On Wed, Oct 17 2007 at 14:28 +0200, James Bottomley <James.Bottomley@xxxxxxxxxxxx> wrote: > On Tue, 2007-10-16 at 14:28 -0700, akpm@xxxxxxxxxxxxxxxxxxxx wrote: >> From: James Bottomley <James.Bottomley@xxxxxxxxxxxx> >> >> On Sun, 2007-10-14 at 12:21 -0700, Andrew Morton wrote: >>> On Sun, 14 Oct 2007 22:45:47 +0400 "Dave Milter" <davemilter@xxxxxxxxx> wrote: >>> >>>> I build linux-2.6.23-mm1 and try to boot it using qemu, >>>> and it crashed with trace like this: >>>> do_page_fault >>>> error_code >>>> lock_acquire >>>> _spin_lock_irqsave >>>> gdth_timeout >>>> run_timer_softirq >>>> __do_softirq >>>> do_softirq >>>> >>>> I have screenshot, but have no idea, is it legal to include it, if I >>>> sent copy to lkml. >>>> config of kernel in attachment, >>>> I apply all three patches from hot-fixes. >>>> >>> The screenshot is here: http://userweb.kernel.org/~akpm/crash.png >>> >>> It would appear that gdth_timeout() is passing a bad pointer into >>> spin_lock_irqsave(). >> There's a bug in the gdth rework in that the instance can be deleted >> from the list before the actual timer is stopped. This can be worked >> around I think by the following patch; although we really should be >> stopping the timer from firing when the list goes empty. >> >> >> >> Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> >> --- >> >> drivers/scsi/gdth.c | 3 +++ >> 1 file changed, 3 insertions(+) >> >> diff -puN drivers/scsi/gdth.c~git-scsi-misc-gdth-fix drivers/scsi/gdth.c >> --- a/drivers/scsi/gdth.c~git-scsi-misc-gdth-fix >> +++ a/drivers/scsi/gdth.c >> @@ -3793,6 +3793,9 @@ static void gdth_timeout(ulong data) >> gdth_ha_str *ha; >> ulong flags; >> >> + if (list_empty(&gdth_instances)) >> + return; >> + >> ha = list_first_entry(&gdth_instances, gdth_ha_str, list); >> spin_lock_irqsave(&ha->smp_lock, flags); >> > > This is almost certainly the wrong fix for real hardware. Although it > kills the timer when the list goes empty, nothing will ever restart it > when the list fills again. > > Boaz, since you touched all of this, you get to fix it. The correct fix > will be to control the timer along with the actual list instead of at > entry/exit time. If you're not going to add this empty check to the > timer routine, make sure you use del_timer_sync() before removing the > last element from the list. > > James > > OK I can see the crash, but than I do not understand why the fix above works. Let me explain gdth is not yet hotplug. the only place that deletes hosts is in the module_exit(gdth_exit) routine. Now, the code does not do a list_del() anywhere it will just deallocate the hosts and only after that delete the timer. Now, suppose a timer gets through between the time hosts are deallocated and the timer is removed. The list is not empty it points to garbage memory, yes. But the check above will not fix it. See my problem. please try below patch, and report if the WARN_ON I put triggers. If it works I will send a proper patch later with a one that does a list_del() too. (this is with out the patch above) ------------- diff --git a/drivers/scsi/gdth.c b/drivers/scsi/gdth.c index 3ac080e..b84a869 100644 --- a/drivers/scsi/gdth.c +++ b/drivers/scsi/gdth.c @@ -3793,6 +3793,11 @@ static void gdth_timeout(ulong data) gdth_ha_str *ha; ulong flags; + if (list_empty(&gdth_instances)) { + WARN_ON(1); + return; + } + ha = list_first_entry(&gdth_instances, gdth_ha_str, list); spin_lock_irqsave(&ha->smp_lock, flags); @@ -5236,12 +5241,13 @@ static void __exit gdth_exit(void) { gdth_ha_str *ha; +#ifdef GDTH_STATISTICS + del_timer_sync(&gdth_timer); +#endif + list_for_each_entry(ha, &gdth_instances, list) gdth_remove_one(ha); -#ifdef GDTH_STATISTICS - del_timer(&gdth_timer); -#endif unregister_chrdev(major,"gdth"); unregister_reboot_notifier(&gdth_notifier); } - To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
