Multiple users have reported device initialization failure due the driver
not receiving legacy PCI interrupts. This is not unique to any particular
controller, but has been observed on multiple platforms.

There have been no issues reported or observed when with message signaled
interrupts, so this patch attempts to use MSI-x during initialization,
falling back to MSI. If that fails, legacy would become the default.

The setup_io_queues error handling had to change as a result: the admin
queue's msix_entry used to be initialized to the legacy IRQ. The case
where nr_io_queues is 0 would fail request_irq when setting up the admin
queue's interrupt since re-enabling MSI-x fails with 0 vectors, leaving
the admin queue's msix_entry invalid. Instead, return success immediately.
Reported-by: NTim Muhlemmer <muhlemmer@gmail.com>
Reported-by: NJon Derrick <jonathan.derrick@intel.com>
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

This fixes a scenario where device is present and being reset, but a
request to unbind the driver occurs.

A previous patch series addressing a device failure removal scenario
flushed reset_work after controller disable to unblock reset_work waiting
on a completion that wouldn't occur. This isn't safe as-is. The broken
scenario can potentially be induced with:

  modprobe nvme && modprobe -r nvme

To fix, the reset work is flushed immediately after setting the controller
removing flag, and any subsequent reset will not proceed with controller
initialization if the flag is set.

The controller status must be polled while active, so the watchdog timer
is also left active until the controller is disabled to cleanup requests
that may be stuck during namespace removal.

[Fixes: ff23a2a1]
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NJens Axboe <axboe@fb.com>

Make sure the CQE phase (validity) is read before the rest of the
structure. The phase bit is the highest address and the CQE
read will happen on most platforms from lower to upper addresses
and will be done by multiple non-atomic loads. If the structure
is updated by PCI during the reads from the processor, the
processor may get a corrupted copy.

The addition of the new nvme_cqe_valid function that verifies
the validity bit also allows refactoring of the other CQE read
sequences.
Signed-off-by: NMarta Rybczynska <marta.rybczynska@kalray.eu>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Signed-off-by: NJon Derrick <jonathan.derrick@intel.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

The NVMe specification does not require discarded blocks return zeroes on
read, but provides that behavior as a possibility. Some applications more
efficiently use an SSD if reads on discarded blocks were deterministically
zero, based on the "discard_zeroes_data" queue attribute.

There is no specification defined way to determine device behavior on
discarded blocks, so the driver always left the queue setting disabled. We
can only know behavior based on individual device models, so this patch
adds a flag to the NVMe "quirk" list that vendors may set if they know
their controller works that way. The patch also sets the new flag for one
such known device.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Suggested-by: NArtur Paszkiewicz <artur.paszkiewicz@intel.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NMartin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

This moves failed queue handling out of the namespace removal path and
into the reset failure path, fixing a hanging condition if the controller
fails or link down during del_gendisk. Previously the driver had to see
the controller as degraded prior to calling del_gendisk to setup the
queues to fail. But, if the controller happened to fail after this,
there was no task to end outstanding requests.

On failure, all namespace states are set to dead. This has capacity
revalidate to 0, and ends all new requests with error status.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

A reset failure schedules the device to unbind from the driver through
the pci driver's remove. This cleans up all intialization, so there is
no need to duplicate the potentially racy cleanup.

To help understand why a reset failed, the status is logged with the
existing warning message.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

This patch makes nvme namespace removal lockless. It is up to the caller
to ensure no active namespace scanning is occuring. To ensure no scan
work occurs, the nvme pci driver adds a removing state to the controller
device to avoid queueing scan work during removal. The work is flushed
after setting the state, so no new scan work can be queued.

The lockless removal allows the driver to cleanup a namespace
request_queue if the controller fails during removal. Previously this
could deadlock trying to acquire the namespace mutex in order to handle
such events.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

Unmapping the registers on reset or shutdown is not necessary. Keeping
the mapping simplifies reset handling.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

Both LighNVM and NVMe over Fabrics need to look at more than just the
status and result field.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NMatias Bj?rling <m@bjorling.me>
Reviewed-by: NJay Freyensee <james.p.freyensee@intel.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

The only work left in the kthread is the periodic health check for each
controller.  There is no need to run this from process context or keep
a thread context around for it, so replace it with a simpler timer.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

There is no reason to do unconditional polling of CQs per the NVMe
spec.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

Use a dedicated work item to submit async event requests instead of the
global kthread.  This simplifies the code and reduces the latencies to
resubmit a request once an even notification happened.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

We don't need to spam the kernel logs with thousands of IO cancelling
messages. We can infer all IO's are being cancelled with fewer, or
even none at all. This patch rate limits the message and uses the debug
log level as it is mainly used for testing purposes.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

A device failure or link down wouldn't have been detected during namespace
removal. This patch keeps the device in the list for polling so that the
thread may see such failure and initiate a reset. The device is removed
from the list after disable, so we can safely flush the reset work as
it can't be requeued when disable completes.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

It's possible a request may get to the driver after the nvme queue was
disabled. This has the request requeue if that happens.

Note the request is still "started" by the driver, but requeuing will
clear the start state for timeout handling.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

NVMe over Fabrics drivers are going to reuse the core,
so splits nvme.ko into 2 modules:

nvme-core.ko: the core part
nvme.ko: the PCI driver

Export symbols from nvme-core.ko.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NMing Lin <ming.l@ssi.samsung.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Split dev_list_lock into one in the core and one in the PCI driver.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NMing Lin <ming.l@ssi.samsung.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

These variables are used by PCI driver and will also be used in the
forthcoming NVMe over Fabrics drivers.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NMing Lin <ming.l@ssi.samsung.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

We don't want to be able to unload the fabric driver when we have
openened referenced to our namespaces. Thus, for each nvme_open we
take a reference on the fabric driver and put it in nvme_release.
This behavior is consistent with the scsi model.

This resolves the panic when unloading a fabric module with
mpath holders.
Signed-off-by: NSagi Grimberg <sagig@mellanox.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NIan Bakshan <ianb@mellanox.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NMing Lin <ming.l@ssi.samsung.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Having the ctrl name "nvmeX" seems much more friendly than
the underlying device name. Also, with other nvme transports
such as the soon to come nvme-loop we don't have an underlying
device so it doesn't makes sense to make up one.

In order to help matching an instance name to a pci function,
we add a info print in nvme_probe.
Signed-off-by: NSagi Grimberg <sagig@mellanox.com>
Acked-by: NKeith Busch <keith.busch@intel.com>

Manually fixed up the hunk in nvme_cancel_queue_ios().
Signed-off-by: NJens Axboe <axboe@fb.com>

This notifies blk-mq when the tag set contains a different number of
queues prior to freeing unused ones that the request queue points to.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

We don't need to shutdown a controller for a reset. A controller in a
shutdown state may take longer to become ready than one that was simply
disabled. This patch has the driver shut down a controller only if the
device is about to be powered off or being removed. When taking the
controller down for a reset reason, the controller will be disabled
instead.

Function names have been updated in this patch to reflect their changed
semantics.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

The nvme driver deletes IO queues asynchronously since this operation
may potentially take an undesirable amount of time with a large number
of queues if done serially.

The driver used to manage coordinating asynchronous deletions. This
patch simplifies that by leveraging the block layer rather than using
kthread workers and chaining more complicated callbacks.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

NVMe submits all commands through the block layer now. This means we
can let requests queue at the blk-mq hardware context since there is no
path that bypasses this anymore so we don't need to freeze the queues
anymore. The driver can simply stop the h/w queues from running during
a reset instead.

This also fixes a WARN in percpu_ref_reinit when the queue was unfrozen
with requeued requests.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

A negative status has the "do not retry" bit set, which makes it not
retryable.  Use a fake status that can potentially be retried on reset.

An aborted command's status is overridden by the timeout handler so
that it won't be retried, which is necessary to keep initialization from
getting into a reset loop.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

The tag set queue depth needs to be one less than the h/w queue depth
so we don't wrap the circular buffer. This conforms to the specification
defined "Full Queue" condition.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Nothing pci specific about them and We'll need them exported
in other transports too.
Signed-off-by: NSagi Grimberg <sagig@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

This patch fixes a lost request discovered during IO + hot removal.

The driver's pci removal deletes gendisks prior to shutting down the
controller to allow dirty data to sync. Dirty data can not be synced on
a surprise removal, though, and would potentially block indefinitely.

The driver previously had marked the queue as dying in this scenario
to prevent new requests from attempting, however it will still block
for requests that already entered the queue. This patch fixes this by
quiescing IO first, then aborting the requeued requests before deleting
disks.
Reported-by: NSujith Pandel <sujith_pandel@dell.com>
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Tested-by: NSujith Pandel <sujith_pandel@dell.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Requests enabling pcie aer support. Shuts down the controller on error
detected with io frozen state prior to requesting slot reset; resumes
controller after reset completes.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

Merge the two per-request structures in the nvme driver into a single
one.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

We need to move freeing of resources to the ->complete handler to ensure
they are also freed when we cancel the command.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Now that all commands are executed as block layer requests we can remove the
internal completion in the NVMe driver.  Note that we can simply call
blk_mq_complete_request to abort commands as the block layer will protect
against double copletions internally.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

AEN requests are different from other requests in that they don't time out
or can easily be cancelled.  Because of that we should not use the blk-mq
infrastructure but just special case them in the completion path.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

And remove the now unused nvme_submit_cmd helper.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

We'll need them in other places later.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

The number in tag_set->queue depth includes the reserved tags.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>

We don't want to allow new references to open on a device that is
removed. This ties the lifetime of these handles to the physical device's
presence rather than to the open reference count.
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@fb.com>