d/freebsd-skq
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add controller reset capability to nvme(4) and ability to explicitly

Browse Source 
invoke it from nvmecontrol(8).

Controller reset will be performed in cases where I/O are repeatedly
timing out, the controller reports an unrecoverable condition, or
when explicitly requested via IOCTL or an nvme consumer.  Since the
controller may be in such a state where it cannot even process queue
deletion requests, we will perform a controller reset without trying
to clean up anything on the controller first.

Sponsored by:	Intel
Reviewed by:	carl
This commit is contained in:
jimharris 2013-03-26 19:50:46 +00:00
parent bd33256583
commit 93fd264895
8 changed files with 265 additions and 101 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
sbin/nvmecontrol/nvmecontrol.8
View File

@ -33,7 +33,7 @@
.\"
.\" $FreeBSD$
.\"
.Dd September 17, 2012
.Dd March 26, 2013
.Dt NVMECONTROL 8
.Os
.Sh NAME
@ -54,7 +54,10 @@
.Op Fl p
.Aq Fl s Ar size_in_bytes
.Aq Fl t Ar time_in_sec
.Aq device id
.Aq namespace id
.Nm
.Ic reset
.Aq controller id
.Sh DESCRIPTION
NVM Express (NVMe) is a storage protocol standard, for SSDs and other
high-speed storage devices over PCI Express.
@ -62,6 +65,7 @@ high-speed storage devices over PCI Express.
.Dl nvmecontrol devlist
.Pp
Display a list of NVMe controllers and namespaces along with their device nodes.
.Pp
.Dl nvmecontrol identify nvme0
.Pp
Display a human-readable summary of the nvme0 IDENTIFY_CONTROLLER data.
@ -76,6 +80,10 @@ Display a hexadecimal dump of the nvme0 IDENTIFY_NAMESPACE data for namespace
Run a performance test on nvme0ns1 using 32 kernel threads for 30 seconds. Each
thread will issue a single 512 byte read command. Results are printed to
stdout when 30 seconds expires.
.Pp
.Dl nvmecontrol reset nvme0
.Pp
Perform a controller-level reset of the nvme0 controller.
.Sh AUTHORS
.An -nosplit
.Nm

41
sbin/nvmecontrol/nvmecontrol.c
View File

@ -56,6 +56,9 @@ __FBSDID("$FreeBSD$");
" <-i intr|wait> [-f refthread] [-p]\n" \
" <namespace id>\n"
#define RESET_USAGE \
" nvmecontrol reset <controller id>\n"
static void perftest_usage(void);
static void
@ -64,6 +67,7 @@ usage(void)
fprintf(stderr, "usage:\n");
fprintf(stderr, DEVLIST_USAGE);
fprintf(stderr, IDENTIFY_USAGE);
fprintf(stderr, RESET_USAGE);
fprintf(stderr, PERFTEST_USAGE);
exit(EX_USAGE);
}
@ -580,6 +584,41 @@ perftest(int argc, char *argv[])
exit(EX_OK);
}
static void
reset_ctrlr(int argc, char *argv[])
{
struct stat devstat;
char path[64];
int ch, fd;
while ((ch = getopt(argc, argv, "")) != -1) {
switch ((char)ch) {
default:
usage();
}
}
sprintf(path, "/dev/%s", argv[optind]);
if (stat(path, &devstat) != 0) {
printf("Invalid device node '%s'.\n", path);
exit(EX_IOERR);
}
fd = open(path, O_RDWR);
if (fd < 0) {
printf("Could not open %s.\n", path);
exit(EX_NOPERM);
}
if (ioctl(fd, NVME_RESET_CONTROLLER) == -1) {
printf("ioctl to %s failed.\n", path);
exit(EX_IOERR);
}
exit(EX_OK);
}
int
main(int argc, char *argv[])
{
@ -593,6 +632,8 @@ main(int argc, char *argv[])
identify(argc-1, &argv[1]);
else if (strcmp(argv[1], "perftest") == 0)
perftest(argc-1, &argv[1]);
else if (strcmp(argv[1], "reset") == 0)
reset_ctrlr(argc-1, &argv[1]);
usage();

6
sys/dev/nvme/nvme.c
View File

@ -255,7 +255,7 @@ nvme_payload_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
}
}
nvme_qpair_submit_cmd(tr->qpair, tr);
nvme_qpair_submit_tracker(tr->qpair, tr);
}
static int
@ -274,11 +274,11 @@ nvme_attach(device_t dev)
* to cc.en==0. This is because we don't really know what status
* the controller was left in when boot handed off to OS.
*/
status = nvme_ctrlr_reset(ctrlr);
status = nvme_ctrlr_hw_reset(ctrlr);
if (status != 0)
return (status);
status = nvme_ctrlr_reset(ctrlr);
status = nvme_ctrlr_hw_reset(ctrlr);
if (status != 0)
return (status);

1
sys/dev/nvme/nvme.h
View File

@ -37,6 +37,7 @@
#define NVME_IDENTIFY_NAMESPACE _IOR('n', 1, struct nvme_namespace_data)
#define NVME_IO_TEST _IOWR('n', 2, struct nvme_io_test)
#define NVME_BIO_TEST _IOWR('n', 4, struct nvme_io_test)
#define NVME_RESET_CONTROLLER _IO('n', 5)
/*
* Use to mark a command to apply to all namespaces, or to retrieve global

64
sys/dev/nvme/nvme_ctrlr.c
View File

@ -405,13 +405,31 @@ nvme_ctrlr_enable(struct nvme_controller *ctrlr)
}
int
nvme_ctrlr_reset(struct nvme_controller *ctrlr)
nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
{
int i;
nvme_admin_qpair_disable(&ctrlr->adminq);
for (i = 0; i < ctrlr->num_io_queues; i++)
nvme_io_qpair_disable(&ctrlr->ioq[i]);
DELAY(100*1000);
nvme_ctrlr_disable(ctrlr);
return (nvme_ctrlr_enable(ctrlr));
}
void
nvme_ctrlr_reset(struct nvme_controller *ctrlr)
{
int status;
status = nvme_ctrlr_hw_reset(ctrlr);
DELAY(100*1000);
if (status == 0)
nvme_ctrlr_start(ctrlr);
}
static int
nvme_ctrlr_identify(struct nvme_controller *ctrlr)
{
@ -626,6 +644,9 @@ void
nvme_ctrlr_start(void *ctrlr_arg)
{
struct nvme_controller *ctrlr = ctrlr_arg;
int i;
nvme_admin_qpair_enable(&ctrlr->adminq);
if (nvme_ctrlr_identify(ctrlr) != 0)
goto err;
@ -642,16 +663,26 @@ nvme_ctrlr_start(void *ctrlr_arg)
nvme_ctrlr_configure_aer(ctrlr);
nvme_ctrlr_configure_int_coalescing(ctrlr);
for (i = 0; i < ctrlr->num_io_queues; i++)
nvme_io_qpair_enable(&ctrlr->ioq[i]);
ctrlr->is_started = TRUE;
err:
/*
* Initialize sysctls, even if controller failed to start, to
* assist with debugging admin queue pair.
*/
nvme_sysctl_initialize_ctrlr(ctrlr);
config_intrhook_disestablish(&ctrlr->config_hook);
if (ctrlr->num_start_attempts == 0) {
/*
* Initialize sysctls, even if controller failed to start, to
* assist with debugging admin queue pair. Only run this
* code on the initial start attempt though, and not
* subsequent start attempts due to controller-level resets.
*
*/
nvme_sysctl_initialize_ctrlr(ctrlr);
config_intrhook_disestablish(&ctrlr->config_hook);
}
ctrlr->num_start_attempts++;
}
static void
@ -730,6 +761,9 @@ nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
return (ENXIO);
memcpy(arg, &ctrlr->cdata, sizeof(ctrlr->cdata));
break;
case NVME_RESET_CONTROLLER:
nvme_ctrlr_reset(ctrlr);
break;
default:
return (ENOTTY);
}
@ -752,6 +786,7 @@ nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
ctrlr->dev = dev;
ctrlr->is_started = FALSE;
ctrlr->num_start_attempts = 0;
status = nvme_ctrlr_allocate_bar(ctrlr);
@ -835,14 +870,10 @@ nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
void
nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
{
struct nvme_namespace *ns;
int i;
for (i = 0; i < NVME_MAX_NAMESPACES; i++) {
ns = &ctrlr->ns[i];
if (ns->cdev)
destroy_dev(ns->cdev);
}
for (i = 0; i < NVME_MAX_NAMESPACES; i++)
nvme_ns_destruct(&ctrlr->ns[i]);
if (ctrlr->cdev)
destroy_dev(ctrlr->cdev);
@ -853,13 +884,6 @@ nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
free(ctrlr->ioq, M_NVME);
/* Manually abort outstanding async event requests. */
for (i = 0; i < ctrlr->num_aers; i++) {
nvme_qpair_manual_abort_request(&ctrlr->adminq,
ctrlr->aer[i].req, NVME_SCT_GENERIC,
NVME_SC_ABORTED_SQ_DELETION, FALSE);
}
nvme_admin_qpair_destroy(&ctrlr->adminq);
if (ctrlr->resource != NULL) {

17
sys/dev/nvme/nvme_ns.c
View File

@ -345,6 +345,13 @@ nvme_ns_construct(struct nvme_namespace *ns, uint16_t id,
if (ctrlr->cdata.vwc.present)
ns->flags |= NVME_NS_FLUSH_SUPPORTED;
/*
* cdev may have already been created, if we are reconstructing the
* namespace after a controller-level reset.
*/
if (ns->cdev != NULL)
return (0);
/*
* MAKEDEV_ETERNAL was added in r210923, for cdevs that will never
* be destroyed. This avoids refcounting on the cdev object.
@ -361,9 +368,15 @@ nvme_ns_construct(struct nvme_namespace *ns, uint16_t id,
device_get_unit(ctrlr->dev), ns->id);
#endif
if (ns->cdev) {
if (ns->cdev != NULL)
ns->cdev->si_drv1 = ns;
}
return (0);
}
void nvme_ns_destruct(struct nvme_namespace *ns)
{
if (ns->cdev != NULL)
destroy_dev(ns->cdev);
}

18
sys/dev/nvme/nvme_private.h
View File

@ -180,6 +180,8 @@ struct nvme_qpair {
struct nvme_tracker **act_tr;
boolean_t is_enabled;
struct mtx lock __aligned(CACHE_LINE_SIZE);
} __aligned(CACHE_LINE_SIZE);
@ -233,6 +235,7 @@ struct nvme_controller {
struct intr_config_hook config_hook;
uint32_t ns_identified;
uint32_t queues_created;
uint32_t num_start_attempts;
/* For shared legacy interrupt. */
int rid;
@ -361,7 +364,8 @@ void nvme_payload_map_uio(void *arg, bus_dma_segment_t *seg, int nseg,
int nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev);
void nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev);
int nvme_ctrlr_reset(struct nvme_controller *ctrlr);
int nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr);
void nvme_ctrlr_reset(struct nvme_controller *ctrlr);
/* ctrlr defined as void * to allow use with config_intrhook. */
void nvme_ctrlr_start(void *ctrlr_arg);
void nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
@ -373,21 +377,23 @@ void nvme_qpair_construct(struct nvme_qpair *qpair, uint32_t id,
uint16_t vector, uint32_t num_entries,
uint32_t num_trackers, uint32_t max_xfer_size,
struct nvme_controller *ctrlr);
void nvme_qpair_submit_cmd(struct nvme_qpair *qpair,
struct nvme_tracker *tr);
void nvme_qpair_submit_tracker(struct nvme_qpair *qpair,
struct nvme_tracker *tr);
void nvme_qpair_process_completions(struct nvme_qpair *qpair);
void nvme_qpair_submit_request(struct nvme_qpair *qpair,
struct nvme_request *req);
void nvme_qpair_manual_abort_request(struct nvme_qpair *qpair,
struct nvme_request *req, uint32_t sct,
uint32_t sc, boolean_t print_on_error);
void nvme_admin_qpair_enable(struct nvme_qpair *qpair);
void nvme_admin_qpair_disable(struct nvme_qpair *qpair);
void nvme_admin_qpair_destroy(struct nvme_qpair *qpair);
void nvme_io_qpair_enable(struct nvme_qpair *qpair);
void nvme_io_qpair_disable(struct nvme_qpair *qpair);
void nvme_io_qpair_destroy(struct nvme_qpair *qpair);
int nvme_ns_construct(struct nvme_namespace *ns, uint16_t id,
struct nvme_controller *ctrlr);
void nvme_ns_destruct(struct nvme_namespace *ns);
int nvme_ns_physio(struct cdev *dev, struct uio *uio, int ioflag);

207
sys/dev/nvme/nvme_qpair.c
View File

@ -87,23 +87,6 @@ nvme_completion_is_retry(const struct nvme_completion *cpl)
}
}
static struct nvme_tracker *
nvme_qpair_find_tracker(struct nvme_qpair *qpair, struct nvme_request *req)
{
struct nvme_tracker *tr;
uint32_t i;
KASSERT(req != NULL, ("%s: called with NULL req\n", __func__));
for (i = 0; i < qpair->num_entries; ++i) {
tr = qpair->act_tr[i];
if (tr != NULL && tr->req == req)
return (tr);
}
return (NULL);
}
static void
nvme_qpair_construct_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr,
uint16_t cid)
@ -147,7 +130,7 @@ nvme_qpair_complete_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr,
callout_stop(&tr->timer);
if (retry)
nvme_qpair_submit_cmd(qpair, tr);
nvme_qpair_submit_tracker(qpair, tr);
else {
if (req->payload_size > 0 || req->uio != NULL)
bus_dmamap_unload(qpair->dma_tag,
@ -169,6 +152,21 @@ nvme_qpair_complete_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr,
mtx_unlock(&qpair->lock);
}
static void
nvme_qpair_manual_complete_tracker(struct nvme_qpair *qpair,
struct nvme_tracker *tr, uint32_t sct, uint32_t sc,
boolean_t print_on_error)
{
struct nvme_completion cpl;
memset(&cpl, 0, sizeof(cpl));
cpl.sqid = qpair->id;
cpl.cid = tr->cid;
cpl.sf_sct = sct;
cpl.sf_sc = sc;
nvme_qpair_complete_tracker(qpair, tr, &cpl, print_on_error);
}
void
nvme_qpair_process_completions(struct nvme_qpair *qpair)
{
@ -177,6 +175,15 @@ nvme_qpair_process_completions(struct nvme_qpair *qpair)
qpair->num_intr_handler_calls++;
if (!qpair->is_enabled)
/*
* qpair is not enabled, likely because a controller reset is
* is in progress. Ignore the interrupt - any I/O that was
* associated with this interrupt will get retried when the
* reset is complete.
*/
return;
while (1) {
cpl = &qpair->cpl[qpair->cq_head];
@ -236,15 +243,6 @@ nvme_qpair_construct(struct nvme_qpair *qpair, uint32_t id,
qpair->max_xfer_size = max_xfer_size;
qpair->ctrlr = ctrlr;
/*
* First time through the completion queue, HW will set phase
* bit on completions to 1. So set this to 1 here, indicating
* we're looking for a 1 to know which entries have completed.
* we'll toggle the bit each time when the completion queue
* rolls over.
*/
qpair->phase = 1;
if (ctrlr->msix_enabled) {
/*
@ -271,7 +269,6 @@ nvme_qpair_construct(struct nvme_qpair *qpair, uint32_t id,
qpair->num_cmds = 0;
qpair->num_intr_handler_calls = 0;
qpair->sq_head = qpair->sq_tail = qpair->cq_head = 0;
/* TODO: error checking on contigmalloc, bus_dmamap_load calls */
qpair->cmd = contigmalloc(qpair->num_entries *
@ -341,10 +338,30 @@ nvme_qpair_destroy(struct nvme_qpair *qpair)
}
}
static void
nvme_admin_qpair_abort_aers(struct nvme_qpair *qpair)
{
struct nvme_tracker *tr;
tr = TAILQ_FIRST(&qpair->outstanding_tr);
while (tr != NULL) {
if (tr->req->cmd.opc == NVME_OPC_ASYNC_EVENT_REQUEST) {
nvme_qpair_manual_complete_tracker(qpair, tr,
NVME_SCT_GENERIC, NVME_SC_ABORTED_SQ_DELETION,
FALSE);
tr = TAILQ_FIRST(&qpair->outstanding_tr);
} else {
tr = TAILQ_NEXT(tr, tailq);
}
}
}
void
nvme_admin_qpair_destroy(struct nvme_qpair *qpair)
{
nvme_admin_qpair_abort_aers(qpair);
/*
* For NVMe, you don't send delete queue commands for the admin
* queue, so we just need to unload and free the cmd and cpl memory.
@ -412,39 +429,6 @@ nvme_io_qpair_destroy(struct nvme_qpair *qpair)
}
}
static void
nvme_qpair_manual_abort_tracker(struct nvme_qpair *qpair,
struct nvme_tracker *tr, uint32_t sct, uint32_t sc,
boolean_t print_on_error)
{
struct nvme_completion cpl;
memset(&cpl, 0, sizeof(cpl));
cpl.sqid = qpair->id;
cpl.cid = tr->cid;
cpl.sf_sct = sct;
cpl.sf_sc = sc;
nvme_qpair_complete_tracker(qpair, tr, &cpl, print_on_error);
}
void
nvme_qpair_manual_abort_request(struct nvme_qpair *qpair,
struct nvme_request *req, uint32_t sct, uint32_t sc,
boolean_t print_on_error)
{
struct nvme_tracker *tr;
tr = nvme_qpair_find_tracker(qpair, req);
if (tr == NULL) {
printf("%s: request not found\n", __func__);
nvme_dump_command(&req->cmd);
return;
}
nvme_qpair_manual_abort_tracker(qpair, tr, sct, sc, print_on_error);
}
static void
nvme_abort_complete(void *arg, const struct nvme_completion *status)
{
@ -463,7 +447,7 @@ nvme_abort_complete(void *arg, const struct nvme_completion *status)
* status, and then complete the I/O's tracker manually.
*/
printf("abort command failed, aborting command manually\n");
nvme_qpair_manual_abort_tracker(tr->qpair, tr,
nvme_qpair_manual_complete_tracker(tr->qpair, tr,
NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, TRUE);
}
}
@ -478,10 +462,12 @@ nvme_timeout(void *arg)
}
void
nvme_qpair_submit_cmd(struct nvme_qpair *qpair, struct nvme_tracker *tr)
nvme_qpair_submit_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr)
{
struct nvme_request *req;
mtx_assert(&qpair->lock, MA_OWNED);
req = tr->req;
req->cmd.cid = tr->cid;
qpair->act_tr[tr->cid] = tr;
@ -517,11 +503,14 @@ _nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
tr = TAILQ_FIRST(&qpair->free_tr);
if (tr == NULL) {
if (tr == NULL || !qpair->is_enabled) {
/*
* No tracker is available. Put the request on the qpair's
* request queue to be processed when a tracker frees up
* via a command completion.
* No tracker is available, or the qpair is disabled due to
* an in-progress controller-level reset.
*
* Put the request on the qpair's request queue to be processed
* when a tracker frees up via a command completion or when
* the controller reset is completed.
*/
STAILQ_INSERT_TAIL(&qpair->queued_req, req, stailq);
return;
@ -540,7 +529,7 @@ _nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
if (err != 0)
panic("bus_dmamap_load returned non-zero!\n");
} else
nvme_qpair_submit_cmd(tr->qpair, tr);
nvme_qpair_submit_tracker(tr->qpair, tr);
} else {
err = bus_dmamap_load_uio(tr->qpair->dma_tag,
tr->payload_dma_map, req->uio,
@ -558,3 +547,85 @@ nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
_nvme_qpair_submit_request(qpair, req);
mtx_unlock(&qpair->lock);
}
static void
nvme_qpair_enable(struct nvme_qpair *qpair)
{
qpair->is_enabled = TRUE;
qpair->sq_head = qpair->sq_tail = qpair->cq_head = 0;
/*
* First time through the completion queue, HW will set phase
* bit on completions to 1. So set this to 1 here, indicating
* we're looking for a 1 to know which entries have completed.
* we'll toggle the bit each time when the completion queue
* rolls over.
*/
qpair->phase = 1;
memset(qpair->cmd, 0,
qpair->num_entries * sizeof(struct nvme_command));
memset(qpair->cpl, 0,
qpair->num_entries * sizeof(struct nvme_completion));
}
void
nvme_admin_qpair_enable(struct nvme_qpair *qpair)
{
nvme_qpair_enable(qpair);
}
void
nvme_io_qpair_enable(struct nvme_qpair *qpair)
{
STAILQ_HEAD(, nvme_request) temp;
struct nvme_tracker *tr;
struct nvme_request *req;
mtx_lock(&qpair->lock);
nvme_qpair_enable(qpair);
TAILQ_FOREACH(tr, &qpair->outstanding_tr, tailq)
nvme_qpair_submit_tracker(qpair, tr);
STAILQ_INIT(&temp);
STAILQ_SWAP(&qpair->queued_req, &temp, nvme_request);
while (!STAILQ_EMPTY(&temp)) {
req = STAILQ_FIRST(&temp);
STAILQ_REMOVE_HEAD(&temp, stailq);
_nvme_qpair_submit_request(qpair, req);
}
mtx_unlock(&qpair->lock);
}
static void
nvme_qpair_disable(struct nvme_qpair *qpair)
{
struct nvme_tracker *tr;
qpair->is_enabled = FALSE;
mtx_lock(&qpair->lock);
TAILQ_FOREACH(tr, &qpair->outstanding_tr, tailq)
callout_stop(&tr->timer);
mtx_unlock(&qpair->lock);
}
void
nvme_admin_qpair_disable(struct nvme_qpair *qpair)
{
nvme_qpair_disable(qpair);
nvme_admin_qpair_abort_aers(qpair);
}
void
nvme_io_qpair_disable(struct nvme_qpair *qpair)
{
nvme_qpair_disable(qpair);
}