xref: /titanic_41/usr/src/uts/common/fs/zfs/vdev_disk.c (revision efe6bf490da92417f844e59983882a3b1c0b4915)
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 (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
24  * Copyright 2013 Nexenta Systems, Inc.  All rights reserved.
25  * Copyright (c) 2013 Joyent, Inc.  All rights reserved.
26  */
27 
28 #include <sys/zfs_context.h>
29 #include <sys/spa_impl.h>
30 #include <sys/refcount.h>
31 #include <sys/vdev_disk.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/fs/zfs.h>
34 #include <sys/zio.h>
35 #include <sys/sunldi.h>
36 #include <sys/efi_partition.h>
37 #include <sys/fm/fs/zfs.h>
38 
39 /*
40  * Virtual device vector for disks.
41  */
42 
43 extern ldi_ident_t zfs_li;
44 
45 static void vdev_disk_close(vdev_t *);
46 
47 typedef struct vdev_disk_ldi_cb {
48 	list_node_t		lcb_next;
49 	ldi_callback_id_t	lcb_id;
50 } vdev_disk_ldi_cb_t;
51 
52 static void
vdev_disk_alloc(vdev_t * vd)53 vdev_disk_alloc(vdev_t *vd)
54 {
55 	vdev_disk_t *dvd;
56 
57 	dvd = vd->vdev_tsd = kmem_zalloc(sizeof (vdev_disk_t), KM_SLEEP);
58 	/*
59 	 * Create the LDI event callback list.
60 	 */
61 	list_create(&dvd->vd_ldi_cbs, sizeof (vdev_disk_ldi_cb_t),
62 	    offsetof(vdev_disk_ldi_cb_t, lcb_next));
63 }
64 
65 static void
vdev_disk_free(vdev_t * vd)66 vdev_disk_free(vdev_t *vd)
67 {
68 	vdev_disk_t *dvd = vd->vdev_tsd;
69 	vdev_disk_ldi_cb_t *lcb;
70 
71 	if (dvd == NULL)
72 		return;
73 
74 	/*
75 	 * We have already closed the LDI handle. Clean up the LDI event
76 	 * callbacks and free vd->vdev_tsd.
77 	 */
78 	while ((lcb = list_head(&dvd->vd_ldi_cbs)) != NULL) {
79 		list_remove(&dvd->vd_ldi_cbs, lcb);
80 		(void) ldi_ev_remove_callbacks(lcb->lcb_id);
81 		kmem_free(lcb, sizeof (vdev_disk_ldi_cb_t));
82 	}
83 	list_destroy(&dvd->vd_ldi_cbs);
84 	kmem_free(dvd, sizeof (vdev_disk_t));
85 	vd->vdev_tsd = NULL;
86 }
87 
88 /* ARGSUSED */
89 static int
vdev_disk_off_notify(ldi_handle_t lh,ldi_ev_cookie_t ecookie,void * arg,void * ev_data)90 vdev_disk_off_notify(ldi_handle_t lh, ldi_ev_cookie_t ecookie, void *arg,
91     void *ev_data)
92 {
93 	vdev_t *vd = (vdev_t *)arg;
94 	vdev_disk_t *dvd = vd->vdev_tsd;
95 
96 	/*
97 	 * Ignore events other than offline.
98 	 */
99 	if (strcmp(ldi_ev_get_type(ecookie), LDI_EV_OFFLINE) != 0)
100 		return (LDI_EV_SUCCESS);
101 
102 	/*
103 	 * All LDI handles must be closed for the state change to succeed, so
104 	 * call on vdev_disk_close() to do this.
105 	 *
106 	 * We inform vdev_disk_close that it is being called from offline
107 	 * notify context so it will defer cleanup of LDI event callbacks and
108 	 * freeing of vd->vdev_tsd to the offline finalize or a reopen.
109 	 */
110 	dvd->vd_ldi_offline = B_TRUE;
111 	vdev_disk_close(vd);
112 
113 	/*
114 	 * Now that the device is closed, request that the spa_async_thread
115 	 * mark the device as REMOVED and notify FMA of the removal.
116 	 */
117 	zfs_post_remove(vd->vdev_spa, vd);
118 	vd->vdev_remove_wanted = B_TRUE;
119 	spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
120 
121 	return (LDI_EV_SUCCESS);
122 }
123 
124 /* ARGSUSED */
125 static void
vdev_disk_off_finalize(ldi_handle_t lh,ldi_ev_cookie_t ecookie,int ldi_result,void * arg,void * ev_data)126 vdev_disk_off_finalize(ldi_handle_t lh, ldi_ev_cookie_t ecookie,
127     int ldi_result, void *arg, void *ev_data)
128 {
129 	vdev_t *vd = (vdev_t *)arg;
130 
131 	/*
132 	 * Ignore events other than offline.
133 	 */
134 	if (strcmp(ldi_ev_get_type(ecookie), LDI_EV_OFFLINE) != 0)
135 		return;
136 
137 	/*
138 	 * We have already closed the LDI handle in notify.
139 	 * Clean up the LDI event callbacks and free vd->vdev_tsd.
140 	 */
141 	vdev_disk_free(vd);
142 
143 	/*
144 	 * Request that the vdev be reopened if the offline state change was
145 	 * unsuccessful.
146 	 */
147 	if (ldi_result != LDI_EV_SUCCESS) {
148 		vd->vdev_probe_wanted = B_TRUE;
149 		spa_async_request(vd->vdev_spa, SPA_ASYNC_PROBE);
150 	}
151 }
152 
153 static ldi_ev_callback_t vdev_disk_off_callb = {
154 	.cb_vers = LDI_EV_CB_VERS,
155 	.cb_notify = vdev_disk_off_notify,
156 	.cb_finalize = vdev_disk_off_finalize
157 };
158 
159 /* ARGSUSED */
160 static void
vdev_disk_dgrd_finalize(ldi_handle_t lh,ldi_ev_cookie_t ecookie,int ldi_result,void * arg,void * ev_data)161 vdev_disk_dgrd_finalize(ldi_handle_t lh, ldi_ev_cookie_t ecookie,
162     int ldi_result, void *arg, void *ev_data)
163 {
164 	vdev_t *vd = (vdev_t *)arg;
165 
166 	/*
167 	 * Ignore events other than degrade.
168 	 */
169 	if (strcmp(ldi_ev_get_type(ecookie), LDI_EV_DEGRADE) != 0)
170 		return;
171 
172 	/*
173 	 * Degrade events always succeed. Mark the vdev as degraded.
174 	 * This status is purely informative for the user.
175 	 */
176 	(void) vdev_degrade(vd->vdev_spa, vd->vdev_guid, 0);
177 }
178 
179 static ldi_ev_callback_t vdev_disk_dgrd_callb = {
180 	.cb_vers = LDI_EV_CB_VERS,
181 	.cb_notify = NULL,
182 	.cb_finalize = vdev_disk_dgrd_finalize
183 };
184 
185 static void
vdev_disk_hold(vdev_t * vd)186 vdev_disk_hold(vdev_t *vd)
187 {
188 	ddi_devid_t devid;
189 	char *minor;
190 
191 	ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
192 
193 	/*
194 	 * We must have a pathname, and it must be absolute.
195 	 */
196 	if (vd->vdev_path == NULL || vd->vdev_path[0] != '/')
197 		return;
198 
199 	/*
200 	 * Only prefetch path and devid info if the device has
201 	 * never been opened.
202 	 */
203 	if (vd->vdev_tsd != NULL)
204 		return;
205 
206 	if (vd->vdev_wholedisk == -1ULL) {
207 		size_t len = strlen(vd->vdev_path) + 3;
208 		char *buf = kmem_alloc(len, KM_SLEEP);
209 
210 		(void) snprintf(buf, len, "%ss0", vd->vdev_path);
211 
212 		(void) ldi_vp_from_name(buf, &vd->vdev_name_vp);
213 		kmem_free(buf, len);
214 	}
215 
216 	if (vd->vdev_name_vp == NULL)
217 		(void) ldi_vp_from_name(vd->vdev_path, &vd->vdev_name_vp);
218 
219 	if (vd->vdev_devid != NULL &&
220 	    ddi_devid_str_decode(vd->vdev_devid, &devid, &minor) == 0) {
221 		(void) ldi_vp_from_devid(devid, minor, &vd->vdev_devid_vp);
222 		ddi_devid_str_free(minor);
223 		ddi_devid_free(devid);
224 	}
225 }
226 
227 static void
vdev_disk_rele(vdev_t * vd)228 vdev_disk_rele(vdev_t *vd)
229 {
230 	ASSERT(spa_config_held(vd->vdev_spa, SCL_STATE, RW_WRITER));
231 
232 	if (vd->vdev_name_vp) {
233 		VN_RELE_ASYNC(vd->vdev_name_vp,
234 		    dsl_pool_vnrele_taskq(vd->vdev_spa->spa_dsl_pool));
235 		vd->vdev_name_vp = NULL;
236 	}
237 	if (vd->vdev_devid_vp) {
238 		VN_RELE_ASYNC(vd->vdev_devid_vp,
239 		    dsl_pool_vnrele_taskq(vd->vdev_spa->spa_dsl_pool));
240 		vd->vdev_devid_vp = NULL;
241 	}
242 }
243 
244 static uint64_t
vdev_disk_get_space(vdev_t * vd,uint64_t capacity,uint_t blksz)245 vdev_disk_get_space(vdev_t *vd, uint64_t capacity, uint_t blksz)
246 {
247 	ASSERT(vd->vdev_wholedisk);
248 
249 	vdev_disk_t *dvd = vd->vdev_tsd;
250 	dk_efi_t dk_ioc;
251 	efi_gpt_t *efi;
252 	uint64_t avail_space = 0;
253 	int efisize = EFI_LABEL_SIZE * 2;
254 
255 	dk_ioc.dki_data = kmem_alloc(efisize, KM_SLEEP);
256 	dk_ioc.dki_lba = 1;
257 	dk_ioc.dki_length = efisize;
258 	dk_ioc.dki_data_64 = (uint64_t)(uintptr_t)dk_ioc.dki_data;
259 	efi = dk_ioc.dki_data;
260 
261 	if (ldi_ioctl(dvd->vd_lh, DKIOCGETEFI, (intptr_t)&dk_ioc,
262 	    FKIOCTL, kcred, NULL) == 0) {
263 		uint64_t efi_altern_lba = LE_64(efi->efi_gpt_AlternateLBA);
264 
265 		if (capacity > efi_altern_lba)
266 			avail_space = (capacity - efi_altern_lba) * blksz;
267 	}
268 	kmem_free(dk_ioc.dki_data, efisize);
269 	return (avail_space);
270 }
271 
272 /*
273  * We want to be loud in DEBUG kernels when DKIOCGMEDIAINFOEXT fails, or when
274  * even a fallback to DKIOCGMEDIAINFO fails.
275  */
276 #ifdef DEBUG
277 #define	VDEV_DEBUG(...)	cmn_err(CE_NOTE, __VA_ARGS__)
278 #else
279 #define	VDEV_DEBUG(...)	/* Nothing... */
280 #endif
281 
282 static int
vdev_disk_open(vdev_t * vd,uint64_t * psize,uint64_t * max_psize,uint64_t * ashift)283 vdev_disk_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
284     uint64_t *ashift)
285 {
286 	spa_t *spa = vd->vdev_spa;
287 	vdev_disk_t *dvd = vd->vdev_tsd;
288 	ldi_ev_cookie_t ecookie;
289 	vdev_disk_ldi_cb_t *lcb;
290 	union {
291 		struct dk_minfo_ext ude;
292 		struct dk_minfo ud;
293 	} dks;
294 	struct dk_minfo_ext *dkmext = &dks.ude;
295 	struct dk_minfo *dkm = &dks.ud;
296 	int error;
297 	dev_t dev;
298 	int otyp;
299 	boolean_t validate_devid = B_FALSE;
300 	ddi_devid_t devid;
301 	uint64_t capacity = 0, blksz = 0, pbsize;
302 
303 	/*
304 	 * We must have a pathname, and it must be absolute.
305 	 */
306 	if (vd->vdev_path == NULL || vd->vdev_path[0] != '/') {
307 		vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
308 		return (SET_ERROR(EINVAL));
309 	}
310 
311 	/*
312 	 * Reopen the device if it's not currently open. Otherwise,
313 	 * just update the physical size of the device.
314 	 */
315 	if (dvd != NULL) {
316 		if (dvd->vd_ldi_offline && dvd->vd_lh == NULL) {
317 			/*
318 			 * If we are opening a device in its offline notify
319 			 * context, the LDI handle was just closed. Clean
320 			 * up the LDI event callbacks and free vd->vdev_tsd.
321 			 */
322 			vdev_disk_free(vd);
323 		} else {
324 			ASSERT(vd->vdev_reopening);
325 			goto skip_open;
326 		}
327 	}
328 
329 	/*
330 	 * Create vd->vdev_tsd.
331 	 */
332 	vdev_disk_alloc(vd);
333 	dvd = vd->vdev_tsd;
334 
335 	/*
336 	 * When opening a disk device, we want to preserve the user's original
337 	 * intent.  We always want to open the device by the path the user gave
338 	 * us, even if it is one of multiple paths to the same device.  But we
339 	 * also want to be able to survive disks being removed/recabled.
340 	 * Therefore the sequence of opening devices is:
341 	 *
342 	 * 1. Try opening the device by path.  For legacy pools without the
343 	 *    'whole_disk' property, attempt to fix the path by appending 's0'.
344 	 *
345 	 * 2. If the devid of the device matches the stored value, return
346 	 *    success.
347 	 *
348 	 * 3. Otherwise, the device may have moved.  Try opening the device
349 	 *    by the devid instead.
350 	 */
351 	if (vd->vdev_devid != NULL) {
352 		if (ddi_devid_str_decode(vd->vdev_devid, &dvd->vd_devid,
353 		    &dvd->vd_minor) != 0) {
354 			vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
355 			return (SET_ERROR(EINVAL));
356 		}
357 	}
358 
359 	error = EINVAL;		/* presume failure */
360 
361 	if (vd->vdev_path != NULL) {
362 
363 		if (vd->vdev_wholedisk == -1ULL) {
364 			size_t len = strlen(vd->vdev_path) + 3;
365 			char *buf = kmem_alloc(len, KM_SLEEP);
366 
367 			(void) snprintf(buf, len, "%ss0", vd->vdev_path);
368 
369 			error = ldi_open_by_name(buf, spa_mode(spa), kcred,
370 			    &dvd->vd_lh, zfs_li);
371 			if (error == 0) {
372 				spa_strfree(vd->vdev_path);
373 				vd->vdev_path = buf;
374 				vd->vdev_wholedisk = 1ULL;
375 			} else {
376 				kmem_free(buf, len);
377 			}
378 		}
379 
380 		/*
381 		 * If we have not yet opened the device, try to open it by the
382 		 * specified path.
383 		 */
384 		if (error != 0) {
385 			error = ldi_open_by_name(vd->vdev_path, spa_mode(spa),
386 			    kcred, &dvd->vd_lh, zfs_li);
387 		}
388 
389 		/*
390 		 * Compare the devid to the stored value.
391 		 */
392 		if (error == 0 && vd->vdev_devid != NULL &&
393 		    ldi_get_devid(dvd->vd_lh, &devid) == 0) {
394 			if (ddi_devid_compare(devid, dvd->vd_devid) != 0) {
395 				error = SET_ERROR(EINVAL);
396 				(void) ldi_close(dvd->vd_lh, spa_mode(spa),
397 				    kcred);
398 				dvd->vd_lh = NULL;
399 			}
400 			ddi_devid_free(devid);
401 		}
402 
403 		/*
404 		 * If we succeeded in opening the device, but 'vdev_wholedisk'
405 		 * is not yet set, then this must be a slice.
406 		 */
407 		if (error == 0 && vd->vdev_wholedisk == -1ULL)
408 			vd->vdev_wholedisk = 0;
409 	}
410 
411 	/*
412 	 * If we were unable to open by path, or the devid check fails, open by
413 	 * devid instead.
414 	 */
415 	if (error != 0 && vd->vdev_devid != NULL) {
416 		error = ldi_open_by_devid(dvd->vd_devid, dvd->vd_minor,
417 		    spa_mode(spa), kcred, &dvd->vd_lh, zfs_li);
418 	}
419 
420 	/*
421 	 * If all else fails, then try opening by physical path (if available)
422 	 * or the logical path (if we failed due to the devid check).  While not
423 	 * as reliable as the devid, this will give us something, and the higher
424 	 * level vdev validation will prevent us from opening the wrong device.
425 	 */
426 	if (error) {
427 		if (vd->vdev_devid != NULL)
428 			validate_devid = B_TRUE;
429 
430 		if (vd->vdev_physpath != NULL &&
431 		    (dev = ddi_pathname_to_dev_t(vd->vdev_physpath)) != NODEV)
432 			error = ldi_open_by_dev(&dev, OTYP_BLK, spa_mode(spa),
433 			    kcred, &dvd->vd_lh, zfs_li);
434 
435 		/*
436 		 * Note that we don't support the legacy auto-wholedisk support
437 		 * as above.  This hasn't been used in a very long time and we
438 		 * don't need to propagate its oddities to this edge condition.
439 		 */
440 		if (error && vd->vdev_path != NULL)
441 			error = ldi_open_by_name(vd->vdev_path, spa_mode(spa),
442 			    kcred, &dvd->vd_lh, zfs_li);
443 	}
444 
445 	if (error) {
446 		vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
447 		return (error);
448 	}
449 
450 	/*
451 	 * Now that the device has been successfully opened, update the devid
452 	 * if necessary.
453 	 */
454 	if (validate_devid && spa_writeable(spa) &&
455 	    ldi_get_devid(dvd->vd_lh, &devid) == 0) {
456 		if (ddi_devid_compare(devid, dvd->vd_devid) != 0) {
457 			char *vd_devid;
458 
459 			vd_devid = ddi_devid_str_encode(devid, dvd->vd_minor);
460 			zfs_dbgmsg("vdev %s: update devid from %s, "
461 			    "to %s", vd->vdev_path, vd->vdev_devid, vd_devid);
462 			spa_strfree(vd->vdev_devid);
463 			vd->vdev_devid = spa_strdup(vd_devid);
464 			ddi_devid_str_free(vd_devid);
465 		}
466 		ddi_devid_free(devid);
467 	}
468 
469 	/*
470 	 * Once a device is opened, verify that the physical device path (if
471 	 * available) is up to date.
472 	 */
473 	if (ldi_get_dev(dvd->vd_lh, &dev) == 0 &&
474 	    ldi_get_otyp(dvd->vd_lh, &otyp) == 0) {
475 		char *physpath, *minorname;
476 
477 		physpath = kmem_alloc(MAXPATHLEN, KM_SLEEP);
478 		minorname = NULL;
479 		if (ddi_dev_pathname(dev, otyp, physpath) == 0 &&
480 		    ldi_get_minor_name(dvd->vd_lh, &minorname) == 0 &&
481 		    (vd->vdev_physpath == NULL ||
482 		    strcmp(vd->vdev_physpath, physpath) != 0)) {
483 			if (vd->vdev_physpath)
484 				spa_strfree(vd->vdev_physpath);
485 			(void) strlcat(physpath, ":", MAXPATHLEN);
486 			(void) strlcat(physpath, minorname, MAXPATHLEN);
487 			vd->vdev_physpath = spa_strdup(physpath);
488 		}
489 		if (minorname)
490 			kmem_free(minorname, strlen(minorname) + 1);
491 		kmem_free(physpath, MAXPATHLEN);
492 	}
493 
494 	/*
495 	 * Register callbacks for the LDI offline event.
496 	 */
497 	if (ldi_ev_get_cookie(dvd->vd_lh, LDI_EV_OFFLINE, &ecookie) ==
498 	    LDI_EV_SUCCESS) {
499 		lcb = kmem_zalloc(sizeof (vdev_disk_ldi_cb_t), KM_SLEEP);
500 		list_insert_tail(&dvd->vd_ldi_cbs, lcb);
501 		(void) ldi_ev_register_callbacks(dvd->vd_lh, ecookie,
502 		    &vdev_disk_off_callb, (void *) vd, &lcb->lcb_id);
503 	}
504 
505 	/*
506 	 * Register callbacks for the LDI degrade event.
507 	 */
508 	if (ldi_ev_get_cookie(dvd->vd_lh, LDI_EV_DEGRADE, &ecookie) ==
509 	    LDI_EV_SUCCESS) {
510 		lcb = kmem_zalloc(sizeof (vdev_disk_ldi_cb_t), KM_SLEEP);
511 		list_insert_tail(&dvd->vd_ldi_cbs, lcb);
512 		(void) ldi_ev_register_callbacks(dvd->vd_lh, ecookie,
513 		    &vdev_disk_dgrd_callb, (void *) vd, &lcb->lcb_id);
514 	}
515 skip_open:
516 	/*
517 	 * Determine the actual size of the device.
518 	 */
519 	if (ldi_get_size(dvd->vd_lh, psize) != 0) {
520 		vd->vdev_stat.vs_aux = VDEV_AUX_OPEN_FAILED;
521 		return (SET_ERROR(EINVAL));
522 	}
523 
524 	*max_psize = *psize;
525 
526 	/*
527 	 * Determine the device's minimum transfer size.
528 	 * If the ioctl isn't supported, assume DEV_BSIZE.
529 	 */
530 	if ((error = ldi_ioctl(dvd->vd_lh, DKIOCGMEDIAINFOEXT,
531 	    (intptr_t)dkmext, FKIOCTL, kcred, NULL)) == 0) {
532 		capacity = dkmext->dki_capacity - 1;
533 		blksz = dkmext->dki_lbsize;
534 		pbsize = dkmext->dki_pbsize;
535 	} else if ((error = ldi_ioctl(dvd->vd_lh, DKIOCGMEDIAINFO,
536 	    (intptr_t)dkm, FKIOCTL, kcred, NULL)) == 0) {
537 		VDEV_DEBUG(
538 		    "vdev_disk_open(\"%s\"): fallback to DKIOCGMEDIAINFO\n",
539 		    vd->vdev_path);
540 		capacity = dkm->dki_capacity - 1;
541 		blksz = dkm->dki_lbsize;
542 		pbsize = blksz;
543 	} else {
544 		VDEV_DEBUG("vdev_disk_open(\"%s\"): "
545 		    "both DKIOCGMEDIAINFO{,EXT} calls failed, %d\n",
546 		    vd->vdev_path, error);
547 		pbsize = DEV_BSIZE;
548 	}
549 
550 	*ashift = highbit64(MAX(pbsize, SPA_MINBLOCKSIZE)) - 1;
551 
552 	if (vd->vdev_wholedisk == 1) {
553 		int wce = 1;
554 
555 		if (error == 0) {
556 			/*
557 			 * If we have the capability to expand, we'd have
558 			 * found out via success from DKIOCGMEDIAINFO{,EXT}.
559 			 * Adjust max_psize upward accordingly since we know
560 			 * we own the whole disk now.
561 			 */
562 			*max_psize += vdev_disk_get_space(vd, capacity, blksz);
563 			zfs_dbgmsg("capacity change: vdev %s, psize %llu, "
564 			    "max_psize %llu", vd->vdev_path, *psize,
565 			    *max_psize);
566 		}
567 
568 		/*
569 		 * Since we own the whole disk, try to enable disk write
570 		 * caching.  We ignore errors because it's OK if we can't do it.
571 		 */
572 		(void) ldi_ioctl(dvd->vd_lh, DKIOCSETWCE, (intptr_t)&wce,
573 		    FKIOCTL, kcred, NULL);
574 	}
575 
576 	/*
577 	 * Clear the nowritecache bit, so that on a vdev_reopen() we will
578 	 * try again.
579 	 */
580 	vd->vdev_nowritecache = B_FALSE;
581 
582 	return (0);
583 }
584 
585 static void
vdev_disk_close(vdev_t * vd)586 vdev_disk_close(vdev_t *vd)
587 {
588 	vdev_disk_t *dvd = vd->vdev_tsd;
589 
590 	if (vd->vdev_reopening || dvd == NULL)
591 		return;
592 
593 	if (dvd->vd_minor != NULL) {
594 		ddi_devid_str_free(dvd->vd_minor);
595 		dvd->vd_minor = NULL;
596 	}
597 
598 	if (dvd->vd_devid != NULL) {
599 		ddi_devid_free(dvd->vd_devid);
600 		dvd->vd_devid = NULL;
601 	}
602 
603 	if (dvd->vd_lh != NULL) {
604 		(void) ldi_close(dvd->vd_lh, spa_mode(vd->vdev_spa), kcred);
605 		dvd->vd_lh = NULL;
606 	}
607 
608 	vd->vdev_delayed_close = B_FALSE;
609 	/*
610 	 * If we closed the LDI handle due to an offline notify from LDI,
611 	 * don't free vd->vdev_tsd or unregister the callbacks here;
612 	 * the offline finalize callback or a reopen will take care of it.
613 	 */
614 	if (dvd->vd_ldi_offline)
615 		return;
616 
617 	vdev_disk_free(vd);
618 }
619 
620 int
vdev_disk_physio(vdev_t * vd,caddr_t data,size_t size,uint64_t offset,int flags,boolean_t isdump)621 vdev_disk_physio(vdev_t *vd, caddr_t data,
622     size_t size, uint64_t offset, int flags, boolean_t isdump)
623 {
624 	vdev_disk_t *dvd = vd->vdev_tsd;
625 
626 	/*
627 	 * If the vdev is closed, it's likely in the REMOVED or FAULTED state.
628 	 * Nothing to be done here but return failure.
629 	 */
630 	if (dvd == NULL || (dvd->vd_ldi_offline && dvd->vd_lh == NULL))
631 		return (EIO);
632 
633 	ASSERT(vd->vdev_ops == &vdev_disk_ops);
634 
635 	/*
636 	 * If in the context of an active crash dump, use the ldi_dump(9F)
637 	 * call instead of ldi_strategy(9F) as usual.
638 	 */
639 	if (isdump) {
640 		ASSERT3P(dvd, !=, NULL);
641 		return (ldi_dump(dvd->vd_lh, data, lbtodb(offset),
642 		    lbtodb(size)));
643 	}
644 
645 	return (vdev_disk_ldi_physio(dvd->vd_lh, data, size, offset, flags));
646 }
647 
648 int
vdev_disk_ldi_physio(ldi_handle_t vd_lh,caddr_t data,size_t size,uint64_t offset,int flags)649 vdev_disk_ldi_physio(ldi_handle_t vd_lh, caddr_t data,
650     size_t size, uint64_t offset, int flags)
651 {
652 	buf_t *bp;
653 	int error = 0;
654 
655 	if (vd_lh == NULL)
656 		return (SET_ERROR(EINVAL));
657 
658 	ASSERT(flags & B_READ || flags & B_WRITE);
659 
660 	bp = getrbuf(KM_SLEEP);
661 	bp->b_flags = flags | B_BUSY | B_NOCACHE | B_FAILFAST;
662 	bp->b_bcount = size;
663 	bp->b_un.b_addr = (void *)data;
664 	bp->b_lblkno = lbtodb(offset);
665 	bp->b_bufsize = size;
666 
667 	error = ldi_strategy(vd_lh, bp);
668 	ASSERT(error == 0);
669 	if ((error = biowait(bp)) == 0 && bp->b_resid != 0)
670 		error = SET_ERROR(EIO);
671 	freerbuf(bp);
672 
673 	return (error);
674 }
675 
676 static void
vdev_disk_io_intr(buf_t * bp)677 vdev_disk_io_intr(buf_t *bp)
678 {
679 	vdev_buf_t *vb = (vdev_buf_t *)bp;
680 	zio_t *zio = vb->vb_io;
681 
682 	/*
683 	 * The rest of the zio stack only deals with EIO, ECKSUM, and ENXIO.
684 	 * Rather than teach the rest of the stack about other error
685 	 * possibilities (EFAULT, etc), we normalize the error value here.
686 	 */
687 	zio->io_error = (geterror(bp) != 0 ? EIO : 0);
688 
689 	if (zio->io_error == 0 && bp->b_resid != 0)
690 		zio->io_error = SET_ERROR(EIO);
691 
692 	kmem_free(vb, sizeof (vdev_buf_t));
693 
694 	zio_interrupt(zio);
695 }
696 
697 static void
vdev_disk_ioctl_free(zio_t * zio)698 vdev_disk_ioctl_free(zio_t *zio)
699 {
700 	kmem_free(zio->io_vsd, sizeof (struct dk_callback));
701 }
702 
703 static const zio_vsd_ops_t vdev_disk_vsd_ops = {
704 	vdev_disk_ioctl_free,
705 	zio_vsd_default_cksum_report
706 };
707 
708 static void
vdev_disk_ioctl_done(void * zio_arg,int error)709 vdev_disk_ioctl_done(void *zio_arg, int error)
710 {
711 	zio_t *zio = zio_arg;
712 
713 	zio->io_error = error;
714 
715 	zio_interrupt(zio);
716 }
717 
718 static void
vdev_disk_io_start(zio_t * zio)719 vdev_disk_io_start(zio_t *zio)
720 {
721 	vdev_t *vd = zio->io_vd;
722 	vdev_disk_t *dvd = vd->vdev_tsd;
723 	vdev_buf_t *vb;
724 	struct dk_callback *dkc;
725 	buf_t *bp;
726 	int error;
727 
728 	/*
729 	 * If the vdev is closed, it's likely in the REMOVED or FAULTED state.
730 	 * Nothing to be done here but return failure.
731 	 */
732 	if (dvd == NULL || (dvd->vd_ldi_offline && dvd->vd_lh == NULL)) {
733 		zio->io_error = ENXIO;
734 		zio_interrupt(zio);
735 		return;
736 	}
737 
738 	if (zio->io_type == ZIO_TYPE_IOCTL) {
739 		/* XXPOLICY */
740 		if (!vdev_readable(vd)) {
741 			zio->io_error = SET_ERROR(ENXIO);
742 			zio_interrupt(zio);
743 			return;
744 		}
745 
746 		switch (zio->io_cmd) {
747 
748 		case DKIOCFLUSHWRITECACHE:
749 
750 			if (zfs_nocacheflush)
751 				break;
752 
753 			if (vd->vdev_nowritecache) {
754 				zio->io_error = SET_ERROR(ENOTSUP);
755 				break;
756 			}
757 
758 			zio->io_vsd = dkc = kmem_alloc(sizeof (*dkc), KM_SLEEP);
759 			zio->io_vsd_ops = &vdev_disk_vsd_ops;
760 
761 			dkc->dkc_callback = vdev_disk_ioctl_done;
762 			dkc->dkc_flag = FLUSH_VOLATILE;
763 			dkc->dkc_cookie = zio;
764 
765 			error = ldi_ioctl(dvd->vd_lh, zio->io_cmd,
766 			    (uintptr_t)dkc, FKIOCTL, kcred, NULL);
767 
768 			if (error == 0) {
769 				/*
770 				 * The ioctl will be done asychronously,
771 				 * and will call vdev_disk_ioctl_done()
772 				 * upon completion.
773 				 */
774 				return;
775 			}
776 
777 			if (error == ENOTSUP || error == ENOTTY) {
778 				/*
779 				 * If we get ENOTSUP or ENOTTY, we know that
780 				 * no future attempts will ever succeed.
781 				 * In this case we set a persistent bit so
782 				 * that we don't bother with the ioctl in the
783 				 * future.
784 				 */
785 				vd->vdev_nowritecache = B_TRUE;
786 			}
787 			zio->io_error = error;
788 
789 			break;
790 
791 		default:
792 			zio->io_error = SET_ERROR(ENOTSUP);
793 		}
794 
795 		zio_execute(zio);
796 		return;
797 	}
798 
799 	vb = kmem_alloc(sizeof (vdev_buf_t), KM_SLEEP);
800 
801 	vb->vb_io = zio;
802 	bp = &vb->vb_buf;
803 
804 	bioinit(bp);
805 	bp->b_flags = B_BUSY | B_NOCACHE |
806 	    (zio->io_type == ZIO_TYPE_READ ? B_READ : B_WRITE);
807 	if (!(zio->io_flags & (ZIO_FLAG_IO_RETRY | ZIO_FLAG_TRYHARD)))
808 		bp->b_flags |= B_FAILFAST;
809 	bp->b_bcount = zio->io_size;
810 	bp->b_un.b_addr = zio->io_data;
811 	bp->b_lblkno = lbtodb(zio->io_offset);
812 	bp->b_bufsize = zio->io_size;
813 	bp->b_iodone = (int (*)())vdev_disk_io_intr;
814 
815 	/* ldi_strategy() will return non-zero only on programming errors */
816 	VERIFY(ldi_strategy(dvd->vd_lh, bp) == 0);
817 }
818 
819 static void
vdev_disk_io_done(zio_t * zio)820 vdev_disk_io_done(zio_t *zio)
821 {
822 	vdev_t *vd = zio->io_vd;
823 
824 	/*
825 	 * If the device returned EIO, then attempt a DKIOCSTATE ioctl to see if
826 	 * the device has been removed.  If this is the case, then we trigger an
827 	 * asynchronous removal of the device. Otherwise, probe the device and
828 	 * make sure it's still accessible.
829 	 */
830 	if (zio->io_error == EIO && !vd->vdev_remove_wanted) {
831 		vdev_disk_t *dvd = vd->vdev_tsd;
832 		int state = DKIO_NONE;
833 
834 		if (ldi_ioctl(dvd->vd_lh, DKIOCSTATE, (intptr_t)&state,
835 		    FKIOCTL, kcred, NULL) == 0 && state != DKIO_INSERTED) {
836 			/*
837 			 * We post the resource as soon as possible, instead of
838 			 * when the async removal actually happens, because the
839 			 * DE is using this information to discard previous I/O
840 			 * errors.
841 			 */
842 			zfs_post_remove(zio->io_spa, vd);
843 			vd->vdev_remove_wanted = B_TRUE;
844 			spa_async_request(zio->io_spa, SPA_ASYNC_REMOVE);
845 		} else if (!vd->vdev_delayed_close) {
846 			vd->vdev_delayed_close = B_TRUE;
847 		}
848 	}
849 }
850 
851 vdev_ops_t vdev_disk_ops = {
852 	vdev_disk_open,
853 	vdev_disk_close,
854 	vdev_default_asize,
855 	vdev_disk_io_start,
856 	vdev_disk_io_done,
857 	NULL,
858 	vdev_disk_hold,
859 	vdev_disk_rele,
860 	VDEV_TYPE_DISK,		/* name of this vdev type */
861 	B_TRUE			/* leaf vdev */
862 };
863 
864 /*
865  * Given the root disk device devid or pathname, read the label from
866  * the device, and construct a configuration nvlist.
867  */
868 int
vdev_disk_read_rootlabel(char * devpath,char * devid,nvlist_t ** config)869 vdev_disk_read_rootlabel(char *devpath, char *devid, nvlist_t **config)
870 {
871 	ldi_handle_t vd_lh;
872 	vdev_label_t *label;
873 	uint64_t s, size;
874 	int l;
875 	ddi_devid_t tmpdevid;
876 	int error = -1;
877 	char *minor_name;
878 
879 	/*
880 	 * Read the device label and build the nvlist.
881 	 */
882 	if (devid != NULL && ddi_devid_str_decode(devid, &tmpdevid,
883 	    &minor_name) == 0) {
884 		error = ldi_open_by_devid(tmpdevid, minor_name,
885 		    FREAD, kcred, &vd_lh, zfs_li);
886 		ddi_devid_free(tmpdevid);
887 		ddi_devid_str_free(minor_name);
888 	}
889 
890 	if (error && (error = ldi_open_by_name(devpath, FREAD, kcred, &vd_lh,
891 	    zfs_li)))
892 		return (error);
893 
894 	if (ldi_get_size(vd_lh, &s)) {
895 		(void) ldi_close(vd_lh, FREAD, kcred);
896 		return (SET_ERROR(EIO));
897 	}
898 
899 	size = P2ALIGN_TYPED(s, sizeof (vdev_label_t), uint64_t);
900 	label = kmem_alloc(sizeof (vdev_label_t), KM_SLEEP);
901 
902 	*config = NULL;
903 	for (l = 0; l < VDEV_LABELS; l++) {
904 		uint64_t offset, state, txg = 0;
905 
906 		/* read vdev label */
907 		offset = vdev_label_offset(size, l, 0);
908 		if (vdev_disk_ldi_physio(vd_lh, (caddr_t)label,
909 		    VDEV_SKIP_SIZE + VDEV_PHYS_SIZE, offset, B_READ) != 0)
910 			continue;
911 
912 		if (nvlist_unpack(label->vl_vdev_phys.vp_nvlist,
913 		    sizeof (label->vl_vdev_phys.vp_nvlist), config, 0) != 0) {
914 			*config = NULL;
915 			continue;
916 		}
917 
918 		if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_STATE,
919 		    &state) != 0 || state >= POOL_STATE_DESTROYED) {
920 			nvlist_free(*config);
921 			*config = NULL;
922 			continue;
923 		}
924 
925 		if (nvlist_lookup_uint64(*config, ZPOOL_CONFIG_POOL_TXG,
926 		    &txg) != 0 || txg == 0) {
927 			nvlist_free(*config);
928 			*config = NULL;
929 			continue;
930 		}
931 
932 		break;
933 	}
934 
935 	kmem_free(label, sizeof (vdev_label_t));
936 	(void) ldi_close(vd_lh, FREAD, kcred);
937 	if (*config == NULL)
938 		error = SET_ERROR(EIDRM);
939 
940 	return (error);
941 }
942