Notes on block I/O data integrity

As various people wanted to know how the various data integrity patches I've send out recently play together here's a small writeup on what issues we have in QEMU and how to fix it:

There are two major aspects of data integrity we need to care in the QEMU block I/O code:

(1) stable data storage - we must be able to force data out of caches onto the stable media, and we must get completion notification for it.
(2) request ordering - we must be able to make sure some I/O request do not get reordered with other in-flight requests before or after it.

Linux uses two related abstractions to implement the this (other operating system are probably similar)

(1) a cache flush request that flushes the whole volatile write cache to stable storage
(2) a barrier request, which
(a) is guaranteed to actually go all the way to stable storage
(b) does not reordered versus any requests before or after it

For disks not using volatile write caches the cache flush is a no-op, and barrier requests are implemented by draining the queue of outstanding requests before the barrier request, and only allowing new requests to proceed after it has finished. Instead of the queue drain tag ordering could be used, but at this point that is not the case in Linux.

For disks using volatile write caches, the cache flush is implemented by a protocol specific request, and the the barrier request are implemented by performing cache flushes before and after the barrier request, in addition to the draining mentioned above. The second cache flush can be replaced by setting the "Force Unit Access" bit on the barrier request on modern disks.

The above is supported by the QEMU emulated disks in the following way:

- The IDE disk emulation implement the ATA WIN_FLUSH_CACHE/ WIN_FLUSH_CACHE_EXT commands to flush the drive cache, but does not indicate a volatile write cache in the ATA IDENTIFY command. Because of that guests do no not actually send down cache flush request. Linux guests do however drain the I/O queues to guarantee ordering in absence of volatile write caches.
- The SCSI disk emulation implements the SCSI SYNCHRONIZE_CACHE command, and also advertises the write cache enabled bit. This means Linux sends down cache flush requests to implement barriers, and provides sufficient queue draining.
- The virtio-blk driver does not implement any cache flush command. And while there is a virtio-blk feature bit for barrier support it is not support by virtio-blk. Due to the lack of a cache flush command it also is insufficient to implement the required data integrity semantics. Currently the virtio-blk Linux does not advertise any form of barrier support, and we don't even get the queue draining required for proper operation in a cache-less environment.

The I/O from these front end drivers maps to different host kernel I/O patterns depending on the cache= drive command line. There are three choices for it:

(a) cache=writethrough
(b) cache=writeback
(c) cache=none

(a) means all writes are synchronous (O_DSYNC), which means the host kernel guarantees us that data is on stable storage once the I/O request has completed.
In cache=writethrough mode the IDE and SCSI drivers are safe because the queue is properly drained to guarantee I/O ordering. Virtio-blk is not safe due to the lack of queue draining.
(b) means we use regular buffered writes and need a fsync/fdatasync to actually guarantee that data is stable on disk.
In data=writeback mode on the SCSI emulation is safe as all others miss the cache flush requests.
(c) means we use direct I/O (O_DIRECT) to bypass the host cache and perform direct dma to/from the I/O buffer in QEMU. While direct I/O bypasses the host cache it does not guarantee flushing of volatile write caches in disks, nor completion of metadata operations in filesystems (e.g. block allocations).
In data=none only the SCSI emulation is entirely safe right now due to the lack of cache flushes in the other drivers.

Action plan for the guest drivers:

- virtio-blk needs to advertise ordered queue by default. This makes cache=writethrough safe on virtio.

Action plan for QEMU:

- IDE needs to set the write cache enabled bit
- virtio needs to implement a cache flush command and advertise it (also needs a small change to the host driver)
- we need to implement an aio_fsync to not stall the vpu on cache flushes
- investigate only advertising a write cache when we really have one to avoid the cache flush requests for cache=writethrough

Notes on disk cache flushes on Linux hosts:

- barrier requests and cache flushes are supported by all local disk filesystem in popular use (btrfs, ext3, ext4, reiserfs, XFS). However unlike the other filesystems ext3 does _NOT_ enable barriers and cache flush requests by default.
- currently O_SYNC writes or fsync on block device nodes does not flush the disk cache.
- currently none of the filesystems nor the direct access to the block device nodes implements flushes of the disk caches when using O_DIRECT|O_DSYNC or using fsync/fdatasync after an O_DIRECT request.

what about LVM? iv'e read somewhere that it used to just eat barriers used by XFS, making it less safe than simple partitions.

Javier

Oh, any additional layers open another by cans of worms. On Linux until very recently using LVM or software raid means only disabled write caches are safe.

I believe that's still true except if there's more than one backing drive, so software RAID still isn't safe. Did that change?

But even with barriers, software RAID may have a consistency problem if one stripe is updated and the system fails before the matching parity stripe is updated.

I've been told that some hardware RAID implementations implement a kind of journalling to deal with this, but Linux software RAID does not.

-- Jamie

Yes, it did change. That beeing said with the amount of bugs in filesystems realted to write barriers my expectation for the RAID and device mapper code is not too high. I will recommend to keep doing what people caring for their data have been doing since these volatile write caches came up: turn them off.

Unfortunately I tried that on a batch of 1000 or so embedded thingies with ext3, and the write performance plummeted. They are the same thingies where I observed lack of barriers resulting in filesystem corruption after power failure. We really need barriers with ATA disks to get decent write performance.

It's a good recommendation though.

Turning off volatile write cache does not provide commit integrity with RAID.

RAID needs barriers to plug, drain and unplug the queues across all backing devices in a coordinated manner quite apart from the volatile write cache. And then there's still that pesky problem of writes which reach one disk and not it's parity disk.

Unfortunately turning off the volatile write caches could actually make the timing window for failure worse, in the case of system crash without power failure.

-- Jamie

Hello Christopher,

thanks a lot vor this overview, it answers a lot of my questions! May I suggest You put it somewhere on the wiki so it doesn't get forgotten in the maillist only?
It also rises few new questions though. We have experienced postgresql database corruptions lately, two times to be exact. First time, I blamed server crash, but lately (freshly created) database got corrupted for the second time and there were no crashes since the initialisation. The server hardware is surely OK. I didn't have much time to look into this yet, but Your mail just poked me to return to the subject. The situation is a bit more complex, as there are additional two layers of storage there:
we're using SATA/SAS drives, network-mirrored by DRBD, clustered LVM on top of those, and finally qemu-kvm using virtio on top of created logical volumes. So there are plenty of possible culprits, but Your mention of virtio unsafeness while using cache=writethrough (which is the default for drive types other then qcow) leads me to suspicion that this might be the reason of the problem. Databases are sensitive for requests reordering, so I guess using virtio for postgres storage was quite stupid from me :(
So my question is, could You please advise me a bit on the storage configuration? virtio performed much better then SCSI, but of course data integrity is crucial, so would You suggest rather using SCSI? DRBD doesn't have problem with barriers, clustered LVM SHOULD not have problems with it, as we're using just striped volumes, but I'll check it to be sure. So is it safe for me to keep cache=writethrough for the database volume?

thanks a lor in advance for any hints!

with best regards

nik

-------------------------------------
Nikola CIPRICH
LinuxBox.cz, s.r.o.
28. rijna 168, 709 01 Ostrava

tel.:   +420 596 603 142
fax:    +420 596 621 273
mobil:  +420 777 093 799

www.linuxbox.cz

mobil servis: +420 737 238 656
email servis: [email protected]
-------------------------------------

Note that LVM does not implement barriers at all, except for simple cases of a single backing device (I'm not sure if that includes dm-crypt).

So your striped volumes may not offer this level of integrity.

-- Jamie

I'll rather try to get the worst issues fixed ASAP.

I'm pretty sure one of the many laters in your setup will not pass through write barriers, so defintively make sure your write caches are disabled. Also right now virtio is not a good idea for data integrity. The guest side fix for a setup with cache=writethrough or cache=none on block device without volatile disk write cache is however a trivial one line patch I've already submitted. I've attached it below for reference:

Subject: [PATCH] virtio-blk: set QUEUE_ORDERED_DRAIN by default
From: Christoph Hellwig <[email protected]>

Currently virtio-blk doesn't set any QUEUE_ORDERED_ flag by default, which means it does not allow filesystems to use barriers. But the typical use case for virtio-blk is to use a backed that uses synchronous I/O, and in that case we can simply set QUEUE_ORDERED_DRAIN to make the block layer drain the request queue around barrier I/O and provide the semantics that the filesystems need. This is what the SCSI disk driver does for disks that have the write cache disabled.

With this patch we incorrectly advertise barrier support if someone configure qemu with write back caching. While this displays wrong information in the guest there is nothing that guest could have done even if we rightfully told it that we do not support any barriers.

Index: linux-2.6/drivers/block/virtio_blk.c
===================================================================
@@ -336,9 +336,16 @@ static int __devinit virtblk_probe(struc
 	vblk->disk->driverfs_dev = &vdev->dev;
 	index++;
 
-	/* If barriers are supported, tell block layer that queue is ordered */
+	/*
+	 * If barriers are supported, tell block layer that queue is ordered.
+	 *
+	 * If no barriers are supported assume the host uses synchronous
+	 * writes and just drain the the queue before and after the barrier.
+	 */
 	if (virtio_has_feature(vdev, VIRTIO_BLK_F_BARRIER))
 		blk_queue_ordered(vblk->disk->queue, QUEUE_ORDERED_TAG, NULL);
+	else
+		blk_queue_ordered(vblk->disk->queue, QUEUE_ORDERED_DRAIN, NULL);
 
 	/* If disk is read-only in the host, the guest should obey */
 	if (virtio_has_feature(vdev, VIRTIO_BLK_F_RO))

Classic mail. Thanks for the massive and coherent clue injection!

may make us safer". Of course, it may not help: how much does it cost to drain the queue?

The bug, IMHO is that we *should* know. And in future I'd like to fix that, either by adding an VIRTIO_BLK_F_ORDERED feature, or a VIRTIO_BLK_F_UNORDERED feature.

So, virtio-blk needs to be enhanced for this as well.

Thanks!
Rusty.

Really, enabling volatile write caches without advertising a cache flush command is a bug in the storage, where in our case qemu is the storage. So I don't really see the need for two feature bits. Here's my plan for virtio-blk:

- add a new VIRTIO_BLK_F_WCACHE feature. If this feature is set we do
(a) implement the prepare_flush queue operation to send a standalone cache flush
(b) set a proper barrier ordering flag on the queue

Now I'm not entirely sure which queue ordering feature we will use. It is not going to be QUEUE_ORDERED_TAG as for VIRTIO_BLK_F_BARRIER as that leaves all the queue draining to the host. Which for everything that uses something resembling Posix I/O as a backed and has more than one outstanding command at a time just means duplicating all the queue management we already do in the guest for no gain.
The easiest one would be QUEUE_ORDERED_DRAIN_FLUSH, in which case the cache flush command really is everything we need. As a slight optimization of it we could make it QUEUE_ORDERED_DRAIN_FUA which still does all the queue draining in the guest, but only sends one explicit cache flush before the barrier and gthen sets the FUA bit on the actual barrier request. In qemu we still would implement this as fdatasync before and after the request, but we would save one protocol roundtrip.

Now the big question is when do we set the VIRTIO_BLK_F_WCACHE feature. The proper thing to do would be to set it for cache=writeback and cache=none, because they do need the fdatasync, and not for cache=writethrough because it does not require it.

Now Avi is a big advocate for the cache=writethrough should mean go fast and loose and don't care about data integrity. There's a certain point to that as I don't really see a good use case for that mode, but I really hate to make something unsafe that doesn't explicitly say so in the option name.

The complex (not to say over engineered) verison would be to split the caching and data integrity setting into two options:

(1) hostcache=on|off
use buffered vs O_DIRECT I/O
(2) integrity=osync|fsync|none
use O_SYNC, use f(data)sync or do not care about data integrity

OK, I buy that. I'll update the virtio_pci spec accordingly, too.

I've applied your previous patch.

Thanks for taking this on. Both this email and the one on linux-fsdevel about Linux behaviour are wonderfully clear summaries of the issues.

With IDE and SCSI, and perhaps virtio-blk, guests should also be able to disable the "write cache enabled" bit, and that should be equivalent to the guest issuing a cache flush command after every write.

At the host it could be implemented as if every write were followed by flush, or by switching to O_DSYNC (cache=writethrough) in response.

The other way around: for guests where integrity isn't required (e.g. disposable guests for testing - or speed during guest OS installs), you might want an option to ignore cache flush commands - just let the guest *think* it's committing to disk, but don't waste time doing that on the host.

For fdatasync (etc), you've probably noticed that it only needs one cache flush by itself, no second request or FUA write.

Less obviously, there are opportunities to merge and reorder around non-barrier flush requests in the elevator, and to eliminate redundant flush requests.

Also you don't need flushes to reach every backing drive on RAID, but knowing which ones to leave out is tricky and needs more hints from the filesystem.

I agree with the whole of your general plan, both in QEMU and in Linux as a host. Spot on!

-- Jamie