1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (c) 2015 Netflix, Inc.
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 *
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25 * SUCH DAMAGE.
26 *
27 * Derived from ata_da.c:
28 * Copyright (c) 2009 Alexander Motin <mav@FreeBSD.org>
29 */
30
31 #include <sys/param.h>
32
33 #ifdef _KERNEL
34 #include <sys/systm.h>
35 #include <sys/kernel.h>
36 #include <sys/bio.h>
37 #include <sys/sysctl.h>
38 #include <sys/taskqueue.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/conf.h>
42 #include <sys/devicestat.h>
43 #include <sys/eventhandler.h>
44 #include <sys/malloc.h>
45 #include <sys/cons.h>
46 #include <sys/power.h>
47 #include <sys/proc.h>
48 #include <sys/reboot.h>
49 #include <sys/sbuf.h>
50 #include <geom/geom.h>
51 #include <geom/geom_disk.h>
52 #endif /* _KERNEL */
53
54 #ifndef _KERNEL
55 #include <stdio.h>
56 #include <string.h>
57 #endif /* _KERNEL */
58
59 #include <cam/cam.h>
60 #include <cam/cam_ccb.h>
61 #include <cam/cam_periph.h>
62 #include <cam/cam_xpt_periph.h>
63 #include <cam/cam_sim.h>
64 #include <cam/cam_iosched.h>
65
66 #include <cam/nvme/nvme_all.h>
67
68 typedef enum {
69 NDA_STATE_NORMAL
70 } nda_state;
71
72 typedef enum {
73 NDA_FLAG_OPEN = 0x0001,
74 NDA_FLAG_DIRTY = 0x0002,
75 NDA_FLAG_SCTX_INIT = 0x0004,
76 NDA_FLAG_RESCAN = 0x0008,
77 } nda_flags;
78 #define NDA_FLAG_STRING \
79 "\020" \
80 "\001OPEN" \
81 "\002DIRTY" \
82 "\003SCTX_INIT" \
83 "\004RESCAN"
84
85 typedef enum {
86 NDA_Q_4K = 0x01,
87 NDA_Q_NONE = 0x00,
88 } nda_quirks;
89
90 #define NDA_Q_BIT_STRING \
91 "\020" \
92 "\001Bit 0"
93
94 typedef enum {
95 NDA_CCB_BUFFER_IO = 0x01,
96 NDA_CCB_DUMP = 0x02,
97 NDA_CCB_TRIM = 0x03,
98 NDA_CCB_PASS = 0x04,
99 NDA_CCB_TYPE_MASK = 0x0F,
100 } nda_ccb_state;
101
102 /* Offsets into our private area for storing information */
103 #define ccb_state ccb_h.ppriv_field0
104 #define ccb_bp ccb_h.ppriv_ptr1 /* For NDA_CCB_BUFFER_IO */
105 #define ccb_trim ccb_h.ppriv_ptr1 /* For NDA_CCB_TRIM */
106
107 struct nda_softc {
108 struct cam_iosched_softc *cam_iosched;
109 int outstanding_cmds; /* Number of active commands */
110 int refcount; /* Active xpt_action() calls */
111 nda_state state;
112 nda_flags flags;
113 nda_quirks quirks;
114 int unmappedio;
115 quad_t deletes;
116 uint32_t nsid; /* Namespace ID for this nda device */
117 struct disk *disk;
118 struct task sysctl_task;
119 struct sysctl_ctx_list sysctl_ctx;
120 struct sysctl_oid *sysctl_tree;
121 uint64_t trim_count;
122 uint64_t trim_ranges;
123 uint64_t trim_lbas;
124 #ifdef CAM_TEST_FAILURE
125 int force_read_error;
126 int force_write_error;
127 int periodic_read_error;
128 int periodic_read_count;
129 #endif
130 #ifdef CAM_IO_STATS
131 struct sysctl_ctx_list sysctl_stats_ctx;
132 struct sysctl_oid *sysctl_stats_tree;
133 u_int timeouts;
134 u_int errors;
135 u_int invalidations;
136 #endif
137 };
138
139 struct nda_trim_request {
140 struct nvme_dsm_range dsm[NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range)];
141 TAILQ_HEAD(, bio) bps;
142 };
143 _Static_assert(NVME_MAX_DSM_TRIM % sizeof(struct nvme_dsm_range) == 0,
144 "NVME_MAX_DSM_TRIM must be an integral number of ranges");
145
146 /* Need quirk table */
147
148 static disk_ioctl_t ndaioctl;
149 static disk_strategy_t ndastrategy;
150 static dumper_t ndadump;
151 static periph_init_t ndainit;
152 static void ndaasync(void *callback_arg, uint32_t code,
153 struct cam_path *path, void *arg);
154 static void ndasysctlinit(void *context, int pending);
155 static int ndaflagssysctl(SYSCTL_HANDLER_ARGS);
156 static periph_ctor_t ndaregister;
157 static periph_dtor_t ndacleanup;
158 static periph_start_t ndastart;
159 static periph_oninv_t ndaoninvalidate;
160 static void ndadone(struct cam_periph *periph,
161 union ccb *done_ccb);
162 static int ndaerror(union ccb *ccb, uint32_t cam_flags,
163 uint32_t sense_flags);
164 static void ndashutdown(void *arg, int howto);
165 static void ndasuspend(void *arg, enum power_stype stype);
166
167 #ifndef NDA_DEFAULT_SEND_ORDERED
168 #define NDA_DEFAULT_SEND_ORDERED 1
169 #endif
170 #ifndef NDA_DEFAULT_TIMEOUT
171 #define NDA_DEFAULT_TIMEOUT 30 /* Timeout in seconds */
172 #endif
173 #ifndef NDA_DEFAULT_RETRY
174 #define NDA_DEFAULT_RETRY 4
175 #endif
176 #ifndef NDA_MAX_TRIM_ENTRIES
177 #define NDA_MAX_TRIM_ENTRIES (NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range))/* Number of DSM trims to use, max 256 */
178 #endif
179
180 static SYSCTL_NODE(_kern_cam, OID_AUTO, nda, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
181 "CAM Direct Access Disk driver");
182
183 //static int nda_retry_count = NDA_DEFAULT_RETRY;
184 static int nda_send_ordered = NDA_DEFAULT_SEND_ORDERED;
185 static int nda_default_timeout = NDA_DEFAULT_TIMEOUT;
186 static int nda_max_trim_entries = NDA_MAX_TRIM_ENTRIES;
187 static int nda_enable_biospeedup = 1;
188 static int nda_nvd_compat = 1;
189 SYSCTL_INT(_kern_cam_nda, OID_AUTO, max_trim, CTLFLAG_RDTUN,
190 &nda_max_trim_entries, NDA_MAX_TRIM_ENTRIES,
191 "Maximum number of BIO_DELETE to send down as a DSM TRIM.");
192 SYSCTL_INT(_kern_cam_nda, OID_AUTO, enable_biospeedup, CTLFLAG_RDTUN,
193 &nda_enable_biospeedup, 0, "Enable BIO_SPEEDUP processing.");
194 SYSCTL_INT(_kern_cam_nda, OID_AUTO, nvd_compat, CTLFLAG_RDTUN,
195 &nda_nvd_compat, 1, "Enable creation of nvd aliases.");
196
197 /*
198 * All NVMe media is non-rotational, so all nvme device instances
199 * share this to implement the sysctl.
200 */
201 static int nda_rotating_media = 0;
202
203 static struct periph_driver ndadriver =
204 {
205 ndainit, "nda",
206 TAILQ_HEAD_INITIALIZER(ndadriver.units), /* generation */ 0
207 };
208
209 PERIPHDRIVER_DECLARE(nda, ndadriver);
210
211 static MALLOC_DEFINE(M_NVMEDA, "nvme_da", "nvme_da buffers");
212
213 /*
214 * nice wrappers. Maybe these belong in nvme_all.c instead of
215 * here, but this is the only place that uses these. Should
216 * we ever grow another NVME periph, we should move them
217 * all there wholesale.
218 */
219
220 static void
nda_nvme_flush(struct nda_softc * softc,struct ccb_nvmeio * nvmeio)221 nda_nvme_flush(struct nda_softc *softc, struct ccb_nvmeio *nvmeio)
222 {
223 cam_fill_nvmeio(nvmeio,
224 0, /* retries */
225 ndadone, /* cbfcnp */
226 CAM_DIR_NONE, /* flags */
227 NULL, /* data_ptr */
228 0, /* dxfer_len */
229 nda_default_timeout * 1000); /* timeout 30s */
230 nvme_ns_flush_cmd(&nvmeio->cmd, softc->nsid);
231 }
232
233 static void
nda_nvme_trim(struct nda_softc * softc,struct ccb_nvmeio * nvmeio,void * payload,uint32_t num_ranges)234 nda_nvme_trim(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
235 void *payload, uint32_t num_ranges)
236 {
237 cam_fill_nvmeio(nvmeio,
238 0, /* retries */
239 ndadone, /* cbfcnp */
240 CAM_DIR_OUT, /* flags */
241 payload, /* data_ptr */
242 num_ranges * sizeof(struct nvme_dsm_range), /* dxfer_len */
243 nda_default_timeout * 1000); /* timeout 30s */
244 nvme_ns_trim_cmd(&nvmeio->cmd, softc->nsid, num_ranges);
245 }
246
247 static void
nda_nvme_write(struct nda_softc * softc,struct ccb_nvmeio * nvmeio,void * payload,uint64_t lba,uint32_t len,uint32_t count)248 nda_nvme_write(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
249 void *payload, uint64_t lba, uint32_t len, uint32_t count)
250 {
251 cam_fill_nvmeio(nvmeio,
252 0, /* retries */
253 ndadone, /* cbfcnp */
254 CAM_DIR_OUT, /* flags */
255 payload, /* data_ptr */
256 len, /* dxfer_len */
257 nda_default_timeout * 1000); /* timeout 30s */
258 nvme_ns_write_cmd(&nvmeio->cmd, softc->nsid, lba, count);
259 }
260
261 static void
nda_nvme_rw_bio(struct nda_softc * softc,struct ccb_nvmeio * nvmeio,struct bio * bp,uint32_t rwcmd)262 nda_nvme_rw_bio(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
263 struct bio *bp, uint32_t rwcmd)
264 {
265 int flags = rwcmd == NVME_OPC_READ ? CAM_DIR_IN : CAM_DIR_OUT;
266 void *payload;
267 uint64_t lba;
268 uint32_t count;
269
270 if (bp->bio_flags & BIO_UNMAPPED) {
271 flags |= CAM_DATA_BIO;
272 payload = bp;
273 } else {
274 payload = bp->bio_data;
275 }
276
277 lba = bp->bio_pblkno;
278 count = bp->bio_bcount / softc->disk->d_sectorsize;
279
280 cam_fill_nvmeio(nvmeio,
281 0, /* retries */
282 ndadone, /* cbfcnp */
283 flags, /* flags */
284 payload, /* data_ptr */
285 bp->bio_bcount, /* dxfer_len */
286 nda_default_timeout * 1000); /* timeout 30s */
287 nvme_ns_rw_cmd(&nvmeio->cmd, rwcmd, softc->nsid, lba, count);
288 }
289
290 static void
ndasetgeom(struct nda_softc * softc,struct cam_periph * periph)291 ndasetgeom(struct nda_softc *softc, struct cam_periph *periph)
292 {
293 struct disk *disk = softc->disk;
294 const struct nvme_namespace_data *nsd;
295 const struct nvme_controller_data *cd;
296 uint8_t flbas_fmt, lbads, vwc_present;
297 u_int flags;
298
299 nsd = nvme_get_identify_ns(periph);
300 cd = nvme_get_identify_cntrl(periph);
301
302 /*
303 * Preserve flags we can't infer that were set before. UNMAPPED comes
304 * from the PIM, so won't change after we set it the first
305 * time. Subsequent times, we have to preserve it.
306 */
307 flags = disk->d_flags & DISKFLAG_UNMAPPED_BIO; /* Need to preserve */
308
309 flbas_fmt = NVMEV(NVME_NS_DATA_FLBAS_FORMAT, nsd->flbas);
310 lbads = NVMEV(NVME_NS_DATA_LBAF_LBADS, nsd->lbaf[flbas_fmt]);
311 disk->d_sectorsize = 1 << lbads;
312 disk->d_mediasize = (off_t)(disk->d_sectorsize * nsd->nsze);
313 disk->d_delmaxsize = disk->d_mediasize;
314 disk->d_flags = DISKFLAG_DIRECT_COMPLETION;
315 if (nvme_ctrlr_has_dataset_mgmt(cd))
316 disk->d_flags |= DISKFLAG_CANDELETE;
317 vwc_present = NVMEV(NVME_CTRLR_DATA_VWC_PRESENT, cd->vwc);
318 if (vwc_present)
319 disk->d_flags |= DISKFLAG_CANFLUSHCACHE;
320 disk->d_flags |= flags;
321 }
322
323 static void
ndaopen_rescan_done(struct cam_periph * periph,union ccb * ccb)324 ndaopen_rescan_done(struct cam_periph *periph, union ccb *ccb)
325 {
326 struct nda_softc *softc;
327
328 softc = (struct nda_softc *)periph->softc;
329
330 cam_periph_assert(periph, MA_OWNED);
331
332 softc->flags &= ~NDA_FLAG_RESCAN;
333 xpt_release_ccb(ccb);
334 wakeup(&softc->disk->d_mediasize);
335 }
336
337
338 static int
ndaopen(struct disk * dp)339 ndaopen(struct disk *dp)
340 {
341 struct cam_periph *periph;
342 struct nda_softc *softc;
343 union ccb *ccb;
344 int error;
345
346 periph = (struct cam_periph *)dp->d_drv1;
347 if (cam_periph_acquire(periph) != 0) {
348 return(ENXIO);
349 }
350
351 cam_periph_lock(periph);
352 if ((error = cam_periph_hold(periph, PRIBIO|PCATCH)) != 0) {
353 cam_periph_unlock(periph);
354 cam_periph_release(periph);
355 return (error);
356 }
357
358 softc = (struct nda_softc *)periph->softc;
359 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
360 ("ndaopen\n"));
361
362 /*
363 * Rescan the lun in case the mediasize or sectorsize has changed since
364 * we probed the device. Format and secure erase operations can do this,
365 * but the nvme standard doesn't require a async notification of that
366 * happening. da/ada do this by restarting their probe, but since
367 * nvme_xpt gets the identify information we need, we just rescan here
368 * since it's the easiest way to notice size changes.
369 *
370 * Not acquiring / releasing for the geom probe -- it's inline
371 */
372 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
373 ccb->ccb_h.func_code = XPT_SCAN_LUN;
374 ccb->ccb_h.cbfcnp = ndaopen_rescan_done;
375 ccb->ccb_h.ppriv_ptr0 = periph;
376 ccb->crcn.flags = 0;
377 xpt_action(ccb);
378
379 softc->flags |= NDA_FLAG_RESCAN;
380 error = 0;
381 while ((softc->flags & NDA_FLAG_RESCAN) != 0 && error == 0)
382 error = cam_periph_sleep(periph, &softc->disk->d_mediasize, PRIBIO,
383 "ndareprobe", 0);
384 if (error != 0)
385 xpt_print(periph->path, "Unable to retrieve capacity data\n");
386 else
387 ndasetgeom(softc, periph);
388
389 softc->flags |= NDA_FLAG_OPEN;
390
391 cam_periph_unhold(periph);
392 cam_periph_unlock(periph);
393 return (0);
394 }
395
396 static int
ndaclose(struct disk * dp)397 ndaclose(struct disk *dp)
398 {
399 struct cam_periph *periph;
400 struct nda_softc *softc;
401 union ccb *ccb;
402 int error;
403
404 periph = (struct cam_periph *)dp->d_drv1;
405 softc = (struct nda_softc *)periph->softc;
406 cam_periph_lock(periph);
407
408 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
409 ("ndaclose\n"));
410
411 if ((softc->flags & NDA_FLAG_DIRTY) != 0 &&
412 (periph->flags & CAM_PERIPH_INVALID) == 0 &&
413 cam_periph_hold(periph, PRIBIO) == 0) {
414 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
415 nda_nvme_flush(softc, &ccb->nvmeio);
416 error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
417 /*sense_flags*/0, softc->disk->d_devstat);
418
419 if (error != 0)
420 xpt_print(periph->path, "Synchronize cache failed\n");
421 else
422 softc->flags &= ~NDA_FLAG_DIRTY;
423 xpt_release_ccb(ccb);
424 cam_periph_unhold(periph);
425 }
426
427 softc->flags &= ~NDA_FLAG_OPEN;
428
429 while (softc->refcount != 0)
430 cam_periph_sleep(periph, &softc->refcount, PRIBIO, "ndaclose", 1);
431 KASSERT(softc->outstanding_cmds == 0,
432 ("nda %d outstanding commands", softc->outstanding_cmds));
433 cam_periph_unlock(periph);
434 cam_periph_release(periph);
435 return (0);
436 }
437
438 static void
ndaschedule(struct cam_periph * periph)439 ndaschedule(struct cam_periph *periph)
440 {
441 struct nda_softc *softc = (struct nda_softc *)periph->softc;
442
443 if (softc->state != NDA_STATE_NORMAL)
444 return;
445
446 cam_iosched_schedule(softc->cam_iosched, periph);
447 }
448
449 static int
ndaioctl(struct disk * dp,u_long cmd,void * data,int fflag,struct thread * td)450 ndaioctl(struct disk *dp, u_long cmd, void *data, int fflag,
451 struct thread *td)
452 {
453 struct cam_periph *periph;
454
455 periph = (struct cam_periph *)dp->d_drv1;
456
457 switch (cmd) {
458 case NVME_IO_TEST:
459 case NVME_BIO_TEST:
460 /*
461 * These don't map well to the underlying CCBs, so
462 * they are usupported via CAM.
463 */
464 return (ENOTTY);
465 case NVME_GET_NSID:
466 {
467 struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)data;
468 struct ccb_pathinq cpi;
469
470 xpt_path_inq(&cpi, periph->path);
471 strncpy(gnsid->cdev, cpi.xport_specific.nvme.dev_name,
472 sizeof(gnsid->cdev));
473 gnsid->nsid = cpi.xport_specific.nvme.nsid;
474 return (0);
475 }
476 case NVME_PASSTHROUGH_CMD:
477 {
478 struct nvme_pt_command *pt;
479 union ccb *ccb;
480 struct cam_periph_map_info mapinfo;
481 u_int maxmap = dp->d_maxsize;
482 int error;
483
484 /*
485 * Create a NVME_IO CCB to do the passthrough command.
486 */
487 pt = (struct nvme_pt_command *)data;
488 ccb = xpt_alloc_ccb();
489 xpt_setup_ccb(&ccb->ccb_h, periph->path, CAM_PRIORITY_NORMAL);
490 ccb->ccb_state = NDA_CCB_PASS;
491 cam_fill_nvmeio(&ccb->nvmeio,
492 0, /* Retries */
493 ndadone,
494 (pt->is_read ? CAM_DIR_IN : CAM_DIR_OUT) | CAM_DATA_VADDR,
495 pt->buf,
496 pt->len,
497 nda_default_timeout * 1000);
498 memcpy(&ccb->nvmeio.cmd, &pt->cmd, sizeof(pt->cmd));
499
500 /*
501 * Wire the user memory in this request for the I/O
502 */
503 memset(&mapinfo, 0, sizeof(mapinfo));
504 error = cam_periph_mapmem(ccb, &mapinfo, maxmap);
505 if (error)
506 goto out;
507
508 /*
509 * Lock the periph and run the command.
510 */
511 cam_periph_lock(periph);
512 cam_periph_runccb(ccb, NULL, CAM_RETRY_SELTO,
513 SF_RETRY_UA | SF_NO_PRINT, NULL);
514
515 /*
516 * Tear down mapping and return status.
517 */
518 cam_periph_unlock(periph);
519 error = cam_periph_unmapmem(ccb, &mapinfo);
520 if (!cam_ccb_success(ccb))
521 error = EIO;
522 out:
523 cam_periph_lock(periph);
524 xpt_release_ccb(ccb);
525 cam_periph_unlock(periph);
526 return (error);
527 }
528 default:
529 break;
530 }
531 return (ENOTTY);
532 }
533
534 /*
535 * Actually translate the requested transfer into one the physical driver
536 * can understand. The transfer is described by a buf and will include
537 * only one physical transfer.
538 */
539 static void
ndastrategy(struct bio * bp)540 ndastrategy(struct bio *bp)
541 {
542 struct cam_periph *periph;
543 struct nda_softc *softc;
544
545 periph = (struct cam_periph *)bp->bio_disk->d_drv1;
546 softc = (struct nda_softc *)periph->softc;
547
548 cam_periph_lock(periph);
549
550 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastrategy(%p)\n", bp));
551
552 /*
553 * If the device has been made invalid, error out
554 */
555 if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
556 cam_periph_unlock(periph);
557 biofinish(bp, NULL, ENXIO);
558 return;
559 }
560
561 if (bp->bio_cmd == BIO_DELETE)
562 softc->deletes++;
563
564 /*
565 * Place it in the queue of disk activities for this disk
566 */
567 cam_iosched_queue_work(softc->cam_iosched, bp);
568
569 /*
570 * Schedule ourselves for performing the work.
571 */
572 ndaschedule(periph);
573 cam_periph_unlock(periph);
574
575 return;
576 }
577
578 static int
ndadump(void * arg,void * virtual,off_t offset,size_t length)579 ndadump(void *arg, void *virtual, off_t offset, size_t length)
580 {
581 struct cam_periph *periph;
582 struct nda_softc *softc;
583 u_int secsize;
584 struct ccb_nvmeio nvmeio;
585 struct disk *dp;
586 uint64_t lba;
587 uint32_t count;
588 int error = 0;
589
590 dp = arg;
591 periph = dp->d_drv1;
592 softc = (struct nda_softc *)periph->softc;
593 secsize = softc->disk->d_sectorsize;
594 lba = offset / secsize;
595 count = length / secsize;
596
597 if ((periph->flags & CAM_PERIPH_INVALID) != 0)
598 return (ENXIO);
599
600 /* xpt_get_ccb returns a zero'd allocation for the ccb, mimic that here */
601 memset(&nvmeio, 0, sizeof(nvmeio));
602 if (length > 0) {
603 xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
604 nvmeio.ccb_state = NDA_CCB_DUMP;
605 nda_nvme_write(softc, &nvmeio, virtual, lba, length, count);
606 error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
607 0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
608 if (error != 0)
609 printf("Aborting dump due to I/O error %d.\n", error);
610
611 return (error);
612 }
613
614 /* Flush */
615 xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
616
617 nvmeio.ccb_state = NDA_CCB_DUMP;
618 nda_nvme_flush(softc, &nvmeio);
619 error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
620 0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
621 if (error != 0)
622 xpt_print(periph->path, "flush cmd failed\n");
623 return (error);
624 }
625
626 static void
ndainit(void)627 ndainit(void)
628 {
629 cam_status status;
630
631 /*
632 * Install a global async callback. This callback will
633 * receive async callbacks like "new device found".
634 */
635 status = xpt_register_async(AC_FOUND_DEVICE, ndaasync, NULL, NULL);
636
637 if (status != CAM_REQ_CMP) {
638 printf("nda: Failed to attach master async callback "
639 "due to status 0x%x!\n", status);
640 } else if (nda_send_ordered) {
641 /* Register our event handlers */
642 if ((EVENTHANDLER_REGISTER(power_suspend, ndasuspend,
643 NULL, EVENTHANDLER_PRI_LAST)) == NULL)
644 printf("ndainit: power event registration failed!\n");
645 if ((EVENTHANDLER_REGISTER(shutdown_post_sync, ndashutdown,
646 NULL, SHUTDOWN_PRI_DEFAULT)) == NULL)
647 printf("ndainit: shutdown event registration failed!\n");
648 }
649 }
650
651 /*
652 * Callback from GEOM, called when it has finished cleaning up its
653 * resources.
654 */
655 static void
ndadiskgonecb(struct disk * dp)656 ndadiskgonecb(struct disk *dp)
657 {
658 struct cam_periph *periph;
659
660 periph = (struct cam_periph *)dp->d_drv1;
661
662 cam_periph_release(periph);
663 }
664
665 static void
ndaoninvalidate(struct cam_periph * periph)666 ndaoninvalidate(struct cam_periph *periph)
667 {
668 struct nda_softc *softc;
669
670 softc = (struct nda_softc *)periph->softc;
671
672 /*
673 * De-register any async callbacks.
674 */
675 xpt_register_async(0, ndaasync, periph, periph->path);
676 #ifdef CAM_IO_STATS
677 softc->invalidations++;
678 #endif
679
680 /*
681 * Return all queued I/O with ENXIO. Transactions may be queued up here
682 * for retry (since we are called while there's other transactions
683 * pending). Any requests in the hardware will drain before ndacleanup
684 * is called.
685 */
686 cam_iosched_flush(softc->cam_iosched, NULL, ENXIO);
687
688 /*
689 * Tell GEOM that we've gone away, we'll get a callback when it is
690 * done cleaning up its resources.
691 */
692 disk_gone(softc->disk);
693 }
694
695 static void
ndacleanup(struct cam_periph * periph)696 ndacleanup(struct cam_periph *periph)
697 {
698 struct nda_softc *softc;
699
700 softc = (struct nda_softc *)periph->softc;
701
702 cam_periph_unlock(periph);
703
704 cam_iosched_fini(softc->cam_iosched);
705
706 /*
707 * If we can't free the sysctl tree, oh well...
708 */
709 if ((softc->flags & NDA_FLAG_SCTX_INIT) != 0) {
710 #ifdef CAM_IO_STATS
711 if (sysctl_ctx_free(&softc->sysctl_stats_ctx) != 0)
712 xpt_print(periph->path,
713 "can't remove sysctl stats context\n");
714 #endif
715 if (sysctl_ctx_free(&softc->sysctl_ctx) != 0)
716 xpt_print(periph->path,
717 "can't remove sysctl context\n");
718 }
719
720 disk_destroy(softc->disk);
721 free(softc, M_DEVBUF);
722 cam_periph_lock(periph);
723 }
724
725 static void
ndaasync(void * callback_arg,uint32_t code,struct cam_path * path,void * arg)726 ndaasync(void *callback_arg, uint32_t code, struct cam_path *path, void *arg)
727 {
728 struct cam_periph *periph = callback_arg;
729 struct nda_softc *softc;
730
731 switch (code) {
732 case AC_FOUND_DEVICE:
733 {
734 struct ccb_getdev *cgd;
735 cam_status status;
736
737 cgd = (struct ccb_getdev *)arg;
738 if (cgd == NULL)
739 break;
740
741 if (cgd->protocol != PROTO_NVME)
742 break;
743
744 /*
745 * Allocate a peripheral instance for
746 * this device and start the probe
747 * process.
748 */
749 status = cam_periph_alloc(ndaregister, ndaoninvalidate,
750 ndacleanup, ndastart,
751 "nda", CAM_PERIPH_BIO,
752 path, ndaasync,
753 AC_FOUND_DEVICE, cgd);
754
755 if (status != CAM_REQ_CMP
756 && status != CAM_REQ_INPROG)
757 printf("ndaasync: Unable to attach to new device "
758 "due to status 0x%x\n", status);
759 break;
760 }
761 case AC_GETDEV_CHANGED:
762 {
763 off_t mediasize;
764 u_int sectorsize;
765
766 softc = periph->softc;
767 mediasize = softc->disk->d_mediasize;
768 sectorsize = softc->disk->d_sectorsize;
769 ndasetgeom(softc, periph);
770 /*
771 * If the sectorsize changed, then it's new media. Otherwise if
772 * the media size changed, resize the existing disk. Otherwise
773 * do nothing.
774 */
775 if (sectorsize != softc->disk->d_sectorsize)
776 disk_media_changed(softc->disk, M_WAITOK);
777 else if (mediasize != softc->disk->d_mediasize)
778 disk_resize(softc->disk, M_WAITOK);
779 break;
780 }
781 case AC_ADVINFO_CHANGED:
782 {
783 uintptr_t buftype;
784
785 /*
786 * Note: In theory, we could send CDAI_TYPE_NVME_* events here,
787 * but instead the rescan code only sends more specific
788 * AC_GETDEV_CHANGED. There's no way to generically get
789 * notifications of changes to these structures from the drive
790 * (though we could notice with memcmp). The automation in NVME
791 * is at a much more granular level, so we leverage that.
792 */
793 softc = periph->softc;
794 buftype = (uintptr_t)arg;
795 if (buftype == CDAI_TYPE_PHYS_PATH) {
796 disk_attr_changed(softc->disk, "GEOM::physpath",
797 M_WAITOK);
798 }
799 break;
800 }
801 case AC_LOST_DEVICE:
802 default:
803 break;
804 }
805 cam_periph_async(periph, code, path, arg);
806 }
807
808 static void
ndasysctlinit(void * context,int pending)809 ndasysctlinit(void *context, int pending)
810 {
811 struct cam_periph *periph;
812 struct nda_softc *softc;
813 char tmpstr[32], tmpstr2[16];
814
815 periph = (struct cam_periph *)context;
816
817 /* periph was held for us when this task was enqueued */
818 if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
819 cam_periph_release(periph);
820 return;
821 }
822
823 softc = (struct nda_softc *)periph->softc;
824 snprintf(tmpstr, sizeof(tmpstr), "CAM NDA unit %d", periph->unit_number);
825 snprintf(tmpstr2, sizeof(tmpstr2), "%d", periph->unit_number);
826
827 sysctl_ctx_init(&softc->sysctl_ctx);
828 softc->flags |= NDA_FLAG_SCTX_INIT;
829 softc->sysctl_tree = SYSCTL_ADD_NODE_WITH_LABEL(&softc->sysctl_ctx,
830 SYSCTL_STATIC_CHILDREN(_kern_cam_nda), OID_AUTO, tmpstr2,
831 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr, "device_index");
832 if (softc->sysctl_tree == NULL) {
833 printf("ndasysctlinit: unable to allocate sysctl tree\n");
834 cam_periph_release(periph);
835 return;
836 }
837
838 SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
839 OID_AUTO, "unmapped_io", CTLFLAG_RD,
840 &softc->unmappedio, 0, "Unmapped I/O leaf");
841
842 SYSCTL_ADD_QUAD(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
843 OID_AUTO, "deletes", CTLFLAG_RD,
844 &softc->deletes, "Number of BIO_DELETE requests");
845
846 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
847 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
848 "trim_count", CTLFLAG_RD, &softc->trim_count,
849 "Total number of unmap/dsm commands sent");
850 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
851 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
852 "trim_ranges", CTLFLAG_RD, &softc->trim_ranges,
853 "Total number of ranges in unmap/dsm commands");
854 SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
855 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
856 "trim_lbas", CTLFLAG_RD, &softc->trim_lbas,
857 "Total lbas in the unmap/dsm commands sent");
858
859 SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
860 OID_AUTO, "rotating", CTLFLAG_RD, &nda_rotating_media, 1,
861 "Rotating media");
862
863 SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
864 OID_AUTO, "flags", CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
865 softc, 0, ndaflagssysctl, "A",
866 "Flags for drive");
867
868 #ifdef CAM_IO_STATS
869 softc->sysctl_stats_tree = SYSCTL_ADD_NODE(&softc->sysctl_stats_ctx,
870 SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, "stats",
871 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Statistics");
872 if (softc->sysctl_stats_tree == NULL) {
873 printf("ndasysctlinit: unable to allocate sysctl tree for stats\n");
874 cam_periph_release(periph);
875 return;
876 }
877 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
878 SYSCTL_CHILDREN(softc->sysctl_stats_tree),
879 OID_AUTO, "timeouts", CTLFLAG_RD,
880 &softc->timeouts, 0,
881 "Device timeouts reported by the SIM");
882 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
883 SYSCTL_CHILDREN(softc->sysctl_stats_tree),
884 OID_AUTO, "errors", CTLFLAG_RD,
885 &softc->errors, 0,
886 "Transport errors reported by the SIM.");
887 SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
888 SYSCTL_CHILDREN(softc->sysctl_stats_tree),
889 OID_AUTO, "pack_invalidations", CTLFLAG_RD,
890 &softc->invalidations, 0,
891 "Device pack invalidations.");
892 #endif
893
894 #ifdef CAM_TEST_FAILURE
895 SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
896 OID_AUTO, "invalidate", CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE,
897 periph, 0, cam_periph_invalidate_sysctl, "I",
898 "Write 1 to invalidate the drive immediately");
899 #endif
900
901 cam_iosched_sysctl_init(softc->cam_iosched, &softc->sysctl_ctx,
902 softc->sysctl_tree);
903
904 cam_periph_release(periph);
905 }
906
907 static int
ndaflagssysctl(SYSCTL_HANDLER_ARGS)908 ndaflagssysctl(SYSCTL_HANDLER_ARGS)
909 {
910 struct sbuf sbuf;
911 struct nda_softc *softc = arg1;
912 int error;
913
914 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
915 if (softc->flags != 0)
916 sbuf_printf(&sbuf, "0x%b", (unsigned)softc->flags, NDA_FLAG_STRING);
917 else
918 sbuf_putc(&sbuf, '0');
919 error = sbuf_finish(&sbuf);
920 sbuf_delete(&sbuf);
921
922 return (error);
923 }
924
925 static int
ndagetattr(struct bio * bp)926 ndagetattr(struct bio *bp)
927 {
928 int ret;
929 struct cam_periph *periph;
930
931 if (g_handleattr_int(bp, "GEOM::canspeedup", nda_enable_biospeedup))
932 return (EJUSTRETURN);
933
934 periph = (struct cam_periph *)bp->bio_disk->d_drv1;
935 cam_periph_lock(periph);
936 ret = xpt_getattr(bp->bio_data, bp->bio_length, bp->bio_attribute,
937 periph->path);
938 cam_periph_unlock(periph);
939 if (ret == 0)
940 bp->bio_completed = bp->bio_length;
941 return ret;
942 }
943
944 static cam_status
ndaregister(struct cam_periph * periph,void * arg)945 ndaregister(struct cam_periph *periph, void *arg)
946 {
947 struct nda_softc *softc;
948 struct disk *disk;
949 struct ccb_pathinq cpi;
950 const struct nvme_namespace_data *nsd;
951 const struct nvme_controller_data *cd;
952 char announce_buf[80];
953 u_int maxio;
954 int quirks;
955
956 nsd = nvme_get_identify_ns(periph);
957 cd = nvme_get_identify_cntrl(periph);
958
959 softc = (struct nda_softc *)malloc(sizeof(*softc), M_DEVBUF,
960 M_NOWAIT | M_ZERO);
961
962 if (softc == NULL) {
963 printf("ndaregister: Unable to probe new device. "
964 "Unable to allocate softc\n");
965 return(CAM_REQ_CMP_ERR);
966 }
967
968 /* ident_data parsing */
969
970 periph->softc = softc;
971 softc->quirks = NDA_Q_NONE;
972 xpt_path_inq(&cpi, periph->path);
973 TASK_INIT(&softc->sysctl_task, 0, ndasysctlinit, periph);
974
975 /*
976 * The name space ID is the lun, save it for later I/O
977 */
978 softc->nsid = (uint32_t)xpt_path_lun_id(periph->path);
979
980 /*
981 * Register this media as a disk
982 */
983 (void)cam_periph_acquire(periph);
984 cam_periph_unlock(periph);
985 snprintf(announce_buf, sizeof(announce_buf),
986 "kern.cam.nda.%d.quirks", periph->unit_number);
987 quirks = softc->quirks;
988 TUNABLE_INT_FETCH(announce_buf, &quirks);
989 softc->quirks = quirks;
990 softc->disk = disk = disk_alloc();
991 disk->d_rotation_rate = DISK_RR_NON_ROTATING;
992 disk->d_open = ndaopen;
993 disk->d_close = ndaclose;
994 disk->d_strategy = ndastrategy;
995 disk->d_ioctl = ndaioctl;
996 disk->d_getattr = ndagetattr;
997 if (cam_sim_pollable(periph->sim))
998 disk->d_dump = ndadump;
999 disk->d_gone = ndadiskgonecb;
1000 disk->d_name = "nda";
1001 disk->d_drv1 = periph;
1002 disk->d_unit = periph->unit_number;
1003 maxio = cpi.maxio; /* Honor max I/O size of SIM */
1004 if (maxio == 0)
1005 maxio = DFLTPHYS; /* traditional default */
1006 else if (maxio > maxphys)
1007 maxio = maxphys; /* for safety */
1008 disk->d_maxsize = maxio;
1009 ndasetgeom(softc, periph);
1010 if ((cpi.hba_misc & PIM_UNMAPPED) != 0) {
1011 disk->d_flags |= DISKFLAG_UNMAPPED_BIO;
1012 softc->unmappedio = 1;
1013 }
1014
1015 /*
1016 * d_ident and d_descr are both far bigger than the length of either
1017 * the serial or model number strings.
1018 */
1019 cam_strvis_flag(disk->d_descr, cd->mn, NVME_MODEL_NUMBER_LENGTH,
1020 sizeof(disk->d_descr), CAM_STRVIS_FLAG_NONASCII_SPC);
1021
1022 cam_strvis_flag(disk->d_ident, cd->sn, NVME_SERIAL_NUMBER_LENGTH,
1023 sizeof(disk->d_ident), CAM_STRVIS_FLAG_NONASCII_SPC);
1024
1025 disk->d_hba_vendor = cpi.hba_vendor;
1026 disk->d_hba_device = cpi.hba_device;
1027 disk->d_hba_subvendor = cpi.hba_subvendor;
1028 disk->d_hba_subdevice = cpi.hba_subdevice;
1029 snprintf(disk->d_attachment, sizeof(disk->d_attachment),
1030 "%s%d", cpi.dev_name, cpi.unit_number);
1031 if (NVMEV(NVME_NS_DATA_NSFEAT_NPVALID, nsd->nsfeat) != 0 &&
1032 nsd->npwg != 0)
1033 disk->d_stripesize = ((nsd->npwg + 1) * disk->d_sectorsize);
1034 else
1035 disk->d_stripesize = nsd->noiob * disk->d_sectorsize;
1036 disk->d_stripeoffset = 0;
1037 disk->d_devstat = devstat_new_entry(periph->periph_name,
1038 periph->unit_number, disk->d_sectorsize,
1039 DEVSTAT_ALL_SUPPORTED,
1040 DEVSTAT_TYPE_DIRECT | XPORT_DEVSTAT_TYPE(cpi.transport),
1041 DEVSTAT_PRIORITY_DISK);
1042
1043 if (cam_iosched_init(&softc->cam_iosched, periph, disk,
1044 ndaschedule) != 0) {
1045 printf("ndaregister: Unable to probe new device. "
1046 "Unable to allocate iosched memory\n");
1047 free(softc, M_DEVBUF);
1048 return(CAM_REQ_CMP_ERR);
1049 }
1050 cam_iosched_set_sort_queue(softc->cam_iosched, 0);
1051
1052 /*
1053 * Add alias for older nvd drives to ease transition.
1054 */
1055 if (nda_nvd_compat)
1056 disk_add_alias(disk, "nvd");
1057
1058 cam_periph_lock(periph);
1059
1060 snprintf(announce_buf, sizeof(announce_buf),
1061 "%juMB (%ju %u byte sectors)",
1062 (uintmax_t)((uintmax_t)disk->d_mediasize / (1024*1024)),
1063 (uintmax_t)disk->d_mediasize / disk->d_sectorsize,
1064 disk->d_sectorsize);
1065 xpt_announce_periph(periph, announce_buf);
1066 xpt_announce_quirks(periph, softc->quirks, NDA_Q_BIT_STRING);
1067
1068 /*
1069 * Create our sysctl variables, now that we know
1070 * we have successfully attached.
1071 */
1072 if (cam_periph_acquire(periph) == 0)
1073 taskqueue_enqueue(taskqueue_thread, &softc->sysctl_task);
1074
1075 /*
1076 * Register for device going away and info about the drive
1077 * changing (though with NVMe, it can't)
1078 */
1079 xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED | AC_GETDEV_CHANGED,
1080 ndaasync, periph, periph->path);
1081
1082 softc->state = NDA_STATE_NORMAL;
1083
1084 /*
1085 * We'll release this reference once GEOM calls us back via
1086 * ndadiskgonecb(), telling us that our provider has been freed.
1087 */
1088 if (cam_periph_acquire(periph) == 0)
1089 disk_create(softc->disk, DISK_VERSION);
1090
1091 cam_periph_release_locked(periph);
1092 return(CAM_REQ_CMP);
1093 }
1094
1095 static void
ndastart(struct cam_periph * periph,union ccb * start_ccb)1096 ndastart(struct cam_periph *periph, union ccb *start_ccb)
1097 {
1098 struct nda_softc *softc = (struct nda_softc *)periph->softc;
1099 struct ccb_nvmeio *nvmeio = &start_ccb->nvmeio;
1100
1101 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart\n"));
1102
1103 switch (softc->state) {
1104 case NDA_STATE_NORMAL:
1105 {
1106 struct bio *bp;
1107
1108 bp = cam_iosched_next_bio(softc->cam_iosched);
1109 CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart: bio %p\n", bp));
1110 if (bp == NULL) {
1111 xpt_release_ccb(start_ccb);
1112 break;
1113 }
1114
1115 switch (bp->bio_cmd) {
1116 case BIO_WRITE:
1117 softc->flags |= NDA_FLAG_DIRTY;
1118 /* FALLTHROUGH */
1119 case BIO_READ:
1120 {
1121 #ifdef CAM_TEST_FAILURE
1122 int fail = 0;
1123
1124 /*
1125 * Support the failure ioctls. If the command is a
1126 * read, and there are pending forced read errors, or
1127 * if a write and pending write errors, then fail this
1128 * operation with EIO. This is useful for testing
1129 * purposes. Also, support having every Nth read fail.
1130 *
1131 * This is a rather blunt tool.
1132 */
1133 if (bp->bio_cmd == BIO_READ) {
1134 if (softc->force_read_error) {
1135 softc->force_read_error--;
1136 fail = 1;
1137 }
1138 if (softc->periodic_read_error > 0) {
1139 if (++softc->periodic_read_count >=
1140 softc->periodic_read_error) {
1141 softc->periodic_read_count = 0;
1142 fail = 1;
1143 }
1144 }
1145 } else {
1146 if (softc->force_write_error) {
1147 softc->force_write_error--;
1148 fail = 1;
1149 }
1150 }
1151 if (fail) {
1152 biofinish(bp, NULL, EIO);
1153 xpt_release_ccb(start_ccb);
1154 ndaschedule(periph);
1155 return;
1156 }
1157 #endif
1158 KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
1159 round_page(bp->bio_bcount + bp->bio_ma_offset) /
1160 PAGE_SIZE == bp->bio_ma_n,
1161 ("Short bio %p", bp));
1162 nda_nvme_rw_bio(softc, &start_ccb->nvmeio, bp, bp->bio_cmd == BIO_READ ?
1163 NVME_OPC_READ : NVME_OPC_WRITE);
1164 break;
1165 }
1166 case BIO_DELETE:
1167 {
1168 struct nvme_dsm_range *dsm_range, *dsm_end;
1169 struct nda_trim_request *trim;
1170 struct bio *bp1;
1171 int ents;
1172 uint32_t totalcount = 0, ranges = 0;
1173
1174 trim = malloc(sizeof(*trim), M_NVMEDA, M_ZERO | M_NOWAIT);
1175 if (trim == NULL) {
1176 /*
1177 * We have to drop the periph lock when
1178 * returning ENOMEM. g_io_deliver treats these
1179 * request differently and will recursively call
1180 * the start routine which causes us to get into
1181 * ndastrategy with the periph lock held,
1182 * leading to a panic when its acquired again.
1183 */
1184 cam_periph_unlock(periph);
1185 biofinish(bp, NULL, ENOMEM);
1186 cam_periph_lock(periph);
1187 xpt_release_ccb(start_ccb);
1188 ndaschedule(periph);
1189 return;
1190 }
1191 TAILQ_INIT(&trim->bps);
1192 bp1 = bp;
1193 ents = min(nitems(trim->dsm), nda_max_trim_entries);
1194 ents = max(ents, 1);
1195 dsm_range = trim->dsm;
1196 dsm_end = dsm_range + ents;
1197 do {
1198 TAILQ_INSERT_TAIL(&trim->bps, bp1, bio_queue);
1199 dsm_range->length =
1200 htole32(bp1->bio_bcount / softc->disk->d_sectorsize);
1201 dsm_range->starting_lba =
1202 htole64(bp1->bio_offset / softc->disk->d_sectorsize);
1203 ranges++;
1204 totalcount += dsm_range->length;
1205 dsm_range++;
1206 if (dsm_range >= dsm_end)
1207 break;
1208 bp1 = cam_iosched_next_trim(softc->cam_iosched);
1209 /* XXX -- Could collapse adjacent ranges, but we don't for now */
1210 /* XXX -- Could limit based on total payload size */
1211 } while (bp1 != NULL);
1212 start_ccb->ccb_trim = trim;
1213 nda_nvme_trim(softc, &start_ccb->nvmeio, trim->dsm,
1214 dsm_range - trim->dsm);
1215 start_ccb->ccb_state = NDA_CCB_TRIM;
1216 softc->trim_count++;
1217 softc->trim_ranges += ranges;
1218 softc->trim_lbas += totalcount;
1219 /*
1220 * Note: We can have multiple TRIMs in flight, so we don't call
1221 * cam_iosched_submit_trim(softc->cam_iosched);
1222 * since that forces the I/O scheduler to only schedule one at a time.
1223 * On NVMe drives, this is a performance disaster.
1224 */
1225 goto out;
1226 }
1227 case BIO_FLUSH:
1228 nda_nvme_flush(softc, nvmeio);
1229 break;
1230 default:
1231 biofinish(bp, NULL, EOPNOTSUPP);
1232 xpt_release_ccb(start_ccb);
1233 ndaschedule(periph);
1234 return;
1235 }
1236 start_ccb->ccb_state = NDA_CCB_BUFFER_IO;
1237 start_ccb->ccb_bp = bp;
1238 out:
1239 start_ccb->ccb_h.flags |= CAM_UNLOCKED;
1240 softc->outstanding_cmds++;
1241 softc->refcount++; /* For submission only */
1242 cam_periph_unlock(periph);
1243 xpt_action(start_ccb);
1244 cam_periph_lock(periph);
1245 softc->refcount--; /* Submission done */
1246
1247 /* May have more work to do, so ensure we stay scheduled */
1248 ndaschedule(periph);
1249 break;
1250 }
1251 }
1252 }
1253
1254 static void
ndadone(struct cam_periph * periph,union ccb * done_ccb)1255 ndadone(struct cam_periph *periph, union ccb *done_ccb)
1256 {
1257 struct nda_softc *softc;
1258 struct ccb_nvmeio *nvmeio = &done_ccb->nvmeio;
1259 struct cam_path *path;
1260 int state;
1261
1262 softc = (struct nda_softc *)periph->softc;
1263 path = done_ccb->ccb_h.path;
1264
1265 CAM_DEBUG(path, CAM_DEBUG_TRACE, ("ndadone\n"));
1266
1267 state = nvmeio->ccb_state & NDA_CCB_TYPE_MASK;
1268 switch (state) {
1269 case NDA_CCB_BUFFER_IO:
1270 case NDA_CCB_TRIM:
1271 {
1272 int error;
1273
1274 cam_periph_lock(periph);
1275 if ((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1276 error = ndaerror(done_ccb, 0, 0);
1277 if (error == ERESTART) {
1278 /* A retry was scheduled, so just return. */
1279 cam_periph_unlock(periph);
1280 return;
1281 }
1282 if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1283 cam_release_devq(path,
1284 /*relsim_flags*/0,
1285 /*reduction*/0,
1286 /*timeout*/0,
1287 /*getcount_only*/0);
1288 } else {
1289 if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1290 panic("REQ_CMP with QFRZN");
1291 error = 0;
1292 }
1293 if (state == NDA_CCB_BUFFER_IO) {
1294 struct bio *bp;
1295
1296 bp = (struct bio *)done_ccb->ccb_bp;
1297 bp->bio_error = error;
1298 if (error != 0) {
1299 bp->bio_resid = bp->bio_bcount;
1300 bp->bio_flags |= BIO_ERROR;
1301 } else {
1302 bp->bio_resid = 0;
1303 }
1304 softc->outstanding_cmds--;
1305
1306 /*
1307 * We need to call cam_iosched before we call biodone so that we
1308 * don't measure any activity that happens in the completion
1309 * routine, which in the case of sendfile can be quite
1310 * extensive.
1311 */
1312 cam_iosched_bio_complete(softc->cam_iosched, bp, done_ccb);
1313 xpt_release_ccb(done_ccb);
1314 ndaschedule(periph);
1315 cam_periph_unlock(periph);
1316 biodone(bp);
1317 } else { /* state == NDA_CCB_TRIM */
1318 struct nda_trim_request *trim;
1319 struct bio *bp1, *bp2;
1320 TAILQ_HEAD(, bio) queue;
1321
1322 trim = nvmeio->ccb_trim;
1323 TAILQ_INIT(&queue);
1324 TAILQ_CONCAT(&queue, &trim->bps, bio_queue);
1325 free(trim, M_NVMEDA);
1326
1327 /*
1328 * Since we can have multiple trims in flight, we don't
1329 * need to call this here.
1330 * cam_iosched_trim_done(softc->cam_iosched);
1331 */
1332 /*
1333 * The the I/O scheduler that we're finishing the I/O
1334 * so we can keep book. The first one we pass in the CCB
1335 * which has the timing information. The rest we pass in NULL
1336 * so we can keep proper counts.
1337 */
1338 bp1 = TAILQ_FIRST(&queue);
1339 cam_iosched_bio_complete(softc->cam_iosched, bp1, done_ccb);
1340 xpt_release_ccb(done_ccb);
1341 softc->outstanding_cmds--;
1342 ndaschedule(periph);
1343 cam_periph_unlock(periph);
1344 while ((bp2 = TAILQ_FIRST(&queue)) != NULL) {
1345 TAILQ_REMOVE(&queue, bp2, bio_queue);
1346 bp2->bio_error = error;
1347 if (error != 0) {
1348 bp2->bio_flags |= BIO_ERROR;
1349 bp2->bio_resid = bp1->bio_bcount;
1350 } else
1351 bp2->bio_resid = 0;
1352 if (bp1 != bp2)
1353 cam_iosched_bio_complete(softc->cam_iosched, bp2, NULL);
1354 biodone(bp2);
1355 }
1356 }
1357 return;
1358 }
1359 case NDA_CCB_DUMP:
1360 /* No-op. We're polling */
1361 return;
1362 case NDA_CCB_PASS:
1363 /* NVME_PASSTHROUGH_CMD runs this CCB and releases it */
1364 return;
1365 default:
1366 break;
1367 }
1368 xpt_release_ccb(done_ccb);
1369 }
1370
1371 static int
ndaerror(union ccb * ccb,uint32_t cam_flags,uint32_t sense_flags)1372 ndaerror(union ccb *ccb, uint32_t cam_flags, uint32_t sense_flags)
1373 {
1374 #ifdef CAM_IO_STATS
1375 struct nda_softc *softc;
1376 struct cam_periph *periph;
1377
1378 periph = xpt_path_periph(ccb->ccb_h.path);
1379 softc = (struct nda_softc *)periph->softc;
1380 #endif
1381
1382 switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1383 case CAM_CMD_TIMEOUT:
1384 #ifdef CAM_IO_STATS
1385 softc->timeouts++;
1386 #endif
1387 break;
1388 case CAM_REQ_CMP_ERR:
1389 case CAM_NVME_STATUS_ERROR:
1390 #ifdef CAM_IO_STATS
1391 softc->errors++;
1392 #endif
1393 break;
1394 default:
1395 break;
1396 }
1397
1398 return(cam_periph_error(ccb, cam_flags, sense_flags));
1399 }
1400
1401 /*
1402 * Step through all NDA peripheral drivers, and if the device is still open,
1403 * sync the disk cache to physical media.
1404 */
1405 static void
ndaflush(void)1406 ndaflush(void)
1407 {
1408 struct cam_periph *periph;
1409 struct nda_softc *softc;
1410 union ccb *ccb;
1411 int error;
1412
1413 CAM_PERIPH_FOREACH(periph, &ndadriver) {
1414 softc = (struct nda_softc *)periph->softc;
1415
1416 if (SCHEDULER_STOPPED()) {
1417 /*
1418 * If we panicked with the lock held or the periph is not
1419 * open, do not recurse. Otherwise, call ndadump since
1420 * that avoids the sleeping cam_periph_getccb does if no
1421 * CCBs are available.
1422 */
1423 if (!cam_periph_owned(periph) &&
1424 (softc->flags & NDA_FLAG_OPEN)) {
1425 ndadump(softc->disk, NULL, 0, 0);
1426 }
1427 continue;
1428 }
1429
1430 /*
1431 * We only sync the cache if the drive is still open
1432 */
1433 cam_periph_lock(periph);
1434 if ((softc->flags & NDA_FLAG_OPEN) == 0) {
1435 cam_periph_unlock(periph);
1436 continue;
1437 }
1438
1439 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
1440 nda_nvme_flush(softc, &ccb->nvmeio);
1441 error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
1442 /*sense_flags*/ SF_NO_RECOVERY | SF_NO_RETRY,
1443 softc->disk->d_devstat);
1444 if (error != 0)
1445 xpt_print(periph->path, "Synchronize cache failed\n");
1446 xpt_release_ccb(ccb);
1447 cam_periph_unlock(periph);
1448 }
1449 }
1450
1451 static void
ndashutdown(void * arg,int howto)1452 ndashutdown(void *arg, int howto)
1453 {
1454
1455 if ((howto & RB_NOSYNC) != 0)
1456 return;
1457
1458 ndaflush();
1459 }
1460
1461 static void
ndasuspend(void * arg,enum power_stype stype)1462 ndasuspend(void *arg, enum power_stype stype)
1463 {
1464
1465 ndaflush();
1466 }
1467