Fix slub_debug on 5.9-rc
Subsystems (2)
cover Fix slub_debug on 5.9-rc
Hi Vlastimil, et al,
Apologies in advance, for I am way out of my element at the moment.
I wanted to use slub_debug for a problem I'm looking at, and since the slab in question is always the same I added this:
slub_debug=FZ,dma-kmalloc-8
But then my system just crashes on bringup (using the 5.9 RCs):
[ 6.712339] Unable to handle kernel pointer dereference in virtual kernel address space [ 6.712344] Failing address: 0000004400000000 TEID: 0000004400000803 [ 6.712346] Fault in home space mode while using kernel ASCE. [ 6.712351] AS:00000000221ec007 R3:0000000000000024 [ 6.712406] Oops: 003b ilc:3 [#1] SMP [ 6.712410] Modules linked in: pkey zcrypt rng_core [ 6.712415] CPU: 6 PID: 8 Comm: kworker/u564:0 Not tainted 5.9.0-rc7 #4 [ 6.712418] Hardware name: IBM 2964 NE1 749 (LPAR)
The same option works fine on 5.8, so I bisected mm/ and ended up blaming commit e17f1dfba37b ("mm, slub: extend slub_debug syntax for multiple blocks"). I didn't bother reverting this from 5.9, because of the neighboring rework that this was a part of, but I did apply it to 5.8, and recreated the problem there.
Looking at the above commit, I note that kmem_cache_flags() now returns slub_debug instead of flags, which leads every kmem_cache (other than one I'm trying to debug) getting its flags set to zero. That wasn't the case previously, and seems quite suspect:
[ 0.709206] s->name=dma-kmalloc-96 flags=4000 s->flags=0 [ 0.709227] s->name=dma-kmalloc-192 flags=4000 s->flags=0 [ 0.709248] s->name=dma-kmalloc-8 flags=4000 s->flags=4500 [ 0.709269] s->name=dma-kmalloc-16 flags=4000 s->flags=0 [ 0.709290] s->name=dma-kmalloc-32 flags=4000 s->flags=0
If I change my slab_list to "dma-kmalloc-*", and all dma slabs get assigned the debug flags in addition to DMA, the panic seen on boot disappears and my system comes up. I didn't leave my system up long enoughu to see if weirdness on the other slabs occurred, but I suspect it would.
I am running on s390, with defconfig, but I don't think that matters here. Either way, the attached patch restores the variable that this routine returns for slabs not affected by slub_debug, and lets my system boot with various combinations of slub_debug parameters that I've tried. Hopefully it's close to what is needed here.
I look forward to hearing your thoughts on this, and am happy to try other things if I'm too far into the weeds. Thanks!
Eric Farman (1):
mm, slub: Restore initial kmem_cache flags
mm/slub.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
Signature · 1 line
2.17.1
PATCH 1/1 mm, slub: Restore initial kmem_cache flags
The routine that applies debug flags to the kmem_cache slabs inadvertantly prevents non-debug flags from being applied to those same objects. That is, if slub_debug=<flag>,<slab> is specified, non-debugged slabs will end up having flags of zero, and the slabs will be unusable. Fix this by returning the input flags for non-matching slabs as was done previously.
--- mm/slub.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
mm/slub.c +1 -1
@@ -1450,7 +1450,7 @@ slab_flags_t kmem_cache_flags(unsigned int object_size, } } -return slub_debug; +return flags; }#else /* !CONFIG_SLUB_DEBUG */ static inline void setup_object_debug(struct kmem_cache *s,
Signature · 1 line
2.17.1
Re: [PATCH 1/1] mm, slub: Restore initial kmem_cache flags
Quoted 6 lines · Eric Farman
The routine that applies debug flags to the
kmem_cacheslabs inadvertantly prevents non-debug flags from being applied to those same objects. That is, if slub_debug=<flag>,<slab> is specified, non-debugged slabs will end up having flags of zero, and the slabs will be unusable. Fix this by returning the input flags for non-matching slabs as was done previously.
Thanks a lot for debugging this and sorry for the trouble!
Quoted 2 lines
Fixes e17f1dfba37b ("mm, slub: extend slub_debug syntax for multiple blocks")Signed-off-by Eric Farman <[email protected]>
But with a small adjustment below:
Quoted 9 lines
--- mm/slub.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)mm/slub.c +0
-0
mm/slub.c slab_flags_t kmem_cache_flags(unsigned int object_size, · line 1453
Quoted 5 lines
mm/slub.c +1
-1@@ -1450,7 +1450,7 @@ slab_flags_t kmem_cache_flags(unsigned int object_size, } } -return slub_debug; +return flags;
To keep supporting the case of "debug flags set for all caches, with exceptions for listed caches", i.e. "slub_debug=FZ;-,zs_handle,zspage", we should return here this:
return flags | slub_debug;
Thanks again!
Quoted 4 lines
mm/slub.c +0
-0}#else /* !CONFIG_SLUB_DEBUG */ static inline void setup_object_debug(struct kmem_cache *s,
Re: [PATCH 1/1] mm, slub: Restore initial kmem_cache flags
Quoted 9 lines · Vlastimil Babka
Quoted 6 lines · Eric Farman
The routine that applies debug flags to the
kmem_cacheslabs inadvertantly prevents non-debug flags from being applied to those same objects. That is, if slub_debug=<flag>,<slab> is specified, non-debugged slabs will end up having flags of zero, and the slabs will be unusable. Fix this by returning the input flags for non-matching slabs as was done previously.Thanks a lot for debugging this and sorry for the trouble!
You're welcome. Just glad I wasn't losing my mind!
Quoted 20 lines
Quoted 3 lines
Fixes e17f1dfba37b ("mm, slub: extend slub_debug syntax for multiple blocks")Signed-off-by Eric Farman <[email protected]>Acked-by Vlastimil Babka <[email protected]>But with a small adjustment below:
Quoted 12 lines
--- mm/slub.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)mm/slub.c +0
-0@@ -1450,7 +1450,7 @@ slab_flags_t kmem_cache_flags(unsigned int object_size, } }
Quoted 8 lines
mm/slub.c slab_flags_t kmem_cache_flags(unsigned int object_size, · line 1453
Quoted 2 lines
mm/slub.c +1
-1-return slub_debug; +return flags;To keep supporting the case of "debug flags set for all caches, with exceptions for listed caches", i.e. "slub_debug=FZ;-,zs_handle,zspage", we should return here this:
return flags |
slub_debug;
Ah, cool... I wondered about that, but didn't go far enough down the combinations. Does it then make sense to strip out the "if (!slub_debug_string)" check at the beginning of the function? As in:
----8<----
mm/slub.c +0 -0
@@ -1413,10 +1413,6 @@ slab_flags_t kmem_cache_flags(unsigned int
object_size,
char *next_block;
slab_flags_t block_flags;- /* If slub_debug = 0, it folds into the if conditional. */
- if (!slub_debug_string)
- return flags | slub_debug;
-
len = strlen(name);
next_block = slub_debug_string;
/* Go through all blocks of debug options, see if any matches
our slab's name */
@@ -1450,7 +1446,7 @@ slab_flags_t kmem_cache_flags(unsigned int
object_size,
}
}- return flags;
+ return flags | slub_debug;
}
#else /* !CONFIG_SLUB_DEBUG */
static inline void setup_object_debug(struct kmem_cache *s,
Quoted 8 lines
Thanks again!
Quoted 4 lines
mm/slub.c +0
-0}#else /* !CONFIG_SLUB_DEBUG */ static inline void setup_object_debug(struct kmem_cache *s,
Re: [PATCH 1/1] mm, slub: Restore initial kmem_cache flags
Quoted 42 lines · Eric Farman
Quoted 9 lines · Vlastimil Babka
Quoted 6 lines · Eric Farman
The routine that applies debug flags to the
kmem_cacheslabs inadvertantly prevents non-debug flags from being applied to those same objects. That is, if slub_debug=<flag>,<slab> is specified, non-debugged slabs will end up having flags of zero, and the slabs will be unusable. Fix this by returning the input flags for non-matching slabs as was done previously.Thanks a lot for debugging this and sorry for the trouble!
You're welcome. Just glad I wasn't losing my mind!
Quoted 27 lines
Quoted 3 lines
Fixes e17f1dfba37b ("mm, slub: extend slub_debug syntax for multiple blocks")Signed-off-by Eric Farman <[email protected]>Acked-by Vlastimil Babka <[email protected]>But with a small adjustment below:
Quoted 9 lines
--- mm/slub.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)mm/slub.c +0
-0mm/slub.c slab_flags_t kmem_cache_flags(unsigned int object_size, · line 1453
Quoted 5 lines
mm/slub.c +1
-1@@ -1450,7 +1450,7 @@ slab_flags_t kmem_cache_flags(unsigned int object_size, } } -return slub_debug; +return flags;To keep supporting the case of "debug flags set for all caches, with exceptions for listed caches", i.e. "slub_debug=FZ;-,zs_handle,zspage", we should return here this:
Quoted 5 lines
Quoted 1 line
return flags |
slub_debug;Ah, cool... I wondered about that, but didn't go far enough down the combinations. Does it then make sense to strip out the "if (!slub_debug_string)" check at the beginning of the function? As in:
Yeah, that makes sense.
Thanks!
PATCH v2 1/1 mm, slub: Restore initial kmem_cache flags
The routine that applies debug flags to the kmem_cache slabs inadvertantly prevents non-debug flags from being applied to those same objects. That is, if slub_debug=<flag>,<slab> is specified, non-debugged slabs will end up having flags of zero, and the slabs may be unusable.
Fix this by including the input flags for non-matching slabs with the contents of slub_debug, so that the caches are created as expected alongside any debugging options that may be requested. With this, we can remove the check for a NULL slub_debug_string, since it's covered by the loop itself.
---
v2: OR return values per Vlastimil, applied his Ack
v1: https://lore.kernel.org/linux-mm/[email protected]/T/#t
--- mm/slub.c | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-)
mm/slub.c +1 -5
@@ -1413,10 +1413,6 @@ slab_flags_t kmem_cache_flags(unsigned int object_size,char *next_block;slab_flags_t block_flags; -/* If slub_debug = 0, it folds into the if conditional. */ -if (!slub_debug_string) -return flags | slub_debug; - len =strlen (name); next_block = slub_debug_string;/* Go through all blocks of debug options, see if any matches our slab's name */
@@ -1450,7 +1446,7 @@ slab_flags_t kmem_cache_flags(unsigned int object_size, } } -return slub_debug; +return flags | slub_debug; }#else /* !CONFIG_SLUB_DEBUG */ static inline void setup_object_debug(struct kmem_cache *s,
Signature · 1 line
2.17.1