The driver has to end unreturned commands at some point even if the
controller has not provided a completion. The driver tried to be safe by
deleting IO queues prior to ending all unreturned commands. That should
cause the controller to internally abort inflight commands, but IO queue
deletion request does not have to be successful, so all bets are off. We
still have to make progress, so to be extra safe, this patch doesn't
clear a queue to release the dma mapping for a command until after the
pci device has been disabled.
This patch removes the special handling during device initialization
so controller recovery can be done all the time. This is possible since
initialization is not inlined with pci probe anymore.
Reported-by: Nilish Choudhury <nilesh.choudhury@oracle.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
This performs the longest parts of nvme device probe in scheduled work.
This speeds up probe significantly when multiple devices are in use.
Signed-off-by: Keith Busch <keith.busch@intel.com>
This creates a new class type for nvme devices to register their
management character devices with. This is so we do not rely on miscdev
to provide enough minors for as many nvme devices some people plan to
use. The previous limit was approximately 60 NVMe controllers, depending
on the platform and kernel. Now the limit is 1M, which ought to be enough
for anybody.
Since we have a new device class, it makes sense to attach the block
devices under this as well, so part of this patch moves the management
handle initialization prior to the namespaces discovery.
Signed-off-by: Keith Busch <keith.busch@intel.com>
The original translation created collisions on Inquiry VPD 83 for many
existing devices. Newer specifications provide other ways to translate
based on the device's version can be used to create unique identifiers.
Version 1.1 provides an EUI64 field that uniquely identifies each
namespace, and 1.2 added the longer NGUID field for the same reason.
Both follow the IEEE EUI format and readily translate to the SCSI device
identification EUI designator type 2h. For devices implementing either,
the translation will use this type, defaulting to the EUI64 8-byte type if
implemented then NGUID's 16 byte version if not. If neither are provided,
the 1.0 translation is used, and is updated to use the SCSI String format
to guarantee a unique identifier.
Knowing when to use the new fields depends on the nvme controller's
revision. The NVME_VS macro was not decoding this correctly, so that is
fixed in this patch and moved to a more appropriate place.
Since the Identify Namespace structure required an update for the NGUID
field, this patch adds the remaining new 1.2 fields to the structure.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Adds support for NVMe metadata formats and exposes block devices for
all namespaces regardless of their format. Namespace formats that are
unusable will have disk capacity set to 0, but a handle to the block
device is created to simplify device management. A namespace is not
usable when the format requires host interleave block and metadata in
single buffer, has no provisioned storage, or has better data but failed
to register with blk integrity.
The namespace has to be scanned in two phases to support separate
metadata formats. The first establishes the sector size and capacity
prior to invoking add_disk. If metadata is required, the capacity will
be temporarilly set to 0 until it can be revalidated and registered with
the integrity extenstions after add_disk completes.
The driver relies on the integrity extensions to provide the metadata
buffer. NVMe requires this be a single physically contiguous region,
so only one integrity segment is allowed per command. If the metadata
is used for T10 PI, the driver provides mappings to save and restore
the reftag physical block translation. The driver provides no-op
functions for generate and verify if metadata is not used for protection
information. This way the setup is always provided by the block layer.
If a request does not supply a required metadata buffer, the command
is failed with bad address. This could only happen if a user manually
disables verify/generate on such a disk. The only exception to where
this is okay is if the controller is capable of stripping/generating
the metadata, which is possible on some types of formats.
The metadata scatter gather list now occupies the spot in the nvme_iod
that used to be used to link retryable IOD's, but we don't do that
anymore, so the field was unused.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Currently we allocate an nvme_iod for each IO, which holds the
sg list, prps, and other IO related info. Set a threshold of
2 pages and/or 8KB of data, below which we can just embed this
in the per-command pdu in blk-mq. For any IO at or below
NVME_INT_PAGES and NVME_INT_BYTES, we save a kmalloc and kfree.
For higher IOPS, this saves up to 1% of CPU time.
Signed-off-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
This converts the NVMe driver to a blk-mq request-based driver.
The NVMe driver is currently bio-based and implements queue logic within
itself. By using blk-mq, a lot of these responsibilities can be moved
and simplified.
The patch is divided into the following blocks:
* Per-command data and cmdid have been moved into the struct request
field. The cmdid_data can be retrieved using blk_mq_rq_to_pdu() and id
maintenance are now handled by blk-mq through the rq->tag field.
* The logic for splitting bio's has been moved into the blk-mq layer.
The driver instead notifies the block layer about limited gap support in
SG lists.
* blk-mq handles timeouts and is reimplemented within nvme_timeout().
This both includes abort handling and command cancelation.
* Assignment of nvme queues to CPUs are replaced with the blk-mq
version. The current blk-mq strategy is to assign the number of
mapped queues and CPUs to provide synergy, while the nvme driver
assign as many nvme hw queues as possible. This can be implemented in
blk-mq if needed.
* NVMe queues are merged with the tags structure of blk-mq.
* blk-mq takes care of setup/teardown of nvme queues and guards invalid
accesses. Therefore, RCU-usage for nvme queues can be removed.
* IO tracing and accounting are handled by blk-mq and therefore removed.
* Queue suspension logic is replaced with the logic from the block
layer.
Contributions in this patch from:
Sam Bradshaw <sbradshaw@micron.com>
Jens Axboe <axboe@fb.com>
Keith Busch <keith.busch@intel.com>
Robert Nelson <rlnelson@google.com>
Acked-by: Keith Busch <keith.busch@intel.com>
Acked-by: Jens Axboe <axboe@fb.com>
Updated for new ->queue_rq() prototype.
Signed-off-by: Jens Axboe <axboe@fb.com>
Adds support for devices with max page size smaller than the host's.
In the case we encounter such a host/device combination, the driver will
split a page into as many PRP entries as necessary for the device's page
size capabilities. If the device's reported minimum page size is greater
than the host's, the driver will not attempt to enable the device and
return an error instead.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Submits NVMe asynchronous event requests, one event up to the controller
maximum or number of possible different event types (8), whichever is
smaller. Events successfully returned by the controller are logged.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
There is a potential dead lock if a cpu event occurs during nvme probe
since it registered with hot cpu notification. This fixes the race by
having the module register with notification outside of probe rather
than have each device register.
The actual work is done in a scheduled work queue instead of in the
notifier since assigning IO queues has the potential to block if the
driver creates additional queues.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
It's positively immoral to have a global variable called 'io_timeout'.
Keep the module parameter called io_timeout, though.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
It is possible a filesystem may send a flush flagged bio with write
data. There is no such composite NVMe command, so the driver sends flush
and write separately.
The device is allowed to execute these commands in any order, so it was
possible the driver ends the bio after the write completes, but while the
flush is still active. We don't want to let a filesystem believe flush
succeeded before it really has; this could cause data corruption on a
power loss between these events. To fix, this patch splits the flush
and write into chained bios.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
This configures an nvme request_queue as flush capable if the device
has a volatile write cache present.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Make the copyright dates accurate and remove the final paragraph that
includes the address of the FSF.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Pull NVMe driver updates from Matthew Wilcox:
"Various updates to the NVMe driver. The most user-visible change is
that drive hotplugging now works and CPU hotplug while an NVMe drive
is installed should also work better"
* git://git.infradead.org/users/willy/linux-nvme:
NVMe: Retry failed commands with non-fatal errors
NVMe: Add getgeo to block ops
NVMe: Start-stop nvme_thread during device add-remove.
NVMe: Make I/O timeout a module parameter
NVMe: CPU hot plug notification
NVMe: per-cpu io queues
NVMe: Replace DEFINE_PCI_DEVICE_TABLE
NVMe: Fix divide-by-zero in nvme_trans_io_get_num_cmds
NVMe: IOCTL path RCU protect queue access
NVMe: RCU protected access to io queues
NVMe: Initialize device reference count earlier
NVMe: Add CONFIG_PM_SLEEP to suspend/resume functions
For commands returned with failed status, queue these for resubmission
and continue retrying them until success or for a limited amount of
time. The final timeout was arbitrarily chosen so requests can't be
retried indefinitely.
Since these are requeued on the nvmeq that submitted the command, the
callbacks have to take an nvmeq instead of an nvme_dev as a parameter
so that we can use the locked queue to append the iod to retry later.
The nvme_iod conviently can be used to track how long we've been trying
to successfully complete an iod request. The nvme_iod also provides the
nvme prp dma mappings, so I had to move a few things around so we can
keep those mappings.
Signed-off-by: Keith Busch <keith.busch@intel.com>
[fixed checkpatch issue with long line]
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Increase the default timeout to 30 seconds to match SCSI.
Signed-off-by: Keith Busch <keith.busch@intel.com>
[use byte instead of ushort]
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Registers with hot cpu notification to rebalance, and potentially allocate
additional, io queues.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The device's IO queues are associated with CPUs, so we can use a per-cpu
variable to map the a qid to a cpu. This provides a convienient way
to optimally assign queues to multiple cpus when the device supports
fewer queues than the host has cpus. The previous implementation may
have assigned these poorly in these situations. This patch addresses
this by sharing queues among cpus that are "close" together and should
have a lower lock contention penalty.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
This adds rcu protected access to a queue in the nvme IOCTL path
to fix potential races between a surprise removal and queue usage in
nvme_submit_sync_cmd. The fix holds the rcu_read_lock() here to prevent
the nvme_queue from freeing while this path is executing so it can't
sleep, and so this path will no longer wait for a available command
id should they all be in use at the time a passthrough IOCTL request
is received.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
This adds rcu protected access to nvme_queue to fix a race between a
surprise removal freeing the queue and a thread with open reference on
a NVMe block device using that queue.
The queues do not need to be rcu protected during the initialization or
shutdown parts, so I've added a helper function for raw deferencing
to get around the sparse errors.
There is still a hole in the IOCTL path for the same problem, which is
fixed in a subsequent patch.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
PREPARE_[DELAYED_]WORK() are being phased out. They have few users
and a nasty surprise in terms of reentrancy guarantee as workqueue
considers work items to be different if they don't have the same work
function.
nvme_dev->reset_work is multiplexed with multiple work functions.
Introduce nvme_reset_workfn() which invokes nvme_dev->reset_workfn and
always use it as the work function and update the users to set the
->reset_workfn field instead of overriding the work function using
PREPARE_WORK().
It would probably be best to route this with other related updates
through the workqueue tree.
Compile tested.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Matthew Wilcox <willy@linux.intel.com>
Cc: linux-nvme@lists.infradead.org
Send nvme abort command to io requests that have timed out on an
initialized device. If the command is not returned after another timeout,
schedule the controller for reset.
Signed-off-by: Keith Busch <keith.busch@intel.com>
[fix endianness issues]
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Schedules a controller reset when it indicates it has a failed status. If
the device does not become ready after a reset, the pci device will be
scheduled for removal.
Signed-off-by: Keith Busch <keith.busch@intel.com>
[fixed checkpatch issue]
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Adds controller error handling on resume power management. If the device
fails to initialize, the device is queued for a reset. If the reset fails,
a thread is spawned to remove the pci device.
If the device resumes as "busy", the device is responding to admin
commands but will not create IO queues. In this case, we need to remove
the gendisks and free the IO queues since they can't be used and may be
holding bios in their lists.
From testing, the dma pools require a pci device so this had to change
the pci driver 'remove' to release the dma resources in line with that
call instead of after all references to the device are released.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
For 32-bit versions of sg3-utils running on a 64-bit system. This is
mostly a copy from the relevent portions of fs/compat_ioctl.c, with
slight modifications for going through block_device_operations.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Vishal Verma <vishal.l.verma@linux.intel.com>
[fixed up CONFIG_COMPAT=n build problems]
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Changes the type of dev->db_stride to unsigned and changes the value
stored there to be 1 << the current value. Then there is less
calculation to be done at completion time.
Signed-off-by: Haiyan Hu <huhaiyan@huawei.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The NVMe spec recommends using the shutdown normal sequence when safely
taking the controller offline instead of hitting CC.EN on the next
start-up to reset the controller. The spec recommends a minimum of 1
second for the shutdown complete. This patch waits 2 seconds to be on
the safe side.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The 'Number of Namespaces' read from the device was being treated as
signed, which would cause us to not scan any namespaces for a device
with more than 2 billion namespaces. That led to noticing that the
namespace ID was also being treated as signed, which could lead to the
result from NVME_IOCTL_ID being treated as an error code.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
To build user programs that call the NVMe ioctls, we need to have a
user header file. Catch up to the new way of doing that by splitting
the header file into kernel and uapi portions.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Add io stats accounting for bio requests so nvme block devices show
useful disk stats.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Add definitions for the three Firmware Activate actions, and change the
SCSI translation code to construct the command into a temporary variable
instead of translating the endianness back-and-forth.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Reviewed-by: Vishal Verma <vishal.l.verma@linux.intel.com>
This adds support for namespaces with separate meta-data formats in the
submit io ioctl. The meta-data buffer has to be a contiguous, so such
a buffer is allocated and the mapped user pages are copied to/from this
buffer for write/read commands.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
We have an nvme device that has a concept of a stripe size. IO requests
that do not transfer data crossing a stripe boundary has greater
performance compared to IO that does cross it. This patch sets the
stripe size for the device if the device and vendor ids match one with
this feature and splits IO requests that cross the stripe boundary.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Registers a miscellaneous device for each nvme controller probed. This
creates character device files as /dev/nvmeN, where N is the device
instance, and supports nvme admin ioctl commands so devices without
namespaces can be managed.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
When constructing the command, dsmgmt needs to be treated as a 32-bit
value, not a 16-bit value. reftag, apptag and appmask all need to be
converted from native-endian to little-endian. Again, sparse's bitwise
warnings caught this problem. Thanks to Keith for pointing out the
correct way to fix the reftag.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Acked-by: Keith Busch <keith.busch@intel.com>
Introduce nvme_block_nr() to help convert sectors to block numbers.
This fixes an integer overflow in the SCSI conversion layer, and it's
slightly less typing than opencoding it.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Acked-by: Keith Busch <keith.busch@intel.com>
Translates SCSI commands in SG_IO ioctl to NVMe commands.
Uses the scsi-nvme translation spec from nvmexpress.org as reference.
Signed-off-by: Vishal Verma <vishal.l.verma@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The SCSI emulation has the ability to send format commands, so we need
to add the definition of the command. Also add a missing error code.
Signed-off-by: Vishal Verma <vishal.l.verma@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
nvme-scsi.c uses several data structures and definitions that were
previously private to nvme-core.c. Move the definitions to nvme.h,
protected by __KERNEL__.
Signed-off-by: Vishal Verma <vishal.l.verma@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
This adds discard support to block queues if the nvme device is capable of
deallocating blocks as indicated by the controller's optional command support.
A discard flagged bio request will submit an NVMe deallocate Data Set
Management command for the requested blocks.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
This data structure is defined in the NVMe specification. It's not used
by the kernel, but is available for use by userspace software.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Set the depth for IO queues to the device's maximum supported queue
entries if the requested depth exceeds the device's capabilities.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Set the max hw sectors in a namespace's request queue if the nvme device
has a max data transfer size.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The driver was still using an old definition of Identify Controller
which only came to light once we started using the 'number of namespaces'
field properly.
Reported-by: Nisheeth Bhat <nisheeth.bhat@intel.com>
Reported-by: Khosrow Panah <Khosrow.Panah@idt.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The doorbell stride allows devices to spread out their doorbells instead
of packing them tightly. This feature was added as part of ECN 003.
This patch also enables support for more than 512 queues :-)
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Remove the special-purpose IDENTIFY, GET_RANGE_TYPE, DOWNLOAD_FIRMWARE
and ACTIVATE_FIRMWARE commands. Replace them with a generic ADMIN_CMD
ioctl that can submit any admin command.
Add a new ID ioctl that returns the namespace ID of the queried device.
It corresponds to the SCSI Idlun ioctl.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
THe device reports (in its capability register) how long it will take
to initialise. If that time elapses before the ready bit becomes set,
conclude the device is broken and refuse to initialise it. Log a nice
error message so the user knows why we did nothing.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The arbitration field was extended by one bit, shifting the shutdown
notification bits by one. Also, the SQ/CQ entry size was made
configurable for future extensions.
Reported-by: Paul Luse <paul.e.luse@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The read and write commands don't define a 'result', so there's no need
to copy it back to userspace.
Remove the ability of the ioctl to submit commands to a different
namespace; it's just asking for trouble, and the use case I have in mind
will be addressed througha different ioctl in the future. That removes
the need for both the block_shift and nsid arguments.
Check that the opcode is one of 'read' or 'write'. Future opcodes may
be added in the future, but we will need a different structure definition
for them.
The nblocks field is redefined to be 0-based. This allows the user to
request the full 65536 blocks.
Don't byteswap the reftag, apptag and appmask. Martin Petersen tells
me these are calculated in big-endian and are transmitted to the device
in big-endian.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Keith Busch