dm switch: factor out switch_region_table_read
[firefly-linux-kernel-4.4.55.git] / drivers / md / dm-switch.c
1 /*
2  * Copyright (C) 2010-2012 by Dell Inc.  All rights reserved.
3  * Copyright (C) 2011-2013 Red Hat, Inc.
4  *
5  * This file is released under the GPL.
6  *
7  * dm-switch is a device-mapper target that maps IO to underlying block
8  * devices efficiently when there are a large number of fixed-sized
9  * address regions but there is no simple pattern to allow for a compact
10  * mapping representation such as dm-stripe.
11  */
12
13 #include <linux/device-mapper.h>
14
15 #include <linux/module.h>
16 #include <linux/init.h>
17 #include <linux/vmalloc.h>
18
19 #define DM_MSG_PREFIX "switch"
20
21 /*
22  * One region_table_slot_t holds <region_entries_per_slot> region table
23  * entries each of which is <region_table_entry_bits> in size.
24  */
25 typedef unsigned long region_table_slot_t;
26
27 /*
28  * A device with the offset to its start sector.
29  */
30 struct switch_path {
31         struct dm_dev *dmdev;
32         sector_t start;
33 };
34
35 /*
36  * Context block for a dm switch device.
37  */
38 struct switch_ctx {
39         struct dm_target *ti;
40
41         unsigned nr_paths;              /* Number of paths in path_list. */
42
43         unsigned region_size;           /* Region size in 512-byte sectors */
44         unsigned long nr_regions;       /* Number of regions making up the device */
45         signed char region_size_bits;   /* log2 of region_size or -1 */
46
47         unsigned char region_table_entry_bits;  /* Number of bits in one region table entry */
48         unsigned char region_entries_per_slot;  /* Number of entries in one region table slot */
49         signed char region_entries_per_slot_bits;       /* log2 of region_entries_per_slot or -1 */
50
51         region_table_slot_t *region_table;      /* Region table */
52
53         /*
54          * Array of dm devices to switch between.
55          */
56         struct switch_path path_list[0];
57 };
58
59 static struct switch_ctx *alloc_switch_ctx(struct dm_target *ti, unsigned nr_paths,
60                                            unsigned region_size)
61 {
62         struct switch_ctx *sctx;
63
64         sctx = kzalloc(sizeof(struct switch_ctx) + nr_paths * sizeof(struct switch_path),
65                        GFP_KERNEL);
66         if (!sctx)
67                 return NULL;
68
69         sctx->ti = ti;
70         sctx->region_size = region_size;
71
72         ti->private = sctx;
73
74         return sctx;
75 }
76
77 static int alloc_region_table(struct dm_target *ti, unsigned nr_paths)
78 {
79         struct switch_ctx *sctx = ti->private;
80         sector_t nr_regions = ti->len;
81         sector_t nr_slots;
82
83         if (!(sctx->region_size & (sctx->region_size - 1)))
84                 sctx->region_size_bits = __ffs(sctx->region_size);
85         else
86                 sctx->region_size_bits = -1;
87
88         sctx->region_table_entry_bits = 1;
89         while (sctx->region_table_entry_bits < sizeof(region_table_slot_t) * 8 &&
90                (region_table_slot_t)1 << sctx->region_table_entry_bits < nr_paths)
91                 sctx->region_table_entry_bits++;
92
93         sctx->region_entries_per_slot = (sizeof(region_table_slot_t) * 8) / sctx->region_table_entry_bits;
94         if (!(sctx->region_entries_per_slot & (sctx->region_entries_per_slot - 1)))
95                 sctx->region_entries_per_slot_bits = __ffs(sctx->region_entries_per_slot);
96         else
97                 sctx->region_entries_per_slot_bits = -1;
98
99         if (sector_div(nr_regions, sctx->region_size))
100                 nr_regions++;
101
102         sctx->nr_regions = nr_regions;
103         if (sctx->nr_regions != nr_regions || sctx->nr_regions >= ULONG_MAX) {
104                 ti->error = "Region table too large";
105                 return -EINVAL;
106         }
107
108         nr_slots = nr_regions;
109         if (sector_div(nr_slots, sctx->region_entries_per_slot))
110                 nr_slots++;
111
112         if (nr_slots > ULONG_MAX / sizeof(region_table_slot_t)) {
113                 ti->error = "Region table too large";
114                 return -EINVAL;
115         }
116
117         sctx->region_table = vmalloc(nr_slots * sizeof(region_table_slot_t));
118         if (!sctx->region_table) {
119                 ti->error = "Cannot allocate region table";
120                 return -ENOMEM;
121         }
122
123         return 0;
124 }
125
126 static void switch_get_position(struct switch_ctx *sctx, unsigned long region_nr,
127                                 unsigned long *region_index, unsigned *bit)
128 {
129         if (sctx->region_entries_per_slot_bits >= 0) {
130                 *region_index = region_nr >> sctx->region_entries_per_slot_bits;
131                 *bit = region_nr & (sctx->region_entries_per_slot - 1);
132         } else {
133                 *region_index = region_nr / sctx->region_entries_per_slot;
134                 *bit = region_nr % sctx->region_entries_per_slot;
135         }
136
137         *bit *= sctx->region_table_entry_bits;
138 }
139
140 static unsigned switch_region_table_read(struct switch_ctx *sctx, unsigned long region_nr)
141 {
142         unsigned long region_index;
143         unsigned bit;
144
145         switch_get_position(sctx, region_nr, &region_index, &bit);
146
147         return (ACCESS_ONCE(sctx->region_table[region_index]) >> bit) &
148                 ((1 << sctx->region_table_entry_bits) - 1);
149 }
150
151 /*
152  * Find which path to use at given offset.
153  */
154 static unsigned switch_get_path_nr(struct switch_ctx *sctx, sector_t offset)
155 {
156         unsigned path_nr;
157         sector_t p;
158
159         p = offset;
160         if (sctx->region_size_bits >= 0)
161                 p >>= sctx->region_size_bits;
162         else
163                 sector_div(p, sctx->region_size);
164
165         path_nr = switch_region_table_read(sctx, p);
166
167         /* This can only happen if the processor uses non-atomic stores. */
168         if (unlikely(path_nr >= sctx->nr_paths))
169                 path_nr = 0;
170
171         return path_nr;
172 }
173
174 static void switch_region_table_write(struct switch_ctx *sctx, unsigned long region_nr,
175                                       unsigned value)
176 {
177         unsigned long region_index;
178         unsigned bit;
179         region_table_slot_t pte;
180
181         switch_get_position(sctx, region_nr, &region_index, &bit);
182
183         pte = sctx->region_table[region_index];
184         pte &= ~((((region_table_slot_t)1 << sctx->region_table_entry_bits) - 1) << bit);
185         pte |= (region_table_slot_t)value << bit;
186         sctx->region_table[region_index] = pte;
187 }
188
189 /*
190  * Fill the region table with an initial round robin pattern.
191  */
192 static void initialise_region_table(struct switch_ctx *sctx)
193 {
194         unsigned path_nr = 0;
195         unsigned long region_nr;
196
197         for (region_nr = 0; region_nr < sctx->nr_regions; region_nr++) {
198                 switch_region_table_write(sctx, region_nr, path_nr);
199                 if (++path_nr >= sctx->nr_paths)
200                         path_nr = 0;
201         }
202 }
203
204 static int parse_path(struct dm_arg_set *as, struct dm_target *ti)
205 {
206         struct switch_ctx *sctx = ti->private;
207         unsigned long long start;
208         int r;
209
210         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
211                           &sctx->path_list[sctx->nr_paths].dmdev);
212         if (r) {
213                 ti->error = "Device lookup failed";
214                 return r;
215         }
216
217         if (kstrtoull(dm_shift_arg(as), 10, &start) || start != (sector_t)start) {
218                 ti->error = "Invalid device starting offset";
219                 dm_put_device(ti, sctx->path_list[sctx->nr_paths].dmdev);
220                 return -EINVAL;
221         }
222
223         sctx->path_list[sctx->nr_paths].start = start;
224
225         sctx->nr_paths++;
226
227         return 0;
228 }
229
230 /*
231  * Destructor: Don't free the dm_target, just the ti->private data (if any).
232  */
233 static void switch_dtr(struct dm_target *ti)
234 {
235         struct switch_ctx *sctx = ti->private;
236
237         while (sctx->nr_paths--)
238                 dm_put_device(ti, sctx->path_list[sctx->nr_paths].dmdev);
239
240         vfree(sctx->region_table);
241         kfree(sctx);
242 }
243
244 /*
245  * Constructor arguments:
246  *   <num_paths> <region_size> <num_optional_args> [<optional_args>...]
247  *   [<dev_path> <offset>]+
248  *
249  * Optional args are to allow for future extension: currently this
250  * parameter must be 0.
251  */
252 static int switch_ctr(struct dm_target *ti, unsigned argc, char **argv)
253 {
254         static struct dm_arg _args[] = {
255                 {1, (KMALLOC_MAX_SIZE - sizeof(struct switch_ctx)) / sizeof(struct switch_path), "Invalid number of paths"},
256                 {1, UINT_MAX, "Invalid region size"},
257                 {0, 0, "Invalid number of optional args"},
258         };
259
260         struct switch_ctx *sctx;
261         struct dm_arg_set as;
262         unsigned nr_paths, region_size, nr_optional_args;
263         int r;
264
265         as.argc = argc;
266         as.argv = argv;
267
268         r = dm_read_arg(_args, &as, &nr_paths, &ti->error);
269         if (r)
270                 return -EINVAL;
271
272         r = dm_read_arg(_args + 1, &as, &region_size, &ti->error);
273         if (r)
274                 return r;
275
276         r = dm_read_arg_group(_args + 2, &as, &nr_optional_args, &ti->error);
277         if (r)
278                 return r;
279         /* parse optional arguments here, if we add any */
280
281         if (as.argc != nr_paths * 2) {
282                 ti->error = "Incorrect number of path arguments";
283                 return -EINVAL;
284         }
285
286         sctx = alloc_switch_ctx(ti, nr_paths, region_size);
287         if (!sctx) {
288                 ti->error = "Cannot allocate redirection context";
289                 return -ENOMEM;
290         }
291
292         r = dm_set_target_max_io_len(ti, region_size);
293         if (r)
294                 goto error;
295
296         while (as.argc) {
297                 r = parse_path(&as, ti);
298                 if (r)
299                         goto error;
300         }
301
302         r = alloc_region_table(ti, nr_paths);
303         if (r)
304                 goto error;
305
306         initialise_region_table(sctx);
307
308         /* For UNMAP, sending the request down any path is sufficient */
309         ti->num_discard_bios = 1;
310
311         return 0;
312
313 error:
314         switch_dtr(ti);
315
316         return r;
317 }
318
319 static int switch_map(struct dm_target *ti, struct bio *bio)
320 {
321         struct switch_ctx *sctx = ti->private;
322         sector_t offset = dm_target_offset(ti, bio->bi_iter.bi_sector);
323         unsigned path_nr = switch_get_path_nr(sctx, offset);
324
325         bio->bi_bdev = sctx->path_list[path_nr].dmdev->bdev;
326         bio->bi_iter.bi_sector = sctx->path_list[path_nr].start + offset;
327
328         return DM_MAPIO_REMAPPED;
329 }
330
331 /*
332  * We need to parse hex numbers in the message as quickly as possible.
333  *
334  * This table-based hex parser improves performance.
335  * It improves a time to load 1000000 entries compared to the condition-based
336  * parser.
337  *              table-based parser      condition-based parser
338  * PA-RISC      0.29s                   0.31s
339  * Opteron      0.0495s                 0.0498s
340  */
341 static const unsigned char hex_table[256] = {
342 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
343 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
344 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
345 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 255, 255, 255, 255, 255, 255,
346 255, 10, 11, 12, 13, 14, 15, 255, 255, 255, 255, 255, 255, 255, 255, 255,
347 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
348 255, 10, 11, 12, 13, 14, 15, 255, 255, 255, 255, 255, 255, 255, 255, 255,
349 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
350 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
351 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
352 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
353 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
354 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
355 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
356 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
357 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255
358 };
359
360 static __always_inline unsigned long parse_hex(const char **string)
361 {
362         unsigned char d;
363         unsigned long r = 0;
364
365         while ((d = hex_table[(unsigned char)**string]) < 16) {
366                 r = (r << 4) | d;
367                 (*string)++;
368         }
369
370         return r;
371 }
372
373 static int process_set_region_mappings(struct switch_ctx *sctx,
374                              unsigned argc, char **argv)
375 {
376         unsigned i;
377         unsigned long region_index = 0;
378
379         for (i = 1; i < argc; i++) {
380                 unsigned long path_nr;
381                 const char *string = argv[i];
382
383                 if (*string == ':')
384                         region_index++;
385                 else {
386                         region_index = parse_hex(&string);
387                         if (unlikely(*string != ':')) {
388                                 DMWARN("invalid set_region_mappings argument: '%s'", argv[i]);
389                                 return -EINVAL;
390                         }
391                 }
392
393                 string++;
394                 if (unlikely(!*string)) {
395                         DMWARN("invalid set_region_mappings argument: '%s'", argv[i]);
396                         return -EINVAL;
397                 }
398
399                 path_nr = parse_hex(&string);
400                 if (unlikely(*string)) {
401                         DMWARN("invalid set_region_mappings argument: '%s'", argv[i]);
402                         return -EINVAL;
403                 }
404                 if (unlikely(region_index >= sctx->nr_regions)) {
405                         DMWARN("invalid set_region_mappings region number: %lu >= %lu", region_index, sctx->nr_regions);
406                         return -EINVAL;
407                 }
408                 if (unlikely(path_nr >= sctx->nr_paths)) {
409                         DMWARN("invalid set_region_mappings device: %lu >= %u", path_nr, sctx->nr_paths);
410                         return -EINVAL;
411                 }
412
413                 switch_region_table_write(sctx, region_index, path_nr);
414         }
415
416         return 0;
417 }
418
419 /*
420  * Messages are processed one-at-a-time.
421  *
422  * Only set_region_mappings is supported.
423  */
424 static int switch_message(struct dm_target *ti, unsigned argc, char **argv)
425 {
426         static DEFINE_MUTEX(message_mutex);
427
428         struct switch_ctx *sctx = ti->private;
429         int r = -EINVAL;
430
431         mutex_lock(&message_mutex);
432
433         if (!strcasecmp(argv[0], "set_region_mappings"))
434                 r = process_set_region_mappings(sctx, argc, argv);
435         else
436                 DMWARN("Unrecognised message received.");
437
438         mutex_unlock(&message_mutex);
439
440         return r;
441 }
442
443 static void switch_status(struct dm_target *ti, status_type_t type,
444                           unsigned status_flags, char *result, unsigned maxlen)
445 {
446         struct switch_ctx *sctx = ti->private;
447         unsigned sz = 0;
448         int path_nr;
449
450         switch (type) {
451         case STATUSTYPE_INFO:
452                 result[0] = '\0';
453                 break;
454
455         case STATUSTYPE_TABLE:
456                 DMEMIT("%u %u 0", sctx->nr_paths, sctx->region_size);
457                 for (path_nr = 0; path_nr < sctx->nr_paths; path_nr++)
458                         DMEMIT(" %s %llu", sctx->path_list[path_nr].dmdev->name,
459                                (unsigned long long)sctx->path_list[path_nr].start);
460                 break;
461         }
462 }
463
464 /*
465  * Switch ioctl:
466  *
467  * Passthrough all ioctls to the path for sector 0
468  */
469 static int switch_ioctl(struct dm_target *ti, unsigned cmd,
470                         unsigned long arg)
471 {
472         struct switch_ctx *sctx = ti->private;
473         struct block_device *bdev;
474         fmode_t mode;
475         unsigned path_nr;
476         int r = 0;
477
478         path_nr = switch_get_path_nr(sctx, 0);
479
480         bdev = sctx->path_list[path_nr].dmdev->bdev;
481         mode = sctx->path_list[path_nr].dmdev->mode;
482
483         /*
484          * Only pass ioctls through if the device sizes match exactly.
485          */
486         if (ti->len + sctx->path_list[path_nr].start != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT)
487                 r = scsi_verify_blk_ioctl(NULL, cmd);
488
489         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
490 }
491
492 static int switch_iterate_devices(struct dm_target *ti,
493                                   iterate_devices_callout_fn fn, void *data)
494 {
495         struct switch_ctx *sctx = ti->private;
496         int path_nr;
497         int r;
498
499         for (path_nr = 0; path_nr < sctx->nr_paths; path_nr++) {
500                 r = fn(ti, sctx->path_list[path_nr].dmdev,
501                          sctx->path_list[path_nr].start, ti->len, data);
502                 if (r)
503                         return r;
504         }
505
506         return 0;
507 }
508
509 static struct target_type switch_target = {
510         .name = "switch",
511         .version = {1, 0, 0},
512         .module = THIS_MODULE,
513         .ctr = switch_ctr,
514         .dtr = switch_dtr,
515         .map = switch_map,
516         .message = switch_message,
517         .status = switch_status,
518         .ioctl = switch_ioctl,
519         .iterate_devices = switch_iterate_devices,
520 };
521
522 static int __init dm_switch_init(void)
523 {
524         int r;
525
526         r = dm_register_target(&switch_target);
527         if (r < 0)
528                 DMERR("dm_register_target() failed %d", r);
529
530         return r;
531 }
532
533 static void __exit dm_switch_exit(void)
534 {
535         dm_unregister_target(&switch_target);
536 }
537
538 module_init(dm_switch_init);
539 module_exit(dm_switch_exit);
540
541 MODULE_DESCRIPTION(DM_NAME " dynamic path switching target");
542 MODULE_AUTHOR("Kevin D. O'Kelley <Kevin_OKelley@dell.com>");
543 MODULE_AUTHOR("Narendran Ganapathy <Narendran_Ganapathy@dell.com>");
544 MODULE_AUTHOR("Jim Ramsay <Jim_Ramsay@dell.com>");
545 MODULE_AUTHOR("Mikulas Patocka <mpatocka@redhat.com>");
546 MODULE_LICENSE("GPL");