1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License Version 1.0 (CDDL-1.0).
6 * You can obtain a copy of the license from the top-level file
7 * "OPENSOLARIS.LICENSE" or at <http://opensource.org/licenses/CDDL-1.0>.
8 * You may not use this file except in compliance with the license.
9 *
10 * CDDL HEADER END
11 */
12
13 /*
14 * Copyright (c) 2016, 2017, Intel Corporation.
15 */
16
17 #ifdef HAVE_LIBUDEV
18
19 #include <errno.h>
20 #include <fcntl.h>
21 #include <libnvpair.h>
22 #include <libudev.h>
23 #include <libzfs.h>
24 #include <libzutil.h>
25 #include <pthread.h>
26 #include <stdlib.h>
27 #include <string.h>
28
29 #include <sys/sysevent/eventdefs.h>
30 #include <sys/sysevent/dev.h>
31
32 #include "zed_log.h"
33 #include "zed_disk_event.h"
34 #include "agents/zfs_agents.h"
35
36 /*
37 * Portions of ZED need to see disk events for disks belonging to ZFS pools.
38 * A libudev monitor is established to monitor block device actions and pass
39 * them on to internal ZED logic modules. Initially, zfs_mod.c is the only
40 * consumer and is the Linux equivalent for the illumos syseventd ZFS SLM
41 * module responsible for handling disk events for ZFS.
42 */
43
44 pthread_t g_mon_tid;
45 struct udev *g_udev;
46 struct udev_monitor *g_mon;
47
48
49 #define DEV_BYID_PATH "/dev/disk/by-id/"
50
51 /* 64MB is minimum usable disk for ZFS */
52 #define MINIMUM_SECTORS 131072ULL
53
54
55 /*
56 * Post disk event to SLM module
57 *
58 * occurs in the context of monitor thread
59 */
60 static void
zed_udev_event(const char * class,const char * subclass,nvlist_t * nvl)61 zed_udev_event(const char *class, const char *subclass, nvlist_t *nvl)
62 {
63 const char *strval;
64 uint64_t numval;
65
66 zed_log_msg(LOG_INFO, "zed_disk_event:");
67 zed_log_msg(LOG_INFO, "\tclass: %s", class);
68 zed_log_msg(LOG_INFO, "\tsubclass: %s", subclass);
69 if (nvlist_lookup_string(nvl, DEV_NAME, &strval) == 0)
70 zed_log_msg(LOG_INFO, "\t%s: %s", DEV_NAME, strval);
71 if (nvlist_lookup_string(nvl, DEV_PATH, &strval) == 0)
72 zed_log_msg(LOG_INFO, "\t%s: %s", DEV_PATH, strval);
73 if (nvlist_lookup_string(nvl, DEV_IDENTIFIER, &strval) == 0)
74 zed_log_msg(LOG_INFO, "\t%s: %s", DEV_IDENTIFIER, strval);
75 if (nvlist_lookup_boolean(nvl, DEV_IS_PART) == B_TRUE)
76 zed_log_msg(LOG_INFO, "\t%s: B_TRUE", DEV_IS_PART);
77 if (nvlist_lookup_string(nvl, DEV_PHYS_PATH, &strval) == 0)
78 zed_log_msg(LOG_INFO, "\t%s: %s", DEV_PHYS_PATH, strval);
79 if (nvlist_lookup_uint64(nvl, DEV_SIZE, &numval) == 0)
80 zed_log_msg(LOG_INFO, "\t%s: %llu", DEV_SIZE, numval);
81 if (nvlist_lookup_uint64(nvl, DEV_PARENT_SIZE, &numval) == 0)
82 zed_log_msg(LOG_INFO, "\t%s: %llu", DEV_PARENT_SIZE, numval);
83 if (nvlist_lookup_uint64(nvl, ZFS_EV_POOL_GUID, &numval) == 0)
84 zed_log_msg(LOG_INFO, "\t%s: %llu", ZFS_EV_POOL_GUID, numval);
85 if (nvlist_lookup_uint64(nvl, ZFS_EV_VDEV_GUID, &numval) == 0)
86 zed_log_msg(LOG_INFO, "\t%s: %llu", ZFS_EV_VDEV_GUID, numval);
87
88 (void) zfs_agent_post_event(class, subclass, nvl);
89 }
90
91 /*
92 * dev_event_nvlist: place event schema into an nv pair list
93 *
94 * NAME VALUE (example)
95 * -------------- --------------------------------------------------------
96 * DEV_NAME /dev/sdl
97 * DEV_PATH /devices/pci0000:00/0000:00:03.0/0000:04:00.0/host0/...
98 * DEV_IDENTIFIER ata-Hitachi_HTS725050A9A362_100601PCG420VLJ37DMC
99 * DEV_PHYS_PATH pci-0000:04:00.0-sas-0x4433221101000000-lun-0
100 * DEV_IS_PART ---
101 * DEV_SIZE 500107862016
102 * ZFS_EV_POOL_GUID 17523635698032189180
103 * ZFS_EV_VDEV_GUID 14663607734290803088
104 */
105 static nvlist_t *
dev_event_nvlist(struct udev_device * dev)106 dev_event_nvlist(struct udev_device *dev)
107 {
108 nvlist_t *nvl;
109 char strval[128];
110 const char *value, *path;
111 uint64_t guid;
112
113 if (nvlist_alloc(&nvl, NV_UNIQUE_NAME, 0) != 0)
114 return (NULL);
115
116 if (zfs_device_get_devid(dev, strval, sizeof (strval)) == 0)
117 (void) nvlist_add_string(nvl, DEV_IDENTIFIER, strval);
118 if (zfs_device_get_physical(dev, strval, sizeof (strval)) == 0)
119 (void) nvlist_add_string(nvl, DEV_PHYS_PATH, strval);
120 if ((path = udev_device_get_devnode(dev)) != NULL)
121 (void) nvlist_add_string(nvl, DEV_NAME, path);
122 if ((value = udev_device_get_devpath(dev)) != NULL)
123 (void) nvlist_add_string(nvl, DEV_PATH, value);
124 value = udev_device_get_devtype(dev);
125 if ((value != NULL && strcmp("partition", value) == 0) ||
126 (udev_device_get_property_value(dev, "ID_PART_ENTRY_NUMBER")
127 != NULL)) {
128 (void) nvlist_add_boolean(nvl, DEV_IS_PART);
129 }
130 if ((value = udev_device_get_sysattr_value(dev, "size")) != NULL) {
131 uint64_t numval = DEV_BSIZE;
132
133 numval *= strtoull(value, NULL, 10);
134 (void) nvlist_add_uint64(nvl, DEV_SIZE, numval);
135
136 /*
137 * If the device has a parent, then get the parent block
138 * device's size as well. For example, /dev/sda1's parent
139 * is /dev/sda.
140 */
141 struct udev_device *parent_dev = udev_device_get_parent(dev);
142 if ((value = udev_device_get_sysattr_value(parent_dev, "size"))
143 != NULL) {
144 uint64_t numval = DEV_BSIZE;
145
146 numval *= strtoull(value, NULL, 10);
147 (void) nvlist_add_uint64(nvl, DEV_PARENT_SIZE, numval);
148 }
149 }
150
151 /*
152 * Grab the pool and vdev guids from blkid cache
153 */
154 value = udev_device_get_property_value(dev, "ID_FS_UUID");
155 if (value != NULL && (guid = strtoull(value, NULL, 10)) != 0)
156 (void) nvlist_add_uint64(nvl, ZFS_EV_POOL_GUID, guid);
157
158 value = udev_device_get_property_value(dev, "ID_FS_UUID_SUB");
159 if (value != NULL && (guid = strtoull(value, NULL, 10)) != 0)
160 (void) nvlist_add_uint64(nvl, ZFS_EV_VDEV_GUID, guid);
161
162 /*
163 * Either a vdev guid or a devid must be present for matching
164 */
165 if (!nvlist_exists(nvl, DEV_IDENTIFIER) &&
166 !nvlist_exists(nvl, ZFS_EV_VDEV_GUID)) {
167 nvlist_free(nvl);
168 return (NULL);
169 }
170
171 return (nvl);
172 }
173
174 /*
175 * Listen for block device uevents
176 */
177 static void *
zed_udev_monitor(void * arg)178 zed_udev_monitor(void *arg)
179 {
180 struct udev_monitor *mon = arg;
181 const char *tmp;
182 char *tmp2;
183
184 zed_log_msg(LOG_INFO, "Waiting for new udev disk events...");
185
186 while (1) {
187 struct udev_device *dev;
188 const char *action, *type, *part, *sectors;
189 const char *bus, *uuid, *devpath;
190 const char *class, *subclass;
191 nvlist_t *nvl;
192 boolean_t is_zfs = B_FALSE;
193
194 /* allow a cancellation while blocked (recvmsg) */
195 pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, NULL);
196
197 /* blocks at recvmsg until an event occurs */
198 if ((dev = udev_monitor_receive_device(mon)) == NULL) {
199 zed_log_msg(LOG_WARNING, "zed_udev_monitor: receive "
200 "device error %d", errno);
201 continue;
202 }
203
204 /* allow all steps to complete before a cancellation */
205 pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, NULL);
206
207 /*
208 * Strongly typed device is the preferred filter
209 */
210 type = udev_device_get_property_value(dev, "ID_FS_TYPE");
211 if (type != NULL && type[0] != '\0') {
212 if (strcmp(type, "zfs_member") == 0) {
213 is_zfs = B_TRUE;
214 } else {
215 /* not ours, so skip */
216 zed_log_msg(LOG_INFO, "zed_udev_monitor: skip "
217 "%s (in use by %s)",
218 udev_device_get_devnode(dev), type);
219 udev_device_unref(dev);
220 continue;
221 }
222 }
223
224 /*
225 * if this is a disk and it is partitioned, then the
226 * zfs label will reside in a DEVTYPE=partition and
227 * we can skip passing this event
228 *
229 * Special case: Blank disks are sometimes reported with
230 * an erroneous 'atari' partition, and should not be
231 * excluded from being used as an autoreplace disk:
232 *
233 * https://github.com/openzfs/zfs/issues/13497
234 */
235 type = udev_device_get_property_value(dev, "DEVTYPE");
236 part = udev_device_get_property_value(dev,
237 "ID_PART_TABLE_TYPE");
238 if (type != NULL && type[0] != '\0' &&
239 strcmp(type, "disk") == 0 &&
240 part != NULL && part[0] != '\0') {
241 const char *devname =
242 udev_device_get_property_value(dev, "DEVNAME");
243
244 if (strcmp(part, "atari") == 0) {
245 zed_log_msg(LOG_INFO,
246 "%s: %s is reporting an atari partition, "
247 "but we're going to assume it's a false "
248 "positive and still use it (issue #13497)",
249 __func__, devname);
250 } else {
251 zed_log_msg(LOG_INFO,
252 "%s: skip %s since it has a %s partition "
253 "already", __func__, devname, part);
254 /* skip and wait for partition event */
255 udev_device_unref(dev);
256 continue;
257 }
258 }
259
260 /*
261 * ignore small partitions
262 */
263 sectors = udev_device_get_property_value(dev,
264 "ID_PART_ENTRY_SIZE");
265 if (sectors == NULL)
266 sectors = udev_device_get_sysattr_value(dev, "size");
267 if (sectors != NULL &&
268 strtoull(sectors, NULL, 10) < MINIMUM_SECTORS) {
269 zed_log_msg(LOG_INFO,
270 "%s: %s sectors %s < %llu (minimum)",
271 __func__,
272 udev_device_get_property_value(dev, "DEVNAME"),
273 sectors, MINIMUM_SECTORS);
274 udev_device_unref(dev);
275 continue;
276 }
277
278 /*
279 * If the blkid probe didn't find ZFS, then a persistent
280 * device id string is required in the message schema
281 * for matching with vdevs. Preflight here for expected
282 * udev information.
283 *
284 * Special case:
285 * NVMe devices don't have ID_BUS set (at least on RHEL 7-8),
286 * but they are valid for autoreplace. Add a special case for
287 * them by searching for "/nvme/" in the udev DEVPATH:
288 *
289 * DEVPATH=/devices/pci0000:00/0000:00:1e.0/nvme/nvme2/nvme2n1
290 */
291 bus = udev_device_get_property_value(dev, "ID_BUS");
292 uuid = udev_device_get_property_value(dev, "DM_UUID");
293 devpath = udev_device_get_devpath(dev);
294 if (!is_zfs && (bus == NULL && uuid == NULL &&
295 strstr(devpath, "/nvme/") == NULL)) {
296 zed_log_msg(LOG_INFO, "zed_udev_monitor: %s no devid "
297 "source", udev_device_get_devnode(dev));
298 udev_device_unref(dev);
299 continue;
300 }
301
302 action = udev_device_get_action(dev);
303 if (strcmp(action, "add") == 0) {
304 class = EC_DEV_ADD;
305 subclass = ESC_DISK;
306 } else if (strcmp(action, "remove") == 0) {
307 class = EC_DEV_REMOVE;
308 subclass = ESC_DISK;
309 } else if (strcmp(action, "change") == 0) {
310 class = EC_DEV_STATUS;
311 subclass = ESC_DEV_DLE;
312 } else {
313 zed_log_msg(LOG_WARNING, "zed_udev_monitor: %s unknown",
314 action);
315 udev_device_unref(dev);
316 continue;
317 }
318
319 /*
320 * Special case an EC_DEV_ADD for multipath devices
321 *
322 * When a multipath device is created, udev reports the
323 * following:
324 *
325 * 1. "add" event of the dm device for the multipath device
326 * (like /dev/dm-3).
327 * 2. "change" event to create the actual multipath device
328 * symlink (like /dev/mapper/mpatha). The event also
329 * passes back the relevant DM vars we care about, like
330 * DM_UUID.
331 * 3. Another "change" event identical to #2 (that we ignore).
332 *
333 * To get the behavior we want, we treat the "change" event
334 * in #2 as a "add" event; as if "/dev/mapper/mpatha" was
335 * a new disk being added.
336 */
337 if (strcmp(class, EC_DEV_STATUS) == 0 &&
338 udev_device_get_property_value(dev, "DM_UUID") &&
339 udev_device_get_property_value(dev, "MPATH_SBIN_PATH")) {
340 tmp = udev_device_get_devnode(dev);
341 tmp2 = zfs_get_underlying_path(tmp);
342 if (tmp && tmp2 && (strcmp(tmp, tmp2) != 0)) {
343 /*
344 * We have a real underlying device, which
345 * means that this multipath "change" event is
346 * an "add" event.
347 *
348 * If the multipath device and the underlying
349 * dev are the same name (i.e. /dev/dm-5), then
350 * there is no real underlying disk for this
351 * multipath device, and so this "change" event
352 * really is a multipath removal.
353 */
354 class = EC_DEV_ADD;
355 subclass = ESC_DISK;
356 } else {
357 tmp = udev_device_get_property_value(dev,
358 "DM_NR_VALID_PATHS");
359 /* treat as a multipath remove */
360 if (tmp != NULL && strcmp(tmp, "0") == 0) {
361 class = EC_DEV_REMOVE;
362 subclass = ESC_DISK;
363 }
364 }
365 free(tmp2);
366 }
367
368 /*
369 * Special case an EC_DEV_ADD for scsi_debug devices
370 *
371 * These devices require a udevadm trigger command after
372 * creation in order to register the vdev_id scsidebug alias
373 * rule (adds a persistent path (phys_path) used for fault
374 * management automated tests in the ZFS test suite.
375 *
376 * After udevadm trigger command, event registers as a "change"
377 * event but needs to instead be handled as another "add" event
378 * to allow for disk labeling and partitioning to occur.
379 */
380 if (strcmp(class, EC_DEV_STATUS) == 0 &&
381 udev_device_get_property_value(dev, "ID_VDEV") &&
382 udev_device_get_property_value(dev, "ID_MODEL")) {
383 const char *id_model, *id_model_sd = "scsi_debug";
384
385 id_model = udev_device_get_property_value(dev,
386 "ID_MODEL");
387 if (strcmp(id_model, id_model_sd) == 0) {
388 class = EC_DEV_ADD;
389 subclass = ESC_DISK;
390 }
391 }
392
393 if ((nvl = dev_event_nvlist(dev)) != NULL) {
394 zed_udev_event(class, subclass, nvl);
395 nvlist_free(nvl);
396 }
397
398 udev_device_unref(dev);
399 }
400
401 return (NULL);
402 }
403
404 int
zed_disk_event_init(void)405 zed_disk_event_init(void)
406 {
407 int fd, fflags;
408
409 if ((g_udev = udev_new()) == NULL) {
410 zed_log_msg(LOG_WARNING, "udev_new failed (%d)", errno);
411 return (-1);
412 }
413
414 /* Set up a udev monitor for block devices */
415 g_mon = udev_monitor_new_from_netlink(g_udev, "udev");
416 udev_monitor_filter_add_match_subsystem_devtype(g_mon, "block", "disk");
417 udev_monitor_filter_add_match_subsystem_devtype(g_mon, "block",
418 "partition");
419 udev_monitor_enable_receiving(g_mon);
420
421 /* Make sure monitoring socket is blocking */
422 fd = udev_monitor_get_fd(g_mon);
423 if ((fflags = fcntl(fd, F_GETFL)) & O_NONBLOCK)
424 (void) fcntl(fd, F_SETFL, fflags & ~O_NONBLOCK);
425
426 /* spawn a thread to monitor events */
427 if (pthread_create(&g_mon_tid, NULL, zed_udev_monitor, g_mon) != 0) {
428 udev_monitor_unref(g_mon);
429 udev_unref(g_udev);
430 zed_log_msg(LOG_WARNING, "pthread_create failed");
431 return (-1);
432 }
433
434 pthread_setname_np(g_mon_tid, "udev monitor");
435 zed_log_msg(LOG_INFO, "zed_disk_event_init");
436
437 return (0);
438 }
439
440 void
zed_disk_event_fini(void)441 zed_disk_event_fini(void)
442 {
443 /* cancel monitor thread at recvmsg() */
444 (void) pthread_cancel(g_mon_tid);
445 (void) pthread_join(g_mon_tid, NULL);
446
447 /* cleanup udev resources */
448 udev_monitor_unref(g_mon);
449 udev_unref(g_udev);
450
451 zed_log_msg(LOG_INFO, "zed_disk_event_fini");
452 }
453
454 #else
455
456 #include "zed_disk_event.h"
457
458 int
zed_disk_event_init(void)459 zed_disk_event_init(void)
460 {
461 return (0);
462 }
463
464 void
zed_disk_event_fini(void)465 zed_disk_event_fini(void)
466 {
467 }
468
469 #endif /* HAVE_LIBUDEV */
470