btrfs: Add ratelimiting to printing facility
diff mbox

Message ID 1468326012-15910-1-git-send-email-kernel@kyup.com
State Superseded
Headers show

Commit Message

kernel@kyup.com July 12, 2016, 12:20 p.m. UTC
Currently the btrfs printk infrastructure doesn't implement any
kind of ratelimiting. Recently I came accross a case where due to
FS corruption an excessive amount of printk caused the softlockup
detector to trigger and reset the server. This patch aims to avoid
two types of issue: 
 * I want to avoid different levels of messages interefere with the
 ratelimiting of oneanother so as to avoid a situation where a 
 flood of INFO messages causes the ratelimit to trigger, 
 potentially leading to supression of more important messages. 

 * Avoid a flood of any type of messages rendering the machine 
 unusable

---
So recently I got the softlockup to trigger due to excessive btrfs 
printing: 

[3211212.563757] NMI watchdog: BUG: soft lockup - CPU#5 stuck for 22s! [btrfs-transacti:6033]
[3211212.563782] Modules linked in: [omitted]
[3211212.563783] BTRFS info (device loop1): no csum found for inode 450 start 2746998784
[3211212.563785] CPU: 5 PID: 6033 Comm: btrfs-transacti Tainted: P W O 4.4.9-clouder1 #20
[3211212.563786] Hardware name: Supermicro X9DRD-7LN4F(-JBOD)/X9DRD-EF/X9DRD-7LN4F, BIOS 3.0a 12/05/2013
[3211212.563787] task: ffff88046d0d8000 ti: ffff8802bcaa8000 task.ti: ffff8802bcaa8000
[3211212.563792] RIP: 0010:[<ffffffff810a1475>] [<ffffffff810a1475>] console_unlock+0x305/0x4b0
[3211212.563793] RSP: 0018:ffff8802bcaab7b8 EFLAGS: 00000282
[3211212.563794] RAX: 0000000000000000 RBX: ffffffff82027e18 RCX: 0000000000000000
[3211212.563795] RDX: ffff8802bcaab6d8 RSI: 0000000000000004 RDI: ffffffff8204f0c8
[3211212.563795] RBP: ffff8802bcaab7f8 R08: 0000000000000000 R09: ffff8800000b8460
[3211212.563796] R10: 0000000000000001 R11: 0000000000000000 R12: 0000000000000020
[3211212.563797] R13: 0000000000000282 R14: 0000000000000048 R15: 0000000000000001
[3211212.563798] FS: 0000000000000000(0000) GS:ffff88047fca0000(0000) knlGS:0000000000000000
[3211212.563799] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[3211212.563799] CR2: ffffffffff600400 CR3: 000000046afd6000 CR4: 00000000001406e0
[3211212.563800] Stack:
[3211212.563802] 0000000000000000 06ffffff00000000 ffff8802bcaab7e8 00000000ffffffff
[3211212.563804] 0000000000000006 0000000000000000 0000000000000024 0000000000000246
[3211212.563805] ffff8802bcaab878 ffffffff810a1f00 ffffffff81e341a2 ffffffff00000024
[3211212.563806] Call Trace:
[3211212.563809] [<ffffffff810a1f00>] vprintk_emit+0x230/0x490
[3211212.563836] [<ffffffffa07a655d>] ? __btrfs_free_extent+0x24d/0xd80 [btrfs]
[3211212.563838] [<ffffffff810a217f>] vprintk_default+0x1f/0x30
[3211212.563841] [<ffffffff8112d753>] printk+0x46/0x48
[3211212.563853] [<ffffffffa07a655d>] ? __btrfs_free_extent+0x24d/0xd80 [btrfs]
[3211212.563857] [<ffffffff810549d3>] warn_slowpath_common+0x43/0xe0
[3211212.563859] [<ffffffff81054b26>] warn_slowpath_fmt+0x46/0x50
[3211212.563870] [<ffffffffa07a655d>] __btrfs_free_extent+0x24d/0xd80 [btrfs]
[3211212.563874] [<ffffffff81185024>] ? kmem_cache_free+0x214/0x230
[3211212.563890] [<ffffffffa080e730>] ? btrfs_delayed_ref_lock+0x40/0x1f0 [btrfs]
[3211212.563902] [<ffffffffa07a7f2e>] __btrfs_run_delayed_refs+0xe9e/0x12f0 [btrfs]
[3211212.563904] [<ffffffff81185024>] ? kmem_cache_free+0x214/0x230
[3211212.563908] [<ffffffff8107c584>] ? finish_task_switch+0xa4/0x240
[3211212.563911] [<ffffffff8108d026>] ? put_prev_task_fair+0x26/0x40
[3211212.563913] [<ffffffff8163e694>] ? __schedule+0x394/0xac0
[3211212.563917] [<ffffffff8103bcce>] ? physflat_send_IPI_mask+0xe/0x10
[3211212.563928] [<ffffffffa07a840e>] btrfs_run_delayed_refs+0x8e/0x2e0 [btrfs]
[3211212.563930] [<ffffffff8107adb2>] ? __might_sleep+0x52/0x90
[3211212.563932] [<ffffffff8118713b>] ? kmem_cache_alloc+0x11b/0x200
[3211212.563947] [<ffffffffa07bcd51>] btrfs_commit_transaction+0x51/0xb70 [btrfs]
[3211212.563960] [<ffffffffa07bdccd>] ? start_transaction+0x9d/0x4e0 [btrfs]
[3211212.563963] [<ffffffff810b2320>] ? lock_timer_base+0x80/0x80
[3211212.563965] [<ffffffff8107adb2>] ? __might_sleep+0x52/0x90
[3211212.563978] [<ffffffffa07b7063>] transaction_kthread+0x233/0x280 [btrfs]
[3211212.563991] [<ffffffffa07b6e30>] ? close_ctree+0x370/0x370 [btrfs]
[3211212.564002] [<ffffffffa07b6e30>] ? close_ctree+0x370/0x370 [btrfs]
[3211212.564004] [<ffffffff81074557>] kthread+0xd7/0xf0
[3211212.564007] [<ffffffff8107d51e>] ? schedule_tail+0x1e/0xd0
[3211212.564009] [<ffffffff81074480>] ? kthread_freezable_should_stop+0x80/0x80
[3211212.564011] [<ffffffff8164322f>] ret_from_fork+0x3f/0x70
[3211212.564013] [<ffffffff81074480>] ? kthread_freezable_should_stop+0x80/0x80
[3211212.564030] Code: 89 15 b8 32 d9 00 48 8b 05 89 32 d9 00 c7 05 ab 32 d9 00 00 00 00 00 48 89 05 98 32 d9 00 e9 d1

The message was: 

[3211212.563115] BTRFS info (device loop1): no csum found for inode 450 start 2746953728
[3211212.563173] BTRFS info (device loop1): no csum found for inode 450 start 2746957824
[3211212.563231] BTRFS info (device loop1): no csum found for inode 450 start 2746961920
[3211212.563289] BTRFS info (device loop1): no csum found for inode 450 start 2746966016
[3211212.563347] BTRFS info (device loop1): no csum found for inode 450 start 2746970112
[3211212.563407] BTRFS info (device loop1): no csum found for inode 450 start 2746974208

 fs/btrfs/super.c | 29 +++++++++++++++++++++++++++--
 1 file changed, 27 insertions(+), 2 deletions(-)

Comments

David Sterba July 12, 2016, 1:20 p.m. UTC | #1
On Tue, Jul 12, 2016 at 03:20:12PM +0300, Nikolay Borisov wrote:
> Currently the btrfs printk infrastructure doesn't implement any
> kind of ratelimiting.

If you count the whole infrastructure, it does. See ctree.h and macros
ending with _rl (btrfs_err_rl), and should be used where the messages
are likely to flood. Otherwise I think "more is better" regarding
messages as this is helpful when debugging issues.

> Recently I came accross a case where due to
> FS corruption an excessive amount of printk caused the softlockup
> detector to trigger and reset the server. This patch aims to avoid
> two types of issue: 
>  * I want to avoid different levels of messages interefere with the
>  ratelimiting of oneanother so as to avoid a situation where a 
>  flood of INFO messages causes the ratelimit to trigger, 
>  potentially leading to supression of more important messages. 

Yeah, that's my concern as well. What if there's a burst of several
error messages that do not fit to the limit and some of them get
dropped.

>  * Avoid a flood of any type of messages rendering the machine 
>  unusable

While I'd rather set a per-message ratelimiting, it's possible that an
unexpected error will start flooding. So some sort of per-level limiting
could be implemented, as you propose, but I'd suggest to set the numbers
higher. That way it would still flood up to certain level but should
avoid the lockups.
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
kernel@kyup.com July 12, 2016, 1:47 p.m. UTC | #2
On 07/12/2016 04:20 PM, David Sterba wrote:
> On Tue, Jul 12, 2016 at 03:20:12PM +0300, Nikolay Borisov wrote:
>> Currently the btrfs printk infrastructure doesn't implement any
>> kind of ratelimiting.
> 
> If you count the whole infrastructure, it does. See ctree.h and macros
> ending with _rl (btrfs_err_rl), and should be used where the messages
> are likely to flood. Otherwise I think "more is better" regarding
> messages as this is helpful when debugging issues.

So I definitely didn't look at those. But now I have and it seems they
implement more or less the same thing. Also, if I'm reading the code
correctly, as it stands now using the _rl versions seem to be more
flexible as the limits is going to be per-message rather than
per-message class as it is in my proposal. So I'd rather move that
particular csum related message to the _rl infrastructure.

> 
>> Recently I came accross a case where due to
>> FS corruption an excessive amount of printk caused the softlockup
>> detector to trigger and reset the server. This patch aims to avoid
>> two types of issue: 
>>  * I want to avoid different levels of messages interefere with the
>>  ratelimiting of oneanother so as to avoid a situation where a 
>>  flood of INFO messages causes the ratelimit to trigger, 
>>  potentially leading to supression of more important messages. 
> 
> Yeah, that's my concern as well. What if there's a burst of several
> error messages that do not fit to the limit and some of them get
> dropped.
> 
>>  * Avoid a flood of any type of messages rendering the machine 
>>  unusable
> 
> While I'd rather set a per-message ratelimiting, it's possible that an
> unexpected error will start flooding. So some sort of per-level limiting
> could be implemented, as you propose, but I'd suggest to set the numbers
> higher. That way it would still flood up to certain level but should
> avoid the lockups.

Sure, I'm happy to set the limits higher and have it act as a safety
net. But then again, what would make a sensible limit - 100 messages in
5 seconds seems reasonable but this is completely arbitrary. Any thoughts?

--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
David Sterba July 12, 2016, 4:01 p.m. UTC | #3
On Tue, Jul 12, 2016 at 04:47:53PM +0300, Nikolay Borisov wrote:
> 
> 
> On 07/12/2016 04:20 PM, David Sterba wrote:
> > On Tue, Jul 12, 2016 at 03:20:12PM +0300, Nikolay Borisov wrote:
> >> Currently the btrfs printk infrastructure doesn't implement any
> >> kind of ratelimiting.
> > 
> > If you count the whole infrastructure, it does. See ctree.h and macros
> > ending with _rl (btrfs_err_rl), and should be used where the messages
> > are likely to flood. Otherwise I think "more is better" regarding
> > messages as this is helpful when debugging issues.
> 
> So I definitely didn't look at those. But now I have and it seems they
> implement more or less the same thing. Also, if I'm reading the code
> correctly, as it stands now using the _rl versions seem to be more
> flexible as the limits is going to be per-message rather than
> per-message class as it is in my proposal. So I'd rather move that
> particular csum related message to the _rl infrastructure.

Ok.

> >> Recently I came accross a case where due to
> >> FS corruption an excessive amount of printk caused the softlockup
> >> detector to trigger and reset the server. This patch aims to avoid
> >> two types of issue: 
> >>  * I want to avoid different levels of messages interefere with the
> >>  ratelimiting of oneanother so as to avoid a situation where a 
> >>  flood of INFO messages causes the ratelimit to trigger, 
> >>  potentially leading to supression of more important messages. 
> > 
> > Yeah, that's my concern as well. What if there's a burst of several
> > error messages that do not fit to the limit and some of them get
> > dropped.
> > 
> >>  * Avoid a flood of any type of messages rendering the machine 
> >>  unusable
> > 
> > While I'd rather set a per-message ratelimiting, it's possible that an
> > unexpected error will start flooding. So some sort of per-level limiting
> > could be implemented, as you propose, but I'd suggest to set the numbers
> > higher. That way it would still flood up to certain level but should
> > avoid the lockups.
> 
> Sure, I'm happy to set the limits higher and have it act as a safety
> net. But then again, what would make a sensible limit - 100 messages in
> 5 seconds seems reasonable but this is completely arbitrary. Any thoughts?

5 seconds "should be" enough for the system to do other work, and 100
messages could mean 5-8 kB of text. We're talking about a situation when
things are really bad, messages are likely to be duplicated so there
should be enough information captured by the time the ratelimiting kicks
in. So, let it be 100.
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch
diff mbox

diff --git a/fs/btrfs/super.c b/fs/btrfs/super.c
index fe609b81dd1b..85b0a4517e30 100644
--- a/fs/btrfs/super.c
+++ b/fs/btrfs/super.c
@@ -188,6 +188,25 @@  static const char * const logtypes[] = {
 	"debug",
 };
 
+static struct ratelimit_state printk_limits[] = {
+	RATELIMIT_STATE_INIT(printk_limits[0], DEFAULT_RATELIMIT_INTERVAL,
+			     DEFAULT_RATELIMIT_BURST),
+	RATELIMIT_STATE_INIT(printk_limits[1], DEFAULT_RATELIMIT_INTERVAL,
+			     DEFAULT_RATELIMIT_BURST),
+	RATELIMIT_STATE_INIT(printk_limits[2], DEFAULT_RATELIMIT_INTERVAL,
+			     DEFAULT_RATELIMIT_BURST),
+	RATELIMIT_STATE_INIT(printk_limits[3], DEFAULT_RATELIMIT_INTERVAL,
+			     DEFAULT_RATELIMIT_BURST),
+	RATELIMIT_STATE_INIT(printk_limits[4], DEFAULT_RATELIMIT_INTERVAL,
+			     DEFAULT_RATELIMIT_BURST),
+	RATELIMIT_STATE_INIT(printk_limits[5], DEFAULT_RATELIMIT_INTERVAL,
+			     DEFAULT_RATELIMIT_BURST),
+	RATELIMIT_STATE_INIT(printk_limits[6], DEFAULT_RATELIMIT_INTERVAL,
+			     DEFAULT_RATELIMIT_BURST),
+	RATELIMIT_STATE_INIT(printk_limits[7], DEFAULT_RATELIMIT_INTERVAL,
+			     DEFAULT_RATELIMIT_BURST),
+};
+
 void btrfs_printk(const struct btrfs_fs_info *fs_info, const char *fmt, ...)
 {
 	struct super_block *sb = fs_info->sb;
@@ -196,6 +215,7 @@  void btrfs_printk(const struct btrfs_fs_info *fs_info, const char *fmt, ...)
 	va_list args;
 	const char *type = logtypes[4];
 	int kern_level;
+	struct ratelimit_state *ratelimit;
 
 	va_start(args, fmt);
 
@@ -206,13 +226,18 @@  void btrfs_printk(const struct btrfs_fs_info *fs_info, const char *fmt, ...)
 		lvl[size] = '\0';
 		fmt += size;
 		type = logtypes[kern_level - '0'];
-	} else
+		ratelimit = &printk_limits[kern_level - '0'];
+	} else {
 		*lvl = '\0';
+		/* Default to debug output */
+		ratelimit = &printk_limits[7];
+	}
 
 	vaf.fmt = fmt;
 	vaf.va = &args;
 
-	printk("%sBTRFS %s (device %s): %pV\n", lvl, type, sb->s_id, &vaf);
+	if (__ratelimit(ratelimit))
+		printk("%sBTRFS %s (device %s): %pV\n", lvl, type, sb->s_id, &vaf);
 
 	va_end(args);
 }