This patch implements persistent grants for the xen-blk{front,back}
mechanism. The effect of this change is to reduce the number of unmap
operations performed, since they cause a (costly) TLB shootdown. This
allows the I/O performance to scale better when a large number of VMs
are performing I/O.
Previously, the blkfront driver was supplied a bvec[] from the request
queue. This was granted to dom0; dom0 performed the I/O and wrote
directly into the grant-mapped memory and unmapped it; blkfront then
removed foreign access for that grant. The cost of unmapping scales
badly with the number of CPUs in Dom0. An experiment showed that when
Dom0 has 24 VCPUs, and guests are performing parallel I/O to a
ramdisk, the IPIs from performing unmap's is a bottleneck at 5 guests
(at which point 650,000 IOPS are being performed in total). If more
than 5 guests are used, the performance declines. By 10 guests, only
400,000 IOPS are being performed.
This patch improves performance by only unmapping when the connection
between blkfront and back is broken.
On startup blkfront notifies blkback that it is using persistent
grants, and blkback will do the same. If blkback is not capable of
persistent mapping, blkfront will still use the same grants, since it
is compatible with the previous protocol, and simplifies the code
complexity in blkfront.
To perform a read, in persistent mode, blkfront uses a separate pool
of pages that it maps to dom0. When a request comes in, blkfront
transmutes the request so that blkback will write into one of these
free pages. Blkback keeps note of which grefs it has already
mapped. When a new ring request comes to blkback, it looks to see if
it has already mapped that page. If so, it will not map it again. If
the page hasn't been previously mapped, it is mapped now, and a record
is kept of this mapping. Blkback proceeds as usual. When blkfront is
notified that blkback has completed a request, it memcpy's from the
shared memory, into the bvec supplied. A record that the {gref, page}
tuple is mapped, and not inflight is kept.
Writes are similar, except that the memcpy is peformed from the
supplied bvecs, into the shared pages, before the request is put onto
the ring.
Blkback stores a mapping of grefs=>{page mapped to by gref} in
a red-black tree. As the grefs are not known apriori, and provide no
guarantees on their ordering, we have to perform a search
through this tree to find the page, for every gref we receive. This
operation takes O(log n) time in the worst case. In blkfront grants
are stored using a single linked list.
The maximum number of grants that blkback will persistenly map is
currently set to RING_SIZE * BLKIF_MAX_SEGMENTS_PER_REQUEST, to
prevent a malicios guest from attempting a DoS, by supplying fresh
grefs, causing the Dom0 kernel to map excessively. If a guest
is using persistent grants and exceeds the maximum number of grants to
map persistenly the newly passed grefs will be mapped and unmaped.
Using this approach, we can have requests that mix persistent and
non-persistent grants, and we need to handle them correctly.
This allows us to set the maximum number of persistent grants to a
lower value than RING_SIZE * BLKIF_MAX_SEGMENTS_PER_REQUEST, although
setting it will lead to unpredictable performance.
In writing this patch, the question arrises as to if the additional
cost of performing memcpys in the guest (to/from the pool of granted
pages) outweigh the gains of not performing TLB shootdowns. The answer
to that question is `no'. There appears to be very little, if any
additional cost to the guest of using persistent grants. There is
perhaps a small saving, from the reduced number of hypercalls
performed in granting, and ending foreign access.
Signed-off-by: Oliver Chick <oliver.chick@citrix.com>
Signed-off-by: Roger Pau Monne <roger.pau@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
[v1: Fixed up the misuse of bool as int]
* Allow a Linux guest to boot as initial domain and as normal guests
on Xen on ARM (specifically ARMv7 with virtualized extensions).
PV console, block and network frontend/backends are working.
Bug-fixes:
* Fix compile linux-next fallout.
* Fix PVHVM bootup crashing.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.12 (GNU/Linux)
iQEcBAABAgAGBQJQbJELAAoJEFjIrFwIi8fJSI4H/32qrQKyF5IIkFKHTN9FYDC1
OxEGc4y47DIQpGUd/PgZ/i6h9Iyhj+I6pb4lCevykwgd0j83noepluZlCIcJnTfL
HVXNiRIQKqFhqKdjTANxVM4APup+7Lqrvqj6OZfUuoxaZ3tSTLhabJ/7UXf2+9xy
g2RfZtbSdQ1sukQ/A2MeGQNT79rh7v7PrYQUYSrqytjSjSLPTqRf75HWQ+eapIAH
X3aVz8Tn6nTixZWvZOK7rAaD4awsFxGP6E46oFekB02f4x9nWHJiCZiXwb35lORb
tz9F9td99f6N4fPJ9LgcYTaCPwzVnceZKqE9hGfip4uT+0WrEqDxq8QmBqI5YtI=
=gxJD
-----END PGP SIGNATURE-----
Merge tag 'stable/for-linus-3.7-arm-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/konrad/xen
Pull ADM Xen support from Konrad Rzeszutek Wilk:
Features:
* Allow a Linux guest to boot as initial domain and as normal guests
on Xen on ARM (specifically ARMv7 with virtualized extensions). PV
console, block and network frontend/backends are working.
Bug-fixes:
* Fix compile linux-next fallout.
* Fix PVHVM bootup crashing.
The Xen-unstable hypervisor (so will be 4.3 in a ~6 months), supports
ARMv7 platforms.
The goal in implementing this architecture is to exploit the hardware
as much as possible. That means use as little as possible of PV
operations (so no PV MMU) - and use existing PV drivers for I/Os
(network, block, console, etc). This is similar to how PVHVM guests
operate in X86 platform nowadays - except that on ARM there is no need
for QEMU. The end result is that we share a lot of the generic Xen
drivers and infrastructure.
Details on how to compile/boot/etc are available at this Wiki:
http://wiki.xen.org/wiki/Xen_ARMv7_with_Virtualization_Extensions
and this blog has links to a technical discussion/presentations on the
overall architecture:
http://blog.xen.org/index.php/2012/09/21/xensummit-sessions-new-pvh-virtualisation-mode-for-arm-cortex-a15arm-servers-and-x86/
* tag 'stable/for-linus-3.7-arm-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/konrad/xen: (21 commits)
xen/xen_initial_domain: check that xen_start_info is initialized
xen: mark xen_init_IRQ __init
xen/Makefile: fix dom-y build
arm: introduce a DTS for Xen unprivileged virtual machines
MAINTAINERS: add myself as Xen ARM maintainer
xen/arm: compile netback
xen/arm: compile blkfront and blkback
xen/arm: implement alloc/free_xenballooned_pages with alloc_pages/kfree
xen/arm: receive Xen events on ARM
xen/arm: initialize grant_table on ARM
xen/arm: get privilege status
xen/arm: introduce CONFIG_XEN on ARM
xen: do not compile manage, balloon, pci, acpi, pcpu and cpu_hotplug on ARM
xen/arm: Introduce xen_ulong_t for unsigned long
xen/arm: Xen detection and shared_info page mapping
docs: Xen ARM DT bindings
xen/arm: empty implementation of grant_table arch specific functions
xen/arm: sync_bitops
xen/arm: page.h definitions
xen/arm: hypercalls
...
If the caller passes a valid kmap_op to m2p_add_override, we use
kmap_op->dev_bus_addr to store the original mfn, but dev_bus_addr is
part of the interface with Xen and if we are batching the hypercalls it
might not have been written by the hypervisor yet. That means that later
on Xen will write to it and we'll think that the original mfn is
actually what Xen has written to it.
Rather than "stealing" struct members from kmap_op, keep using
page->index to store the original mfn and add another parameter to
m2p_remove_override to get the corresponding kmap_op instead.
It is now responsibility of the caller to keep track of which kmap_op
corresponds to a particular page in the m2p_override (gntdev, the only
user of this interface that passes a valid kmap_op, is already doing that).
CC: stable@kernel.org
Reported-and-Tested-By: Sander Eikelenboom <linux@eikelenboom.it>
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
The only reason for the distinction was for the special case of
'file' (which is assumed to be loopback device), was to reach inside
the loopback device, find the underlaying file, and call fallocate on it.
Fortunately "xen-blkback: convert hole punching to discard request on
loop devices" removes that use-case and we now based the discard
support based on blk_queue_discard(q) and extract all appropriate
parameters from the 'struct request_queue'.
CC: Li Dongyang <lidongyang@novell.com>
Acked-by: Jan Beulich <JBeulich@suse.com>
[v1: Dropping pointless initializer and keeping blank line]
[v2: Remove the kfree as it is not used anymore]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
As of dfaa2ef68e, loop devices support
discard request now. We could just issue a discard request, and
the loop driver will punch the hole for us, so we don't need to touch
the internals of loop device and punch the hole ourselves, Thanks.
V0->V1: rebased on devel/for-jens-3.3
Signed-off-by: Li Dongyang <lidongyang@novell.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Part of the blkdev_issue_discard(xx) operation is that it can also
issue a secure discard operation that will permanantly remove the
sectors in question. We advertise that we can support that via the
'discard-secure' attribute and on the request, if the 'secure' bit
is set, we will attempt to pass in REQ_DISCARD | REQ_SECURE.
CC: Li Dongyang <lidongyang@novell.com>
[v1: Used 'flag' instead of 'secure:1' bit]
[v2: Use 'reserved' uint8_t instead of adding a new value]
[v3: Check for nseg when mapping instead of operation]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
In a union type structure to deal with the overlapping
attributes in a easier manner.
Suggested-by: Ian Campbell <Ian.Campbell@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
* 'for-3.2/drivers' of git://git.kernel.dk/linux-block: (30 commits)
virtio-blk: use ida to allocate disk index
hpsa: add small delay when using PCI Power Management to reset for kump
cciss: add small delay when using PCI Power Management to reset for kump
xen/blkback: Fix two races in the handling of barrier requests.
xen/blkback: Check for proper operation.
xen/blkback: Fix the inhibition to map pages when discarding sector ranges.
xen/blkback: Report VBD_WSECT (wr_sect) properly.
xen/blkback: Support 'feature-barrier' aka old-style BARRIER requests.
xen-blkfront: plug device number leak in xlblk_init() error path
xen-blkfront: If no barrier or flush is supported, use invalid operation.
xen-blkback: use kzalloc() in favor of kmalloc()+memset()
xen-blkback: fixed indentation and comments
xen-blkfront: fix a deadlock while handling discard response
xen-blkfront: Handle discard requests.
xen-blkback: Implement discard requests ('feature-discard')
xen-blkfront: add BLKIF_OP_DISCARD and discard request struct
drivers/block/loop.c: remove unnecessary bdev argument from loop_clr_fd()
drivers/block/loop.c: emit uevent on auto release
drivers/block/cpqarray.c: use pci_dev->revision
loop: always allow userspace partitions and optionally support automatic scanning
...
Fic up trivial header file includsion conflict in drivers/block/loop.c
There are two windows of opportunity to cause a race when
processing a barrier request. This patch fixes this.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
The patch titled: "xen/blkback: Fix the inhibition to map pages
when discarding sector ranges." had the right idea except that
it used the wrong comparison operator. It had == instead of !=.
This fixes the bug where all (except discard) operations would
have been ignored.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
The 'operation' parameters are the ones provided to the bio layer while
the req->operation are the ones passed in between the backend and
frontend. We used the wrong 'operation' value to squash the
call to map pages when processing the discard operation resulting
in an hypercall that did nothing. Lets guard against going in the
mapping function by checking for the proper operation type.
CC: Li Dongyang <lidongyang@novell.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
We did not increment the amount of sectors written to disk
b/c we tested for the == WRITE which is incorrect - as the
operations are more of WRITE_FLUSH, WRITE_ODIRECT. This patch
fixes it by doing a & WRITE check.
CC: stable@kernel.org
Reported-by: Andy Burns <xen.lists@burns.me.uk>
Suggested-by: Ian Campbell <Ian.Campbell@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
We emulate the barrier requests by draining the outstanding bio's
and then sending the WRITE_FLUSH command. To drain the I/Os
we use the refcnt that is used during disconnect to wait for all
the I/Os before disconnecting from the frontend. We latch on its
value and if it reaches either the threshold for disconnect or when
there are no more outstanding I/Os, then we have drained all I/Os.
Suggested-by: Christopher Hellwig <hch@infradead.org>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
This fixes the problem of three of those four memset()-s having
improper size arguments passed: Sizeof a pointer-typed expression
returns the size of the pointer, not that of the pointed to data.
It also reverts using kmalloc() instead of kzalloc() for the allocation
of the pending grant handles array, as that array gets fully
initialized in a subsequent loop.
Reported-by: Julia Lawall <julia@diku.dk>
Signed-off-by: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
..aka ATA TRIM/SCSI UNMAP command to be passed through the frontend
and used as appropiately by the backend. We also advertise
certain granulity parameters to the frontend so it can plug them in.
If the backend is a realy device - we just end up using
'blkdev_issue_discard' while for loopback devices - we just punch
a hole in the image file.
Signed-off-by: Li Dongyang <lidongyang@novell.com>
[v1: Fixed up pr_debug and commit description]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
If we want to use granted pages for AIO, changing the mappings of a user
vma and the corresponding p2m is not enough, we also need to update the
kernel mappings accordingly.
Currently this is only needed for pages that are created for user usages
through /dev/xen/gntdev. As in, pages that have been in use by the
kernel and use the P2M will not need this special mapping.
However there are no guarantees that in the future the kernel won't
start accessing pages through the 1:1 even for internal usage.
In order to avoid the complexity of dealing with highmem, we allocated
the pages lowmem.
We issue a HYPERVISOR_grant_table_op right away in
m2p_add_override and we remove the mappings using another
HYPERVISOR_grant_table_op in m2p_remove_override.
Considering that m2p_add_override and m2p_remove_override are called
once per page we use multicalls and hypercall batching.
Use the kmap_op pointer directly as argument to do the mapping as it is
guaranteed to be present up until the unmapping is done.
Before issuing any unmapping multicalls, we need to make sure that the
mapping has already being done, because we need the kmap->handle to be
set correctly.
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
[v1: Removed GRANT_FRAME_BIT usage]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Add xen-backend:vbd module alias to the xen-blkback module. This allows
automatic loading of the module.
Signed-off-by: Bastian Blank <waldi@debian.org>
Acked-by: Ian Campbell <ian.campbell@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Running RING_FINAL_CHECK_FOR_REQUESTS from make_response is a bad
idea. It means that in-flight I/O is essentially blocking continued
batches. This essentially kills throughput on frontends which unplug
(or even just notify) early and rightfully assume addtional requests
will be picked up on time, not synchronously.
Signed-off-by: Daniel Stodden <daniel.stodden@citrix.com>
[v1: Rebased and fixed compile problems]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
blkbk->pending_pages can be NULL here so I added a check for it.
Signed-off-by: Dan Carpenter <error27@gmail.com>
[v1: Redid the loop a bit]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
The sector number on empty barrier requests may (will?) be -1, which,
given that it's being treated as unsigned 64-bit quantity, will almost
always exceed the actual (virtual) disk's size.
Inspired by Konrad's "When writting barriers set the sector number to
zero...".
While at it also add overflow checking to the math in vbd_translate().
Signed-off-by: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
vbd_resize() up_read()'s xs_state.suspend_mutex twice in a row via double
xenbus_transaction_end() calls. The next down_read() in
xenbus_transaction_start() (at eg. the next resize attempt) hangs.
Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=618317
Acked-by: Jan Beulich <jbeulich@novell.com>
Acked-by: Ian Campbell <ian.campbell@citrix.com>
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
We do a check for the operations right before calling dispatch_rw_block_io.
And then we do the same check in dispatch_rw_block_io. This patch
squashes those checks into the 'dispatch_rw_block_io' function.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
We drop the support for 'feature-barrier' and add in the support
for the 'feature-flush-cache' if the real backend storage supports
flushing.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
If one runs a simple fio request with random read/write with a
20%/80% ratio, the numbers are incredibly bad when using the CFQ scheduler.
IOmeter | | | |
64K, randrw | NOOP | CFQ | deadline |
randrwmix=80 | | | |
--------------+-------+------+----------+
blkback |103/27 |32/10 | 102/27 |
--------------+-------+------+----------+
QEMU qdisk |103/27 |102/27| 102/27 |
The problem as explained by Vivek Goyal was:
".. that difference is that sync vs async requests. In the case of
a kernel thread submitting IO, [..] all the WRITES might be being
considered as async and will go in a different queue. If you mix those
with some READS, they are always sync and will go in differnet queue.
In presence of sync queue, CFQ will idle and choke up WRITES in
an attempt to improve latencies of READs.
In case of AIO [note: this is what QEMU qdisk is doing] , [..]
it is direct IO and both READS and WRITES will be considered SYNC
and will go in a single queue and no choking of WRITES will take place."
The solution is quite simple, tack on REQ_SYNC (which is
what the WRITE_ODIRECT macro points to) and the numbers go
back up.
Suggested-by: Vivek Goyal <vgoyal@redhat.com
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
We used to the plug/unplug on the submit_bio. But that means
if within a stream of WRITE, WRITE, WRITE,...,WRITE we have
one READ, it could stall the pipeline (as the 'submio_bio'
could trigger the unplug_fnc to be called and stall/sync
when doing the READ). Instead we want to move the unplugging
when the whole (or as a much as possible) ring buffer has been
processed. This also eliminates us doing plug/unplug for
each request.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
And also shorten the name if it has blkback to blkbk.
This results in the symbol table (if compiled in the kernel)
to be much shorter, prettier, and also easier to search for.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Daniel Stodden suggested to eliminate vbd.c and interface.c, inlining the
critical bits where they belong, respectively.
Leaving only blkback.c for the data- and xenbus.c for the control path.
Suggested-by: Daniel Stodden <daniel.stodden@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Roger Pau Monne