This provides unified readq()/writeq() helper functions for 32-bit
drivers.
For some cases, readq/writeq without atomicity is harmful, and order of
io access has to be specified explicitly. So in this patch, new two
header files which contain non-atomic readq/writeq are added.
- <asm-generic/io-64-nonatomic-lo-hi.h> provides non-atomic readq/
writeq with the order of lower address -> higher address
- <asm-generic/io-64-nonatomic-hi-lo.h> provides non-atomic readq/
writeq with reversed order
This allows us to remove some readq()s that were added drivers when the
default non-atomic ones were removed in commit dbee8a0aff ("x86:
remove 32-bit versions of readq()/writeq()")
The drivers which need readq/writeq but can do with the non-atomic ones
must add the line:
#include <asm-generic/io-64-nonatomic-lo-hi.h> /* or hi-lo.h */
But this will be nop in 64-bit environments, and no other #ifdefs are
required. So I believe that this patch can solve the problem of
1. driver-specific readq/writeq
2. atomicity and order of io access
This patch is tested with building allyesconfig and allmodconfig as
ARCH=x86 and ARCH=i386 on top of tip/master.
Cc: Kashyap Desai <Kashyap.Desai@lsi.com>
Cc: Len Brown <lenb@kernel.org>
Cc: Ravi Anand <ravi.anand@qlogic.com>
Cc: Vikas Chaudhary <vikas.chaudhary@qlogic.com>
Cc: Matthew Garrett <mjg@redhat.com>
Cc: Jason Uhlenkott <juhlenko@akamai.com>
Cc: James Bottomley <James.Bottomley@parallels.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Roland Dreier <roland@purestorage.com>
Cc: James Bottomley <jbottomley@parallels.com>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: Matthew Wilcox <matthew.r.wilcox@intel.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Hitoshi Mitake <h.mitake@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The type of 'make_request_fn' changed in 5a7bbad27a ("block: remove
support for bio remapping from ->make_request"), but the merge of the
nvme driver didn't take that into account, and as a result the driver
would compile with a warning:
drivers/block/nvme.c: In function 'nvme_alloc_ns':
drivers/block/nvme.c:1336:2: warning: passing argument 2 of 'blk_queue_make_request' from incompatible pointer type [enabled by default]
include/linux/blkdev.h:830:13: note: expected 'void (*)(struct request_queue *, struct bio *)' but argument is of type 'int (*)(struct request_queue *, struct bio *)'
It's benign, but the warning is annoying.
Reported-by: Stephen Rothwell <sfr@canb.auug.org>
Cc: Matthew Wilcox <matthew.r.wilcox@intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The number of submission & completion queues should be set by calling
Set Features, not Get Features.
Reported-by: Kwok Kong <Kwok.Kong@idt.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
QUEUE_FLAG_* are flags (other than QUEUE_FLAG_DEFAULT), so they cannot
be ORed together. Set the queue flags using queue_flag_set_unlocked().
Reported-by: Donald Wood <donald.e.wood@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
By using the iod->nents field (the same way other I/O paths do), we can
avoid recalculating the number of sg entries at unmap time, and make
nvme_unmap_user_pages() easier to call.
Also, use the 'write' parameter instead of assuming DMA_FROM_DEVICE.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
We were always mapping as DMA_FROM_DEVICE then unmapping with
DMA_TO_DEVICE which was clearly not correct. Follow the same pattern as
nvme_submit_io() and key off the bottom bit of the opcode to determine
whether this is a read or a write.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
IO_TIMEOUT is a little too generic and might be used by other parts of
the kernel in the future.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The new merged data structure is called nvme_iod. This improves performance
for mid-sized I/Os (in the 16k range) since we save a memory allocation.
It is also a slightly simpler interface to use.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The queue is only needed for some rare occasions, and it's more consistent
to pass the device around.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Upcoming patches require calling get_nvmeq when we don't have a namespace.
Some callers already have the device in a local variable anyway.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Instead of encoding the handler type in the bottom two bits of the
per-completion context pointer, store the handler function as well
as the context pointer. This gives us more flexibility and the code
is clearer. It comes at the cost of an extra 8k of memory per queue,
but this feels like a reasonable price to pay.
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>
ECN 001 documented that namespace 0 is not valid. Sending an Identify
with CNS of 0 and Namespace of 0 is an undefined command.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The existing calculation underestimated the number of pages required
as it did not take into account the pointer at the end of each page.
The replacement calculation may overestimate the number of pages required
if the last page in the PRP List is entirely full. By using ->npages
as a counter as we fill in the pages, we ensure that we don't try to
free a page that was never allocated.
Signed-off-by: Nisheeth Bhat <nisheeth.bhat@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Instead of open-coding calls to nvme_submit_admin_cmd, these
small wrappers are simpler to use (the patch removes 14 lines from
nvme_dev_add() for example).
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
dma_unmap_sg() must be called with the same 'nents' passed to
dma_map_sg(), not the number returned from dma_map_sg().
Signed-off-by: Nisheeth Bhat <nisheeth.bhat@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Our SG list was constructed to always fill the entire first page, even
if that was more than the length of the I/O. This is probably harmless,
but some IOMMUs might do something bad.
Correcting the first call to sg_set_page() made it look a lot closer to
the sg_set_page() in the loop, so fold the first call to sg_set_page()
into the loop.
Reported-by: Nisheeth Bhat <nisheeth.bhat@intel.com>
Signed-off-by: Matthew Wilcox <willy@linux.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>
If the I/O was not completed by a single NVMe command, we add the
bio to the congestion list and wake up the kthread to resubmit it.
But the kthread calls remove_wait_queue() unconditionally, which
will oops if it's not on the wait queue. So add the kthread to
the wait queue before waking it up.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
nvme_setup_io_queues() was assuming that a NULL return from
nvme_create_queue() was an out-of-memory error. That's not necessarily
true; the adapter might return -EIO, for example. Change the calling
convention to return an ERR_PTR on failure instead of NULL.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
For the benefit of reviewers, add comments to a few functions describing
their calling context
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
If any of the memory allocations in nvme_setup_prps fail, handle it by
modifying the passed-in data length to reflect the number of bytes we are
actually able to send. Also allow the caller to specify the GFP flags
they need; for user-initiated commands, we can use GFP_KERNEL allocations.
The various callers are updated to handle this possibility; the main
I/O path is already prepared for this possibility (as it may happen
due to nvme_map_bio being unable to map all the segments of the I/O).
The other callers return -ENOMEM instead of doing partial I/Os.
Reported-by: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The current approach of using the namespace ID as the minor number
doesn't work when there are multiple adapters in the machine. Rather
than statically partitioning the number of namespaces between adapters,
dynamically allocate minor numbers to namespaces as they are detected.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The trailing '_data' on the end was annoying and inconsistent. Also, make
it actually return the data since this is needed for timing out commands.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
When an I/O completed with an error, we would call bio_endio twice
(once with -EIO and once with 0). Found by inspection.
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>
We need to clear the affinity mask before calling free_irq()
Reported-by: Shane Michael Matthews <shane.matthews@intel.com>
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>
Make ioctls work for 32-bit applications on 64-bit kernels. The structures
are defined to be the same for both 32- and 64-bit applications, so
we can use the same handler for both.
Reported-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Fill in all the num_possible_cpus() entries with duplicate pointers.
This reduces the complexity of the frequently-called get_nvmeq(), as
well as avoiding a bug in it when there are fewer queues than CPUs.
Reported-by: Shane Michael Matthews <shane.matthews@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Once there are no more bios on the congestion list, we can stop waking
up the nvme kthread every time a completion happens.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
If the last element in the PRP list fits on the end of the page, there's
no need to allocate an extra page to put that single element in. It can
fit on the end of the page.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The spec says this is a 0s based value. We don't need to handle the
maximal value because it's reserved to mean "every namespace".
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The head can never overrun the tail since we won't allocate enough command
IDs to let that happen. The status codes are in sync with the spec.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
The spec says we're not allowed to completely fill the submission queue.
Solve this by reducing the number of allocatable cmdids by 1.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
When we submit subsequent portions of the I/O, we need to access the
updated block, not start reading again from the original position.
This was showing up as miscompares in the XFS randholes testcase.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
NVMe scatterlists must be virtually contiguous, like almost all I/Os.
However, when the filesystem lays out files with a hole, it can be that
adjacent LBAs map to non-adjacent virtual addresses. Handle this by
submitting one NVMe command at a time for each virtually discontiguous
range.
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Hitoshi Mitake
Linus Torvalds
Matthew Wilcox
Nisheeth Bhat