panic on bootup due to __GFP_ZERO patch
Subsystems (3)
I'm getting a panic during pidmap_init with a backtrace that looks something like:
buffered_rmqueue__alloc_pagesget_zeroed_pagepidmap_initstart_kernel
Reverting the __GFP_ZERO patch fixes the issue, haven't drilled down any deeper yet to see what in the patch is causing the problem. This is x86 w/out HIGHMEM (and no NUMA).
Any ideas?
thanks,
-chris
Quoted 9 lines · Chris Wright
I'm getting a panic during
pidmap_initwith a backtrace that looks something like:
buffered_rmqueue__alloc_pagesget_zeroed_pagepidmap_initstart_kernel
Quoted 3 lines
Reverting the
__GFP_ZEROpatch fixes the issue, haven't drilled down any deeper yet to see what in the patch is causing the problem. This is x86 w/out HIGHMEM (and no NUMA).
ACK, there has been a number of folks hit by this since I updated the Fedora rawhide kernel to snapshots including this change.
https://bugzilla.redhat.com/beta/show_bug.cgi?id=144480
I've also hit in on my test box that has 256MB. The pattern so far does seem to be 'no highmem', though I've not actually tried a recent snapshot on my highmem box.
Dave
Quoted 10 lines · Chris Wright
I'm getting a panic during
pidmap_initwith a backtrace that looks something like:
buffered_rmqueue__alloc_pagesget_zeroed_pagepidmap_initstart_kernel
Quoted 4 lines
Reverting the
__GFP_ZEROpatch fixes the issue, haven't drilled down any deeper yet to see what in the patch is causing the problem. This is x86 w/out HIGHMEM (and no NUMA).
Well it's doing clear_highpage() before __alloc_pages() has called kernel_map_pages(), so CONFIG_DEBUG_PAGEALLOC is quite kaput.
So the current __GFP_ZERO buglist is:
1: Breaks CONFIG_DEBUG_PAGEALLOC
2: Breaks the cache aliasing protection for anonymous pages
3: prep_zero_page() uses KM_USER0 so __GFP_ZERO from IRQ context will
cause rare memory corruption.
PATCH 1/1 Fixes for prep_zero_page
Quoted 11 lines · Andrew Morton
Well it's doing
clear_highpage()before__alloc_pages()has calledkernel_map_pages(), soCONFIG_DEBUG_PAGEALLOCis quite kaput.So the current
__GFP_ZERObuglist is:1: Breaks
CONFIG_DEBUG_PAGEALLOC2: Breaks the cache aliasing protection for anonymous pages
3:
prep_zero_page()usesKM_USER0so__GFP_ZEROfrom IRQ context will
cause rare memory corruption.
The following should take care of 1 and 3. I opted to unmap the pages again after the clear page so that it remains isolated and we don't have to make additional checks to see if we should unmap the pages.
Index: linux-2.6.10-mm2/include/linux/highmem.h
===================================================================
RCS file: /home/cvsroot/linux-2.6.10-mm2/include/linux/highmem.h,v retrieving revision 1.1.1.1
include/linux/highmem.h +12 -0
@@ -50,6 +50,18 @@ static inline void clear_highpage(structkunmap_atomic (kaddr,KM_USER0 ); } +static inline void clear_irq_highpage (struct page *page) +{ +char *kaddr; +unsigned long flags; + +local_irq_save (flags); + kaddr =kmap_atomic (page,KM_IRQ0 ); +clear_page (kaddr); +kunmap_atomic (kaddr,KM_IRQ0 ); +local_irq_restore (flags); +} +/* * Same but also flushes aliased cache contents to RAM. */
Index: linux-2.6.10-mm2/mm/page_alloc.c
===================================================================
RCS file: /home/cvsroot/linux-2.6.10-mm2/mm/page_alloc.c,v retrieving revision 1.1.1.1
mm/page_alloc.c +10 -3
@@ -691,10 +691,17 @@ perthread_pages_alloc(void)*/ static inline void prep_zero_page (struct page *page,int order) { -int i; +int i, nr_pages = (1 << order); +int context =in_interrupt (); -for (i =0 ; i < (1 << order); i++) -clear_highpage (page + i); +kernel_map_pages (page, nr_pages,1 ); +for (i =0 ; i < nr_pages; i++) { +if (likely (!context)) +clear_highpage (page + i); +else +clear_irq_highpage (page + i); + } +kernel_map_pages (page, nr_pages,0 ); } static struct page *
Re: [PATCH] Fixes for prep_zero_page
Quoted 34 lines · Zwane Mwaikambo
Quoted 11 lines · Andrew Morton
Well it's doing
clear_highpage()before__alloc_pages()has calledkernel_map_pages(), soCONFIG_DEBUG_PAGEALLOCis quite kaput.So the current
__GFP_ZERObuglist is:1: Breaks
CONFIG_DEBUG_PAGEALLOC2: Breaks the cache aliasing protection for anonymous pages
3:
prep_zero_page()usesKM_USER0so__GFP_ZEROfrom IRQ context will
cause rare memory corruption.The following should take care of 1 and 3. I opted to unmap the pages again after the clear page so that it remains isolated and we don't have to make additional checks to see if we should unmap the pages.
Signed-off-by Zwane Mwaikambo <[email protected]>Index:
linux-2.6.10-mm2/include/linux/highmem.h
===================================================================
RCS file:/home/cvsroot/linux-2.6.10-mm2/include/linux/highmem.h,vretrieving revision 1.1.1.1
diff -u -p -B -r1.1.1.1highmem.hinclude/linux/highmem.h +2
-0@@ -50,6 +50,18 @@ static inline void clear_highpage(structkunmap_atomic (kaddr,KM_USER0 ); } +static inline void clear_irq_highpage (struct page *page) +{
include/linux/highmem.h static inline void clear_highpage(struct · line 64
Quoted 10 lines
include/linux/highmem.h +10
-0+char *kaddr; +unsigned long flags; + +local_irq_save (flags); + kaddr =kmap_atomic (page,KM_IRQ0 ); +clear_page (kaddr); +kunmap_atomic (kaddr,KM_IRQ0 ); +local_irq_restore (flags); +} +
This won't work right if someone tries to allocate memory while holding a KM_IRQ0 atomic kmap.
It would be quite bizarre for anyone to be allocating highmem pages from IRQ context anyway, but as a generic mechanism this really should work as expected in all contexts. That means a new kmap slot.
Quoted 19 lines
include/linux/highmem.h +0
-0/* * Same but also flushes aliased cache contents to RAM. */ Index:
linux-2.6.10-mm2/mm/page_alloc.c
===================================================================
RCS file:/home/cvsroot/linux-2.6.10-mm2/mm/page_alloc.c,vretrieving revision 1.1.1.1
diff -u -p -B -r1.1.1.1page_alloc.cmm/page_alloc.c +2
-2@@ -691,10 +691,17 @@ perthread_pages_alloc(void)*/ static inline void prep_zero_page (struct page *page,int order) { -int i; +int i, nr_pages = (1 << order); +int context =in_interrupt (); -for (i =0 ; i < (1 << order); i++)
mm/page_alloc.c perthread_pages_alloc(void) · line 705
Quoted 10 lines
mm/page_alloc.c +8
-1-clear_highpage (page + i); +kernel_map_pages (page, nr_pages,1 ); +for (i =0 ; i < nr_pages; i++) { +if (likely (!context)) +clear_highpage (page + i); +else +clear_irq_highpage (page + i); + } +kernel_map_pages (page, nr_pages,0 ); }
Can't we simply move the page zeroing to the very end of __alloc_pages()?
Re: [PATCH] Fixes for prep_zero_page
Quoted 6 lines · Andrew Morton
This won't work right if someone tries to allocate memory while holding a
KM_IRQ0atomic kmap.It would be quite bizarre for anyone to be allocating highmem pages from IRQ context anyway, but as a generic mechanism this really should work as expected in all contexts. That means a new kmap slot.
I was trying to find users of nested KM_IRQ to verify against so i just went with the first slot. The problem with sharing a slot with irq context is that you have to exclude interrupts whilst doing the zeroing too, unless we maybe create two slots.
Quoted 1 line
Can't we simply move the page zeroing to the very end of
__alloc_pages()?
Ok, i've changed that bit to something like;
Index: linux-2.6.10-mm2/mm/page_alloc.c
===================================================================
RCS file: /home/cvsroot/linux-2.6.10-mm2/mm/page_alloc.c,v retrieving revision 1.1.1.1
mm/page_alloc.c +2 -3
@@ -732,9 +740,6 @@ buffered_rmqueue(struct zone *zone, intmod_page_state_zone (zone, pgalloc,1 << order);prep_new_page (page, order); -if (gfp_flags & __GFP_ZERO) -prep_zero_page (page, order); -if (order && (gfp_flags & __GFP_COMP))prep_compound_page (page, order); }
@@ -935,6 +940,8 @@ nopage: got_pg:zone_statistics (zonelist, z);kernel_map_pages (page,1 << order,1 ); +if (gfp_mask & __GFP_ZERO) +prep_zero_page (page, order);return page; }
Re: [PATCH] Fixes for prep_zero_page
Quoted 4 lines · * Zwane Mwaikambo ([email protected]) wrote:
Quoted 1 line · Andrew Morton
Can't we simply move the page zeroing to the very end of
__alloc_pages()?Ok, i've changed that bit to something like;
I did it the other way around, and moved kernel_map_pages to prep_new_page so it's called before zeroing to keep that with the other prep bits in buffered_rmqueue. Made sense to me that kernel_map_pages is part of prepping a new page, but this isn't my area, so I could be way off ;-) It works for me with DEBUG_PAGEALLOC enabled.
===== mm/page_alloc.c 1.251 vs edited =====
mm/page_alloc.c +1 -1
@@ -413,6 +413,7 @@ static void prep_new_page(struct page *p 1 << PG_checked | 1 << PG_mappedtodisk); page->private =0 ;set_page_refs (page, order); +kernel_map_pages (page,1 << order,1 ); }/*
@@ -823,7 +824,6 @@ nopage:return NULL ; got_pg:zone_statistics (zonelist, z); -kernel_map_pages (page,1 << order,1 );return page; }
Re: [PATCH] Fixes for prep_zero_page
Quoted 11 lines · Chris Wright
Quoted 4 lines · * Zwane Mwaikambo ([email protected]) wrote:
Quoted 1 line · Andrew Morton
Can't we simply move the page zeroing to the very end of
__alloc_pages()?Ok, i've changed that bit to something like;
I did it the other way around, and moved
kernel_map_pagestoprep_new_pageso it's called before zeroing to keep that with the other prep bits inbuffered_rmqueue. Made sense to me thatkernel_map_pagesis part of prepping a new page, but this isn't my area, so I could be way off ;-) It works for me withDEBUG_PAGEALLOCenabled.
A lot more digestible than my offering ;)
Re: [PATCH] Fixes for prep_zero_page
It looks like it's still not happy with CONFIG_DEBUG_PAGEALLOC under load.
Unable to handle kernel paging request at virtual address ec5d97f4 printing eip: c014a882 *pde = 0083e067 Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC Modules linked in: CPU: 0 EIP: 0060:[<c014a882>] Not tainted VLI EFLAGS: 00010002 (2.6.10-mm2)
EIP is at check_slabuse+0x52/0xf0
eax: ec5d97f4 ebx: c18dd180 ecx: ec5b685c edx: ec5d9000
esi: c18dd180 edi: ec5d9000 ebp: c19d1efc esp: c19d1ed4
ds: 007b es: 007b ss: 0068
Process events/0 (pid: 6, threadinfo=c19d0000 task=c19abac0)
Stack: c19d1efc c0149c64 c19d1f0c c0149b0e c19d1ef0 00000000 ec5b685c ec5b685c
c18dd180 c18dd1a0 c19d1f24 c014aa19 c18dd2a0 c19d1f24 c014acec 00000000
c18dd180 c18dd2a0 c18dd180 00000005 c19d1f48 c014af91 c18dd2c8 c18dd2a0
Call Trace:
[<c0103fda>] show_stack+0x7a/0x90
[<c0104166>] show_registers+0x156/0x1c0
[<c0104380>] die+0x100/0x190
[<c0117159>] do_page_fault+0x369/0x65f
[<c0103c6f>] error_code+0x2b/0x30
[<c014aa19>] check_redzone+0xf9/0x140
[<c014af91>] cache_reap+0x221/0x230
[<c0130b2b>] worker_thread+0x17b/0x210
[<c0135605>] kthread+0xa5/0xb0
[<c0101415>] kernel_thread_helper+0x5/0x10
Code: 00 00 8d bc 27 00 00 00 00 83 c4 1c 5b 5e 5f 5d c3 8b 43 38 8b 7d ec 8b 4d f0 0f af f8 8b 410c 01 c7 89 fa 89 d8 e8 ee d3 ff ff <8b> 30 89 fa 89 d8 e8 13 d4 ff ff 81 fe 71 f0 2c 5a 8b 00 74 7b
<3>Debug: sleeping function called from invalid context atinclude/linux/rwsem.h:43
in_atomic():1, irqs_disabled():0
[<c0104007>] dump_stack+0x17/0x20 [<c011cb7c>] __might_sleep+0xac/0xc0 [<c0120753>] profile_task_exit+0x23/0x60 [<c012297c>] do_exit+0x1c/0x440 [<c0104408>] die+0x188/0x190 [<c0117159>] do_page_fault+0x369/0x65f [<c0103c6f>] error_code+0x2b/0x30 [<c014aa19>] check_redzone+0xf9/0x140 [<c014af91>] cache_reap+0x221/0x230 [<c0130b2b>] worker_thread+0x17b/0x210 [<c0135605>] kthread+0xa5/0xb0 [<c0101415>] kernel_thread_helper+0x5/0x10
note: events/0[6] exited with preempt_count 1
Re: [PATCH] Fixes for prep_zero_page
Quoted 3 lines · * Zwane Mwaikambo ([email protected]) wrote:
It looks like it's still not happy with
CONFIG_DEBUG_PAGEALLOCunder load.Unable to handle kernel paging request at virtual address ec5d97f4
Is that in vmalloc space?
Quoted 10 lines
printing eip:
c014a882
*pde =0083e067Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC Modules linked in: CPU: 0 EIP: 0060:[<c014a882>] Not tainted VLI EFLAGS: 00010002 (2.6.10-mm2)EIP is at
check_slabuse+0x52/0xf0
Hmm, isn't that from Manfred's patch to periodically scan? Doesn't look to me like it's related to the page prep fixup. What kind of load, etc?
thanks,
-chris
Re: [PATCH] Fixes for prep_zero_page
Quoted 6 lines · Chris Wright
Quoted 3 lines · * Zwane Mwaikambo ([email protected]) wrote:
It looks like it's still not happy with
CONFIG_DEBUG_PAGEALLOCunder load.Unable to handle kernel paging request at virtual address ec5d97f4Is that in vmalloc space?
Hmm it looks like.
Quoted 11 lines
Quoted 10 lines
printing eip:
c014a882
*pde =0083e067Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC Modules linked in: CPU: 0 EIP: 0060:[<c014a882>] Not tainted VLI EFLAGS: 00010002 (2.6.10-mm2)EIP is at
check_slabuse+0x52/0xf0
Quoted 2 lines
Hmm, isn't that from Manfred's patch to periodically scan? Doesn't look to me like it's related to the page prep fixup. What kind of load, etc?
I had an email exchange with Christopher and we came to the conclusion that it's breakage elsewhere with CONFIG_DEBUG_PAGEALLOC.