sdp->use_192_bytes_for_3f)
2635                                 first_len = 192;
2636                         dbd = 0;
2637                 }
2638         } else if (sdp->type == TYPE_RBC) {
2639                 modepage = 6;
2640                 dbd = 8;
2641         } else {
2642                 modepage = 8;
2643                 dbd = 0;
2644         }
2645
2646         /* cautiously ask */
2647         res = sd_do_mode_sense(sdp, dbd, modepage, buffer, first_len,
2648                         &data, &sshdr);
2649
2650         if (!scsi_status_is_good(res))
2651                 goto bad_sense;
2652
2653         if (!data.header_length) {
2654                 modepage = 6;
2655                 first_len = 0;
2656                 sd_first_printk(KERN_ERR, sdkp,
2657                                 "Missing header in MODE_SENSE response\n");
2658         }
2659
2660         /* that went OK, now ask for the proper length */
2661         len = data.length;
2662
2663         /*
2664          * We're only interested in the first three bytes, actually.
2665          * But the data cache page is defined for the first 20.
2666          */
2667         if (len < 3)
2668                 goto bad_sense;
2669         else if (len > SD_BUF_SIZE) {
2670                 sd_first_printk(KERN_NOTICE, sdkp, "Truncating mode parameter "
2671                           "data from %d to %d bytes\n", len, SD_BUF_SIZE);
2672                 len = SD_BUF_SIZE;
2673         }
2674         if (modepage == 0x3F && sdp->use_192_bytes_for_3f)
2675                 len = 192;
2676
2677         /* Get the data */
2678         if (len > first_len)
2679                 res = sd_do_mode_sense(sdp, dbd, modepage, buffer, len,
2680                                 &data, &sshdr);
2681
2682         if (scsi_status_is_good(res)) {
2683                 int offset = data.header_length + data.block_descriptor_length;
2684
2685                 while (offset < len) {
2686                         u8 page_code = buffer[offset] & 0x3F;
2687                         u8 spf       = buffer[offset] & 0x40;
2688
2689                         if (page_code == 8 || page_code == 6) {
2690                                 /* We're interested only in the first 3 bytes.
2691                                  */
2692                                 if (len - offset <= 2) {
2693                                         sd_first_printk(KERN_ERR, sdkp,
2694                                                 "Incomplete mode parameter "
2695                                                         "data\n");
2696                                         goto defaults;
2697                                 } else {
2698                                         modepage = page_code;
2699                                         goto Page_found;
2700                                 }
2701                         } else {
2702                                 /* Go to the next page */
2703                                 if (spf && len - offset > 3)
2704                                         offset += 4 + (buffer[offset+2] << 8) +
2705                                                 buffer[offset+3];
2706                                 else if (!spf && len - offset > 1)
2707                                         offset += 2 + buffer[offset+1];
2708                                 else {
2709                                         sd_first_printk(KERN_ERR, sdkp,
2710                                                         "Incomplete mode "
2711                                                         "parameter data\n");
2712                                         goto defaults;
2713                                 }
2714                         }
2715                 }
2716
2717                 sd_first_printk(KERN_ERR, sdkp, "No Caching mode page found\n");
2718                 goto defaults;
2719
2720         Page_found:
2721                 if (modepage == 8) {
2722                         sdkp->WCE = ((buffer[offset + 2] & 0x04) != 0);
2723                         sdkp->RCD = ((buffer[offset + 2] & 0x01) != 0);
2724                 } else {
2725                         sdkp->WCE = ((buffer[offset + 2] & 0x01) == 0);
2726                         sdkp->RCD = 0;
2727                 }
2728
2729                 sdkp->DPOFUA = (data.device_specific & 0x10) != 0;
2730                 if (sdp->broken_fua) {
2731                         sd_first_printk(KERN_NOTICE, sdkp, "Disabling FUA\n");
2732                         sdkp->DPOFUA = 0;
2733                 } else if (sdkp->DPOFUA && !sdkp->device->use_10_for_rw &&
2734                            !sdkp->device->use_16_for_rw) {
2735                         sd_first_printk(KERN_NOTICE, sdkp,
2736                                   "Uses READ/WRITE(6), disabling FUA\n");
2737                         sdkp->DPOFUA = 0;
2738                 }
2739
2740                 /* No cache flush allowed for write protected devices */
2741                 if (sdkp->WCE && sdkp->write_prot)
2742                         sdkp->WCE = 0;
2743
2744                 if (sdkp->first_scan || old_wce != sdkp->WCE ||
2745                     old_rcd != sdkp->RCD || old_dpofua != sdkp->DPOFUA)
2746                         sd_printk(KERN_NOTICE, sdkp,
2747                                   "Write cache: %s, read cache: %s, %s\n",
2748                                   sdkp->WCE ? "enabled" : "disabled",
2749                                   sdkp->RCD ? "disabled" : "enabled",
2750                                   sdkp->DPOFUA ? "supports DPO and FUA"
2751                                   : "doesn't support DPO or FUA");
2752
2753                 return;
2754         }
2755
2756 bad_sense:
2757         if (scsi_sense_valid(&sshdr) &&
2758             sshdr.sense_key == ILLEGAL_REQUEST &&
2759             sshdr.asc == 0x24 && sshdr.ascq == 0x0)
2760                 /* Invalid field in CDB */
2761                 sd_first_printk(KERN_NOTICE, sdkp, "Cache data unavailable\n");
2762         else
2763                 sd_first_printk(KERN_ERR, sdkp,
2764                                 "Asking for cache data failed\n");
2765
2766 defaults:
2767         if (sdp->wce_default_on) {
2768                 sd_first_printk(KERN_NOTICE, sdkp,
2769                                 "Assuming drive cache: write back\n");
2770                 sdkp->WCE = 1;
2771         } else {
2772                 sd_first_printk(KERN_ERR, sdkp,
2773                                 "Assuming drive cache: write through\n");
2774                 sdkp->WCE = 0;
2775         }
2776         sdkp->RCD = 0;
2777         sdkp->DPOFUA = 0;
2778 }
2779
2780 /*
2781  * The ATO bit indicates whether the DIF application tag is available
2782  * for use by the operating system.
2783  */
2784 static void sd_read_app_tag_own(struct scsi_disk *sdkp, unsigned char *buffer)
2785 {
2786         int res, offset;
2787         struct scsi_device *sdp = sdkp->device;
2788         struct scsi_mode_data data;
2789         struct scsi_sense_hdr sshdr;
2790
2791         if (sdp->type != TYPE_DISK && sdp->type != TYPE_ZBC)
2792                 return;
2793
2794         if (sdkp->protection_type == 0)
2795                 return;
2796
2797         res = scsi_mode_sense(sdp, 1, 0x0a, buffer, 36, SD_TIMEOUT,
2798                               SD_MAX_RETRIES, &data, &sshdr);
2799
2800         if (!scsi_status_is_good(res) || !data.header_length ||
2801             data.length < 6) {
2802                 sd_first_printk(KERN_WARNING, sdkp,
2803                           "getting Control mode page failed, assume no ATO\n");
2804
2805                 if (scsi_sense_valid(&sshdr))
2806                         sd_print_sense_hdr(sdkp, &sshdr);
2807
2808                 return;
2809         }
2810
2811         offset = data.header_length + data.block_descriptor_length;
2812
2813         if ((buffer[offset] & 0x3f) != 0x0a) {
2814                 sd_first_printk(KERN_ERR, sdkp, "ATO Got wrong page\n");
2815                 return;
2816         }
2817
2818         if ((buffer[offset + 5] & 0x80) == 0)
2819                 return;
2820
2821         sdkp->ATO = 1;
2822
2823         return;
2824 }
2825
2826 /**
2827  * sd_read_block_limits - Query disk device for preferred I/O sizes.
2828  * @sdkp: disk to query
2829  */
2830 static void sd_read_block_limits(struct scsi_disk *sdkp)
2831 {
2832         unsigned int sector_sz = sdkp->device->sector_size;
2833         const int vpd_len = 64;
2834         unsigned char *buffer = kmalloc(vpd_len, GFP_KERNEL);
2835
2836         if (!buffer ||
2837             /* Block Limits VPD */
2838             scsi_get_vpd_page(sdkp->device, 0xb0, buffer, vpd_len))
2839                 goto out;
2840
2841         blk_queue_io_min(sdkp->disk->queue,
2842                          get_unaligned_be16(&buffer[6]) * sector_sz);
2843
2844         sdkp->max_xfer_blocks = get_unaligned_be32(&buffer[8]);
2845         sdkp->opt_xfer_blocks = get_unaligned_be32(&buffer[12]);
2846
2847         if (buffer[3] == 0x3c) {
2848                 unsigned int lba_count, desc_count;
2849
2850                 sdkp->max_ws_blocks = (u32)get_unaligned_be64(&buffer[36]);
2851
2852                 if (!sdkp->lbpme)
2853                         goto out;
2854
2855                 lba_count = get_unaligned_be32(&buffer[20]);
2856                 desc_count = get_unaligned_be32(&buffer[24]);
2857
2858                 if (lba_count && desc_count)
2859                         sdkp->max_unmap_blocks = lba_count;
2860
2861                 sdkp->unmap_granularity = get_unaligned_be32(&buffer[28]);
2862
2863                 if (buffer[32] & 0x80)
2864                         sdkp->unmap_alignment =
2865                                 get_unaligned_be32(&buffer[32]) & ~(1 << 31);
2866
2867                 if (!sdkp->lbpvpd) { /* LBP VPD page not provided */
2868
2869                         if (sdkp->max_unmap_blocks)
2870                                 sd_config_discard(sdkp, SD_LBP_UNMAP);
2871                         else
2872                                 sd_config_discard(sdkp, SD_LBP_WS16);
2873
2874                 } else {        /* LBP VPD page tells us what to use */
2875                         if (sdkp->lbpu && sdkp->max_unmap_blocks)
2876                                 sd_config_discard(sdkp, SD_LBP_UNMAP);
2877                         else if (sdkp->lbpws)
2878                                 sd_config_discard(sdkp, SD_LBP_WS16);
2879                         else if (sdkp->lbpws10)
2880                                 sd_config_discard(sdkp, SD_LBP_WS10);
2881                         else if (sdkp->lbpu && sdkp->max_unmap_blocks)
2882                                 sd_config_discard(sdkp, SD_LBP_UNMAP);
2883                         else
2884                                 sd_config_discard(sdkp, SD_LBP_DISABLE);
2885                 }
2886         }
2887
2888  out:
2889         kfree(buffer);
2890 }
2891
2892 /**
2893  * sd_read_block_characteristics - Query block dev. characteristics
2894  * @sdkp: disk to query
2895  */
2896 static void sd_read_block_characteristics(struct scsi_disk *sdkp)
2897 {
2898         struct request_queue *q = sdkp->disk->queue;
2899         unsigned char *buffer;
2900         u16 rot;
2901         const int vpd_len = 64;
2902
2903         buffer = kmalloc(vpd_len, GFP_KERNEL);
2904
2905         if (!buffer ||
2906             /* Block Device Characteristics VPD */
2907             scsi_get_vpd_page(sdkp->device, 0xb1, buffer, vpd_len))
2908                 goto out;
2909
2910         rot = get_unaligned_be16(&buffer[4]);
2911
2912         if (rot == 1) {
2913                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
2914                 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, q);
2915         }
2916
2917         if (sdkp->device->type == TYPE_ZBC) {
2918                 /* Host-managed */
2919                 q->limits.zoned = BLK_ZONED_HM;
2920         } else {
2921                 sdkp->zoned = (buffer[8] >> 4) & 3;
2922                 if (sdkp->zoned == 1)
2923                         /* Host-aware */
2924                         q->limits.zoned = BLK_ZONED_HA;
2925                 else
2926                         /*
2927                          * Treat drive-managed devices as
2928                          * regular block devices.
2929                          */
2930                         q->limits.zoned = BLK_ZONED_NONE;
2931         }
2932         if (blk_queue_is_zoned(q) && sdkp->first_scan)
2933                 sd_printk(KERN_NOTICE, sdkp, "Host-%s zoned block device\n",
2934                       q->limits.zoned == BLK_ZONED_HM ? "managed" : "aware");
2935
2936  out:
2937         kfree(buffer);
2938 }
2939
2940 /**
2941  * sd_read_block_provisioning - Query provisioning VPD page
2942  * @sdkp: disk to query
2943  */
2944 static void sd_read_block_provisioning(struct scsi_disk *sdkp)
2945 {
2946         unsigned char *buffer;
2947         const int vpd_len = 8;
2948
2949         if (sdkp->lbpme == 0)
2950                 return;
2951
2952         buffer = kmalloc(vpd_len, GFP_KERNEL);
2953
2954         if (!buffer || scsi_get_vpd_page(sdkp->device, 0xb2, buffer, vpd_len))
2955                 goto out;
2956
2957         sdkp->lbpvpd    = 1;
2958         sdkp->lbpu      = (buffer[5] >> 7) & 1; /* UNMAP */
2959         sdkp->lbpws     = (buffer[5] >> 6) & 1; /* WRITE SAME(16) with UNMAP */
2960         sdkp->lbpws10   = (buffer[5] >> 5) & 1; /* WRITE SAME(10) with UNMAP */
2961
2962  out:
2963         kfree(buffer);
2964 }
2965
2966 static void sd_read_write_same(struct scsi_disk *sdkp, unsigned char *buffer)
2967 {
2968         struct scsi_device *sdev = sdkp->device;
2969
2970         if (sdev->host->no_write_same) {
2971                 sdev->no_write_same = 1;
2972
2973                 return;
2974         }
2975
2976         if (scsi_report_opcode(sdev, buffer, SD_BUF_SIZE, INQUIRY) < 0) {
2977                 /* too large values might cause issues with arcmsr */
2978                 int vpd_buf_len = 64;
2979
2980                 sdev->no_report_opcodes = 1;
2981
2982                 /* Disable WRITE SAME if REPORT SUPPORTED OPERATION
2983                  * CODES is unsupported and the device has an ATA
2984                  * Information VPD page (SAT).
2985                  */
2986                 if (!scsi_get_vpd_page(sdev, 0x89, buffer, vpd_buf_len))
2987                         sdev->no_write_same = 1;
2988         }
2989
2990         if (scsi_report_opcode(sdev, buffer, SD_BUF_SIZE, WRITE_SAME_16) == 1)
2991                 sdkp->ws16 = 1;
2992
2993         if (scsi_report_opcode(sdev, buffer, SD_BUF_SIZE, WRITE_SAME) == 1)
2994                 sdkp->ws10 = 1;
2995 }
2996
2997 /**
2998  *      sd_revalidate_disk - called the first time a new disk is seen,
2999  *      performs disk spin up, read_capacity, etc.
3000  *      @disk: struct gendisk we care about
3001  **/
3002 static int sd_revalidate_disk(struct gendisk *disk)
3003 {
3004         struct scsi_disk *sdkp = scsi_disk(disk);
3005         struct scsi_device *sdp = sdkp->device;
3006         struct request_queue *q = sdkp->disk->queue;
3007         sector_t old_capacity = sdkp->capacity;
3008         unsigned char *buffer;
3009         unsigned int dev_max, rw_max;
3010
3011         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp,
3012                                       "sd_revalidate_disk\n"));
3013
3014         /*
3015          * If the device is offline, don't try and read capacity or any
3016          * of the other niceties.
3017          */
3018         if (!scsi_device_online(sdp))
3019                 goto out;
3020
3021         buffer = kmalloc(SD_BUF_SIZE, GFP_KERNEL);
3022         if (!buffer) {
3023                 sd_printk(KERN_WARNING, sdkp, "sd_revalidate_disk: Memory "
3024                           "allocation failure.\n");
3025                 goto out;
3026         }
3027
3028         sd_spinup_disk(sdkp);
3029
3030         /*
3031          * Without media there is no reason to ask; moreover, some devices
3032          * react badly if we do.
3033          */
3034         if (sdkp->media_present) {
3035                 sd_read_capacity(sdkp, buffer);
3036
3037                 if (scsi_device_supports_vpd(sdp)) {
3038                         sd_read_block_provisioning(sdkp);
3039                         sd_read_block_limits(sdkp);
3040                         sd_read_block_characteristics(sdkp);
3041                         sd_zbc_read_zones(sdkp, buffer);
3042                 }
3043
3044                 sd_print_capacity(sdkp, old_capacity);
3045
3046                 sd_read_write_protect_flag(sdkp, buffer);
3047                 sd_read_cache_type(sdkp, buffer);
3048                 sd_read_app_tag_own(sdkp, buffer);
3049                 sd_read_write_same(sdkp, buffer);
3050         }
3051
3052         sdkp->first_scan = 0;
3053
3054         /*
3055          * We now have all cache related info, determine how we deal
3056          * with flush requests.
3057          */
3058         sd_set_flush_flag(sdkp);
3059
3060         /* Initial block count limit based on CDB TRANSFER LENGTH field size. */
3061         dev_max = sdp->use_16_for_rw ? SD_MAX_XFER_BLOCKS : SD_DEF_XFER_BLOCKS;
3062
3063         /* Some devices report a maximum block count for READ/WRITE requests. */
3064         dev_max = min_not_zero(dev_max, sdkp->max_xfer_blocks);
3065         q->limits.max_dev_sectors = logical_to_sectors(sdp, dev_max);
3066
3067         /*
3068          * Use the device's preferred I/O size for reads and writes
3069          * unless the reported value is unreasonably small, large, or
3070          * garbage.
3071          */
3072         if (sdkp->opt_xfer_blocks &&
3073             sdkp->opt_xfer_blocks <= dev_max &&
3074             sdkp->opt_xfer_blocks <= SD_DEF_XFER_BLOCKS &&
3075             logical_to_bytes(sdp, sdkp->opt_xfer_blocks) >= PAGE_SIZE) {
3076                 q->limits.io_opt = logical_to_bytes(sdp, sdkp->opt_xfer_blocks);
3077                 rw_max = logical_to_sectors(sdp, sdkp->opt_xfer_blocks);
3078         } else
3079                 rw_max = min_not_zero(logical_to_sectors(sdp, dev_max),
3080                                       (sector_t)BLK_DEF_MAX_SECTORS);
3081
3082         /* Combine with controller limits */
3083         q->limits.max_sectors = min(rw_max, queue_max_hw_sectors(q));
3084
3085         set_capacity(disk, logical_to_sectors(sdp, sdkp->capacity));
3086         sd_config_write_same(sdkp);
3087         kfree(buffer);
3088
3089  out:
3090         return 0;
3091 }
3092
3093 /**
3094  *      sd_unlock_native_capacity - unlock native capacity
3095  *      @disk: struct gendisk to set capacity for
3096  *
3097  *      Block layer calls this function if it detects that partitions
3098  *      on @disk reach beyond the end of the device.  If the SCSI host
3099  *      implements ->unlock_native_capacity() method, it's invoked to
3100  *      give it a chance to adjust the device capacity.
3101  *
3102  *      CONTEXT:
3103  *      Defined by block layer.  Might sleep.
3104  */
3105 static void sd_unlock_native_capacity(struct gendisk *disk)
3106 {
3107         struct scsi_device *sdev = scsi_disk(disk)->device;
3108
3109         if (sdev->host->hostt->unlock_native_capacity)
3110                 sdev->host->hostt->unlock_native_capacity(sdev);
3111 }
3112
3113 /**
3114  *      sd_format_disk_name - format disk name
3115  *      @prefix: name prefix - ie. "sd" for SCSI disks
3116  *      @index: index of the disk to format name for
3117  *      @buf: output buffer
3118  *      @buflen: length of the output buffer
3119  *
3120  *      SCSI disk names starts at sda.  The 26th device is sdz and the
3121  *      27th is sdaa.  The last one for two lettered suffix is sdzz
3122  *      which is followed by sdaaa.
3123  *
3124  *      This is basically 26 base counting with one extra 'nil' entry
3125  *      at the beginning from the second digit on and can be
3126  *      determined using similar method as 26 base conversion with the
3127  *      index shifted -1 after each digit is computed.
3128  *
3129  *      CONTEXT:
3130  *      Don't care.
3131  *
3132  *      RETURNS:
3133  *      0 on success, -errno on failure.
3134  */
3135 static int sd_format_disk_name(char *prefix, int index, char *buf, int buflen)
3136 {
3137         const int base = 'z' - 'a' + 1;
3138         char *begin = buf + strlen(prefix);
3139         char *end = buf + buflen;
3140         char *p;
3141         int unit;
3142
3143         p = end - 1;
3144         *p = '\0';
3145         unit = base;
3146         do {
3147                 if (p == begin)
3148                         return -EINVAL;
3149                 *--p = 'a' + (index % unit);
3150                 index = (index / unit) - 1;
3151         } while (index >= 0);
3152
3153         memmove(begin, p, end - p);
3154         memcpy(buf, prefix, strlen(prefix));
3155
3156         return 0;
3157 }
3158
3159 /*
3160  * The asynchronous part of sd_probe
3161  */
3162 static void sd_probe_async(void *data, async_cookie_t cookie)
3163 {
3164         struct scsi_disk *sdkp = data;
3165         struct scsi_device *sdp;
3166         struct gendisk *gd;
3167         u32 index;
3168         struct device *dev;
3169
3170         sdp = sdkp->device;
3171         gd = sdkp->disk;
3172         index = sdkp->index;
3173         dev = &sdp->sdev_gendev;
3174
3175         gd->major = sd_major((index & 0xf0) >> 4);
3176         gd->first_minor = ((index & 0xf) << 4) | (index & 0xfff00);
3177         gd->minors = SD_MINORS;
3178
3179         gd->fops = &sd_fops;
3180         gd->private_data = &sdkp->driver;
3181         gd->queue = sdkp->device->request_queue;
3182
3183         /* defaults, until the device tells us otherwise */
3184         sdp->sector_size = 512;
3185         sdkp->capacity = 0;
3186         sdkp->media_present = 1;
3187         sdkp->write_prot = 0;
3188         sdkp->cache_override = 0;
3189         sdkp->WCE = 0;
3190         sdkp->RCD = 0;
3191         sdkp->ATO = 0;
3192         sdkp->first_scan = 1;
3193         sdkp->max_medium_access_timeouts = SD_MAX_MEDIUM_TIMEOUTS;
3194
3195         sd_revalidate_disk(gd);
3196
3197         gd->flags = GENHD_FL_EXT_DEVT;
3198         if (sdp->removable) {
3199                 gd->flags |= GENHD_FL_REMOVABLE;
3200                 gd->events |= DISK_EVENT_MEDIA_CHANGE;
3201         }
3202
3203         blk_pm_runtime_init(sdp->request_queue, dev);
3204         device_add_disk(dev, gd);
3205         if (sdkp->capacity)
3206                 sd_dif_config_host(sdkp);
3207
3208         sd_revalidate_disk(gd);
3209
3210         sd_printk(KERN_NOTICE, sdkp, "Attached SCSI %sdisk\n",
3211                   sdp->removable ? "removable " : "");
3212         scsi_autopm_put_device(sdp);
3213         put_device(&sdkp->dev);
3214 }
3215
3216 /**
3217  *      sd_probe - called during driver initialization and whenever a
3218  *      new scsi device is attached to the system. It is called once
3219  *      for each scsi device (not just disks) present.
3220  *      @dev: pointer to device object
3221  *
3222  *      Returns 0 if successful (or not interested in this scsi device 
3223  *      (e.g. scanner)); 1 when there is an error.
3224  *
3225  *      Note: this function is invoked from the scsi mid-level.
3226  *      This function sets up the mapping between a given 
3227  *      <host,channel,id,lun> (found in sdp) and new device name 
3228  *      (e.g. /dev/sda). More precisely it is the block device major 
3229  *      and minor number that is chosen here.
3230  *
3231  *      Assume sd_probe is not re-entrant (for time being)
3232  *      Also think about sd_probe() and sd_remove() running coincidentally.
3233  **/
3234 static int sd_probe(struct device *dev)
3235 {
3236         struct scsi_device *sdp = to_scsi_device(dev);
3237         struct scsi_disk *sdkp;
3238         struct gendisk *gd;
3239         int index;
3240         int error;
3241
3242         scsi_autopm_get_device(sdp);
3243         error = -ENODEV;
3244         if (sdp->type != TYPE_DISK &&
3245             sdp->type != TYPE_ZBC &&
3246             sdp->type != TYPE_MOD &&
3247             sdp->type != TYPE_RBC)
3248                 goto out;
3249
3250 #ifndef CONFIG_BLK_DEV_ZONED
3251         if (sdp->type == TYPE_ZBC)
3252                 goto out;
3253 #endif
3254         SCSI_LOG_HLQUEUE(3, sdev_printk(KERN_INFO, sdp,
3255                                         "sd_probe\n"));
3256
3257         error = -ENOMEM;
3258         sdkp = kzalloc(sizeof(*sdkp), GFP_KERNEL);
3259         if (!sdkp)
3260                 goto out;
3261
3262         gd = alloc_disk(SD_MINORS);
3263         if (!gd)
3264                 goto out_free;
3265
3266         do {
3267                 if (!ida_pre_get(&sd_index_ida, GFP_KERNEL))
3268                         goto out_put;
3269
3270                 spin_lock(&sd_index_lock);
3271                 error = ida_get_new(&sd_index_ida, &index);
3272                 spin_unlock(&sd_index_lock);
3273         } while (error == -EAGAIN);
3274
3275         if (error) {
3276                 sdev_printk(KERN_WARNING, sdp, "sd_probe: memory exhausted.\n");
3277                 goto out_put;
3278         }
3279
3280         error = sd_format_disk_name("sd", index, gd->disk_name, DISK_NAME_LEN);
3281         if (error) {
3282                 sdev_printk(KERN_WARNING, sdp, "SCSI disk (sd) name length exceeded.\n");
3283                 goto out_free_index;
3284         }
3285
3286         sdkp->device = sdp;
3287         sdkp->driver = &sd_template;
3288         sdkp->disk = gd;
3289         sdkp->index = index;
3290         atomic_set(&sdkp->openers, 0);
3291         atomic_set(&sdkp->device->ioerr_cnt, 0);
3292
3293         if (!sdp->request_queue->rq_timeout) {
3294                 if (sdp->type != TYPE_MOD)
3295                         blk_queue_rq_timeout(sdp->request_queue, SD_TIMEOUT);
3296                 else
3297                         blk_queue_rq_timeout(sdp->request_queue,
3298                                              SD_MOD_TIMEOUT);
3299         }
3300
3301         device_initialize(&sdkp->dev);
3302         sdkp->dev.parent = dev;
3303         sdkp->dev.class = &sd_disk_class;
3304         dev_set_name(&sdkp->dev, "%s", dev_name(dev));
3305
3306         error = device_add(&sdkp->dev);
3307         if (error)
3308                 goto out_free_index;
3309
3310         get_device(dev);
3311         dev_set_drvdata(dev, sdkp);
3312
3313         get_device(&sdkp->dev); /* prevent release before async_schedule */
3314         async_schedule_domain(sd_probe_async, sdkp, &scsi_sd_probe_domain);
3315
3316         return 0;
3317
3318  out_free_index:
3319         spin_lock(&sd_index_lock);
3320         ida_remove(&sd_index_ida, index);
3321         spin_unlock(&sd_index_lock);
3322  out_put:
3323         put_disk(gd);
3324  out_free:
3325         kfree(sdkp);
3326  out:
3327         scsi_autopm_put_device(sdp);
3328         return error;
3329 }
3330
3331 /**
3332  *      sd_remove - called whenever a scsi disk (previously recognized by
3333  *      sd_probe) is detached from the system. It is called (potentially
3334  *      multiple times) during sd module unload.
3335  *      @dev: pointer to device object
3336  *
3337  *      Note: this function is invoked from the scsi mid-level.
3338  *      This function potentially frees up a device name (e.g. /dev/sdc)
3339  *      that could be re-used by a subsequent sd_probe().
3340  *      This function is not called when the built-in sd driver is "exit-ed".
3341  **/
3342 static int sd_remove(struct device *dev)
3343 {
3344         struct scsi_disk *sdkp;
3345         dev_t devt;
3346
3347         sdkp = dev_get_drvdata(dev);
3348         devt = disk_devt(sdkp->disk);
3349         scsi_autopm_get_device(sdkp->device);
3350
3351         async_synchronize_full_domain(&scsi_sd_pm_domain);
3352         async_synchronize_full_domain(&scsi_sd_probe_domain);
3353         device_del(&sdkp->dev);
3354         del_gendisk(sdkp->disk);
3355         sd_shutdown(dev);
3356
3357         sd_zbc_remove(sdkp);
3358
3359         blk_register_region(devt, SD_MINORS, NULL,
3360                             sd_default_probe, NULL, NULL);
3361
3362         mutex_lock(&sd_ref_mutex);
3363         dev_set_drvdata(dev, NULL);
3364         put_device(&sdkp->dev);
3365         mutex_unlock(&sd_ref_mutex);
3366
3367         return 0;
3368 }
3369
3370 /**
3371  *      scsi_disk_release - Called to free the scsi_disk structure
3372  *      @dev: pointer to embedded class device
3373  *
3374  *      sd_ref_mutex must be held entering this routine.  Because it is
3375  *      called on last put, you should always use the scsi_disk_get()
3376  *      scsi_disk_put() helpers which manipulate the semaphore directly
3377  *      and never do a direct put_device.
3378  **/
3379 static void scsi_disk_release(struct device *dev)
3380 {
3381         struct scsi_disk *sdkp = to_scsi_disk(dev);
3382         struct gendisk *disk = sdkp->disk;
3383         
3384         spin_lock(&sd_index_lock);
3385         ida_remove(&sd_index_ida, sdkp->index);
3386         spin_unlock(&sd_index_lock);
3387
3388         disk->private_data = NULL;
3389         put_disk(disk);
3390         put_device(&sdkp->device->sdev_gendev);
3391
3392         kfree(sdkp);
3393 }
3394
3395 static int sd_start_stop_device(struct scsi_disk *sdkp, int start)
3396 {
3397         unsigned char cmd[6] = { START_STOP };  /* START_VALID */
3398         struct scsi_sense_hdr sshdr;
3399         struct scsi_device *sdp = sdkp->device;
3400         int res;
3401
3402         if (start)
3403                 cmd[4] |= 1;    /* START */
3404
3405         if (sdp->start_stop_pwr_cond)
3406                 cmd[4] |= start ? 1 << 4 : 3 << 4;      /* Active or Standby */
3407
3408         if (!scsi_device_online(sdp))
3409                 return -ENODEV;
3410
3411         res = scsi_execute(sdp, cmd, DMA_NONE, NULL, 0, NULL, &sshdr,
3412                         SD_TIMEOUT, SD_MAX_RETRIES, 0, RQF_PM, NULL);
3413         if (res) {
3414                 sd_print_result(sdkp, "Start/Stop Unit failed", res);
3415                 if (driver_byte(res) & DRIVER_SENSE)
3416                         sd_print_sense_hdr(sdkp, &sshdr);
3417                 if (scsi_sense_valid(&sshdr) &&
3418                         /* 0x3a is medium not present */
3419                         sshdr.asc == 0x3a)
3420                         res = 0;
3421         }
3422
3423         /* SCSI error codes must not go to the generic layer */
3424         if (res)
3425                 return -EIO;
3426
3427         return 0;
3428 }
3429
3430 /*
3431  * Send a SYNCHRONIZE CACHE instruction down to the device through
3432  * the normal SCSI command structure.  Wait for the command to
3433  * complete.
3434  */
3435 static void sd_shutdown(struct device *dev)
3436 {
3437         struct scsi_disk *sdkp = dev_get_drvdata(dev);
3438
3439         if (!sdkp)
3440                 return;         /* this can happen */
3441
3442         if (pm_runtime_suspended(dev))
3443                 return;
3444
3445         if (sdkp->WCE && sdkp->media_present) {
3446                 sd_printk(KERN_NOTICE, sdkp, "Synchronizing SCSI cache\n");
3447                 sd_sync_cache(sdkp);
3448         }
3449
3450         if (system_state != SYSTEM_RESTART && sdkp->device->manage_start_stop) {
3451                 sd_printk(KERN_NOTICE, sdkp, "Stopping disk\n");
3452                 sd_start_stop_device(sdkp, 0);
3453         }
3454 }
3455
3456 static int sd_suspend_common(struct device *dev, bool ignore_stop_errors)
3457 {
3458         struct scsi_disk *sdkp = dev_get_drvdata(dev);
3459         int ret = 0;
3460
3461         if (!sdkp)      /* E.g.: runtime suspend following sd_remove() */
3462                 return 0;
3463
3464         if (sdkp->WCE && sdkp->media_present) {
3465                 sd_printk(KERN_NOTICE, sdkp, "Synchronizing SCSI cache\n");
3466                 ret = sd_sync_cache(sdkp);
3467                 if (ret) {
3468                         /* ignore OFFLINE device */
3469                         if (ret == -ENODEV)
3470                                 ret = 0;
3471                         goto done;
3472                 }
3473         }
3474
3475         if (sdkp->device->manage_start_stop) {
3476                 sd_printk(KERN_NOTICE, sdkp, "Stopping disk\n");
3477                 /* an error is not worth aborting a system sleep */
3478                 ret = sd_start_stop_device(sdkp, 0);
3479                 if (ignore_stop_errors)
3480                         ret = 0;
3481         }
3482
3483 done:
3484         return ret;
3485 }
3486
3487 static int sd_suspend_system(struct device *dev)
3488 {
3489         return sd_suspend_common(dev, true);
3490 }
3491
3492 static int sd_suspend_runtime(struct device *dev)
3493 {
3494         return sd_suspend_common(dev, false);
3495 }
3496
3497 static int sd_resume(struct device *dev)
3498 {
3499         struct scsi_disk *sdkp = dev_get_drvdata(dev);
3500
3501         if (!sdkp)      /* E.g.: runtime resume at the start of sd_probe() */
3502                 return 0;
3503
3504         if (!sdkp->device->manage_start_stop)
3505                 return 0;
3506
3507         sd_printk(KERN_NOTICE, sdkp, "Starting disk\n");
3508         return sd_start_stop_device(sdkp, 1);
3509 }
3510
3511 /**
3512  *      init_sd - entry point for this driver (both when built in or when
3513  *      a module).
3514  *
3515  *      Note: this function registers this driver with the scsi mid-level.
3516  **/
3517 static int __init init_sd(void)
3518 {
3519         int majors = 0, i, err;
3520
3521         SCSI_LOG_HLQUEUE(3, printk("init_sd: sd driver entry point\n"));
3522
3523         for (i = 0; i < SD_MAJORS; i++) {
3524                 if (register_blkdev(sd_major(i), "sd") != 0)
3525                         continue;
3526                 majors++;
3527                 blk_register_region(sd_major(i), SD_MINORS, NULL,
3528                                     sd_default_probe, NULL, NULL);
3529         }
3530
3531         if (!majors)
3532                 return -ENODEV;
3533
3534         err = class_register(&sd_disk_class);
3535         if (err)
3536                 goto err_out;
3537
3538         sd_cdb_cache = kmem_cache_create("sd_ext_cdb", SD_EXT_CDB_SIZE,
3539                                          0, 0, NULL);
3540         if (!sd_cdb_cache) {
3541                 printk(KERN_ERR "sd: can't init extended cdb cache\n");
3542                 err = -ENOMEM;
3543                 goto err_out_class;
3544         }
3545
3546         sd_cdb_pool = mempool_create_slab_pool(SD_MEMPOOL_SIZE, sd_cdb_cache);
3547         if (!sd_cdb_pool) {
3548                 printk(KERN_ERR "sd: can't init extended cdb pool\n");
3549                 err = -ENOMEM;
3550                 goto err_out_cache;
3551         }
3552
3553         err = scsi_register_driver(&sd_template.gendrv);
3554         if (err)
3555                 goto err_out_driver;
3556
3557         return 0;
3558
3559 err_out_driver:
3560         mempool_destroy(sd_cdb_pool);
3561
3562 err_out_cache:
3563         kmem_cache_destroy(sd_cdb_cache);
3564
3565 err_out_class:
3566         class_unregister(&sd_disk_class);
3567 err_out:
3568         for (i = 0; i < SD_MAJORS; i++)
3569                 unregister_blkdev(sd_major(i), "sd");
3570         return err;
3571 }
3572
3573 /**
3574  *      exit_sd - exit point for this driver (when it is a module).
3575  *
3576  *      Note: this function unregisters this driver from the scsi mid-level.
3577  **/
3578 static void __exit exit_sd(void)
3579 {
3580         int i;
3581
3582         SCSI_LOG_HLQUEUE(3, printk("exit_sd: exiting sd driver\n"));
3583
3584         scsi_unregister_driver(&sd_template.gendrv);
3585         mempool_destroy(sd_cdb_pool);
3586         kmem_cache_destroy(sd_cdb_cache);
3587
3588         class_unregister(&sd_disk_class);
3589
3590         for (i = 0; i < SD_MAJORS; i++) {
3591                 blk_unregister_region(sd_major(i), SD_MINORS);
3592                 unregister_blkdev(sd_major(i), "sd");
3593         }
3594 }
3595
3596 module_init(init_sd);
3597 module_exit(exit_sd);
3598
3599 static void sd_print_sense_hdr(struct scsi_disk *sdkp,
3600                                struct scsi_sense_hdr *sshdr)
3601 {
3602         scsi_print_sense_hdr(sdkp->device,
3603                              sdkp->disk ? sdkp->disk->disk_name : NULL, sshdr);
3604 }
3605
3606 static void sd_print_result(const struct scsi_disk *sdkp, const char *msg,
3607                             int result)
3608 {
3609         const char *hb_string = scsi_hostbyte_string(result);
3610         const char *db_string = scsi_driverbyte_string(result);
3611
3612         if (hb_string || db_string)
3613                 sd_printk(KERN_INFO, sdkp,
3614                           "%s: Result: hostbyte=%s driverbyte=%s\n", msg,
3615                           hb_string ? hb_string : "invalid",
3616                           db_string ? db_string : "invalid");
3617         else
3618                 sd_printk(KERN_INFO, sdkp,
3619                           "%s: Result: hostbyte=0x%02x driverbyte=0x%02x\n",
3620                           msg, host_byte(result), driver_byte(result));
3621 }
3622