Merge tag 'libnvdimm-for-5.5' of git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm...
[sfrench/cifs-2.6.git] / drivers / nvdimm / bus.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright(c) 2013-2015 Intel Corporation. All rights reserved.
4  */
5 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
6 #include <linux/libnvdimm.h>
7 #include <linux/sched/mm.h>
8 #include <linux/vmalloc.h>
9 #include <linux/uaccess.h>
10 #include <linux/module.h>
11 #include <linux/blkdev.h>
12 #include <linux/fcntl.h>
13 #include <linux/async.h>
14 #include <linux/genhd.h>
15 #include <linux/ndctl.h>
16 #include <linux/sched.h>
17 #include <linux/slab.h>
18 #include <linux/cpu.h>
19 #include <linux/fs.h>
20 #include <linux/io.h>
21 #include <linux/mm.h>
22 #include <linux/nd.h>
23 #include "nd-core.h"
24 #include "nd.h"
25 #include "pfn.h"
26
27 int nvdimm_major;
28 static int nvdimm_bus_major;
29 struct class *nd_class;
30 static DEFINE_IDA(nd_ida);
31
32 static int to_nd_device_type(struct device *dev)
33 {
34         if (is_nvdimm(dev))
35                 return ND_DEVICE_DIMM;
36         else if (is_memory(dev))
37                 return ND_DEVICE_REGION_PMEM;
38         else if (is_nd_blk(dev))
39                 return ND_DEVICE_REGION_BLK;
40         else if (is_nd_dax(dev))
41                 return ND_DEVICE_DAX_PMEM;
42         else if (is_nd_region(dev->parent))
43                 return nd_region_to_nstype(to_nd_region(dev->parent));
44
45         return 0;
46 }
47
48 static int nvdimm_bus_uevent(struct device *dev, struct kobj_uevent_env *env)
49 {
50         return add_uevent_var(env, "MODALIAS=" ND_DEVICE_MODALIAS_FMT,
51                         to_nd_device_type(dev));
52 }
53
54 static struct module *to_bus_provider(struct device *dev)
55 {
56         /* pin bus providers while regions are enabled */
57         if (is_nd_region(dev)) {
58                 struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
59
60                 return nvdimm_bus->nd_desc->module;
61         }
62         return NULL;
63 }
64
65 static void nvdimm_bus_probe_start(struct nvdimm_bus *nvdimm_bus)
66 {
67         nvdimm_bus_lock(&nvdimm_bus->dev);
68         nvdimm_bus->probe_active++;
69         nvdimm_bus_unlock(&nvdimm_bus->dev);
70 }
71
72 static void nvdimm_bus_probe_end(struct nvdimm_bus *nvdimm_bus)
73 {
74         nvdimm_bus_lock(&nvdimm_bus->dev);
75         if (--nvdimm_bus->probe_active == 0)
76                 wake_up(&nvdimm_bus->wait);
77         nvdimm_bus_unlock(&nvdimm_bus->dev);
78 }
79
80 static int nvdimm_bus_probe(struct device *dev)
81 {
82         struct nd_device_driver *nd_drv = to_nd_device_driver(dev->driver);
83         struct module *provider = to_bus_provider(dev);
84         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
85         int rc;
86
87         if (!try_module_get(provider))
88                 return -ENXIO;
89
90         dev_dbg(&nvdimm_bus->dev, "START: %s.probe(%s)\n",
91                         dev->driver->name, dev_name(dev));
92
93         nvdimm_bus_probe_start(nvdimm_bus);
94         debug_nvdimm_lock(dev);
95         rc = nd_drv->probe(dev);
96         debug_nvdimm_unlock(dev);
97
98         if ((rc == 0 || rc == -EOPNOTSUPP) &&
99                         dev->parent && is_nd_region(dev->parent))
100                 nd_region_advance_seeds(to_nd_region(dev->parent), dev);
101         nvdimm_bus_probe_end(nvdimm_bus);
102
103         dev_dbg(&nvdimm_bus->dev, "END: %s.probe(%s) = %d\n", dev->driver->name,
104                         dev_name(dev), rc);
105
106         if (rc != 0)
107                 module_put(provider);
108         return rc;
109 }
110
111 static int nvdimm_bus_remove(struct device *dev)
112 {
113         struct nd_device_driver *nd_drv = to_nd_device_driver(dev->driver);
114         struct module *provider = to_bus_provider(dev);
115         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
116         int rc = 0;
117
118         if (nd_drv->remove) {
119                 debug_nvdimm_lock(dev);
120                 rc = nd_drv->remove(dev);
121                 debug_nvdimm_unlock(dev);
122         }
123
124         dev_dbg(&nvdimm_bus->dev, "%s.remove(%s) = %d\n", dev->driver->name,
125                         dev_name(dev), rc);
126         module_put(provider);
127         return rc;
128 }
129
130 static void nvdimm_bus_shutdown(struct device *dev)
131 {
132         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
133         struct nd_device_driver *nd_drv = NULL;
134
135         if (dev->driver)
136                 nd_drv = to_nd_device_driver(dev->driver);
137
138         if (nd_drv && nd_drv->shutdown) {
139                 nd_drv->shutdown(dev);
140                 dev_dbg(&nvdimm_bus->dev, "%s.shutdown(%s)\n",
141                                 dev->driver->name, dev_name(dev));
142         }
143 }
144
145 void nd_device_notify(struct device *dev, enum nvdimm_event event)
146 {
147         nd_device_lock(dev);
148         if (dev->driver) {
149                 struct nd_device_driver *nd_drv;
150
151                 nd_drv = to_nd_device_driver(dev->driver);
152                 if (nd_drv->notify)
153                         nd_drv->notify(dev, event);
154         }
155         nd_device_unlock(dev);
156 }
157 EXPORT_SYMBOL(nd_device_notify);
158
159 void nvdimm_region_notify(struct nd_region *nd_region, enum nvdimm_event event)
160 {
161         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(&nd_region->dev);
162
163         if (!nvdimm_bus)
164                 return;
165
166         /* caller is responsible for holding a reference on the device */
167         nd_device_notify(&nd_region->dev, event);
168 }
169 EXPORT_SYMBOL_GPL(nvdimm_region_notify);
170
171 struct clear_badblocks_context {
172         resource_size_t phys, cleared;
173 };
174
175 static int nvdimm_clear_badblocks_region(struct device *dev, void *data)
176 {
177         struct clear_badblocks_context *ctx = data;
178         struct nd_region *nd_region;
179         resource_size_t ndr_end;
180         sector_t sector;
181
182         /* make sure device is a region */
183         if (!is_memory(dev))
184                 return 0;
185
186         nd_region = to_nd_region(dev);
187         ndr_end = nd_region->ndr_start + nd_region->ndr_size - 1;
188
189         /* make sure we are in the region */
190         if (ctx->phys < nd_region->ndr_start
191                         || (ctx->phys + ctx->cleared) > ndr_end)
192                 return 0;
193
194         sector = (ctx->phys - nd_region->ndr_start) / 512;
195         badblocks_clear(&nd_region->bb, sector, ctx->cleared / 512);
196
197         if (nd_region->bb_state)
198                 sysfs_notify_dirent(nd_region->bb_state);
199
200         return 0;
201 }
202
203 static void nvdimm_clear_badblocks_regions(struct nvdimm_bus *nvdimm_bus,
204                 phys_addr_t phys, u64 cleared)
205 {
206         struct clear_badblocks_context ctx = {
207                 .phys = phys,
208                 .cleared = cleared,
209         };
210
211         device_for_each_child(&nvdimm_bus->dev, &ctx,
212                         nvdimm_clear_badblocks_region);
213 }
214
215 static void nvdimm_account_cleared_poison(struct nvdimm_bus *nvdimm_bus,
216                 phys_addr_t phys, u64 cleared)
217 {
218         if (cleared > 0)
219                 badrange_forget(&nvdimm_bus->badrange, phys, cleared);
220
221         if (cleared > 0 && cleared / 512)
222                 nvdimm_clear_badblocks_regions(nvdimm_bus, phys, cleared);
223 }
224
225 long nvdimm_clear_poison(struct device *dev, phys_addr_t phys,
226                 unsigned int len)
227 {
228         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
229         struct nvdimm_bus_descriptor *nd_desc;
230         struct nd_cmd_clear_error clear_err;
231         struct nd_cmd_ars_cap ars_cap;
232         u32 clear_err_unit, mask;
233         unsigned int noio_flag;
234         int cmd_rc, rc;
235
236         if (!nvdimm_bus)
237                 return -ENXIO;
238
239         nd_desc = nvdimm_bus->nd_desc;
240         /*
241          * if ndctl does not exist, it's PMEM_LEGACY and
242          * we want to just pretend everything is handled.
243          */
244         if (!nd_desc->ndctl)
245                 return len;
246
247         memset(&ars_cap, 0, sizeof(ars_cap));
248         ars_cap.address = phys;
249         ars_cap.length = len;
250         noio_flag = memalloc_noio_save();
251         rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_ARS_CAP, &ars_cap,
252                         sizeof(ars_cap), &cmd_rc);
253         memalloc_noio_restore(noio_flag);
254         if (rc < 0)
255                 return rc;
256         if (cmd_rc < 0)
257                 return cmd_rc;
258         clear_err_unit = ars_cap.clear_err_unit;
259         if (!clear_err_unit || !is_power_of_2(clear_err_unit))
260                 return -ENXIO;
261
262         mask = clear_err_unit - 1;
263         if ((phys | len) & mask)
264                 return -ENXIO;
265         memset(&clear_err, 0, sizeof(clear_err));
266         clear_err.address = phys;
267         clear_err.length = len;
268         noio_flag = memalloc_noio_save();
269         rc = nd_desc->ndctl(nd_desc, NULL, ND_CMD_CLEAR_ERROR, &clear_err,
270                         sizeof(clear_err), &cmd_rc);
271         memalloc_noio_restore(noio_flag);
272         if (rc < 0)
273                 return rc;
274         if (cmd_rc < 0)
275                 return cmd_rc;
276
277         nvdimm_account_cleared_poison(nvdimm_bus, phys, clear_err.cleared);
278
279         return clear_err.cleared;
280 }
281 EXPORT_SYMBOL_GPL(nvdimm_clear_poison);
282
283 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv);
284
285 static struct bus_type nvdimm_bus_type = {
286         .name = "nd",
287         .uevent = nvdimm_bus_uevent,
288         .match = nvdimm_bus_match,
289         .probe = nvdimm_bus_probe,
290         .remove = nvdimm_bus_remove,
291         .shutdown = nvdimm_bus_shutdown,
292 };
293
294 static void nvdimm_bus_release(struct device *dev)
295 {
296         struct nvdimm_bus *nvdimm_bus;
297
298         nvdimm_bus = container_of(dev, struct nvdimm_bus, dev);
299         ida_simple_remove(&nd_ida, nvdimm_bus->id);
300         kfree(nvdimm_bus);
301 }
302
303 static const struct device_type nvdimm_bus_dev_type = {
304         .release = nvdimm_bus_release,
305         .groups = nvdimm_bus_attribute_groups,
306 };
307
308 bool is_nvdimm_bus(struct device *dev)
309 {
310         return dev->type == &nvdimm_bus_dev_type;
311 }
312
313 struct nvdimm_bus *walk_to_nvdimm_bus(struct device *nd_dev)
314 {
315         struct device *dev;
316
317         for (dev = nd_dev; dev; dev = dev->parent)
318                 if (is_nvdimm_bus(dev))
319                         break;
320         dev_WARN_ONCE(nd_dev, !dev, "invalid dev, not on nd bus\n");
321         if (dev)
322                 return to_nvdimm_bus(dev);
323         return NULL;
324 }
325
326 struct nvdimm_bus *to_nvdimm_bus(struct device *dev)
327 {
328         struct nvdimm_bus *nvdimm_bus;
329
330         nvdimm_bus = container_of(dev, struct nvdimm_bus, dev);
331         WARN_ON(!is_nvdimm_bus(dev));
332         return nvdimm_bus;
333 }
334 EXPORT_SYMBOL_GPL(to_nvdimm_bus);
335
336 struct nvdimm_bus *nvdimm_to_bus(struct nvdimm *nvdimm)
337 {
338         return to_nvdimm_bus(nvdimm->dev.parent);
339 }
340 EXPORT_SYMBOL_GPL(nvdimm_to_bus);
341
342 struct nvdimm_bus *nvdimm_bus_register(struct device *parent,
343                 struct nvdimm_bus_descriptor *nd_desc)
344 {
345         struct nvdimm_bus *nvdimm_bus;
346         int rc;
347
348         nvdimm_bus = kzalloc(sizeof(*nvdimm_bus), GFP_KERNEL);
349         if (!nvdimm_bus)
350                 return NULL;
351         INIT_LIST_HEAD(&nvdimm_bus->list);
352         INIT_LIST_HEAD(&nvdimm_bus->mapping_list);
353         init_waitqueue_head(&nvdimm_bus->wait);
354         nvdimm_bus->id = ida_simple_get(&nd_ida, 0, 0, GFP_KERNEL);
355         if (nvdimm_bus->id < 0) {
356                 kfree(nvdimm_bus);
357                 return NULL;
358         }
359         mutex_init(&nvdimm_bus->reconfig_mutex);
360         badrange_init(&nvdimm_bus->badrange);
361         nvdimm_bus->nd_desc = nd_desc;
362         nvdimm_bus->dev.parent = parent;
363         nvdimm_bus->dev.type = &nvdimm_bus_dev_type;
364         nvdimm_bus->dev.groups = nd_desc->attr_groups;
365         nvdimm_bus->dev.bus = &nvdimm_bus_type;
366         nvdimm_bus->dev.of_node = nd_desc->of_node;
367         dev_set_name(&nvdimm_bus->dev, "ndbus%d", nvdimm_bus->id);
368         rc = device_register(&nvdimm_bus->dev);
369         if (rc) {
370                 dev_dbg(&nvdimm_bus->dev, "registration failed: %d\n", rc);
371                 goto err;
372         }
373
374         return nvdimm_bus;
375  err:
376         put_device(&nvdimm_bus->dev);
377         return NULL;
378 }
379 EXPORT_SYMBOL_GPL(nvdimm_bus_register);
380
381 void nvdimm_bus_unregister(struct nvdimm_bus *nvdimm_bus)
382 {
383         if (!nvdimm_bus)
384                 return;
385         device_unregister(&nvdimm_bus->dev);
386 }
387 EXPORT_SYMBOL_GPL(nvdimm_bus_unregister);
388
389 static int child_unregister(struct device *dev, void *data)
390 {
391         /*
392          * the singular ndctl class device per bus needs to be
393          * "device_destroy"ed, so skip it here
394          *
395          * i.e. remove classless children
396          */
397         if (dev->class)
398                 return 0;
399
400         if (is_nvdimm(dev)) {
401                 struct nvdimm *nvdimm = to_nvdimm(dev);
402                 bool dev_put = false;
403
404                 /* We are shutting down. Make state frozen artificially. */
405                 nvdimm_bus_lock(dev);
406                 set_bit(NVDIMM_SECURITY_FROZEN, &nvdimm->sec.flags);
407                 if (test_and_clear_bit(NDD_WORK_PENDING, &nvdimm->flags))
408                         dev_put = true;
409                 nvdimm_bus_unlock(dev);
410                 cancel_delayed_work_sync(&nvdimm->dwork);
411                 if (dev_put)
412                         put_device(dev);
413         }
414         nd_device_unregister(dev, ND_SYNC);
415
416         return 0;
417 }
418
419 static void free_badrange_list(struct list_head *badrange_list)
420 {
421         struct badrange_entry *bre, *next;
422
423         list_for_each_entry_safe(bre, next, badrange_list, list) {
424                 list_del(&bre->list);
425                 kfree(bre);
426         }
427         list_del_init(badrange_list);
428 }
429
430 static int nd_bus_remove(struct device *dev)
431 {
432         struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev);
433
434         mutex_lock(&nvdimm_bus_list_mutex);
435         list_del_init(&nvdimm_bus->list);
436         mutex_unlock(&nvdimm_bus_list_mutex);
437
438         wait_event(nvdimm_bus->wait,
439                         atomic_read(&nvdimm_bus->ioctl_active) == 0);
440
441         nd_synchronize();
442         device_for_each_child(&nvdimm_bus->dev, NULL, child_unregister);
443
444         spin_lock(&nvdimm_bus->badrange.lock);
445         free_badrange_list(&nvdimm_bus->badrange.list);
446         spin_unlock(&nvdimm_bus->badrange.lock);
447
448         nvdimm_bus_destroy_ndctl(nvdimm_bus);
449
450         return 0;
451 }
452
453 static int nd_bus_probe(struct device *dev)
454 {
455         struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev);
456         int rc;
457
458         rc = nvdimm_bus_create_ndctl(nvdimm_bus);
459         if (rc)
460                 return rc;
461
462         mutex_lock(&nvdimm_bus_list_mutex);
463         list_add_tail(&nvdimm_bus->list, &nvdimm_bus_list);
464         mutex_unlock(&nvdimm_bus_list_mutex);
465
466         /* enable bus provider attributes to look up their local context */
467         dev_set_drvdata(dev, nvdimm_bus->nd_desc);
468
469         return 0;
470 }
471
472 static struct nd_device_driver nd_bus_driver = {
473         .probe = nd_bus_probe,
474         .remove = nd_bus_remove,
475         .drv = {
476                 .name = "nd_bus",
477                 .suppress_bind_attrs = true,
478                 .bus = &nvdimm_bus_type,
479                 .owner = THIS_MODULE,
480                 .mod_name = KBUILD_MODNAME,
481         },
482 };
483
484 static int nvdimm_bus_match(struct device *dev, struct device_driver *drv)
485 {
486         struct nd_device_driver *nd_drv = to_nd_device_driver(drv);
487
488         if (is_nvdimm_bus(dev) && nd_drv == &nd_bus_driver)
489                 return true;
490
491         return !!test_bit(to_nd_device_type(dev), &nd_drv->type);
492 }
493
494 static ASYNC_DOMAIN_EXCLUSIVE(nd_async_domain);
495
496 void nd_synchronize(void)
497 {
498         async_synchronize_full_domain(&nd_async_domain);
499 }
500 EXPORT_SYMBOL_GPL(nd_synchronize);
501
502 static void nd_async_device_register(void *d, async_cookie_t cookie)
503 {
504         struct device *dev = d;
505
506         if (device_add(dev) != 0) {
507                 dev_err(dev, "%s: failed\n", __func__);
508                 put_device(dev);
509         }
510         put_device(dev);
511         if (dev->parent)
512                 put_device(dev->parent);
513 }
514
515 static void nd_async_device_unregister(void *d, async_cookie_t cookie)
516 {
517         struct device *dev = d;
518
519         /* flush bus operations before delete */
520         nvdimm_bus_lock(dev);
521         nvdimm_bus_unlock(dev);
522
523         device_unregister(dev);
524         put_device(dev);
525 }
526
527 void __nd_device_register(struct device *dev)
528 {
529         if (!dev)
530                 return;
531
532         /*
533          * Ensure that region devices always have their NUMA node set as
534          * early as possible. This way we are able to make certain that
535          * any memory associated with the creation and the creation
536          * itself of the region is associated with the correct node.
537          */
538         if (is_nd_region(dev))
539                 set_dev_node(dev, to_nd_region(dev)->numa_node);
540
541         dev->bus = &nvdimm_bus_type;
542         if (dev->parent) {
543                 get_device(dev->parent);
544                 if (dev_to_node(dev) == NUMA_NO_NODE)
545                         set_dev_node(dev, dev_to_node(dev->parent));
546         }
547         get_device(dev);
548
549         async_schedule_dev_domain(nd_async_device_register, dev,
550                                   &nd_async_domain);
551 }
552
553 void nd_device_register(struct device *dev)
554 {
555         device_initialize(dev);
556         __nd_device_register(dev);
557 }
558 EXPORT_SYMBOL(nd_device_register);
559
560 void nd_device_unregister(struct device *dev, enum nd_async_mode mode)
561 {
562         bool killed;
563
564         switch (mode) {
565         case ND_ASYNC:
566                 /*
567                  * In the async case this is being triggered with the
568                  * device lock held and the unregistration work needs to
569                  * be moved out of line iff this is thread has won the
570                  * race to schedule the deletion.
571                  */
572                 if (!kill_device(dev))
573                         return;
574
575                 get_device(dev);
576                 async_schedule_domain(nd_async_device_unregister, dev,
577                                 &nd_async_domain);
578                 break;
579         case ND_SYNC:
580                 /*
581                  * In the sync case the device is being unregistered due
582                  * to a state change of the parent. Claim the kill state
583                  * to synchronize against other unregistration requests,
584                  * or otherwise let the async path handle it if the
585                  * unregistration was already queued.
586                  */
587                 nd_device_lock(dev);
588                 killed = kill_device(dev);
589                 nd_device_unlock(dev);
590
591                 if (!killed)
592                         return;
593
594                 nd_synchronize();
595                 device_unregister(dev);
596                 break;
597         }
598 }
599 EXPORT_SYMBOL(nd_device_unregister);
600
601 /**
602  * __nd_driver_register() - register a region or a namespace driver
603  * @nd_drv: driver to register
604  * @owner: automatically set by nd_driver_register() macro
605  * @mod_name: automatically set by nd_driver_register() macro
606  */
607 int __nd_driver_register(struct nd_device_driver *nd_drv, struct module *owner,
608                 const char *mod_name)
609 {
610         struct device_driver *drv = &nd_drv->drv;
611
612         if (!nd_drv->type) {
613                 pr_debug("driver type bitmask not set (%ps)\n",
614                                 __builtin_return_address(0));
615                 return -EINVAL;
616         }
617
618         if (!nd_drv->probe) {
619                 pr_debug("%s ->probe() must be specified\n", mod_name);
620                 return -EINVAL;
621         }
622
623         drv->bus = &nvdimm_bus_type;
624         drv->owner = owner;
625         drv->mod_name = mod_name;
626
627         return driver_register(drv);
628 }
629 EXPORT_SYMBOL(__nd_driver_register);
630
631 int nvdimm_revalidate_disk(struct gendisk *disk)
632 {
633         struct device *dev = disk_to_dev(disk)->parent;
634         struct nd_region *nd_region = to_nd_region(dev->parent);
635         int disk_ro = get_disk_ro(disk);
636
637         /*
638          * Upgrade to read-only if the region is read-only preserve as
639          * read-only if the disk is already read-only.
640          */
641         if (disk_ro || nd_region->ro == disk_ro)
642                 return 0;
643
644         dev_info(dev, "%s read-only, marking %s read-only\n",
645                         dev_name(&nd_region->dev), disk->disk_name);
646         set_disk_ro(disk, 1);
647
648         return 0;
649
650 }
651 EXPORT_SYMBOL(nvdimm_revalidate_disk);
652
653 static ssize_t modalias_show(struct device *dev, struct device_attribute *attr,
654                 char *buf)
655 {
656         return sprintf(buf, ND_DEVICE_MODALIAS_FMT "\n",
657                         to_nd_device_type(dev));
658 }
659 static DEVICE_ATTR_RO(modalias);
660
661 static ssize_t devtype_show(struct device *dev, struct device_attribute *attr,
662                 char *buf)
663 {
664         return sprintf(buf, "%s\n", dev->type->name);
665 }
666 static DEVICE_ATTR_RO(devtype);
667
668 static struct attribute *nd_device_attributes[] = {
669         &dev_attr_modalias.attr,
670         &dev_attr_devtype.attr,
671         NULL,
672 };
673
674 /*
675  * nd_device_attribute_group - generic attributes for all devices on an nd bus
676  */
677 const struct attribute_group nd_device_attribute_group = {
678         .attrs = nd_device_attributes,
679 };
680
681 static ssize_t numa_node_show(struct device *dev,
682                 struct device_attribute *attr, char *buf)
683 {
684         return sprintf(buf, "%d\n", dev_to_node(dev));
685 }
686 static DEVICE_ATTR_RO(numa_node);
687
688 static int nvdimm_dev_to_target_node(struct device *dev)
689 {
690         struct device *parent = dev->parent;
691         struct nd_region *nd_region = NULL;
692
693         if (is_nd_region(dev))
694                 nd_region = to_nd_region(dev);
695         else if (parent && is_nd_region(parent))
696                 nd_region = to_nd_region(parent);
697
698         if (!nd_region)
699                 return NUMA_NO_NODE;
700         return nd_region->target_node;
701 }
702
703 static ssize_t target_node_show(struct device *dev,
704                 struct device_attribute *attr, char *buf)
705 {
706         return sprintf(buf, "%d\n", nvdimm_dev_to_target_node(dev));
707 }
708 static DEVICE_ATTR_RO(target_node);
709
710 static struct attribute *nd_numa_attributes[] = {
711         &dev_attr_numa_node.attr,
712         &dev_attr_target_node.attr,
713         NULL,
714 };
715
716 static umode_t nd_numa_attr_visible(struct kobject *kobj, struct attribute *a,
717                 int n)
718 {
719         struct device *dev = container_of(kobj, typeof(*dev), kobj);
720
721         if (!IS_ENABLED(CONFIG_NUMA))
722                 return 0;
723
724         if (a == &dev_attr_target_node.attr &&
725                         nvdimm_dev_to_target_node(dev) == NUMA_NO_NODE)
726                 return 0;
727
728         return a->mode;
729 }
730
731 /*
732  * nd_numa_attribute_group - NUMA attributes for all devices on an nd bus
733  */
734 const struct attribute_group nd_numa_attribute_group = {
735         .attrs = nd_numa_attributes,
736         .is_visible = nd_numa_attr_visible,
737 };
738
739 int nvdimm_bus_create_ndctl(struct nvdimm_bus *nvdimm_bus)
740 {
741         dev_t devt = MKDEV(nvdimm_bus_major, nvdimm_bus->id);
742         struct device *dev;
743
744         dev = device_create(nd_class, &nvdimm_bus->dev, devt, nvdimm_bus,
745                         "ndctl%d", nvdimm_bus->id);
746
747         if (IS_ERR(dev))
748                 dev_dbg(&nvdimm_bus->dev, "failed to register ndctl%d: %ld\n",
749                                 nvdimm_bus->id, PTR_ERR(dev));
750         return PTR_ERR_OR_ZERO(dev);
751 }
752
753 void nvdimm_bus_destroy_ndctl(struct nvdimm_bus *nvdimm_bus)
754 {
755         device_destroy(nd_class, MKDEV(nvdimm_bus_major, nvdimm_bus->id));
756 }
757
758 static const struct nd_cmd_desc __nd_cmd_dimm_descs[] = {
759         [ND_CMD_IMPLEMENTED] = { },
760         [ND_CMD_SMART] = {
761                 .out_num = 2,
762                 .out_sizes = { 4, 128, },
763         },
764         [ND_CMD_SMART_THRESHOLD] = {
765                 .out_num = 2,
766                 .out_sizes = { 4, 8, },
767         },
768         [ND_CMD_DIMM_FLAGS] = {
769                 .out_num = 2,
770                 .out_sizes = { 4, 4 },
771         },
772         [ND_CMD_GET_CONFIG_SIZE] = {
773                 .out_num = 3,
774                 .out_sizes = { 4, 4, 4, },
775         },
776         [ND_CMD_GET_CONFIG_DATA] = {
777                 .in_num = 2,
778                 .in_sizes = { 4, 4, },
779                 .out_num = 2,
780                 .out_sizes = { 4, UINT_MAX, },
781         },
782         [ND_CMD_SET_CONFIG_DATA] = {
783                 .in_num = 3,
784                 .in_sizes = { 4, 4, UINT_MAX, },
785                 .out_num = 1,
786                 .out_sizes = { 4, },
787         },
788         [ND_CMD_VENDOR] = {
789                 .in_num = 3,
790                 .in_sizes = { 4, 4, UINT_MAX, },
791                 .out_num = 3,
792                 .out_sizes = { 4, 4, UINT_MAX, },
793         },
794         [ND_CMD_CALL] = {
795                 .in_num = 2,
796                 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, },
797                 .out_num = 1,
798                 .out_sizes = { UINT_MAX, },
799         },
800 };
801
802 const struct nd_cmd_desc *nd_cmd_dimm_desc(int cmd)
803 {
804         if (cmd < ARRAY_SIZE(__nd_cmd_dimm_descs))
805                 return &__nd_cmd_dimm_descs[cmd];
806         return NULL;
807 }
808 EXPORT_SYMBOL_GPL(nd_cmd_dimm_desc);
809
810 static const struct nd_cmd_desc __nd_cmd_bus_descs[] = {
811         [ND_CMD_IMPLEMENTED] = { },
812         [ND_CMD_ARS_CAP] = {
813                 .in_num = 2,
814                 .in_sizes = { 8, 8, },
815                 .out_num = 4,
816                 .out_sizes = { 4, 4, 4, 4, },
817         },
818         [ND_CMD_ARS_START] = {
819                 .in_num = 5,
820                 .in_sizes = { 8, 8, 2, 1, 5, },
821                 .out_num = 2,
822                 .out_sizes = { 4, 4, },
823         },
824         [ND_CMD_ARS_STATUS] = {
825                 .out_num = 3,
826                 .out_sizes = { 4, 4, UINT_MAX, },
827         },
828         [ND_CMD_CLEAR_ERROR] = {
829                 .in_num = 2,
830                 .in_sizes = { 8, 8, },
831                 .out_num = 3,
832                 .out_sizes = { 4, 4, 8, },
833         },
834         [ND_CMD_CALL] = {
835                 .in_num = 2,
836                 .in_sizes = { sizeof(struct nd_cmd_pkg), UINT_MAX, },
837                 .out_num = 1,
838                 .out_sizes = { UINT_MAX, },
839         },
840 };
841
842 const struct nd_cmd_desc *nd_cmd_bus_desc(int cmd)
843 {
844         if (cmd < ARRAY_SIZE(__nd_cmd_bus_descs))
845                 return &__nd_cmd_bus_descs[cmd];
846         return NULL;
847 }
848 EXPORT_SYMBOL_GPL(nd_cmd_bus_desc);
849
850 u32 nd_cmd_in_size(struct nvdimm *nvdimm, int cmd,
851                 const struct nd_cmd_desc *desc, int idx, void *buf)
852 {
853         if (idx >= desc->in_num)
854                 return UINT_MAX;
855
856         if (desc->in_sizes[idx] < UINT_MAX)
857                 return desc->in_sizes[idx];
858
859         if (nvdimm && cmd == ND_CMD_SET_CONFIG_DATA && idx == 2) {
860                 struct nd_cmd_set_config_hdr *hdr = buf;
861
862                 return hdr->in_length;
863         } else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2) {
864                 struct nd_cmd_vendor_hdr *hdr = buf;
865
866                 return hdr->in_length;
867         } else if (cmd == ND_CMD_CALL) {
868                 struct nd_cmd_pkg *pkg = buf;
869
870                 return pkg->nd_size_in;
871         }
872
873         return UINT_MAX;
874 }
875 EXPORT_SYMBOL_GPL(nd_cmd_in_size);
876
877 u32 nd_cmd_out_size(struct nvdimm *nvdimm, int cmd,
878                 const struct nd_cmd_desc *desc, int idx, const u32 *in_field,
879                 const u32 *out_field, unsigned long remainder)
880 {
881         if (idx >= desc->out_num)
882                 return UINT_MAX;
883
884         if (desc->out_sizes[idx] < UINT_MAX)
885                 return desc->out_sizes[idx];
886
887         if (nvdimm && cmd == ND_CMD_GET_CONFIG_DATA && idx == 1)
888                 return in_field[1];
889         else if (nvdimm && cmd == ND_CMD_VENDOR && idx == 2)
890                 return out_field[1];
891         else if (!nvdimm && cmd == ND_CMD_ARS_STATUS && idx == 2) {
892                 /*
893                  * Per table 9-276 ARS Data in ACPI 6.1, out_field[1] is
894                  * "Size of Output Buffer in bytes, including this
895                  * field."
896                  */
897                 if (out_field[1] < 4)
898                         return 0;
899                 /*
900                  * ACPI 6.1 is ambiguous if 'status' is included in the
901                  * output size. If we encounter an output size that
902                  * overshoots the remainder by 4 bytes, assume it was
903                  * including 'status'.
904                  */
905                 if (out_field[1] - 4 == remainder)
906                         return remainder;
907                 return out_field[1] - 8;
908         } else if (cmd == ND_CMD_CALL) {
909                 struct nd_cmd_pkg *pkg = (struct nd_cmd_pkg *) in_field;
910
911                 return pkg->nd_size_out;
912         }
913
914
915         return UINT_MAX;
916 }
917 EXPORT_SYMBOL_GPL(nd_cmd_out_size);
918
919 void wait_nvdimm_bus_probe_idle(struct device *dev)
920 {
921         struct nvdimm_bus *nvdimm_bus = walk_to_nvdimm_bus(dev);
922
923         do {
924                 if (nvdimm_bus->probe_active == 0)
925                         break;
926                 nvdimm_bus_unlock(dev);
927                 nd_device_unlock(dev);
928                 wait_event(nvdimm_bus->wait,
929                                 nvdimm_bus->probe_active == 0);
930                 nd_device_lock(dev);
931                 nvdimm_bus_lock(dev);
932         } while (true);
933 }
934
935 static int nd_pmem_forget_poison_check(struct device *dev, void *data)
936 {
937         struct nd_cmd_clear_error *clear_err =
938                 (struct nd_cmd_clear_error *)data;
939         struct nd_btt *nd_btt = is_nd_btt(dev) ? to_nd_btt(dev) : NULL;
940         struct nd_pfn *nd_pfn = is_nd_pfn(dev) ? to_nd_pfn(dev) : NULL;
941         struct nd_dax *nd_dax = is_nd_dax(dev) ? to_nd_dax(dev) : NULL;
942         struct nd_namespace_common *ndns = NULL;
943         struct nd_namespace_io *nsio;
944         resource_size_t offset = 0, end_trunc = 0, start, end, pstart, pend;
945
946         if (nd_dax || !dev->driver)
947                 return 0;
948
949         start = clear_err->address;
950         end = clear_err->address + clear_err->cleared - 1;
951
952         if (nd_btt || nd_pfn || nd_dax) {
953                 if (nd_btt)
954                         ndns = nd_btt->ndns;
955                 else if (nd_pfn)
956                         ndns = nd_pfn->ndns;
957                 else if (nd_dax)
958                         ndns = nd_dax->nd_pfn.ndns;
959
960                 if (!ndns)
961                         return 0;
962         } else
963                 ndns = to_ndns(dev);
964
965         nsio = to_nd_namespace_io(&ndns->dev);
966         pstart = nsio->res.start + offset;
967         pend = nsio->res.end - end_trunc;
968
969         if ((pstart >= start) && (pend <= end))
970                 return -EBUSY;
971
972         return 0;
973
974 }
975
976 static int nd_ns_forget_poison_check(struct device *dev, void *data)
977 {
978         return device_for_each_child(dev, data, nd_pmem_forget_poison_check);
979 }
980
981 /* set_config requires an idle interleave set */
982 static int nd_cmd_clear_to_send(struct nvdimm_bus *nvdimm_bus,
983                 struct nvdimm *nvdimm, unsigned int cmd, void *data)
984 {
985         struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc;
986
987         /* ask the bus provider if it would like to block this request */
988         if (nd_desc->clear_to_send) {
989                 int rc = nd_desc->clear_to_send(nd_desc, nvdimm, cmd, data);
990
991                 if (rc)
992                         return rc;
993         }
994
995         /* require clear error to go through the pmem driver */
996         if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR)
997                 return device_for_each_child(&nvdimm_bus->dev, data,
998                                 nd_ns_forget_poison_check);
999
1000         if (!nvdimm || cmd != ND_CMD_SET_CONFIG_DATA)
1001                 return 0;
1002
1003         /* prevent label manipulation while the kernel owns label updates */
1004         wait_nvdimm_bus_probe_idle(&nvdimm_bus->dev);
1005         if (atomic_read(&nvdimm->busy))
1006                 return -EBUSY;
1007         return 0;
1008 }
1009
1010 static int __nd_ioctl(struct nvdimm_bus *nvdimm_bus, struct nvdimm *nvdimm,
1011                 int read_only, unsigned int ioctl_cmd, unsigned long arg)
1012 {
1013         struct nvdimm_bus_descriptor *nd_desc = nvdimm_bus->nd_desc;
1014         const struct nd_cmd_desc *desc = NULL;
1015         unsigned int cmd = _IOC_NR(ioctl_cmd);
1016         struct device *dev = &nvdimm_bus->dev;
1017         void __user *p = (void __user *) arg;
1018         char *out_env = NULL, *in_env = NULL;
1019         const char *cmd_name, *dimm_name;
1020         u32 in_len = 0, out_len = 0;
1021         unsigned int func = cmd;
1022         unsigned long cmd_mask;
1023         struct nd_cmd_pkg pkg;
1024         int rc, i, cmd_rc;
1025         void *buf = NULL;
1026         u64 buf_len = 0;
1027
1028         if (nvdimm) {
1029                 desc = nd_cmd_dimm_desc(cmd);
1030                 cmd_name = nvdimm_cmd_name(cmd);
1031                 cmd_mask = nvdimm->cmd_mask;
1032                 dimm_name = dev_name(&nvdimm->dev);
1033         } else {
1034                 desc = nd_cmd_bus_desc(cmd);
1035                 cmd_name = nvdimm_bus_cmd_name(cmd);
1036                 cmd_mask = nd_desc->cmd_mask;
1037                 dimm_name = "bus";
1038         }
1039
1040         if (cmd == ND_CMD_CALL) {
1041                 if (copy_from_user(&pkg, p, sizeof(pkg)))
1042                         return -EFAULT;
1043         }
1044
1045         if (!desc || (desc->out_num + desc->in_num == 0) ||
1046                         !test_bit(cmd, &cmd_mask))
1047                 return -ENOTTY;
1048
1049         /* fail write commands (when read-only) */
1050         if (read_only)
1051                 switch (cmd) {
1052                 case ND_CMD_VENDOR:
1053                 case ND_CMD_SET_CONFIG_DATA:
1054                 case ND_CMD_ARS_START:
1055                 case ND_CMD_CLEAR_ERROR:
1056                 case ND_CMD_CALL:
1057                         dev_dbg(dev, "'%s' command while read-only.\n",
1058                                         nvdimm ? nvdimm_cmd_name(cmd)
1059                                         : nvdimm_bus_cmd_name(cmd));
1060                         return -EPERM;
1061                 default:
1062                         break;
1063                 }
1064
1065         /* process an input envelope */
1066         in_env = kzalloc(ND_CMD_MAX_ENVELOPE, GFP_KERNEL);
1067         if (!in_env)
1068                 return -ENOMEM;
1069         for (i = 0; i < desc->in_num; i++) {
1070                 u32 in_size, copy;
1071
1072                 in_size = nd_cmd_in_size(nvdimm, cmd, desc, i, in_env);
1073                 if (in_size == UINT_MAX) {
1074                         dev_err(dev, "%s:%s unknown input size cmd: %s field: %d\n",
1075                                         __func__, dimm_name, cmd_name, i);
1076                         rc = -ENXIO;
1077                         goto out;
1078                 }
1079                 if (in_len < ND_CMD_MAX_ENVELOPE)
1080                         copy = min_t(u32, ND_CMD_MAX_ENVELOPE - in_len, in_size);
1081                 else
1082                         copy = 0;
1083                 if (copy && copy_from_user(&in_env[in_len], p + in_len, copy)) {
1084                         rc = -EFAULT;
1085                         goto out;
1086                 }
1087                 in_len += in_size;
1088         }
1089
1090         if (cmd == ND_CMD_CALL) {
1091                 func = pkg.nd_command;
1092                 dev_dbg(dev, "%s, idx: %llu, in: %u, out: %u, len %llu\n",
1093                                 dimm_name, pkg.nd_command,
1094                                 in_len, out_len, buf_len);
1095         }
1096
1097         /* process an output envelope */
1098         out_env = kzalloc(ND_CMD_MAX_ENVELOPE, GFP_KERNEL);
1099         if (!out_env) {
1100                 rc = -ENOMEM;
1101                 goto out;
1102         }
1103
1104         for (i = 0; i < desc->out_num; i++) {
1105                 u32 out_size = nd_cmd_out_size(nvdimm, cmd, desc, i,
1106                                 (u32 *) in_env, (u32 *) out_env, 0);
1107                 u32 copy;
1108
1109                 if (out_size == UINT_MAX) {
1110                         dev_dbg(dev, "%s unknown output size cmd: %s field: %d\n",
1111                                         dimm_name, cmd_name, i);
1112                         rc = -EFAULT;
1113                         goto out;
1114                 }
1115                 if (out_len < ND_CMD_MAX_ENVELOPE)
1116                         copy = min_t(u32, ND_CMD_MAX_ENVELOPE - out_len, out_size);
1117                 else
1118                         copy = 0;
1119                 if (copy && copy_from_user(&out_env[out_len],
1120                                         p + in_len + out_len, copy)) {
1121                         rc = -EFAULT;
1122                         goto out;
1123                 }
1124                 out_len += out_size;
1125         }
1126
1127         buf_len = (u64) out_len + (u64) in_len;
1128         if (buf_len > ND_IOCTL_MAX_BUFLEN) {
1129                 dev_dbg(dev, "%s cmd: %s buf_len: %llu > %d\n", dimm_name,
1130                                 cmd_name, buf_len, ND_IOCTL_MAX_BUFLEN);
1131                 rc = -EINVAL;
1132                 goto out;
1133         }
1134
1135         buf = vmalloc(buf_len);
1136         if (!buf) {
1137                 rc = -ENOMEM;
1138                 goto out;
1139         }
1140
1141         if (copy_from_user(buf, p, buf_len)) {
1142                 rc = -EFAULT;
1143                 goto out;
1144         }
1145
1146         nd_device_lock(dev);
1147         nvdimm_bus_lock(dev);
1148         rc = nd_cmd_clear_to_send(nvdimm_bus, nvdimm, func, buf);
1149         if (rc)
1150                 goto out_unlock;
1151
1152         rc = nd_desc->ndctl(nd_desc, nvdimm, cmd, buf, buf_len, &cmd_rc);
1153         if (rc < 0)
1154                 goto out_unlock;
1155
1156         if (!nvdimm && cmd == ND_CMD_CLEAR_ERROR && cmd_rc >= 0) {
1157                 struct nd_cmd_clear_error *clear_err = buf;
1158
1159                 nvdimm_account_cleared_poison(nvdimm_bus, clear_err->address,
1160                                 clear_err->cleared);
1161         }
1162
1163         if (copy_to_user(p, buf, buf_len))
1164                 rc = -EFAULT;
1165
1166 out_unlock:
1167         nvdimm_bus_unlock(dev);
1168         nd_device_unlock(dev);
1169 out:
1170         kfree(in_env);
1171         kfree(out_env);
1172         vfree(buf);
1173         return rc;
1174 }
1175
1176 enum nd_ioctl_mode {
1177         BUS_IOCTL,
1178         DIMM_IOCTL,
1179 };
1180
1181 static int match_dimm(struct device *dev, void *data)
1182 {
1183         long id = (long) data;
1184
1185         if (is_nvdimm(dev)) {
1186                 struct nvdimm *nvdimm = to_nvdimm(dev);
1187
1188                 return nvdimm->id == id;
1189         }
1190
1191         return 0;
1192 }
1193
1194 static long nd_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1195                 enum nd_ioctl_mode mode)
1196
1197 {
1198         struct nvdimm_bus *nvdimm_bus, *found = NULL;
1199         long id = (long) file->private_data;
1200         struct nvdimm *nvdimm = NULL;
1201         int rc, ro;
1202
1203         ro = ((file->f_flags & O_ACCMODE) == O_RDONLY);
1204         mutex_lock(&nvdimm_bus_list_mutex);
1205         list_for_each_entry(nvdimm_bus, &nvdimm_bus_list, list) {
1206                 if (mode == DIMM_IOCTL) {
1207                         struct device *dev;
1208
1209                         dev = device_find_child(&nvdimm_bus->dev,
1210                                         file->private_data, match_dimm);
1211                         if (!dev)
1212                                 continue;
1213                         nvdimm = to_nvdimm(dev);
1214                         found = nvdimm_bus;
1215                 } else if (nvdimm_bus->id == id) {
1216                         found = nvdimm_bus;
1217                 }
1218
1219                 if (found) {
1220                         atomic_inc(&nvdimm_bus->ioctl_active);
1221                         break;
1222                 }
1223         }
1224         mutex_unlock(&nvdimm_bus_list_mutex);
1225
1226         if (!found)
1227                 return -ENXIO;
1228
1229         nvdimm_bus = found;
1230         rc = __nd_ioctl(nvdimm_bus, nvdimm, ro, cmd, arg);
1231
1232         if (nvdimm)
1233                 put_device(&nvdimm->dev);
1234         if (atomic_dec_and_test(&nvdimm_bus->ioctl_active))
1235                 wake_up(&nvdimm_bus->wait);
1236
1237         return rc;
1238 }
1239
1240 static long bus_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1241 {
1242         return nd_ioctl(file, cmd, arg, BUS_IOCTL);
1243 }
1244
1245 static long dimm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1246 {
1247         return nd_ioctl(file, cmd, arg, DIMM_IOCTL);
1248 }
1249
1250 static int nd_open(struct inode *inode, struct file *file)
1251 {
1252         long minor = iminor(inode);
1253
1254         file->private_data = (void *) minor;
1255         return 0;
1256 }
1257
1258 static const struct file_operations nvdimm_bus_fops = {
1259         .owner = THIS_MODULE,
1260         .open = nd_open,
1261         .unlocked_ioctl = bus_ioctl,
1262         .compat_ioctl = compat_ptr_ioctl,
1263         .llseek = noop_llseek,
1264 };
1265
1266 static const struct file_operations nvdimm_fops = {
1267         .owner = THIS_MODULE,
1268         .open = nd_open,
1269         .unlocked_ioctl = dimm_ioctl,
1270         .compat_ioctl = compat_ptr_ioctl,
1271         .llseek = noop_llseek,
1272 };
1273
1274 int __init nvdimm_bus_init(void)
1275 {
1276         int rc;
1277
1278         rc = bus_register(&nvdimm_bus_type);
1279         if (rc)
1280                 return rc;
1281
1282         rc = register_chrdev(0, "ndctl", &nvdimm_bus_fops);
1283         if (rc < 0)
1284                 goto err_bus_chrdev;
1285         nvdimm_bus_major = rc;
1286
1287         rc = register_chrdev(0, "dimmctl", &nvdimm_fops);
1288         if (rc < 0)
1289                 goto err_dimm_chrdev;
1290         nvdimm_major = rc;
1291
1292         nd_class = class_create(THIS_MODULE, "nd");
1293         if (IS_ERR(nd_class)) {
1294                 rc = PTR_ERR(nd_class);
1295                 goto err_class;
1296         }
1297
1298         rc = driver_register(&nd_bus_driver.drv);
1299         if (rc)
1300                 goto err_nd_bus;
1301
1302         return 0;
1303
1304  err_nd_bus:
1305         class_destroy(nd_class);
1306  err_class:
1307         unregister_chrdev(nvdimm_major, "dimmctl");
1308  err_dimm_chrdev:
1309         unregister_chrdev(nvdimm_bus_major, "ndctl");
1310  err_bus_chrdev:
1311         bus_unregister(&nvdimm_bus_type);
1312
1313         return rc;
1314 }
1315
1316 void nvdimm_bus_exit(void)
1317 {
1318         driver_unregister(&nd_bus_driver.drv);
1319         class_destroy(nd_class);
1320         unregister_chrdev(nvdimm_bus_major, "ndctl");
1321         unregister_chrdev(nvdimm_major, "dimmctl");
1322         bus_unregister(&nvdimm_bus_type);
1323         ida_destroy(&nd_ida);
1324 }