random i/o error without error in dmesg

random i/o error without error in dmesg

[Szalma László]

Hi,

I have this error for a time, It's not easy to reproduce, i write 
everything i know at the moment.

I maintain some servers running xen (4.5.1) and gentoo dom0 with recent 
kernels (3.18.*, 4.1.6, 4.2.3, 4.2.4). I use gentoo-sources patchset.
Running xen domu s, for www and mysql.
I have mysql servers in domu with high load (lots of read write). These 
systems are identical in term of configuration and kernel.

Sometimes I got mysql errors randomly (sometimes more than one at a day, 
sometimes one at a week), but it is more frequent on high load.

The mysql errors are because the file cannot be read from the 
filesystem. If i try to run md5sum on it it shows io error.

At this point mysql stop && umount && mount && mysql start solves the 
problem.

calling
echo 3 > /proc/sys/vm/drop_caches
sometimes solves the io error, but not every time. The problem rarely 
randomly fixed without remount.

The problem seems to have no connection to the dom0 kernel and the xen 
version. I have this problem for example on these dom0 -s:

kernel: 3.19.3  xen 4.5.0
kernel: 4.2.3 xen 4.5.1

The problem seems to have started with the kernel 4.0 series, but I am 
not sure. In the summer the load was low, and the problem occured very 
rarely.

In this case of io error:
btrfs scrub finds no error.
no memory or hdd/ssd hardware error (smart, memtest, etc) (not only one 
physical server is affected) and no errors in dmesg at all.
tried different kernel configs, but I don't think I have anything 
extraordinary.
I use deadline scheduler.
I use these mount options:

/dev/xvdb1 on /mnt/mysql_naplo_b2 type btrfs 
(rw,noatime,compress=zlib,nossd,noacl,space_cache,subvolid=5,subvol=/)

I tried to reformat the filesystem with recent btrfs-progs: (and olders 
before)
btrfs-progs v4.2.2
I use default mkfs options (skinny extents)
After format the problem was disappeared for some days. (it seems 
correlation with the age of the filesystem?)
I do manual defragment on the filesystem with a script simply 
recursively check "filefrag" for count the fragmentation and defrag if 
it is more than 50 and the file is larger than 64kbyte. (this sometimes 
lowers the frequency of the problem)
The files unreadable are usually small files, for example:

filefrag:
/mnt/mysql_naplo_b2/mozanaplo_boly_altisk_2015/n_helyettesites.MYD: 2 
extents found
ls -l:
-rw-rw---- 1 mysql mysql 8092 okt   22 08.24 
/mnt/mysql_naplo_b2/mozanaplo_boly_altisk_2015/n_helyettesites.MYD

There is no error in dmesg, no io errors, no kernel panic, etc at all.

The (virtual) servers has 3-4GB of memory, and I use a 2GB tmpfs for the 
temporary tables (this way the physical memory usage is somewhat hectic).

The filesystem has no snapshots, but sometimes (for rebuilding 
replication) I take on, and delete it. (but the problem happens on 
filesystems with no snapshot created ever)

I did not try downgrading the kernel (for 3.18), but I always try to 
upgrade.

I guess this problem has some connection to the memory usage (but there 
is no out of memory).

I am able to try any debug mode if you suggest one, but it's not 
reproducable, it happens randomly. I think there should be some errors 
in the dmesg if I encounter io errors, but I am not sure if this error 
has direct connection for btrfs at all. I didn't try other filesystems. 
The problem was occured with kernel versions: 4.0.1, 4.0.4, 4.1.6, 
4.2.1, 4.2.3, 4.2.4.

I checked the bugzilla, and google for similar problem, but I couldn't 
find any similar.

This problem sometimes (i think it is the same) happen on a www server 
too, with apache log files (they are fragmented heavily), but very 
rarely. I don't have any problem with this configuration on other 
servers even mysql servers with lower load.

I welcome any suggestion:

László Szalma

Re: random i/o error without error in dmesg

[Marc Joliet]

[-- Attachment #1: Type: text/plain, Size: 3398 bytes --]

Hi

FWIW, this sounds like what I've been seeing with dovecot.  In case it's 
relevant, I'll try to explain.

After some uptime, I'll see log messages like this:

Okt 26 12:05:46 thetick dovecot[467]: imap(marcec): Error: pread() failed with 
file /home/marcec/.mdbox/mailboxes/BTRFS/dbox-Mails/dovecot.index.log: 
Input/output error

Occasionally they go away by themselves, but usually I have to reboot to make 
them go away.  This happens when getmail attempts to fetch mail, which fails 
due to the above error.  After the reboot getmail succeeds again.

As in Szalma's case, btrfs-scrub never reports anything wrong.

I use LZO compression on the relevant file system, so I wanted to wait until 
kernel 4.1.11 before reporting this, but that hasn't hit Gentoo yet (and 
neither has 4.1.10, for some reason).  I don't use quotas.

According to the what I see in the systemd journal, the errors started on 
2015-06-01 with kernel 3.19.8.  Note that, strangely enough, I had been using 
that same version since 2015-05-23, so for more than a week before the error 
cropped up.  I checked whether I made any changes to the configuration, and 
found this:

diff --git a/kernels/kernel-config-3.19.8-gentoo b/kernels/kernel-
config-3.19.8-gentoo
index b061b31..8cf8eba 100644
--- a/kernels/kernel-config-3.19.8-gentoo
+++ b/kernels/kernel-config-3.19.8-gentoo
@@ -64,7 +64,7 @@ CONFIG_INIT_ENV_ARG_LIMIT=32
 CONFIG_CROSS_COMPILE=""
 # CONFIG_COMPILE_TEST is not set
 CONFIG_LOCALVERSION=""
-CONFIG_LOCALVERSION_AUTO=y
+# CONFIG_LOCALVERSION_AUTO is not set
 CONFIG_HAVE_KERNEL_GZIP=y
 CONFIG_HAVE_KERNEL_BZIP2=y
 CONFIG_HAVE_KERNEL_LZMA=y
@@ -73,8 +73,8 @@ CONFIG_HAVE_KERNEL_LZO=y
 CONFIG_HAVE_KERNEL_LZ4=y
 # CONFIG_KERNEL_GZIP is not set
 # CONFIG_KERNEL_BZIP2 is not set
-CONFIG_KERNEL_LZMA=y
-# CONFIG_KERNEL_XZ is not set
+# CONFIG_KERNEL_LZMA is not set
+CONFIG_KERNEL_XZ=y
 # CONFIG_KERNEL_LZO is not set
 # CONFIG_KERNEL_LZ4 is not set
 CONFIG_DEFAULT_HOSTNAME="(none)"
@@ -132,7 +132,7 @@ CONFIG_TICK_CPU_ACCOUNTING=y
 # CONFIG_VIRT_CPU_ACCOUNTING_GEN is not set
 # CONFIG_IRQ_TIME_ACCOUNTING is not set
 CONFIG_BSD_PROCESS_ACCT=y
-# CONFIG_BSD_PROCESS_ACCT_V3 is not set
+CONFIG_BSD_PROCESS_ACCT_V3=y
 CONFIG_TASKSTATS=y
 CONFIG_TASK_DELAY_ACCT=y
 CONFIG_TASK_XACCT=y

The only change I can think of that might affect anything is 
CONFIG_BSD_PROCESS_ACCT_V3=y (I don't remember why exactly I set it).  I can 
try without it set, but maybe the kernel configuration is a red herring?

Anyway, the current state of the system is:

# uname -r 
4.1.9-gentoo-r1
# btrfs filesystem show / 
Label: 'MARCEC_ROOT'  uuid: 0267d8b3-a074-460a-832d-5d5fd36bae64
        Total devices 1 FS bytes used 74.40GiB
        devid    1 size 107.79GiB used 105.97GiB path /dev/sda1

btrfs-progs v4.2.2
# btrfs filesystem df /
Data, single: total=98.94GiB, used=72.30GiB
System, single: total=32.00MiB, used=20.00KiB
Metadata, single: total=7.00GiB, used=2.10GiB
GlobalReserve, single: total=512.00MiB, used=0.00B

The filesystem is mounted as (leaving out subvolume mounts which use the same 
mount options):
/dev/sda1 on / type btrfs (rw,noatime,compress=lzo,ssd,discard,space_cache)

Greetings,
-- 
Marc Joliet
--
"People who think they know everything really annoy those of us who know we
don't" - Bjarne Stroustrup

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 819 bytes --]

Re: random i/o error without error in dmesg

[Duncan]

Marc Joliet posted on Mon, 26 Oct 2015 15:23:39 +0100 as excerpted:

> Occasionally they go away by themselves, but usually I have to reboot to
> make them go away.  This happens when getmail attempts to fetch mail,
> which fails due to the above error.  After the reboot getmail succeeds
> again.

Just out of curiosity, does a remount,ro, followed by a remount,rw, solve 
the problem?

The ro/rw cycle should force anything in memory to device, so if that 
eliminates the problem, it could well be some sort of sync issue.  If it 
doesn't, then it's more likely an in-memory filesystem state issue, 
that's cleared by the reboot.

And if the ro/rw cycle doesn't clear the problem, what about a full 
unmount/mount cycle, at least of that subvolume?

If you're running multiple subvolumes with root being one of them, you 
can't of course unmount the entire filesystem, but you could go down to 
emergency mode (systemctl emergency), try unmounting everything but /, 
mounting / ro, and then switching back to normal mode (from emergency 
mode, simply exiting should return you to normal multi-user or gui 
target, remounting filesystems as necessary, etc).

IOW, does it take a full reboot to clear the problem, or is a simple ro/rw 
mount cycle enough, or an unmount/remount?

Finally, assuming root itself isn't btrfs, if you have btrfs configured 
as a module, you could try unmounting all btrfs and then unloading the 
module, then reloading and remounting.  That should entirely clear all in-
memory btrfs state, so if that doesn't solve the problem, while rebooting 
does, then the problem's very possibly outside of btrfs scope.  Of course 
if root is btrfs, you can't really check that.

-- 
Duncan - List replies preferred.   No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master."  Richard Stallman

Re: random i/o error without error in dmesg

[Marc Joliet]

[-- Attachment #1: Type: text/plain, Size: 2126 bytes --]

On Tuesday 27 October 2015 06:23:06 Duncan wrote:
>Marc Joliet posted on Mon, 26 Oct 2015 15:23:39 +0100 as excerpted:
>> Occasionally they go away by themselves, but usually I have to reboot to
>> make them go away.  This happens when getmail attempts to fetch mail,
>> which fails due to the above error.  After the reboot getmail succeeds
>> again.
>
>Just out of curiosity, does a remount,ro, followed by a remount,rw, solve
>the problem?
>
>The ro/rw cycle should force anything in memory to device, so if that
>eliminates the problem, it could well be some sort of sync issue.  If it
>doesn't, then it's more likely an in-memory filesystem state issue,
>that's cleared by the reboot.
>
>And if the ro/rw cycle doesn't clear the problem, what about a full
>unmount/mount cycle, at least of that subvolume?
>
>If you're running multiple subvolumes with root being one of them, you
>can't of course unmount the entire filesystem, but you could go down to
>emergency mode (systemctl emergency), try unmounting everything but /,
>mounting / ro, and then switching back to normal mode (from emergency
>mode, simply exiting should return you to normal multi-user or gui
>target, remounting filesystems as necessary, etc).
>
>IOW, does it take a full reboot to clear the problem, or is a simple ro/rw
>mount cycle enough, or an unmount/remount?
>
>Finally, assuming root itself isn't btrfs, if you have btrfs configured
>as a module, you could try unmounting all btrfs and then unloading the
>module, then reloading and remounting.  That should entirely clear all in-
>memory btrfs state, so if that doesn't solve the problem, while rebooting
>does, then the problem's very possibly outside of btrfs scope.  Of course
>if root is btrfs, you can't really check that.

Thanks for the hints.  I just upgraded to gentoo-sources 4.1.11 and will see 
if that changes anything.  If not, I'll have to try unmounting /home from 
emergency mode (it's a subvolume mount).

-- 
Marc Joliet
--
"People who think they know everything really annoy those of us who know we
don't" - Bjarne Stroustrup

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 819 bytes --]

Re: random i/o error without error in dmesg

[Szalma László]

I wait for the problem to occur, i will try the remount,ro remount,rw. 
But I know for sure, in my case umount / mount solved the problem every 
time!

Reboot was not needed. (I think reboot is simpler because the fs is 
used, and the service has to be stopped etc.)

László Szalma

2015-10-27 07:23 keltezéssel, Duncan írta:
> Marc Joliet posted on Mon, 26 Oct 2015 15:23:39 +0100 as excerpted:
>
>> Occasionally they go away by themselves, but usually I have to reboot to
>> make them go away.  This happens when getmail attempts to fetch mail,
>> which fails due to the above error.  After the reboot getmail succeeds
>> again.
> Just out of curiosity, does a remount,ro, followed by a remount,rw, solve
> the problem?
>
> The ro/rw cycle should force anything in memory to device, so if that
> eliminates the problem, it could well be some sort of sync issue.  If it
> doesn't, then it's more likely an in-memory filesystem state issue,
> that's cleared by the reboot.
>
> And if the ro/rw cycle doesn't clear the problem, what about a full
> unmount/mount cycle, at least of that subvolume?
>
> If you're running multiple subvolumes with root being one of them, you
> can't of course unmount the entire filesystem, but you could go down to
> emergency mode (systemctl emergency), try unmounting everything but /,
> mounting / ro, and then switching back to normal mode (from emergency
> mode, simply exiting should return you to normal multi-user or gui
> target, remounting filesystems as necessary, etc).
>
> IOW, does it take a full reboot to clear the problem, or is a simple ro/rw
> mount cycle enough, or an unmount/remount?
>
> Finally, assuming root itself isn't btrfs, if you have btrfs configured
> as a module, you could try unmounting all btrfs and then unloading the
> module, then reloading and remounting.  That should entirely clear all in-
> memory btrfs state, so if that doesn't solve the problem, while rebooting
> does, then the problem's very possibly outside of btrfs scope.  Of course
> if root is btrfs, you can't really check that.
>

Re: random i/o error without error in dmesg

[Marc Joliet]

[-- Attachment #1: Type: text/plain, Size: 2737 bytes --]

OK, upgrading to gentoo-sources 4.1.11 didn't help, so I tried these steps.  
More inline below.

On Tuesday 27 October 2015 06:23:06 Duncan wrote:
>Marc Joliet posted on Mon, 26 Oct 2015 15:23:39 +0100 as excerpted:
>> Occasionally they go away by themselves, but usually I have to reboot to
>> make them go away.  This happens when getmail attempts to fetch mail,
>> which fails due to the above error.  After the reboot getmail succeeds
>> again.
>
>Just out of curiosity, does a remount,ro, followed by a remount,rw, solve
>the problem?
>
>The ro/rw cycle should force anything in memory to device, so if that
>eliminates the problem, it could well be some sort of sync issue.  If it
>doesn't, then it's more likely an in-memory filesystem state issue,
>that's cleared by the reboot.

Didn't try this directly, but...

>And if the ro/rw cycle doesn't clear the problem, what about a full
>unmount/mount cycle, at least of that subvolume?

...this didn't work, after which...

>If you're running multiple subvolumes with root being one of them, you
>can't of course unmount the entire filesystem, but you could go down to
>emergency mode (systemctl emergency), try unmounting everything but /,
>mounting / ro, and then switching back to normal mode (from emergency
>mode, simply exiting should return you to normal multi-user or gui
>target, remounting filesystems as necessary, etc).

...I tried most of this.  I unmounted everything except for /var and /, 
neither of which I could mount read-only.  It didn't help, either.

>IOW, does it take a full reboot to clear the problem, or is a simple ro/rw
>mount cycle enough, or an unmount/remount?

Seems that a full reboot is needed, but I would expect that it would have the 
same effect if I were to pivot back into the initramfs, unmount / from there, 
then boot back into the system.  Because quite frankly, I can't think of any 
reason why a power cycle to the SSD should make a difference here.  I vaguely 
remember that systemd can do that, so I'll see if I can find out how.

>Finally, assuming root itself isn't btrfs, if you have btrfs configured
>as a module, you could try unmounting all btrfs and then unloading the
>module, then reloading and remounting.  That should entirely clear all in-
>memory btrfs state, so if that doesn't solve the problem, while rebooting
>does, then the problem's very possibly outside of btrfs scope.  Of course
>if root is btrfs, you can't really check that.

Nope, btrfs is built-in (though it doesn't have to be, what with me using an 
initramfs).

Thanks
-- 
Marc Joliet
--
"People who think they know everything really annoy those of us who know we
don't" - Bjarne Stroustrup

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 819 bytes --]

Re: random i/o error without error in dmesg

[Duncan]

Marc Joliet posted on Tue, 27 Oct 2015 21:54:40 +0100 as excerpted:

>>IOW, does it take a full reboot to clear the problem, or is a simple
>>ro/rw mount cycle enough, or an unmount/remount?
> 
> Seems that a full reboot is needed, but I would expect that it would
> have the same effect if I were to pivot back into the initramfs, unmount
> / from there,
> then boot back into the system.  Because quite frankly, I can't think of
> any reason why a power cycle to the SSD should make a difference here. 
> I vaguely remember that systemd can do that, so I'll see if I can find
> out how.

Agree with both the systemd returning to the initr* point (which I 
actually had in mind while writing the above but don't remember the 
details either, so chose to omit in the interest of limiting the size of 
the reply and research necessary to generate it), and the ssd power-cycle 
point.

>>Finally, assuming root itself isn't btrfs, if you have btrfs configured
>>as a module, you could try unmounting all btrfs and then unloading the
>>module, then reloading and remounting.  That should entirely clear all
>>in-memory btrfs state, so if that doesn't solve the problem, while
>>rebooting does, then the problem's very possibly outside of btrfs scope.
>> Of course if root is btrfs, you can't really check that.
> 
> Nope, btrfs is built-in (though it doesn't have to be, what with me
> using an initramfs).

Same here, also gentoo as I guess you know from previous exchanges.  But 
unfortunately, if your initr* is anything like mine, and your kernel 
monolithic as mine, making btrfs a module with a btrfs root isn't the 
easy thing it might seem to those who run ordinary distro supplied binary 
kernels with pretty much everything modularized, as doing so involves a 
whole new set of research on how to get that module properly included in 
the initr* and loaded there, as well as installing and building the whole 
module-handling infrastructure (modprobe and friends) again, as it's not 
actually installed on the system at all at this point, because with the 
kernel entirely monolithic, module-handling tools are unnecessary and 
thus just another unnecessary package to have to keep building updates 
for, if they remain installed.

So I definitely sympathize with the feeling that such a stone is better 
left unturned, if overturning it is at all a possibility that can be 
avoided, as it is here, this whole exercise being simply one of better 
pinning the bug down, not yet actually trying to solve it.  And given 
that unturned stone, there are certainly easier ways.

And one of those easier ways is investigating that whole systemd return 
to initr* idea, since we both remember reading something about it, but 
aren't familiar with the details.  In addition to addressing the problem 
headon if anyone offers a way to do so, that's the path I'd be looking at 
right now.


Meanwhile, addressing something in the snipped content, if a mount 
refuses to remount read-only, it's normally due to deleted but still open 
files.  Often, in many cases the overwhelming majority of the time, 
that's due to updates that have been done, removing for instance old 
library files that are still loaded into running executables.  The 
generalized solution is to quit or restart the affected executables, thus 
releasing the last open references to the otherwise already deleted 
library and other files, so the filesystem can finish deleting them.  
Once this is done, the filesystem can normally be remounted ro (tho bugs 
like the one of this thread may prevent that).

The problem lies in actually figuring out what still running executables 
are still holding open file references to otherwise deleted files.  /proc/ 
actually makes this sort of data available in its collection of files for 
a running process, as there's one (as the saying goes, finding which one 
is an exercise left for the reader, as I could look it up or check, but 
so can you if you're /that/ interested) that tells which files the 
process has open, and deleted files are clearly marked.

But it's still a hassle to go thru the appropriate /proc/ files for each 
process, at least if it's done manually, so various helper tools have 
been developed to automate the process.  The one I use here is a python-
based script, packaged on gentoo as...

app-admin/lib_users

Run as root, it'll tell you which executables are running that still hold 
references to deleted files, so you can restart them.  Of course when a 
library was updated for security reasons doing this restart is a good 
idea from that perspective as well.  Some tools actually have a table of 
executables and the services they belong to, and will automate the 
restart, but of course these sorts of tools tend to be somewhat complex 
and distro specific, while lib_users simply tells you the executables and 
lets you decide what to do with that information, that being both simpler 
and without the distro-specifics of the lists that the fully automated 
tools must use.

If you do your updates while an X-based session is running and run 
lib_users before shutting it down, you'll often find a whole slew of x-
based executables listed, so much so that I don't even bother until I've 
shut down X, these days.  That leaves only services and longer running CLI 
executables, perhaps including the shell itself, in the list.

If it's only a couple higher-level services listed, it's generally 
simplest to just restart them.  If it's many services or lower level 
services such that restarting them will disturb the services depending on 
them[1], from my experience, it's often easier to simply do a systemctl 
emergency, perhaps even hitting ctrl-D straight from the resulting root 
password prompt, to go immediately back to multi-user mode, as even 
without entering the password, the drop to emergency mode will have 
restarted all normal services.

What's left in the lib_user list after a dip to emergency mode, whether 
you've returned to multi-user mode or not, is often only systemd and 
perhaps its journald service, itself.  Journald can be restarted manually 
in the usual way (systemctl restart journald.service or whatever, I 
usually use tab completion so don't pay all that much attention to the 
specific name), if necessary, while systemd itself can be "reexecuted" 
using the systemctl daemon-reexec (tab-completion again) command.

With a drop out of X mode, a dip to emergency mode and reexecing systemd 
itself, at least in my experience with what I'm running here, lib_users 
normally reports nothing further, and a remount read-only succeeds.

Of course, if lib_users reports nothing further still holding references 
to deleted files, and a remount read-only STILL fails, that's a major 
note of trouble and an important finding in itself.


Meanwhile, as explained in the systemd docs (specifically the systemd for 
administrators series, IIRC), systemd dropping back to the initr* is 
actually its way of automatically doing effectively the same thing we 
were using lib_users and all those restarts to do, getting rid of all 
possible still running on root executables, including systemd itself, by 
reexecing systemd itself back in the initr*, as a way to help eliminate 
*everything* running on root, so it can not only be remounted read-only, 
but actually unmounted the same as any other filesystem, as userspace is 
now actually running from the initr* once again.  That's a far *FAR* 
safer reboot after upgrade than traditional sysvinit solutions were able 
to do. =:^)

So the whole systemd switchroot back to initr* thing happens to be 
actually linked in here too.  Tho in this case we'd be using it for more, 
and indeed, there are systemd docs on how to get the reboot process to 
stop in in the initr*[1].  While I don't recall actually seeing anything 
about short-circuiting the reboot and restarting from the initr*, it 
should certainly be possible to configure systemd and the initr* to allow 
it.

And I might have it configured that way here too, to short-circuit the 
reboot, were it not for the two facts that I (1) often use a dip to 
emergency mode (and reexec of systemd if necessary) for many of the 
reasons one might otherwise dip to initr*, and (2) generally track pre-
release kernels, so a lot of the time when I reboot, it's to start a 
freshly built kernel, and the dip to initr* wouldn't help with that.

---
[1] Low-level services:  Low-level services aren't so much a problem with 
systemd, since it generally keeps thinks like sockets available and hooks 
them back up to the restarted service, without disturbing the higher 
level services that would otherwise need a restart too, at all.  But the 
shear number of otherwise manual restarts necessary can still make it 
easier to simply dip to emergency mode temporarily, before going normal 
multi-user again.

[2] Stopping the reboot process back in the initr*, before final 
restart:  I did actually try to get this working at one point, but due to 
some particulars specific to my own rather strange layout, certain initr* 
related systemd service file symlinks were orphaned at the time, and it 
didn't work as it was designed and documented to work.  I've since fixed 
those symlink issues, but I've not tried again to stop in the initr* on 
the way down, so I've never actually seen it do that, tho if I'm fast 
enough and the reboot slow enough, I can at least see the return to initr* 
stuff scrolling by now before it blinks out, and it wasn't doing that 
before, so I expect I could actually get it to stop in in the initr* on 
the way down, now, if I tried, by following the documentation.  Anyway, 
that's why I'm not offering specific help on that aspect, because it was 
broken for me when I tried it, and while I believe it fixed now, I've not 
tried it since the fix, so I don't have the actual benefit of experience 
to offer.
-- 
Duncan - List replies preferred.   No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master."  Richard Stallman

Re: random i/o error without error in dmesg

[Austin S Hemmelgarn]

[-- Attachment #1: Type: text/plain, Size: 3685 bytes --]

On 2015-10-28 01:21, Duncan wrote:
> Marc Joliet posted on Tue, 27 Oct 2015 21:54:40 +0100 as excerpted:
>
>>> IOW, does it take a full reboot to clear the problem, or is a simple
>>> ro/rw mount cycle enough, or an unmount/remount?
>>
>> Seems that a full reboot is needed, but I would expect that it would
>> have the same effect if I were to pivot back into the initramfs, unmount
>> / from there,
>> then boot back into the system.  Because quite frankly, I can't think of
>> any reason why a power cycle to the SSD should make a difference here.
>> I vaguely remember that systemd can do that, so I'll see if I can find
>> out how.
>
> Agree with both the systemd returning to the initr* point (which I
> actually had in mind while writing the above but don't remember the
> details either, so chose to omit in the interest of limiting the size of
> the reply and research necessary to generate it), and the ssd power-cycle
> point.
>
>>> Finally, assuming root itself isn't btrfs, if you have btrfs configured
>>> as a module, you could try unmounting all btrfs and then unloading the
>>> module, then reloading and remounting.  That should entirely clear all
>>> in-memory btrfs state, so if that doesn't solve the problem, while
>>> rebooting does, then the problem's very possibly outside of btrfs scope.
>>> Of course if root is btrfs, you can't really check that.
>>
>> Nope, btrfs is built-in (though it doesn't have to be, what with me
>> using an initramfs).
>
> Same here, also gentoo as I guess you know from previous exchanges.  But
> unfortunately, if your initr* is anything like mine, and your kernel
> monolithic as mine, making btrfs a module with a btrfs root isn't the
> easy thing it might seem to those who run ordinary distro supplied binary
> kernels with pretty much everything modularized, as doing so involves a
> whole new set of research on how to get that module properly included in
> the initr* and loaded there, as well as installing and building the whole
> module-handling infrastructure (modprobe and friends) again, as it's not
> actually installed on the system at all at this point, because with the
> kernel entirely monolithic, module-handling tools are unnecessary and
> thus just another unnecessary package to have to keep building updates
> for, if they remain installed.
FWIW, I'm pretty sure that genkernel-next (and possibly genkernel too at 
this point, but that's never had working userland support for BTRFS 
AFAICT) includes btrfs.ko and it's dependencies in any initr* it 
generates when it sees those modules on the system, although I'm not 
100% certain because I always build the driver for my root filesystem 
into the kernel directly.
>
[ snip ]
> What's left in the lib_user list after a dip to emergency mode, whether
> you've returned to multi-user mode or not, is often only systemd and
> perhaps its journald service, itself.  Journald can be restarted manually
> in the usual way (systemctl restart journald.service or whatever, I
> usually use tab completion so don't pay all that much attention to the
> specific name), if necessary, while systemd itself can be "reexecuted"
> using the systemctl daemon-reexec (tab-completion again) command.
Udev might also be listed (for some reason, a lot of distros don't stop 
it when going to emergency mode, I don't know about what Gentoo does in 
this case though because I don't use systemd), and at least the last 
time I checked, dropping to emergency mode on Debian and Fedora testing 
versions also keeps any dhcp clients or other stuff needed for 
maintaining the network connection running as well.



[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 3019 bytes --]

Re: random i/o error without error in dmesg

[Marc Joliet]

[-- Attachment #1: Type: text/plain, Size: 5659 bytes --]

On Wednesday 28 October 2015 05:21:13 Duncan wrote:
>Marc Joliet posted on Tue, 27 Oct 2015 21:54:40 +0100 as excerpted:
>>>IOW, does it take a full reboot to clear the problem, or is a simple
>>>ro/rw mount cycle enough, or an unmount/remount?
>>>
>> Seems that a full reboot is needed, but I would expect that it would
>> have the same effect if I were to pivot back into the initramfs, unmount
>> / from there,
>> then boot back into the system.  Because quite frankly, I can't think of
>> any reason why a power cycle to the SSD should make a difference here.
>> I vaguely remember that systemd can do that, so I'll see if I can find
>> out how.
>
>Agree with both the systemd returning to the initr* point (which I
>actually had in mind while writing the above but don't remember the
>details either, so chose to omit in the interest of limiting the size of
>the reply and research necessary to generate it), and the ssd power-cycle
>point.

I haven't found any single command that lets you do that, but I can try one of 
the special targets as detailed in bootup(7) (e.g., initrd.target) when I have 
a chance.

>>>Finally, assuming root itself isn't btrfs, if you have btrfs configured
>>>as a module, you could try unmounting all btrfs and then unloading the
>>>module, then reloading and remounting.  That should entirely clear all
>>>in-memory btrfs state, so if that doesn't solve the problem, while
>>>rebooting does, then the problem's very possibly outside of btrfs scope.
>>>
>>> Of course if root is btrfs, you can't really check that.
>> 
>> Nope, btrfs is built-in (though it doesn't have to be, what with me
>> using an initramfs).
>
>Same here, also gentoo as I guess you know from previous exchanges.  But
>unfortunately, if your initr* is anything like mine, and your kernel
>monolithic as mine, making btrfs a module with a btrfs root isn't the
>easy thing it might seem to those who run ordinary distro supplied binary
>kernels with pretty much everything modularized, as doing so involves a
>whole new set of research on how to get that module properly included in
>the initr* and loaded there, as well as installing and building the whole
>module-handling infrastructure (modprobe and friends) again, as it's not
>actually installed on the system at all at this point, because with the
>kernel entirely monolithic, module-handling tools are unnecessary and
>thus just another unnecessary package to have to keep building updates
>for, if they remain installed.

My kernel is fairly modular, and I use dracut to make my initramfs, so I 
wouldn't be surprised if it works.  For me, personally, I just don't see any 
point in making btrfs a module.

(And yes, of course I know you run Gentoo ;-) .)

>So I definitely sympathize with the feeling that such a stone is better
>left unturned, if overturning it is at all a possibility that can be
>avoided, as it is here, this whole exercise being simply one of better
>pinning the bug down, not yet actually trying to solve it.  And given
>that unturned stone, there are certainly easier ways.
>
>And one of those easier ways is investigating that whole systemd return
>to initr* idea, since we both remember reading something about it, but
>aren't familiar with the details.  In addition to addressing the problem
>headon if anyone offers a way to do so, that's the path I'd be looking at
>right now.

Like I said above, I'll try it out when I have a moment where I have a more 
"steady hand" so-to-speak.

[snip deleted files stuff]
>app-admin/lib_users
[snip the rest of the deleted files stuff]

I use that to find processes that need restarting after upgrades, though I'll 
sometimes check to see if it's really a library that's causing it to show up, 
since often a process is listed because of stuff like the font cache, or, in 
the case of the FISH shell, it's own history file.

But yeah, didn't think of running that, but in rescue mode there were at most 
a dozen processes running, so there's not much to choose from, anyway.  I did 
have to kill two remaining user processes first (pulseaudio and... I forgot 
the other one).  I didn't try the same with / and /var because I was eager to 
get back to a normally running system ;-) .

>Of course, if lib_users reports nothing further still holding references
>to deleted files, and a remount read-only STILL fails, that's a major
>note of trouble and an important finding in itself.

I don't expect that, but I'll make note of it if I encounter it.

>Meanwhile, as explained in the systemd docs (specifically the systemd for
>administrators series, IIRC), systemd dropping back to the initr* is
>actually its way of automatically doing effectively the same thing we
>were using lib_users and all those restarts to do, getting rid of all
>possible still running on root executables, including systemd itself, by
>reexecing systemd itself back in the initr*, as a way to help eliminate
>*everything* running on root, so it can not only be remounted read-only,
>but actually unmounted the same as any other filesystem, as userspace is
>now actually running from the initr* once again.  That's a far *FAR*
>safer reboot after upgrade than traditional sysvinit solutions were able
>to do. =:^)

Yeah, the ability to do that is a nice plus of using an initramfs.  Although 
I've never been clear on why it's *safer*.  Is it because the remount might 
fail?  Or are there other reasons, too?

[...]

-- 
Marc Joliet
--
"People who think they know everything really annoy those of us who know we
don't" - Bjarne Stroustrup

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 819 bytes --]

Re: random i/o error without error in dmesg

[Duncan]

Marc Joliet posted on Thu, 29 Oct 2015 22:10:24 +0100 as excerpted:

>>Meanwhile, as explained in the systemd docs (specifically the systemd
>>for administrators series, IIRC), systemd dropping back to the initr* is
>>actually its way of automatically doing effectively the same thing we
>>were using lib_users and all those restarts to do, getting rid of all
>>possible still running on root executables, including systemd itself, by
>>reexecing systemd itself back in the initr*, as a way to help eliminate
>>*everything* running on root, so it can not only be remounted read-only,
>>but actually unmounted the same as any other filesystem, as userspace is
>>now actually running from the initr* once again.  That's a far *FAR*
>>safer reboot after upgrade than traditional sysvinit solutions were able
>>to do. =:^)
> 
> Yeah, the ability to do that is a nice plus of using an initramfs.
> Although I've never been clear on why it's *safer*.  Is it because the
> remount might fail?  Or are there other reasons, too?

While I don't claim anything but informed admin level authority on the 
problem...

It's first worth noting that the problem a return to initramfs helps 
solve is in practice reasonably rare and obscure, since if it weren't, 
people would have been experiencing it in serious numbers on sysvinit-
based systems all along, and something would have been done to solve it 
long before systemd came along.  So it's a relatively narrow issue that 
in practice can only affect a few users, a relatively small portion of 
the time.

>From my read of the systemd docs, it's more pointing out a theoretical 
issue than a practical one, pointing out that systemd is in fact a more 
theoretically correct solution to the (implicitly mostly theoretical) 
problem.

In that context, I believe the (mostly theoretical) point is as much that 
we were treating / (and perhaps another mount or two) special, remounting 
it read-only instead of unmounting it because in practice there wasn't 
any other choice, and that now that systemd offers the choice, it can in 
fact be treated just like any other filesystem, fully unmounting it 
before shutdown.

Since exceptions to rules are nice places for bugs to hide, in theory at 
least (the remount-ro root being such a universal exception that in 
practice it's a rule of its own, and bugs couldn't long hide in that 
exception /because/ of its universalness), being able to treat / like any 
other filesystem and unmount it is a "purer and more correct" solution.

IOW, it's a nice counter to the "systemd isn't unixy enough" point, as 
here, it's more "unixy" than sysvinit ever was.

That said, I expect that over the years there have been plenty of 
otherwise nice implementations of various useful things that ran into a 
shutdown/reboot-time problem due to root's remount-ro exception, that 
either limited them to non-root-filesystem deployment or sent them back 
for a workaround, if not causing them to rejected outright as unworkable, 
that in this new return-to-initr*-and-unmount-root environment will see 
faster deployment without the workarounds that heretofore were required.  
Of course that'll end up being a limitation on deployment on non-initr* 
direct-to-root boot sequences, but in this primarily prebuild binary 
distro with prebuild by-necessity-modular-kernel-and-initr* environment, 
that's unlikely to slow down wide deployment by much, and anyone wanting 
to do direct-to-root boots and/or non-systemd-based deployments will just 
have to find their own workarounds, which may ultimately be incorporated 
into upstream, or not, depending on upstream's whims.

Which, bringing it all back to the btrfs list title topic, is already 
where multi-device btrfs as / filesystem is in terms of initr*, since 
that's basically broken without an initr* to assemble it.  And of course 
the same thing goes for / on LVM, since it too requires userspace to 
activate, which means initr* if / is on it.

-- 
Duncan - List replies preferred.   No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master."  Richard Stallman

Re: random i/o error without error in dmesg

[Szalma László]

Ok, I had a chance to try some things.

1.: the error

md5sum xyz
md5sum: xyz: Input/output error

(no any errors in dmesg)

2.: mount -o remount,ro /mnt/x

(could not do, it is used)
mysql stop && mount -o remount,ro /mnt/x
problem persists: io error.
mount -o remount,rw /mnt/x
still io error
umount /mnt/x
mount /mnt/x
NO io error, md5sum works!

The umount/mount ALWAYS solved the problem for me, mount -o remount,ro 
was tried for the first time, but it was not enought. Reboot was not needed.
(kernel 4.2.4)

László Szalma


2015-10-27 07:23 keltezéssel, Duncan írta:
> Marc Joliet posted on Mon, 26 Oct 2015 15:23:39 +0100 as excerpted:
>
>> Occasionally they go away by themselves, but usually I have to reboot to
>> make them go away.  This happens when getmail attempts to fetch mail,
>> which fails due to the above error.  After the reboot getmail succeeds
>> again.
> Just out of curiosity, does a remount,ro, followed by a remount,rw, solve
> the problem?
>
> The ro/rw cycle should force anything in memory to device, so if that
> eliminates the problem, it could well be some sort of sync issue.  If it
> doesn't, then it's more likely an in-memory filesystem state issue,
> that's cleared by the reboot.
>
> And if the ro/rw cycle doesn't clear the problem, what about a full
> unmount/mount cycle, at least of that subvolume?
>
> If you're running multiple subvolumes with root being one of them, you
> can't of course unmount the entire filesystem, but you could go down to
> emergency mode (systemctl emergency), try unmounting everything but /,
> mounting / ro, and then switching back to normal mode (from emergency
> mode, simply exiting should return you to normal multi-user or gui
> target, remounting filesystems as necessary, etc).
>
> IOW, does it take a full reboot to clear the problem, or is a simple ro/rw
> mount cycle enough, or an unmount/remount?
>
> Finally, assuming root itself isn't btrfs, if you have btrfs configured
> as a module, you could try unmounting all btrfs and then unloading the
> module, then reloading and remounting.  That should entirely clear all in-
> memory btrfs state, so if that doesn't solve the problem, while rebooting
> does, then the problem's very possibly outside of btrfs scope.  Of course
> if root is btrfs, you can't really check that.
>

Re: random i/o error without error in dmesg

[Duncan]

Szalma László posted on Wed, 28 Oct 2015 09:44:12 +0100 as excerpted:

> The umount/mount ALWAYS solved the problem for me, mount -o remount,ro
> was tried for the first time, but it was not enought. Reboot was not
> needed.
> (kernel 4.2.4)

So that means it's filesystem state that's tracked thru a remount, not 
something like deleted/orphan files that a remount should square away.  
But a full unmount clears the state, without having to unload the btrfs 
kernel module or reboot.

Which at least limits the active zone of the problem.  Unfortunately, I'm 
not a dev and don't have a clue where to go from there... unless it's 
related to the recent delayed-refs bug, in which case I believe a late 
4.3 rc, or 4.3.0 when released, should fix it.  That fix should be CCed 
to stable and thus appear there eventually, but I normally track pre-
releases, not stable, so don't know how long it might be...  And again, I 
don't know it's even related, but it does appear to be the active bug of 
the moment, so in the absence of knowing, one can hope.

-- 
Duncan - List replies preferred.   No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master."  Richard Stallman

Re: random i/o error without error in dmesg

[Szalma László]

2015-10-28 09:44 keltezéssel, Szalma László írta:
> Ok, I had a chance to try some things.
>
> 1.: the error
>
> md5sum xyz
> md5sum: xyz: Input/output error
>
> (no any errors in dmesg)
>
> 2.: mount -o remount,ro /mnt/x
>
> (could not do, it is used)
> mysql stop && mount -o remount,ro /mnt/x
> problem persists: io error.
> mount -o remount,rw /mnt/x
> still io error
> umount /mnt/x
> mount /mnt/x
> NO io error, md5sum works!
>
> The umount/mount ALWAYS solved the problem for me, mount -o remount,ro 
> was tried for the first time, but it was not enought. Reboot was not 
> needed.
> (kernel 4.2.4)
>
> László Szalma
>

Unfortunately the problem with kernel 4.3.0 still exists.

László Szalma

Re: random i/o error without error in dmesg

[Philipp Serr]

On Mon, 2 Nov 2015 19:26:01 +0100, Szalma László wrote:
> Unfortunately the problem with kernel 4.3.0 still exists.


Any news on that issue? I'm currently @ 4.3.3 and the issue persists for 
me, too.

There's one more information I might be able to contribute: In my case 
always the same single file is affected. The issue is triggered mostly 
when backing up a read-only snapshot of the volume using rsync. Once the 
issue occures, the file within the snapshot volume as well as the file 
in the main volume can't be read anymore until reboot (didn't try 
remounting the affected root partition).

Re: random i/o error without error in dmesg

[Marc Joliet]

[-- Attachment #1: Type: text/plain, Size: 5005 bytes --]

Am Friday 22 April 2016
schrieb Marc Joliet <marcec@gmx.de>

>Hi
>
>FWIW, this sounds like what I've been seeing with dovecot.  In case it's
>relevant, I'll try to explain.
>
>After some uptime, I'll see log messages like this:
>
>Okt 26 12:05:46 thetick dovecot[467]: imap(marcec): Error: pread() failed
>with file /home/marcec/.mdbox/mailboxes/BTRFS/dbox-Mails/dovecot.index.log:
>Input/output error
>
>Occasionally they go away by themselves, but usually I have to reboot to make
>them go away.  This happens when getmail attempts to fetch mail, which fails
>due to the above error.  After the reboot getmail succeeds again.
>
>As in Szalma's case, btrfs-scrub never reports anything wrong.

FWIW, this still happens.  I tried various things, in order:

- set mmap_disable=yes (as mentioned in [0] and [1]),
- stop mounting with compress=lzo, and
- upgrade dovecot 2.2.19 to 2.2.23,

none of which helped.

I found very little on the web: [0] (in German) is fairly recent, but the 
error apparently just went away.  [1] is pretty old (2007) and the reporter 
was using unionfs.  I also found a russian forum entry [2], which relates to 
btrfs again, but if Google Translate is accurate enough, the user "solved" the 
problem by using XFS instead.

My current attempt is to use the maildir mailbox format (instead of mdbox) in 
the hope that it doesn't trigger this bug, while reverting the changes above 
(except for the dovecot upgrade).  It's "just" a home server, so I don't 
expect any major differences in performance (also since KMail probably uses 
akonadi for searching).

[...]
>Anyway, the current state of the system is:
>
># uname -r
>4.1.9-gentoo-r1
># btrfs filesystem show /
>Label: 'MARCEC_ROOT'  uuid: 0267d8b3-a074-460a-832d-5d5fd36bae64
>        Total devices 1 FS bytes used 74.40GiB
>        devid    1 size 107.79GiB used 105.97GiB path /dev/sda1
>
>btrfs-progs v4.2.2
># btrfs filesystem df /
>Data, single: total=98.94GiB, used=72.30GiB
>System, single: total=32.00MiB, used=20.00KiB
>Metadata, single: total=7.00GiB, used=2.10GiB
>GlobalReserve, single: total=512.00MiB, used=0.00B
>
>The filesystem is mounted as (leaving out subvolume mounts which use the same
>mount options):
>/dev/sda1 on / type btrfs (rw,noatime,compress=lzo,ssd,discard,space_cache)

I have since obtained an old Mac Mini and have been using it as a home server 
since early April, onto which I moved dovecot (using doveadm-sync(1)).  Its 
current state is:

diefledermaus etc # uname -r
4.4.8-gentoo
diefledermaus etc # btrfs --version
btrfs-progs v4.5.1
diefledermaus etc # btrfs fi show
Label: 'MACROOT'  uuid: 8b2b0242-5816-4529-9e88-6c82ffff2eaf
        Total devices 1 FS bytes used 12.29GiB
        devid    1 size 251.20GiB used 15.09GiB path /dev/sda3

Label: 'MARCEC_BACKUP'  uuid: f97b3cda-15e8-418b-bb9b-235391ef2a38
        Total devices 1 FS bytes used 797.14GiB
        devid    1 size 976.56GiB used 819.12GiB path /dev/sdb2

diefledermaus etc # btrfs fi df /
Data, single: total=12.01GiB, used=11.38GiB
System, DUP: total=40.00MiB, used=16.00KiB
Metadata, DUP: total=1.50GiB, used=926.77MiB
GlobalReserve, single: total=288.00MiB, used=0.00B
diefledermaus etc # mount | grep root
/dev/sda3 on / type btrfs 
(rw,noatime,compress=lzo,space_cache,autodefrag,subvolid=257,subvol=/rootfs)
/dev/sda3 on /mnt/rootfs type btrfs 
(rw,noatime,compress=lzo,space_cache,autodefrag,subvolid=5,subvol=/)

Note that the file system was *freshly* created using btrfs-progs 4.3.1 and 
gentoo-sources 4.1.15-r1 (the current stable versions in Gentoo; -r1 is a 
Gentoo-specific revision number).

References:

[0] https://debianforum.de/forum/viewtopic.php?f=27&t=157374
[1] http://www.dovecot.org/list/dovecot/2007-November/026551.html
[2] https://www.linux.org.ru/forum/general/11598803

-----------------------------------------------------------

I'd like to add that dovecot is the *only* software I use that has this sort 
of problem with btrfs -- and it's *only* the transaction *.log files; the few 
times the index files had problems, they were automatically fixed.  However, 
once the problem shows up, no other software can read the files, either (e.g., 
cat fails).  Furthermore, restarting dovecot doesn't help, only rebooting, 
i.e., unmounting the entire file system, not just the subvolume the mailbox 
resides on (as discussed previously in this sub-thread).  So I'm not sure 
what, exactly, is at fault here: dovecot, btrfs, or both?

So given all the above, does anybody have any further suggestions on how to 
proceed?  Because regardless of whether the switch to maildir ends up 
resolving the issue for me, something is going wrong.  I'm thinking of filing 
a bug report with dovecot; perhaps none of their devs test with btrfs?

Greetings
-- 
Marc Joliet
--
"People who think they know everything really annoy those of us who know we
don't" - Bjarne Stroustrup

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 819 bytes --]

Re: random i/o error without error in dmesg

[Marc Joliet]

[-- Attachment #1: Type: text/plain, Size: 822 bytes --]

Am Saturday 07 May 2016
schrieb Marc Joliet <marcec@gmx.de>

>I'm thinking of filing 
>a bug report with dovecot; perhaps none of their devs test with btrfs?

So I did this, and got a little bit of feedback.  Quoting from the reply I 
got:

"*.index.log files are always appended to using O_APPEND flag. Maybe this is 
relevant.

Also when a new .log file is created it's opened without the O_APPEND flag and 
the O_APPEND is added later. This was causing a bug recently in unionfs, which 
ignored the flag change and caused log file corruption."

The other bit of advise was to stress test dovecot using imaptest, but I'll 
have to do that when I have more time.

Greetings
-- 
Marc Joliet
--
"People who think they know everything really annoy those of us who know we
don't" - Bjarne Stroustrup

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 819 bytes --]