Re: [PATCH] x86_64: dynamic MCE poll interval (try 2)
On Fri, Apr 27, 2007 at 02:34:11PM -0700, Tim Hockin wrote: > From: Tim Hockin <[EMAIL PROTECTED]> > > Background: > We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, > especially when we are trying really hard to stress the system out. The > current MCE poller uses a static interval which does not care whether it > has or has not found MCEs recently. Added thanks -Andi - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
[PATCH] x86_64: dynamic MCE poll interval (try 2)
From: Tim Hockin <[EMAIL PROTECTED]> Background: We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, especially when we are trying really hard to stress the system out. The current MCE poller uses a static interval which does not care whether it has or has not found MCEs recently. Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. The "Machine check events logged" printk() is rate limited to the check_interval, which should be identical behavior to the old functionality. Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Alternatives: I considered simply reducing the polling interval to 10 ms immediately and keeping it there as long as we continue to find errors. This felt a bit heavy handed, but does perform significantly better for the default check_interval of 5 minutes (we're using a few seconds when testing for DRAM errors). I could be convinced to go with this, if anyone felt it was not too aggressive. Testing: I used an error-injecting DIMM to create lots of correctable DRAM errors and verified that the polling interval accelerates. The printk() only happens once per check_interval seconds. Patch: This patch is against 2.6.21-rc7. Signed-Off-By: Tim Hockin <[EMAIL PROTECTED]> --- This is the fourth version of this patch. The only change from the prior version is to not initialize a static var to 0. This patch should come through cleanly, I hope. diff -pruN linux-2.6.20/Documentation/x86_64/machinecheck linux-2.6.20+th/Documentation/x86_64/machinecheck --- linux-2.6.20/Documentation/x86_64/machinecheck 2007-04-24 23:36:03.0 -0700 +++ linux-2.6.20+th/Documentation/x86_64/machinecheck 2007-04-27 10:11:10.0 -0700 @@ -36,7 +36,12 @@ between all CPUs. check_interval How often to poll for corrected machine check errors, in seconds - (Note output is hexademical). Default 5 minutes. + (Note output is hexademical). Default 5 minutes. When the poller + finds MCEs it triggers an exponential speedup (poll more often) on + the polling interval. When the poller stops finding MCEs, it + triggers an exponential backoff (poll less often) on the polling + interval. The check_interval variable is both the initial and + maximum polling interval. tolerant Tolerance level. When a machine check exception occurs for a non diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c linux-2.6.20+th/arch/x86_64/kernel/mce.c --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-27 10:01:02.0 -0700 +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c2007-04-27 10:41:02.0 -0700 @@ -323,10 +323,13 @@ void mce_log_therm_throt_event(unsigned #endif /* CONFIG_X86_MCE_INTEL */ /* - * Periodic polling timer for "silent" machine check errors. + * Periodic polling timer for "silent" machine check errors. If the + * poller finds an MCE, poll 2x faster. When the poller finds no more + * errors, poll 2x slower (up to check_interval seconds). */ static int check_interval = 5 * 60; /* 5 minutes */ +static int next_interval; /* in jiffies */ static void mcheck_timer(struct work_struct *work); static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer); @@ -339,7 +342,6 @@ static void mcheck_check_cpu(void *info) static void mcheck_timer(struct work_struct *work) { on_each_cpu(mcheck_check_cpu, NULL, 1, 1); - schedule_delayed_work(_work, check_interval * HZ); /* * It's ok to read stale data here for notify_user and @@ -349,17 +351,30 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user && console_logged) { + static unsigned long last_print; + unsigned long now = jiffies; + + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, _logged); - printk(KERN_INFO "Machine check events logged\n"); + if (time_after_eq(now, last_print + (check_interval*HZ))) { + last_print = now; + printk(KERN_INFO "Machine check events logged\n"); + } + } else { + next_interval = min(next_interval*2, check_interval*HZ); } + + schedule_delayed_work(_work, next_interval); } static __init int periodic_mcheck_init(void) { - if (check_interval) - schedule_delayed_work(_work, check_interval*HZ); + next_interval = check_interval * HZ; + if (next_interval) +
Re: [PATCH] x86_64: dynamic MCE poll interval
On Fri, 27 Apr 2007 13:51:28 -0700 Tim Hockin wrote: > This is the third version of this patch. The only change from the prior > version is to use time_after_eq(). > > diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c > linux-2.6.20+th/arch/x86_64/kernel/mce.c > --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-27 10:01:02.0 > -0700 > +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c 2007-04-27 10:41:02.0 > -0700 > @@ -349,17 +351,30 @@ static void mcheck_timer(struct work_str >* writes. >*/ > if (notify_user && console_logged) { > + static unsigned long last_print = 0; Don't need to init statics to 0... > + unsigned long now = jiffies; > + > + /* if we logged an MCE, reduce the polling interval */ > + next_interval = max(next_interval/2, HZ/100); > notify_user = 0; > clear_bit(0, _logged); > - printk(KERN_INFO "Machine check events logged\n"); > + if (time_after_eq(now, last_print + (check_interval*HZ))) { > + last_print = now; > + printk(KERN_INFO "Machine check events logged\n"); > + } > + } else { > + next_interval = min(next_interval*2, check_interval*HZ); > } > + > + schedule_delayed_work(_work, next_interval); > } --- ~Randy *** Remember to use Documentation/SubmitChecklist when testing your code *** - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
Sorry, Gmail mangles whitespace unless you do just the right thing. Let me work around it. Proper patch coming. On 4/27/07, Tim Hockin <[EMAIL PROTECTED]> wrote: From: Tim Hockin <[EMAIL PROTECTED]> Background: We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, especially when we are trying really hard to stress the system out. The current MCE poller uses a static interval which does not care whether it has or has not found MCEs recently. Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. The "Machine check events logged" printk() is rate limited to the check_interval, which should be identical behavior to the old functionality. Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Alternatives: I considered simply reducing the polling interval to 10 ms immediately and keeping it there as long as we continue to find errors. This felt a bit heavy handed, but does perform significantly better for the default check_interval of 5 minutes (we're using a few seconds when testing for DRAM errors). I could be convinced to go with this, if anyone felt it was not too aggressive. Testing: I used an error-injecting DIMM to create lots of correctable DRAM errors and verified that the polling interval accelerates. The printk() only happens once per check_interval seconds. Patch: This patch is against 2.6.21-rc7. Signed-Off-By: Tim Hockin <[EMAIL PROTECTED]> --- This is the third version of this patch. The only change from the prior version is to use time_after_eq(). diff -pruN linux-2.6.20/Documentation/x86_64/machinecheck linux-2.6.20+th/Documentation/x86_64/machinecheck --- linux-2.6.20/Documentation/x86_64/machinecheck 2007-04-24 23:36:03.0 -0700 +++ linux-2.6.20+th/Documentation/x86_64/machinecheck 2007-04-27 10:11:10.0 -0700 @@ -36,7 +36,12 @@ between all CPUs. check_interval How often to poll for corrected machine check errors, in seconds - (Note output is hexademical). Default 5 minutes. + (Note output is hexademical). Default 5 minutes. When the poller + finds MCEs it triggers an exponential speedup (poll more often) on + the polling interval. When the poller stops finding MCEs, it + triggers an exponential backoff (poll less often) on the polling + interval. The check_interval variable is both the initial and + maximum polling interval. tolerant Tolerance level. When a machine check exception occurs for a non diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c linux-2.6.20+th/arch/x86_64/kernel/mce.c --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-27 10:01:02.0 -0700 +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c2007-04-27 10:41:02.0 -0700 @@ -323,10 +323,13 @@ void mce_log_therm_throt_event(unsigned #endif /* CONFIG_X86_MCE_INTEL */ /* - * Periodic polling timer for "silent" machine check errors. + * Periodic polling timer for "silent" machine check errors. If the + * poller finds an MCE, poll 2x faster. When the poller finds no more + * errors, poll 2x slower (up to check_interval seconds). */ static int check_interval = 5 * 60; /* 5 minutes */ +static int next_interval; /* in jiffies */ static void mcheck_timer(struct work_struct *work); static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer); @@ -339,7 +342,6 @@ static void mcheck_check_cpu(void *info) static void mcheck_timer(struct work_struct *work) { on_each_cpu(mcheck_check_cpu, NULL, 1, 1); - schedule_delayed_work(_work, check_interval * HZ); /* * It's ok to read stale data here for notify_user and @@ -349,17 +351,30 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user && console_logged) { + static unsigned long last_print = 0; + unsigned long now = jiffies; + + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, _logged); - printk(KERN_INFO "Machine check events logged\n"); + if (time_after_eq(now, last_print + (check_interval*HZ))) { + last_print = now; + printk(KERN_INFO "Machine check events logged\n"); + } + } else { + next_interval = min(next_interval*2, check_interval*HZ); } + + schedule_delayed_work(_work, next_interval); } static __init int periodic_mcheck_init(void) { - if (check_interval) -
Re: [PATCH] x86_64: dynamic MCE poll interval
From: Tim Hockin <[EMAIL PROTECTED]> Background: We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, especially when we are trying really hard to stress the system out. The current MCE poller uses a static interval which does not care whether it has or has not found MCEs recently. Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. The "Machine check events logged" printk() is rate limited to the check_interval, which should be identical behavior to the old functionality. Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Alternatives: I considered simply reducing the polling interval to 10 ms immediately and keeping it there as long as we continue to find errors. This felt a bit heavy handed, but does perform significantly better for the default check_interval of 5 minutes (we're using a few seconds when testing for DRAM errors). I could be convinced to go with this, if anyone felt it was not too aggressive. Testing: I used an error-injecting DIMM to create lots of correctable DRAM errors and verified that the polling interval accelerates. The printk() only happens once per check_interval seconds. Patch: This patch is against 2.6.21-rc7. Signed-Off-By: Tim Hockin <[EMAIL PROTECTED]> --- This is the third version of this patch. The only change from the prior version is to use time_after_eq(). diff -pruN linux-2.6.20/Documentation/x86_64/machinecheck linux-2.6.20+th/Documentation/x86_64/machinecheck --- linux-2.6.20/Documentation/x86_64/machinecheck 2007-04-24 23:36:03.0 -0700 +++ linux-2.6.20+th/Documentation/x86_64/machinecheck 2007-04-27 10:11:10.0 -0700 @@ -36,7 +36,12 @@ between all CPUs. check_interval How often to poll for corrected machine check errors, in seconds - (Note output is hexademical). Default 5 minutes. + (Note output is hexademical). Default 5 minutes. When the poller + finds MCEs it triggers an exponential speedup (poll more often) on + the polling interval. When the poller stops finding MCEs, it + triggers an exponential backoff (poll less often) on the polling + interval. The check_interval variable is both the initial and + maximum polling interval. tolerant Tolerance level. When a machine check exception occurs for a non diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c linux-2.6.20+th/arch/x86_64/kernel/mce.c --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-27 10:01:02.0 -0700 +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c2007-04-27 10:41:02.0 -0700 @@ -323,10 +323,13 @@ void mce_log_therm_throt_event(unsigned #endif /* CONFIG_X86_MCE_INTEL */ /* - * Periodic polling timer for "silent" machine check errors. + * Periodic polling timer for "silent" machine check errors. If the + * poller finds an MCE, poll 2x faster. When the poller finds no more + * errors, poll 2x slower (up to check_interval seconds). */ static int check_interval = 5 * 60; /* 5 minutes */ +static int next_interval; /* in jiffies */ static void mcheck_timer(struct work_struct *work); static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer); @@ -339,7 +342,6 @@ static void mcheck_check_cpu(void *info) static void mcheck_timer(struct work_struct *work) { on_each_cpu(mcheck_check_cpu, NULL, 1, 1); - schedule_delayed_work(_work, check_interval * HZ); /* * It's ok to read stale data here for notify_user and @@ -349,17 +351,30 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user && console_logged) { + static unsigned long last_print = 0; + unsigned long now = jiffies; + + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, _logged); - printk(KERN_INFO "Machine check events logged\n"); + if (time_after_eq(now, last_print + (check_interval*HZ))) { + last_print = now; + printk(KERN_INFO "Machine check events logged\n"); + } + } else { + next_interval = min(next_interval*2, check_interval*HZ); } + + schedule_delayed_work(_work, next_interval); } static __init int periodic_mcheck_init(void) { - if (check_interval) - schedule_delayed_work(_work, check_interval*HZ); + next_interval = check_interval * HZ; + if (next_interval) + schedule_delayed_work(_work, next_interval);
Re: [PATCH] x86_64: dynamic MCE poll interval
> - printk(KERN_INFO "Machine check events logged\n"); > + if ((now - last_print) >= check_interval*HZ) { Shouldn't this use time_after() to handle wrapping? -Andi - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
From: Tim Hockin <[EMAIL PROTECTED]> Background: We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, especially when we are trying really hard to stress the system out. The current MCE poller uses a static interval which does not care whether it has or has not found MCEs recently. Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. The "Machine check events logged" printk() is rate limited to the check_interval, which should be identical behavior to the old functionality. Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Alternatives: I considered simply reducing the polling interval to 10 ms immediately and keeping it there as long as we continue to find errors. This felt a bit heavy handed, but does perform significantly better for the default check_interval of 5 minutes (we're using a few seconds when testing for DRAM errors). I could be convinced to go with this, if anyone felt it was not too aggressive. Testing: I used an error-injecting DIMM to create lots of correctable DRAM errors and verified that the polling interval accelerates. The printk() only happens once per check_interval seconds. Patch: This patch is against 2.6.21-rc7. Signed-Off-By: Tim Hockin <[EMAIL PROTECTED]> --- This is the second version of this patch. The more I think about the alternative described above, the more I like it, but I recognize that it is heavy-handed. If correctable errors matter that much to you, you can start with a small check_interval, I suppose. diff -pruN linux-2.6.20/Documentation/x86_64/machinecheck linux-2.6.20+th/Documentation/x86_64/machinecheck --- linux-2.6.20/Documentation/x86_64/machinecheck 2007-04-24 23:36:03.0 -0700 +++ linux-2.6.20+th/Documentation/x86_64/machinecheck 2007-04-27 10:11:10.0 -0700 @@ -36,7 +36,12 @@ between all CPUs. check_interval How often to poll for corrected machine check errors, in seconds - (Note output is hexademical). Default 5 minutes. + (Note output is hexademical). Default 5 minutes. When the poller + finds MCEs it triggers an exponential speedup (poll more often) on + the polling interval. When the poller stops finding MCEs, it + triggers an exponential backoff (poll less often) on the polling + interval. The check_interval variable is both the initial and + maximum polling interval. tolerant Tolerance level. When a machine check exception occurs for a non diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c linux-2.6.20+th/arch/x86_64/kernel/mce.c --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-27 10:01:02.0 -0700 +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c2007-04-27 10:41:02.0 -0700 @@ -323,10 +323,13 @@ void mce_log_therm_throt_event(unsigned #endif /* CONFIG_X86_MCE_INTEL */ /* - * Periodic polling timer for "silent" machine check errors. + * Periodic polling timer for "silent" machine check errors. If the + * poller finds an MCE, poll 2x faster. When the poller finds no more + * errors, poll 2x slower (up to check_interval seconds). */ static int check_interval = 5 * 60; /* 5 minutes */ +static int next_interval; /* in jiffies */ static void mcheck_timer(struct work_struct *work); static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer); @@ -339,7 +342,6 @@ static void mcheck_check_cpu(void *info) static void mcheck_timer(struct work_struct *work) { on_each_cpu(mcheck_check_cpu, NULL, 1, 1); - schedule_delayed_work(_work, check_interval * HZ); /* * It's ok to read stale data here for notify_user and @@ -349,17 +351,30 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user && console_logged) { + static unsigned long last_print = 0; + unsigned long now = jiffies; + + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, _logged); - printk(KERN_INFO "Machine check events logged\n"); + if ((now - last_print) >= check_interval*HZ) { + last_print = now; + printk(KERN_INFO "Machine check events logged\n"); + } + } else { + next_interval = min(next_interval*2, check_interval*HZ); } + + schedule_delayed_work(_work, next_interval); } static __init int periodic_mcheck_init(void) { - if (check_interval) - schedule_delayed_work(_work, check_interval*HZ); +
Re: [PATCH] x86_64: dynamic MCE poll interval
On Fri, Apr 27, 2007 at 10:05:58AM -0700, Tim Hockin wrote: > On 27 Apr 2007 19:02:30 +0200, Andi Kleen <[EMAIL PROTECTED]> wrote: > >On Fri, Apr 27, 2007 at 09:58:14AM -0700, Tim Hockin wrote: > >> On 27 Apr 2007 11:09:17 +0200, Andi Kleen <[EMAIL PROTECTED]> wrote: > >> >On Thu, Apr 26, 2007 at 06:02:52PM -0700, Tim Hockin wrote: > >> >> Description: > >> >> This patch makes the MCE poller adjust the polling interval > >dynamically. > >> >> If we find an MCE, poll 2x faster (down to 10 ms). When we stop > >finding > >> >> MCEs, poll 2x slower (up to check_interval seconds). The > >check_interval > >> >> tunable becomes the max polling interval. > >> > > >> >Can you please fix the documentation then? > >> > >> Which documentation, specifically? :) > > > >Documentation/x86_64/{boot-options.txt,machinecheck} > > What needs to be changed in boot-options? Nothing for your patch sorry. Only in machinecheck -Andi - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
On 27 Apr 2007 19:02:30 +0200, Andi Kleen <[EMAIL PROTECTED]> wrote: On Fri, Apr 27, 2007 at 09:58:14AM -0700, Tim Hockin wrote: > On 27 Apr 2007 11:09:17 +0200, Andi Kleen <[EMAIL PROTECTED]> wrote: > >On Thu, Apr 26, 2007 at 06:02:52PM -0700, Tim Hockin wrote: > >> Description: > >> This patch makes the MCE poller adjust the polling interval dynamically. > >> If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding > >> MCEs, poll 2x slower (up to check_interval seconds). The check_interval > >> tunable becomes the max polling interval. > > > >Can you please fix the documentation then? > > Which documentation, specifically? :) Documentation/x86_64/{boot-options.txt,machinecheck} What needs to be changed in boot-options? > > >> Result: > >> If you start to take a lot of correctable errors (not exceptions), you > >> log them faster and more accurately (less chance of overflowing the MCA > >> registers). If you don't take a lot of errors, you will see no change. > > > >Makes sense. > > > >AMD RevF can do this using the threshold interrupts too for DIMM errors > >too without any delays -- perhaps it would also make sense to configure > >this by default that it always triggers on all DIMM errors. > >Right now it is just an option in /sys > > Can I look at this as a followon patch? I have a number of mce Sure. -Andi - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
On Fri, Apr 27, 2007 at 09:58:14AM -0700, Tim Hockin wrote: > On 27 Apr 2007 11:09:17 +0200, Andi Kleen <[EMAIL PROTECTED]> wrote: > >On Thu, Apr 26, 2007 at 06:02:52PM -0700, Tim Hockin wrote: > >> Description: > >> This patch makes the MCE poller adjust the polling interval dynamically. > >> If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding > >> MCEs, poll 2x slower (up to check_interval seconds). The check_interval > >> tunable becomes the max polling interval. > > > >Can you please fix the documentation then? > > Which documentation, specifically? :) Documentation/x86_64/{boot-options.txt,machinecheck} > > >> Result: > >> If you start to take a lot of correctable errors (not exceptions), you > >> log them faster and more accurately (less chance of overflowing the MCA > >> registers). If you don't take a lot of errors, you will see no change. > > > >Makes sense. > > > >AMD RevF can do this using the threshold interrupts too for DIMM errors > >too without any delays -- perhaps it would also make sense to configure > >this by default that it always triggers on all DIMM errors. > >Right now it is just an option in /sys > > Can I look at this as a followon patch? I have a number of mce Sure. -Andi - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
On 27 Apr 2007 11:09:17 +0200, Andi Kleen <[EMAIL PROTECTED]> wrote: On Thu, Apr 26, 2007 at 06:02:52PM -0700, Tim Hockin wrote: > Description: > This patch makes the MCE poller adjust the polling interval dynamically. > If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding > MCEs, poll 2x slower (up to check_interval seconds). The check_interval > tunable becomes the max polling interval. Can you please fix the documentation then? Which documentation, specifically? :) > Result: > If you start to take a lot of correctable errors (not exceptions), you > log them faster and more accurately (less chance of overflowing the MCA > registers). If you don't take a lot of errors, you will see no change. Makes sense. AMD RevF can do this using the threshold interrupts too for DIMM errors too without any delays -- perhaps it would also make sense to configure this by default that it always triggers on all DIMM errors. Right now it is just an option in /sys Can I look at this as a followon patch? I have a number of mce related patches in the pipeline, and I am trying to keep them small for testing sanity - they are hard enough to test :) The printk should not happen too often. Can you add some hardcoded limit there than it doesn't happen more often than every hour or so (or perhaps use a exponential backoff here too?) It is only to tell users to check mcelog output. Sure. I'll fix it up and hit you again today. - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
On Thu, Apr 26, 2007 at 06:02:52PM -0700, Tim Hockin wrote: > Description: > This patch makes the MCE poller adjust the polling interval dynamically. > If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding > MCEs, poll 2x slower (up to check_interval seconds). The check_interval > tunable becomes the max polling interval. Can you please fix the documentation then? > > Result: > If you start to take a lot of correctable errors (not exceptions), you > log them faster and more accurately (less chance of overflowing the MCA > registers). If you don't take a lot of errors, you will see no change. Makes sense. AMD RevF can do this using the threshold interrupts too for DIMM errors too without any delays -- perhaps it would also make sense to configure this by default that it always triggers on all DIMM errors. Right now it is just an option in /sys > @@ -349,17 +349,24 @@ static void mcheck_timer(struct work_str >* writes. >*/ > if (notify_user && console_logged) { > + /* if we logged an MCE, reduce the polling interval */ > + next_interval = max(next_interval/2, HZ/100); > notify_user = 0; > clear_bit(0, _logged); > printk(KERN_INFO "Machine check events logged\n"); The printk should not happen too often. Can you add some hardcoded limit there than it doesn't happen more often than every hour or so (or perhaps use a exponential backoff here too?) It is only to tell users to check mcelog output. -Andi - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
On Thu, Apr 26, 2007 at 06:02:52PM -0700, Tim Hockin wrote: Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. Can you please fix the documentation then? Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Makes sense. AMD RevF can do this using the threshold interrupts too for DIMM errors too without any delays -- perhaps it would also make sense to configure this by default that it always triggers on all DIMM errors. Right now it is just an option in /sys @@ -349,17 +349,24 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user console_logged) { + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, console_logged); printk(KERN_INFO Machine check events logged\n); The printk should not happen too often. Can you add some hardcoded limit there than it doesn't happen more often than every hour or so (or perhaps use a exponential backoff here too?) It is only to tell users to check mcelog output. -Andi - To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
On 27 Apr 2007 11:09:17 +0200, Andi Kleen [EMAIL PROTECTED] wrote: On Thu, Apr 26, 2007 at 06:02:52PM -0700, Tim Hockin wrote: Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. Can you please fix the documentation then? Which documentation, specifically? :) Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Makes sense. AMD RevF can do this using the threshold interrupts too for DIMM errors too without any delays -- perhaps it would also make sense to configure this by default that it always triggers on all DIMM errors. Right now it is just an option in /sys Can I look at this as a followon patch? I have a number of mce related patches in the pipeline, and I am trying to keep them small for testing sanity - they are hard enough to test :) The printk should not happen too often. Can you add some hardcoded limit there than it doesn't happen more often than every hour or so (or perhaps use a exponential backoff here too?) It is only to tell users to check mcelog output. Sure. I'll fix it up and hit you again today. - To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
On Fri, Apr 27, 2007 at 09:58:14AM -0700, Tim Hockin wrote: On 27 Apr 2007 11:09:17 +0200, Andi Kleen [EMAIL PROTECTED] wrote: On Thu, Apr 26, 2007 at 06:02:52PM -0700, Tim Hockin wrote: Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. Can you please fix the documentation then? Which documentation, specifically? :) Documentation/x86_64/{boot-options.txt,machinecheck} Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Makes sense. AMD RevF can do this using the threshold interrupts too for DIMM errors too without any delays -- perhaps it would also make sense to configure this by default that it always triggers on all DIMM errors. Right now it is just an option in /sys Can I look at this as a followon patch? I have a number of mce Sure. -Andi - To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
On 27 Apr 2007 19:02:30 +0200, Andi Kleen [EMAIL PROTECTED] wrote: On Fri, Apr 27, 2007 at 09:58:14AM -0700, Tim Hockin wrote: On 27 Apr 2007 11:09:17 +0200, Andi Kleen [EMAIL PROTECTED] wrote: On Thu, Apr 26, 2007 at 06:02:52PM -0700, Tim Hockin wrote: Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. Can you please fix the documentation then? Which documentation, specifically? :) Documentation/x86_64/{boot-options.txt,machinecheck} What needs to be changed in boot-options? Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Makes sense. AMD RevF can do this using the threshold interrupts too for DIMM errors too without any delays -- perhaps it would also make sense to configure this by default that it always triggers on all DIMM errors. Right now it is just an option in /sys Can I look at this as a followon patch? I have a number of mce Sure. -Andi - To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
On Fri, Apr 27, 2007 at 10:05:58AM -0700, Tim Hockin wrote: On 27 Apr 2007 19:02:30 +0200, Andi Kleen [EMAIL PROTECTED] wrote: On Fri, Apr 27, 2007 at 09:58:14AM -0700, Tim Hockin wrote: On 27 Apr 2007 11:09:17 +0200, Andi Kleen [EMAIL PROTECTED] wrote: On Thu, Apr 26, 2007 at 06:02:52PM -0700, Tim Hockin wrote: Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. Can you please fix the documentation then? Which documentation, specifically? :) Documentation/x86_64/{boot-options.txt,machinecheck} What needs to be changed in boot-options? Nothing for your patch sorry. Only in machinecheck -Andi - To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
From: Tim Hockin [EMAIL PROTECTED] Background: We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, especially when we are trying really hard to stress the system out. The current MCE poller uses a static interval which does not care whether it has or has not found MCEs recently. Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. The Machine check events logged printk() is rate limited to the check_interval, which should be identical behavior to the old functionality. Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Alternatives: I considered simply reducing the polling interval to 10 ms immediately and keeping it there as long as we continue to find errors. This felt a bit heavy handed, but does perform significantly better for the default check_interval of 5 minutes (we're using a few seconds when testing for DRAM errors). I could be convinced to go with this, if anyone felt it was not too aggressive. Testing: I used an error-injecting DIMM to create lots of correctable DRAM errors and verified that the polling interval accelerates. The printk() only happens once per check_interval seconds. Patch: This patch is against 2.6.21-rc7. Signed-Off-By: Tim Hockin [EMAIL PROTECTED] --- This is the second version of this patch. The more I think about the alternative described above, the more I like it, but I recognize that it is heavy-handed. If correctable errors matter that much to you, you can start with a small check_interval, I suppose. diff -pruN linux-2.6.20/Documentation/x86_64/machinecheck linux-2.6.20+th/Documentation/x86_64/machinecheck --- linux-2.6.20/Documentation/x86_64/machinecheck 2007-04-24 23:36:03.0 -0700 +++ linux-2.6.20+th/Documentation/x86_64/machinecheck 2007-04-27 10:11:10.0 -0700 @@ -36,7 +36,12 @@ between all CPUs. check_interval How often to poll for corrected machine check errors, in seconds - (Note output is hexademical). Default 5 minutes. + (Note output is hexademical). Default 5 minutes. When the poller + finds MCEs it triggers an exponential speedup (poll more often) on + the polling interval. When the poller stops finding MCEs, it + triggers an exponential backoff (poll less often) on the polling + interval. The check_interval variable is both the initial and + maximum polling interval. tolerant Tolerance level. When a machine check exception occurs for a non diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c linux-2.6.20+th/arch/x86_64/kernel/mce.c --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-27 10:01:02.0 -0700 +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c2007-04-27 10:41:02.0 -0700 @@ -323,10 +323,13 @@ void mce_log_therm_throt_event(unsigned #endif /* CONFIG_X86_MCE_INTEL */ /* - * Periodic polling timer for silent machine check errors. + * Periodic polling timer for silent machine check errors. If the + * poller finds an MCE, poll 2x faster. When the poller finds no more + * errors, poll 2x slower (up to check_interval seconds). */ static int check_interval = 5 * 60; /* 5 minutes */ +static int next_interval; /* in jiffies */ static void mcheck_timer(struct work_struct *work); static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer); @@ -339,7 +342,6 @@ static void mcheck_check_cpu(void *info) static void mcheck_timer(struct work_struct *work) { on_each_cpu(mcheck_check_cpu, NULL, 1, 1); - schedule_delayed_work(mcheck_work, check_interval * HZ); /* * It's ok to read stale data here for notify_user and @@ -349,17 +351,30 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user console_logged) { + static unsigned long last_print = 0; + unsigned long now = jiffies; + + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, console_logged); - printk(KERN_INFO Machine check events logged\n); + if ((now - last_print) = check_interval*HZ) { + last_print = now; + printk(KERN_INFO Machine check events logged\n); + } + } else { + next_interval = min(next_interval*2, check_interval*HZ); } + + schedule_delayed_work(mcheck_work, next_interval); } static __init int periodic_mcheck_init(void) { - if (check_interval) - schedule_delayed_work(mcheck_work,
Re: [PATCH] x86_64: dynamic MCE poll interval
- printk(KERN_INFO Machine check events logged\n); + if ((now - last_print) = check_interval*HZ) { Shouldn't this use time_after() to handle wrapping? -Andi - To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] x86_64: dynamic MCE poll interval
From: Tim Hockin [EMAIL PROTECTED] Background: We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, especially when we are trying really hard to stress the system out. The current MCE poller uses a static interval which does not care whether it has or has not found MCEs recently. Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. The Machine check events logged printk() is rate limited to the check_interval, which should be identical behavior to the old functionality. Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Alternatives: I considered simply reducing the polling interval to 10 ms immediately and keeping it there as long as we continue to find errors. This felt a bit heavy handed, but does perform significantly better for the default check_interval of 5 minutes (we're using a few seconds when testing for DRAM errors). I could be convinced to go with this, if anyone felt it was not too aggressive. Testing: I used an error-injecting DIMM to create lots of correctable DRAM errors and verified that the polling interval accelerates. The printk() only happens once per check_interval seconds. Patch: This patch is against 2.6.21-rc7. Signed-Off-By: Tim Hockin [EMAIL PROTECTED] --- This is the third version of this patch. The only change from the prior version is to use time_after_eq(). diff -pruN linux-2.6.20/Documentation/x86_64/machinecheck linux-2.6.20+th/Documentation/x86_64/machinecheck --- linux-2.6.20/Documentation/x86_64/machinecheck 2007-04-24 23:36:03.0 -0700 +++ linux-2.6.20+th/Documentation/x86_64/machinecheck 2007-04-27 10:11:10.0 -0700 @@ -36,7 +36,12 @@ between all CPUs. check_interval How often to poll for corrected machine check errors, in seconds - (Note output is hexademical). Default 5 minutes. + (Note output is hexademical). Default 5 minutes. When the poller + finds MCEs it triggers an exponential speedup (poll more often) on + the polling interval. When the poller stops finding MCEs, it + triggers an exponential backoff (poll less often) on the polling + interval. The check_interval variable is both the initial and + maximum polling interval. tolerant Tolerance level. When a machine check exception occurs for a non diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c linux-2.6.20+th/arch/x86_64/kernel/mce.c --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-27 10:01:02.0 -0700 +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c2007-04-27 10:41:02.0 -0700 @@ -323,10 +323,13 @@ void mce_log_therm_throt_event(unsigned #endif /* CONFIG_X86_MCE_INTEL */ /* - * Periodic polling timer for silent machine check errors. + * Periodic polling timer for silent machine check errors. If the + * poller finds an MCE, poll 2x faster. When the poller finds no more + * errors, poll 2x slower (up to check_interval seconds). */ static int check_interval = 5 * 60; /* 5 minutes */ +static int next_interval; /* in jiffies */ static void mcheck_timer(struct work_struct *work); static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer); @@ -339,7 +342,6 @@ static void mcheck_check_cpu(void *info) static void mcheck_timer(struct work_struct *work) { on_each_cpu(mcheck_check_cpu, NULL, 1, 1); - schedule_delayed_work(mcheck_work, check_interval * HZ); /* * It's ok to read stale data here for notify_user and @@ -349,17 +351,30 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user console_logged) { + static unsigned long last_print = 0; + unsigned long now = jiffies; + + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, console_logged); - printk(KERN_INFO Machine check events logged\n); + if (time_after_eq(now, last_print + (check_interval*HZ))) { + last_print = now; + printk(KERN_INFO Machine check events logged\n); + } + } else { + next_interval = min(next_interval*2, check_interval*HZ); } + + schedule_delayed_work(mcheck_work, next_interval); } static __init int periodic_mcheck_init(void) { - if (check_interval) - schedule_delayed_work(mcheck_work, check_interval*HZ); + next_interval = check_interval * HZ; + if (next_interval) + schedule_delayed_work(mcheck_work,
Re: [PATCH] x86_64: dynamic MCE poll interval
Sorry, Gmail mangles whitespace unless you do just the right thing. Let me work around it. Proper patch coming. On 4/27/07, Tim Hockin [EMAIL PROTECTED] wrote: From: Tim Hockin [EMAIL PROTECTED] Background: We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, especially when we are trying really hard to stress the system out. The current MCE poller uses a static interval which does not care whether it has or has not found MCEs recently. Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. The Machine check events logged printk() is rate limited to the check_interval, which should be identical behavior to the old functionality. Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Alternatives: I considered simply reducing the polling interval to 10 ms immediately and keeping it there as long as we continue to find errors. This felt a bit heavy handed, but does perform significantly better for the default check_interval of 5 minutes (we're using a few seconds when testing for DRAM errors). I could be convinced to go with this, if anyone felt it was not too aggressive. Testing: I used an error-injecting DIMM to create lots of correctable DRAM errors and verified that the polling interval accelerates. The printk() only happens once per check_interval seconds. Patch: This patch is against 2.6.21-rc7. Signed-Off-By: Tim Hockin [EMAIL PROTECTED] --- This is the third version of this patch. The only change from the prior version is to use time_after_eq(). diff -pruN linux-2.6.20/Documentation/x86_64/machinecheck linux-2.6.20+th/Documentation/x86_64/machinecheck --- linux-2.6.20/Documentation/x86_64/machinecheck 2007-04-24 23:36:03.0 -0700 +++ linux-2.6.20+th/Documentation/x86_64/machinecheck 2007-04-27 10:11:10.0 -0700 @@ -36,7 +36,12 @@ between all CPUs. check_interval How often to poll for corrected machine check errors, in seconds - (Note output is hexademical). Default 5 minutes. + (Note output is hexademical). Default 5 minutes. When the poller + finds MCEs it triggers an exponential speedup (poll more often) on + the polling interval. When the poller stops finding MCEs, it + triggers an exponential backoff (poll less often) on the polling + interval. The check_interval variable is both the initial and + maximum polling interval. tolerant Tolerance level. When a machine check exception occurs for a non diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c linux-2.6.20+th/arch/x86_64/kernel/mce.c --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-27 10:01:02.0 -0700 +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c2007-04-27 10:41:02.0 -0700 @@ -323,10 +323,13 @@ void mce_log_therm_throt_event(unsigned #endif /* CONFIG_X86_MCE_INTEL */ /* - * Periodic polling timer for silent machine check errors. + * Periodic polling timer for silent machine check errors. If the + * poller finds an MCE, poll 2x faster. When the poller finds no more + * errors, poll 2x slower (up to check_interval seconds). */ static int check_interval = 5 * 60; /* 5 minutes */ +static int next_interval; /* in jiffies */ static void mcheck_timer(struct work_struct *work); static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer); @@ -339,7 +342,6 @@ static void mcheck_check_cpu(void *info) static void mcheck_timer(struct work_struct *work) { on_each_cpu(mcheck_check_cpu, NULL, 1, 1); - schedule_delayed_work(mcheck_work, check_interval * HZ); /* * It's ok to read stale data here for notify_user and @@ -349,17 +351,30 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user console_logged) { + static unsigned long last_print = 0; + unsigned long now = jiffies; + + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, console_logged); - printk(KERN_INFO Machine check events logged\n); + if (time_after_eq(now, last_print + (check_interval*HZ))) { + last_print = now; + printk(KERN_INFO Machine check events logged\n); + } + } else { + next_interval = min(next_interval*2, check_interval*HZ); } + + schedule_delayed_work(mcheck_work, next_interval); } static __init int periodic_mcheck_init(void) { - if (check_interval) -
Re: [PATCH] x86_64: dynamic MCE poll interval
On Fri, 27 Apr 2007 13:51:28 -0700 Tim Hockin wrote: This is the third version of this patch. The only change from the prior version is to use time_after_eq(). diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c linux-2.6.20+th/arch/x86_64/kernel/mce.c --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-27 10:01:02.0 -0700 +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c 2007-04-27 10:41:02.0 -0700 @@ -349,17 +351,30 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user console_logged) { + static unsigned long last_print = 0; Don't need to init statics to 0... + unsigned long now = jiffies; + + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, console_logged); - printk(KERN_INFO Machine check events logged\n); + if (time_after_eq(now, last_print + (check_interval*HZ))) { + last_print = now; + printk(KERN_INFO Machine check events logged\n); + } + } else { + next_interval = min(next_interval*2, check_interval*HZ); } + + schedule_delayed_work(mcheck_work, next_interval); } --- ~Randy *** Remember to use Documentation/SubmitChecklist when testing your code *** - To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
[PATCH] x86_64: dynamic MCE poll interval (try 2)
From: Tim Hockin [EMAIL PROTECTED] Background: We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, especially when we are trying really hard to stress the system out. The current MCE poller uses a static interval which does not care whether it has or has not found MCEs recently. Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. The Machine check events logged printk() is rate limited to the check_interval, which should be identical behavior to the old functionality. Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Alternatives: I considered simply reducing the polling interval to 10 ms immediately and keeping it there as long as we continue to find errors. This felt a bit heavy handed, but does perform significantly better for the default check_interval of 5 minutes (we're using a few seconds when testing for DRAM errors). I could be convinced to go with this, if anyone felt it was not too aggressive. Testing: I used an error-injecting DIMM to create lots of correctable DRAM errors and verified that the polling interval accelerates. The printk() only happens once per check_interval seconds. Patch: This patch is against 2.6.21-rc7. Signed-Off-By: Tim Hockin [EMAIL PROTECTED] --- This is the fourth version of this patch. The only change from the prior version is to not initialize a static var to 0. This patch should come through cleanly, I hope. diff -pruN linux-2.6.20/Documentation/x86_64/machinecheck linux-2.6.20+th/Documentation/x86_64/machinecheck --- linux-2.6.20/Documentation/x86_64/machinecheck 2007-04-24 23:36:03.0 -0700 +++ linux-2.6.20+th/Documentation/x86_64/machinecheck 2007-04-27 10:11:10.0 -0700 @@ -36,7 +36,12 @@ between all CPUs. check_interval How often to poll for corrected machine check errors, in seconds - (Note output is hexademical). Default 5 minutes. + (Note output is hexademical). Default 5 minutes. When the poller + finds MCEs it triggers an exponential speedup (poll more often) on + the polling interval. When the poller stops finding MCEs, it + triggers an exponential backoff (poll less often) on the polling + interval. The check_interval variable is both the initial and + maximum polling interval. tolerant Tolerance level. When a machine check exception occurs for a non diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c linux-2.6.20+th/arch/x86_64/kernel/mce.c --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-27 10:01:02.0 -0700 +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c2007-04-27 10:41:02.0 -0700 @@ -323,10 +323,13 @@ void mce_log_therm_throt_event(unsigned #endif /* CONFIG_X86_MCE_INTEL */ /* - * Periodic polling timer for silent machine check errors. + * Periodic polling timer for silent machine check errors. If the + * poller finds an MCE, poll 2x faster. When the poller finds no more + * errors, poll 2x slower (up to check_interval seconds). */ static int check_interval = 5 * 60; /* 5 minutes */ +static int next_interval; /* in jiffies */ static void mcheck_timer(struct work_struct *work); static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer); @@ -339,7 +342,6 @@ static void mcheck_check_cpu(void *info) static void mcheck_timer(struct work_struct *work) { on_each_cpu(mcheck_check_cpu, NULL, 1, 1); - schedule_delayed_work(mcheck_work, check_interval * HZ); /* * It's ok to read stale data here for notify_user and @@ -349,17 +351,30 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user console_logged) { + static unsigned long last_print; + unsigned long now = jiffies; + + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, console_logged); - printk(KERN_INFO Machine check events logged\n); + if (time_after_eq(now, last_print + (check_interval*HZ))) { + last_print = now; + printk(KERN_INFO Machine check events logged\n); + } + } else { + next_interval = min(next_interval*2, check_interval*HZ); } + + schedule_delayed_work(mcheck_work, next_interval); } static __init int periodic_mcheck_init(void) { - if (check_interval) - schedule_delayed_work(mcheck_work, check_interval*HZ); + next_interval = check_interval * HZ; + if (next_interval) +
Re: [PATCH] x86_64: dynamic MCE poll interval (try 2)
On Fri, Apr 27, 2007 at 02:34:11PM -0700, Tim Hockin wrote: From: Tim Hockin [EMAIL PROTECTED] Background: We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, especially when we are trying really hard to stress the system out. The current MCE poller uses a static interval which does not care whether it has or has not found MCEs recently. Added thanks -Andi - To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
[PATCH] x86_64: dynamic MCE poll interval
From: Tim Hockin <[EMAIL PROTECTED]> Background: We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, especially when we are trying really hard to stress the system out. The current MCE poller uses a static interval which does not care whether it has or has not found MCEs recently. Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Alternatives: I considered simply reducing the polling interval to 10 ms immediately and keeping it there as long as we continue to find errors. This felt a bit heavy handed, but does perform significantly better for the default check_interval of 5 minutes (we're using a few seconds when testing for DRAM errors). Patch: This patch is against 2.6.21-rc7. Signed-Off-By: Tim Hockin <[EMAIL PROTECTED]>o --- diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c linux-2.6.20+th/arch/x86_64/kernel/mce.c --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-24 23:36:04.0 -0700 +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c2007-04-26 10:40:29.0 -0700 @@ -327,6 +327,7 @@ void mce_log_therm_throt_event(unsigned */ static int check_interval = 5 * 60; /* 5 minutes */ +static int next_interval; /* in jiffies */ static void mcheck_timer(struct work_struct *work); static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer); @@ -339,7 +340,6 @@ static void mcheck_check_cpu(void *info) static void mcheck_timer(struct work_struct *work) { on_each_cpu(mcheck_check_cpu, NULL, 1, 1); - schedule_delayed_work(_work, check_interval * HZ); /* * It's ok to read stale data here for notify_user and @@ -349,17 +349,24 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user && console_logged) { + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, _logged); printk(KERN_INFO "Machine check events logged\n"); + } else { + next_interval = min(next_interval*2, check_interval*HZ); } + + schedule_delayed_work(_work, next_interval); } static __init int periodic_mcheck_init(void) { - if (check_interval) - schedule_delayed_work(_work, check_interval*HZ); + next_interval = check_interval * HZ; + if (next_interval) + schedule_delayed_work(_work, next_interval); return 0; } __initcall(periodic_mcheck_init); @@ -597,12 +604,13 @@ static int mce_resume(struct sys_device /* Reinit MCEs after user configuration changes */ static void mce_restart(void) { - if (check_interval) + if (next_interval) cancel_delayed_work(_work); /* Timer race is harmless here */ on_each_cpu(mce_init, NULL, 1, 1); - if (check_interval) - schedule_delayed_work(_work, check_interval*HZ); + next_interval = check_interval * HZ; + if (next_interval) + schedule_delayed_work(_work, next_interval); } static struct sysdev_class mce_sysclass = { - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
[PATCH] x86_64: dynamic MCE poll interval
From: Tim Hockin [EMAIL PROTECTED] Background: We've found that MCEs (specifically DRAM SBEs) tend to come in bunches, especially when we are trying really hard to stress the system out. The current MCE poller uses a static interval which does not care whether it has or has not found MCEs recently. Description: This patch makes the MCE poller adjust the polling interval dynamically. If we find an MCE, poll 2x faster (down to 10 ms). When we stop finding MCEs, poll 2x slower (up to check_interval seconds). The check_interval tunable becomes the max polling interval. Result: If you start to take a lot of correctable errors (not exceptions), you log them faster and more accurately (less chance of overflowing the MCA registers). If you don't take a lot of errors, you will see no change. Alternatives: I considered simply reducing the polling interval to 10 ms immediately and keeping it there as long as we continue to find errors. This felt a bit heavy handed, but does perform significantly better for the default check_interval of 5 minutes (we're using a few seconds when testing for DRAM errors). Patch: This patch is against 2.6.21-rc7. Signed-Off-By: Tim Hockin [EMAIL PROTECTED]o --- diff -pruN linux-2.6.20/arch/x86_64/kernel/mce.c linux-2.6.20+th/arch/x86_64/kernel/mce.c --- linux-2.6.20/arch/x86_64/kernel/mce.c 2007-04-24 23:36:04.0 -0700 +++ linux-2.6.20+th/arch/x86_64/kernel/mce.c2007-04-26 10:40:29.0 -0700 @@ -327,6 +327,7 @@ void mce_log_therm_throt_event(unsigned */ static int check_interval = 5 * 60; /* 5 minutes */ +static int next_interval; /* in jiffies */ static void mcheck_timer(struct work_struct *work); static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer); @@ -339,7 +340,6 @@ static void mcheck_check_cpu(void *info) static void mcheck_timer(struct work_struct *work) { on_each_cpu(mcheck_check_cpu, NULL, 1, 1); - schedule_delayed_work(mcheck_work, check_interval * HZ); /* * It's ok to read stale data here for notify_user and @@ -349,17 +349,24 @@ static void mcheck_timer(struct work_str * writes. */ if (notify_user console_logged) { + /* if we logged an MCE, reduce the polling interval */ + next_interval = max(next_interval/2, HZ/100); notify_user = 0; clear_bit(0, console_logged); printk(KERN_INFO Machine check events logged\n); + } else { + next_interval = min(next_interval*2, check_interval*HZ); } + + schedule_delayed_work(mcheck_work, next_interval); } static __init int periodic_mcheck_init(void) { - if (check_interval) - schedule_delayed_work(mcheck_work, check_interval*HZ); + next_interval = check_interval * HZ; + if (next_interval) + schedule_delayed_work(mcheck_work, next_interval); return 0; } __initcall(periodic_mcheck_init); @@ -597,12 +604,13 @@ static int mce_resume(struct sys_device /* Reinit MCEs after user configuration changes */ static void mce_restart(void) { - if (check_interval) + if (next_interval) cancel_delayed_work(mcheck_work); /* Timer race is harmless here */ on_each_cpu(mce_init, NULL, 1, 1); - if (check_interval) - schedule_delayed_work(mcheck_work, check_interval*HZ); + next_interval = check_interval * HZ; + if (next_interval) + schedule_delayed_work(mcheck_work, next_interval); } static struct sysdev_class mce_sysclass = { - To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/