On 2016-01-29 17:06, Henk Slager wrote:
On Fri, Jan 29, 2016 at 9:40 PM, Austin S. Hemmelgarn
<ahferroin7@xxxxxxxxx> wrote:
On 2016-01-29 15:27, Henk Slager wrote:
On Fri, Jan 29, 2016 at 1:14 PM, Austin S. Hemmelgarn
<ahferroin7@xxxxxxxxx> wrote:
On 2016-01-28 18:01, Chris Murphy wrote:
On Thu, Jan 28, 2016 at 1:44 PM, Austin S. Hemmelgarn
<ahferroin7@xxxxxxxxx> wrote:
Interesting, I figured a umount should include telling the drive to
flush the write cache; but maybe not, if the drive or connection (i.e.
USB enclosure) doesn't support FUA?
It's supposed to send an FUA, but depending on the hardware, this may
either
disappear on the way to the disk, or more likely just be a no-op. A
lot
of
cheap older HDD's just ignore it, and I've seen a lot of USB enclosures
that
just eat the command and don't pass anything to the disk, so sometimes
you
have to get creative to actually flush the cache. It's worth noting
that
most such disks are not safe to use BTRFS on anyway though, because FUA
is
part of what's used to force write barriers.
Err. Really?
[ 0.833452] scsi 0:0:0:0: Direct-Access ATA Samsung SSD
840 DB6Q PQ: 0 ANSI: 5
[ 0.835810] ata3.00: ACPI cmd ef/10:03:00:00:00:a0 (SET FEATURES)
filtered out
[ 0.835827] ata3.00: configured for UDMA/100
[ 0.838010] usb 1-1: new high-speed USB device number 2 using
ehci-pci
[ 0.839785] sd 0:0:0:0: Attached scsi generic sg0 type 0
[ 0.839810] sd 0:0:0:0: [sda] 488397168 512-byte logical blocks:
(250 GB/233 GiB)
[ 0.840381] sd 0:0:0:0: [sda] Write Protect is off
[ 0.840393] sd 0:0:0:0: [sda] Mode Sense: 00 3a 00 00
[ 0.840634] sd 0:0:0:0: [sda] Write cache: enabled, read cache:
enabled, doesn't support DPO or FUA
This is not a cheap or old HDD. It's not in an enclosure. I get the
same message for a new Toshiba 1TiB drive I just stuck in a new Intel
NUC. So now what?
Well, depending on how the kernel talks to the device, there are ways
around
this, but most of them are slow (like waiting for the write cache to
drain).
Just like SCT ERC, most drives marketed for 'desktop' usage don't
actually
support FUA, but they report this fact correctly, so the kernel can often
work around it. Most of the older drives that have issues actually
report
that they support it, but just treat it like a no-op. Last I checked,
Seagate's 'NAS' drives and whatever they've re-branded their other
enterprise line as, as well as WD's 'Red' drives support both SCT ERC and
FUA, but I don't know about any other brands (most of the Hitachi,
Toshiba,
and Samsung drives I've seen do not support FUA). This is in-fact part
of
the reason I'm saving up to get good NAS rated drives for my home server,
because those almost always support both SCT ERC and FUA.
[ 0.895207] sd 2:0:0:0: [sdc] Write cache: enabled, read cache:
enabled, doesn't support DPO or FUA
SCT ERC is supported though.
This is a 4TB (64MB buffer size) WD40EFRX-68WT0N0 FirmWare 82.00A82
and sold as 'NAS' drive.
That is at the same time troubling and not all that surprising (
SSD's don't implement it so why should we?' I hate marketing idiocy...). I
was apparently misinformed about WD's disks (although given the apparent
insanity of the firmware on some of their drives, that really doesn't
surprise me either).
How long do you think data will stay dirty in the drives writebuffer
(average/min/max)?
That depends on a huge number of factors, and I don't really have a good
answer. The 1TB 7200RPM single platter Seagate drives I'm using right now
(which have a 64MB cache) take less than 0.1 second for streaming writes,
and less than 0.5 on average for scattered writes, so it's not too bad most
of the time, but it's still a performance hit, and I do get marginally
better performance by turning off the on-disk write-cache (I've got a very
atypical workload though, so YMMV).
I think you refer to the transfer from PC main RAM via SATA to 64MB buffer.
What I try to estimate is the transfer-time from 64MB buffer to the
platter(s). Indeed a huge number of factors and without insight in the
drives ASIC/firmware design, just assumptions, but anyhow I am giving
it a try:
Ah, I misunderstood what you meant, sorry for the confusion.
- min: assume complete 64MB dirty and 1 sequential datablock in outer
cyl, no seek done, then 64 / 150 = ~0.5s
- max: assume only 1 physical sector sized max scattered (all non
sequential) datablocks, 150MB/s outer cyl write speed, 75MB/s inner
cyl write speed, 4ms avg seektime, no merging writes per head
position, 1 (side) platter, then
( 4k / 150M ) * 8k = ~200ms +
( 4k / 75M ) * 8k = ~400ms +
16k * 4ms = ~64s,
so in total more than 1 minute in this very simple and worst-case
model. Drive firmware can't be so inefficient, so seeks are probably
mostly mitigated, so then it is likely around 1s or a few seconds.
This all would mean that after default 30s commit in btrfs, the
drive's powersupply must not fail for 0.5s..few seconds.
If there is powerloss in this timeframe, the fs can get corrupt, but
AFAIU, there is previous roots, generations etc that can be used, such
that btrfs fs can restart without mount failure etc, just possibly 30
+ few seconds dataloss.
In theory, that entirely depends on how the drive batches and possibly
reorders writes in cache. If things get reordered poorly, then it's
fully possible to have all your SB's pointing at invalid tree roots.
The point of FUA as used by most filesystems is to act as a _very_
strong write barrier (it's supposed to flush the write-cache, thus
anything before it can't get reordered after it).
So if those calculations make sense, I am concluding that I am not
that worried about lack of FUA in normal (non-SMR) spinning drives.
Understandable, the failure modes it's supposed to be protecting against
are relatively rare, so unless you are working with data that you can't
afford to need to restore from backup or are using a system that
absolutely has to come back online without administrative intervention
after an unclean shutdown, it's usually not needed.
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
