| Message ID | 80c767a5d5927c099aea5178fbf2c897b459fa90.1732106544.git.geert@linux-m68k.org (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | slab: Fix too strict alignment check in create_cache() | expand |
On Wed, Nov 20, 2024 at 01:46:21PM +0100, Geert Uytterhoeven wrote: > On m68k, where the minimum alignment of unsigned long is 2 bytes: > > Kernel panic - not syncing: __kmem_cache_create_args: Failed to create slab 'io_kiocb'. Error -22 > CPU: 0 UID: 0 PID: 1 Comm: swapper Not tainted 6.12.0-atari-03776-g7eaa1f99261a #1783 > Stack from 0102fe5c: > 0102fe5c 00514a2b 00514a2b ffffff00 00000001 0051f5ed 00425e78 00514a2b > 0041eb74 ffffffea 00000310 0051f5ed ffffffea ffffffea 00601f60 00000044 > 0102ff20 000e7a68 0051ab8e 004383b8 0051f5ed ffffffea 000000b8 00000007 > 01020c00 00000000 000e77f0 0041e5f0 005f67c0 0051f5ed 000000b6 0102fef4 > 00000310 0102fef4 00000000 00000016 005f676c 0060a34c 00000010 00000004 > 00000038 0000009a 01000000 000000b8 005f668e 0102e000 00001372 0102ff88 > Call Trace: [<00425e78>] dump_stack+0xc/0x10 > [<0041eb74>] panic+0xd8/0x26c > [<000e7a68>] __kmem_cache_create_args+0x278/0x2e8 > [<000e77f0>] __kmem_cache_create_args+0x0/0x2e8 > [<0041e5f0>] memset+0x0/0x8c > [<005f67c0>] io_uring_init+0x54/0xd2 > > The minimal alignment of an integral type may differ from its size, > hence is not safe to assume that an arbitrary freeptr_t (which is > basically an unsigned long) is always aligned to 4 or 8 bytes. > > As nothing seems to require the additional alignment, it is safe to fix > this by relaxing the check to the actual minimum alignment of freeptr_t. > > Fixes: aaa736b186239b7d ("io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache") > Fixes: d345bd2e9834e2da ("mm: add kmem_cache_create_rcu()") > Reported-by: Guenter Roeck <linux@roeck-us.net> > Closes: https://lore.kernel.org/37c588d4-2c32-4aad-a19e-642961f200d7@roeck-us.net > Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org> On m68k: Tested-by: Guenter Roeck <linux@roeck-us.net>
On 11/20/24 5:49 AM, Geert Uytterhoeven wrote: > On m68k, where the minimum alignment of unsigned long is 2 bytes: > > Kernel panic - not syncing: __kmem_cache_create_args: Failed to create slab 'io_kiocb'. Error -22 > CPU: 0 UID: 0 PID: 1 Comm: swapper Not tainted 6.12.0-atari-03776-g7eaa1f99261a #1783 > Stack from 0102fe5c: > 0102fe5c 00514a2b 00514a2b ffffff00 00000001 0051f5ed 00425e78 00514a2b > 0041eb74 ffffffea 00000310 0051f5ed ffffffea ffffffea 00601f60 00000044 > 0102ff20 000e7a68 0051ab8e 004383b8 0051f5ed ffffffea 000000b8 00000007 > 01020c00 00000000 000e77f0 0041e5f0 005f67c0 0051f5ed 000000b6 0102fef4 > 00000310 0102fef4 00000000 00000016 005f676c 0060a34c 00000010 00000004 > 00000038 0000009a 01000000 000000b8 005f668e 0102e000 00001372 0102ff88 > Call Trace: [<00425e78>] dump_stack+0xc/0x10 > [<0041eb74>] panic+0xd8/0x26c > [<000e7a68>] __kmem_cache_create_args+0x278/0x2e8 > [<000e77f0>] __kmem_cache_create_args+0x0/0x2e8 > [<0041e5f0>] memset+0x0/0x8c > [<005f67c0>] io_uring_init+0x54/0xd2 > > The minimal alignment of an integral type may differ from its size, > hence is not safe to assume that an arbitrary freeptr_t (which is > basically an unsigned long) is always aligned to 4 or 8 bytes. > > As nothing seems to require the additional alignment, it is safe to fix > this by relaxing the check to the actual minimum alignment of freeptr_t. > > Fixes: aaa736b186239b7d ("io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache") > Fixes: d345bd2e9834e2da ("mm: add kmem_cache_create_rcu()") > Reported-by: Guenter Roeck <linux@roeck-us.net> > Closes: https://lore.kernel.org/37c588d4-2c32-4aad-a19e-642961f200d7@roeck-us.net > Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org> > --- > mm/slab_common.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/mm/slab_common.c b/mm/slab_common.c > index 893d320599151845..f2f201d865c108bd 100644 > --- a/mm/slab_common.c > +++ b/mm/slab_common.c > @@ -230,7 +230,7 @@ static struct kmem_cache *create_cache(const char *name, > if (args->use_freeptr_offset && > (args->freeptr_offset >= object_size || > !(flags & SLAB_TYPESAFE_BY_RCU) || > - !IS_ALIGNED(args->freeptr_offset, sizeof(freeptr_t)))) > + !IS_ALIGNED(args->freeptr_offset, __alignof(freeptr_t)))) > goto out; > > err = -ENOMEM; This looks much better, thanks. Reviewed-by: Jens Axboe <axboe@kernel.dk>
On 11/20/24 13:49, Geert Uytterhoeven wrote: > On m68k, where the minimum alignment of unsigned long is 2 bytes: > > Kernel panic - not syncing: __kmem_cache_create_args: Failed to create slab 'io_kiocb'. Error -22 > CPU: 0 UID: 0 PID: 1 Comm: swapper Not tainted 6.12.0-atari-03776-g7eaa1f99261a #1783 > Stack from 0102fe5c: > 0102fe5c 00514a2b 00514a2b ffffff00 00000001 0051f5ed 00425e78 00514a2b > 0041eb74 ffffffea 00000310 0051f5ed ffffffea ffffffea 00601f60 00000044 > 0102ff20 000e7a68 0051ab8e 004383b8 0051f5ed ffffffea 000000b8 00000007 > 01020c00 00000000 000e77f0 0041e5f0 005f67c0 0051f5ed 000000b6 0102fef4 > 00000310 0102fef4 00000000 00000016 005f676c 0060a34c 00000010 00000004 > 00000038 0000009a 01000000 000000b8 005f668e 0102e000 00001372 0102ff88 > Call Trace: [<00425e78>] dump_stack+0xc/0x10 > [<0041eb74>] panic+0xd8/0x26c > [<000e7a68>] __kmem_cache_create_args+0x278/0x2e8 > [<000e77f0>] __kmem_cache_create_args+0x0/0x2e8 > [<0041e5f0>] memset+0x0/0x8c > [<005f67c0>] io_uring_init+0x54/0xd2 > > The minimal alignment of an integral type may differ from its size, > hence is not safe to assume that an arbitrary freeptr_t (which is > basically an unsigned long) is always aligned to 4 or 8 bytes. > > As nothing seems to require the additional alignment, it is safe to fix > this by relaxing the check to the actual minimum alignment of freeptr_t. > > Fixes: aaa736b186239b7d ("io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache") > Fixes: d345bd2e9834e2da ("mm: add kmem_cache_create_rcu()") > Reported-by: Guenter Roeck <linux@roeck-us.net> > Closes: https://lore.kernel.org/37c588d4-2c32-4aad-a19e-642961f200d7@roeck-us.net > Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org> Thanks, will add it to slab pull for 6.13. > --- > mm/slab_common.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/mm/slab_common.c b/mm/slab_common.c > index 893d320599151845..f2f201d865c108bd 100644 > --- a/mm/slab_common.c > +++ b/mm/slab_common.c > @@ -230,7 +230,7 @@ static struct kmem_cache *create_cache(const char *name, > if (args->use_freeptr_offset && > (args->freeptr_offset >= object_size || > !(flags & SLAB_TYPESAFE_BY_RCU) || > - !IS_ALIGNED(args->freeptr_offset, sizeof(freeptr_t)))) > + !IS_ALIGNED(args->freeptr_offset, __alignof(freeptr_t)))) Seems only bunch of places uses __alignof but many use __alignoff__ and this also is what seems to be documented? > goto out; > > err = -ENOMEM;
On 11/20/24 07:03, Vlastimil Babka wrote: > On 11/20/24 13:49, Geert Uytterhoeven wrote: >> On m68k, where the minimum alignment of unsigned long is 2 bytes: >> >> Kernel panic - not syncing: __kmem_cache_create_args: Failed to create slab 'io_kiocb'. Error -22 >> CPU: 0 UID: 0 PID: 1 Comm: swapper Not tainted 6.12.0-atari-03776-g7eaa1f99261a #1783 >> Stack from 0102fe5c: >> 0102fe5c 00514a2b 00514a2b ffffff00 00000001 0051f5ed 00425e78 00514a2b >> 0041eb74 ffffffea 00000310 0051f5ed ffffffea ffffffea 00601f60 00000044 >> 0102ff20 000e7a68 0051ab8e 004383b8 0051f5ed ffffffea 000000b8 00000007 >> 01020c00 00000000 000e77f0 0041e5f0 005f67c0 0051f5ed 000000b6 0102fef4 >> 00000310 0102fef4 00000000 00000016 005f676c 0060a34c 00000010 00000004 >> 00000038 0000009a 01000000 000000b8 005f668e 0102e000 00001372 0102ff88 >> Call Trace: [<00425e78>] dump_stack+0xc/0x10 >> [<0041eb74>] panic+0xd8/0x26c >> [<000e7a68>] __kmem_cache_create_args+0x278/0x2e8 >> [<000e77f0>] __kmem_cache_create_args+0x0/0x2e8 >> [<0041e5f0>] memset+0x0/0x8c >> [<005f67c0>] io_uring_init+0x54/0xd2 >> >> The minimal alignment of an integral type may differ from its size, >> hence is not safe to assume that an arbitrary freeptr_t (which is >> basically an unsigned long) is always aligned to 4 or 8 bytes. >> >> As nothing seems to require the additional alignment, it is safe to fix >> this by relaxing the check to the actual minimum alignment of freeptr_t. >> >> Fixes: aaa736b186239b7d ("io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache") >> Fixes: d345bd2e9834e2da ("mm: add kmem_cache_create_rcu()") >> Reported-by: Guenter Roeck <linux@roeck-us.net> >> Closes: https://lore.kernel.org/37c588d4-2c32-4aad-a19e-642961f200d7@roeck-us.net >> Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org> > > Thanks, will add it to slab pull for 6.13. > >> --- >> mm/slab_common.c | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/mm/slab_common.c b/mm/slab_common.c >> index 893d320599151845..f2f201d865c108bd 100644 >> --- a/mm/slab_common.c >> +++ b/mm/slab_common.c >> @@ -230,7 +230,7 @@ static struct kmem_cache *create_cache(const char *name, >> if (args->use_freeptr_offset && >> (args->freeptr_offset >= object_size || >> !(flags & SLAB_TYPESAFE_BY_RCU) || >> - !IS_ALIGNED(args->freeptr_offset, sizeof(freeptr_t)))) >> + !IS_ALIGNED(args->freeptr_offset, __alignof(freeptr_t)))) > > Seems only bunch of places uses __alignof but many use __alignoff__ and this > also is what seems to be documented? __alignoff__ -> __alignof__ Guenter
On 11/20/24 16:14, Guenter Roeck wrote: > On 11/20/24 07:03, Vlastimil Babka wrote: >> On 11/20/24 13:49, Geert Uytterhoeven wrote: >>> On m68k, where the minimum alignment of unsigned long is 2 bytes: >>> >>> Kernel panic - not syncing: __kmem_cache_create_args: Failed to create slab 'io_kiocb'. Error -22 >>> CPU: 0 UID: 0 PID: 1 Comm: swapper Not tainted 6.12.0-atari-03776-g7eaa1f99261a #1783 >>> Stack from 0102fe5c: >>> 0102fe5c 00514a2b 00514a2b ffffff00 00000001 0051f5ed 00425e78 00514a2b >>> 0041eb74 ffffffea 00000310 0051f5ed ffffffea ffffffea 00601f60 00000044 >>> 0102ff20 000e7a68 0051ab8e 004383b8 0051f5ed ffffffea 000000b8 00000007 >>> 01020c00 00000000 000e77f0 0041e5f0 005f67c0 0051f5ed 000000b6 0102fef4 >>> 00000310 0102fef4 00000000 00000016 005f676c 0060a34c 00000010 00000004 >>> 00000038 0000009a 01000000 000000b8 005f668e 0102e000 00001372 0102ff88 >>> Call Trace: [<00425e78>] dump_stack+0xc/0x10 >>> [<0041eb74>] panic+0xd8/0x26c >>> [<000e7a68>] __kmem_cache_create_args+0x278/0x2e8 >>> [<000e77f0>] __kmem_cache_create_args+0x0/0x2e8 >>> [<0041e5f0>] memset+0x0/0x8c >>> [<005f67c0>] io_uring_init+0x54/0xd2 >>> >>> The minimal alignment of an integral type may differ from its size, >>> hence is not safe to assume that an arbitrary freeptr_t (which is >>> basically an unsigned long) is always aligned to 4 or 8 bytes. >>> >>> As nothing seems to require the additional alignment, it is safe to fix >>> this by relaxing the check to the actual minimum alignment of freeptr_t. >>> >>> Fixes: aaa736b186239b7d ("io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache") >>> Fixes: d345bd2e9834e2da ("mm: add kmem_cache_create_rcu()") >>> Reported-by: Guenter Roeck <linux@roeck-us.net> >>> Closes: https://lore.kernel.org/37c588d4-2c32-4aad-a19e-642961f200d7@roeck-us.net >>> Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org> >> >> Thanks, will add it to slab pull for 6.13. >> >>> --- >>> mm/slab_common.c | 2 +- >>> 1 file changed, 1 insertion(+), 1 deletion(-) >>> >>> diff --git a/mm/slab_common.c b/mm/slab_common.c >>> index 893d320599151845..f2f201d865c108bd 100644 >>> --- a/mm/slab_common.c >>> +++ b/mm/slab_common.c >>> @@ -230,7 +230,7 @@ static struct kmem_cache *create_cache(const char *name, >>> if (args->use_freeptr_offset && >>> (args->freeptr_offset >= object_size || >>> !(flags & SLAB_TYPESAFE_BY_RCU) || >>> - !IS_ALIGNED(args->freeptr_offset, sizeof(freeptr_t)))) >>> + !IS_ALIGNED(args->freeptr_offset, __alignof(freeptr_t)))) >> >> Seems only bunch of places uses __alignof but many use __alignoff__ and this >> also is what seems to be documented? > > __alignoff__ -> __alignof__ Yeah I meant __alignof__ Will chage it locally then. > Guenter >
Hi Vlastimil, On Wed, Nov 20, 2024 at 4:44 PM Vlastimil Babka <vbabka@suse.cz> wrote: > On 11/20/24 16:14, Guenter Roeck wrote: > > On 11/20/24 07:03, Vlastimil Babka wrote: > >> On 11/20/24 13:49, Geert Uytterhoeven wrote: > >>> On m68k, where the minimum alignment of unsigned long is 2 bytes: > >>> > >>> Kernel panic - not syncing: __kmem_cache_create_args: Failed to create slab 'io_kiocb'. Error -22 > >>> CPU: 0 UID: 0 PID: 1 Comm: swapper Not tainted 6.12.0-atari-03776-g7eaa1f99261a #1783 > >>> Stack from 0102fe5c: > >>> 0102fe5c 00514a2b 00514a2b ffffff00 00000001 0051f5ed 00425e78 00514a2b > >>> 0041eb74 ffffffea 00000310 0051f5ed ffffffea ffffffea 00601f60 00000044 > >>> 0102ff20 000e7a68 0051ab8e 004383b8 0051f5ed ffffffea 000000b8 00000007 > >>> 01020c00 00000000 000e77f0 0041e5f0 005f67c0 0051f5ed 000000b6 0102fef4 > >>> 00000310 0102fef4 00000000 00000016 005f676c 0060a34c 00000010 00000004 > >>> 00000038 0000009a 01000000 000000b8 005f668e 0102e000 00001372 0102ff88 > >>> Call Trace: [<00425e78>] dump_stack+0xc/0x10 > >>> [<0041eb74>] panic+0xd8/0x26c > >>> [<000e7a68>] __kmem_cache_create_args+0x278/0x2e8 > >>> [<000e77f0>] __kmem_cache_create_args+0x0/0x2e8 > >>> [<0041e5f0>] memset+0x0/0x8c > >>> [<005f67c0>] io_uring_init+0x54/0xd2 > >>> > >>> The minimal alignment of an integral type may differ from its size, > >>> hence is not safe to assume that an arbitrary freeptr_t (which is > >>> basically an unsigned long) is always aligned to 4 or 8 bytes. > >>> > >>> As nothing seems to require the additional alignment, it is safe to fix > >>> this by relaxing the check to the actual minimum alignment of freeptr_t. > >>> > >>> Fixes: aaa736b186239b7d ("io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache") > >>> Fixes: d345bd2e9834e2da ("mm: add kmem_cache_create_rcu()") > >>> Reported-by: Guenter Roeck <linux@roeck-us.net> > >>> Closes: https://lore.kernel.org/37c588d4-2c32-4aad-a19e-642961f200d7@roeck-us.net > >>> Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org> > >> > >> Thanks, will add it to slab pull for 6.13. > >> > >>> --- > >>> mm/slab_common.c | 2 +- > >>> 1 file changed, 1 insertion(+), 1 deletion(-) > >>> > >>> diff --git a/mm/slab_common.c b/mm/slab_common.c > >>> index 893d320599151845..f2f201d865c108bd 100644 > >>> --- a/mm/slab_common.c > >>> +++ b/mm/slab_common.c > >>> @@ -230,7 +230,7 @@ static struct kmem_cache *create_cache(const char *name, > >>> if (args->use_freeptr_offset && > >>> (args->freeptr_offset >= object_size || > >>> !(flags & SLAB_TYPESAFE_BY_RCU) || > >>> - !IS_ALIGNED(args->freeptr_offset, sizeof(freeptr_t)))) > >>> + !IS_ALIGNED(args->freeptr_offset, __alignof(freeptr_t)))) > >> > >> Seems only bunch of places uses __alignof but many use __alignoff__ and this > >> also is what seems to be documented? > > > > __alignoff__ -> __alignof__ > > Yeah I meant __alignof__ > Will chage it locally then. Thank you! Gr{oetje,eeting}s, Geert
On Wed, 20 Nov 2024, Vlastimil Babka wrote: > > > > Fixes: aaa736b186239b7d ("io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache") > > Fixes: d345bd2e9834e2da ("mm: add kmem_cache_create_rcu()") > > Reported-by: Guenter Roeck <linux@roeck-us.net> > > Closes: https://lore.kernel.org/37c588d4-2c32-4aad-a19e-642961f200d7@roeck-us.net > > Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org> > > Thanks, will add it to slab pull for 6.13. Note that there are widespread assumptions in kernel code that the alignment of scalars is the "natural alignment". Other portions of the kernel may break. The compiler actually goes along with this?? How do you deal with torn reads/writes in such a scenario? Is this UP only?
On Wed, Nov 20, 2024 at 09:50:47AM -0800, Christoph Lameter (Ampere) wrote: > On Wed, 20 Nov 2024, Vlastimil Babka wrote: > > > > > > > Fixes: aaa736b186239b7d ("io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache") > > > Fixes: d345bd2e9834e2da ("mm: add kmem_cache_create_rcu()") > > > Reported-by: Guenter Roeck <linux@roeck-us.net> > > > Closes: https://lore.kernel.org/37c588d4-2c32-4aad-a19e-642961f200d7@roeck-us.net > > > Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org> > > > > Thanks, will add it to slab pull for 6.13. > > Note that there are widespread assumptions in kernel code that the > alignment of scalars is the "natural alignment". Other portions of the > kernel may break. The compiler actually goes along with this?? u64s aren't aligned on x86-32. it's caused some problems over the years, but things work ok in general. > How do you deal with torn reads/writes in such a scenario? Is this UP > only? there were never a lot of smp m68k. not sure i can think of one, tbh. sun3 and hp300/400 seem like the obvious people who might have done an smp m68k, but neither did.
Hi Christoph, On Wed, Nov 20, 2024 at 6:50 PM Christoph Lameter (Ampere) <cl@gentwo.org> wrote: > On Wed, 20 Nov 2024, Vlastimil Babka wrote: > > > Fixes: aaa736b186239b7d ("io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache") > > > Fixes: d345bd2e9834e2da ("mm: add kmem_cache_create_rcu()") > > > Reported-by: Guenter Roeck <linux@roeck-us.net> > > > Closes: https://lore.kernel.org/37c588d4-2c32-4aad-a19e-642961f200d7@roeck-us.net > > > Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org> > > > > Thanks, will add it to slab pull for 6.13. > > Note that there are widespread assumptions in kernel code that the > alignment of scalars is the "natural alignment". Other portions of the > kernel may break. The compiler actually goes along with this?? Linux has supported m68k since last century. Any new such assumptions are fixed quickly (at least in the kernel). If you need a specific alignment, make sure to use __aligned and/or appropriate padding in structures. And yes, the compiler knows, and provides __alignof__. > How do you deal with torn reads/writes in such a scenario? Is this UP > only? Linux does not support (rate) SMP m68k machines. Gr{oetje,eeting}s, Geert
On Wed, Nov 20, 2024 at 01:46:21PM +0100, Geert Uytterhoeven wrote: > On m68k, where the minimum alignment of unsigned long is 2 bytes: > > Kernel panic - not syncing: __kmem_cache_create_args: Failed to create slab 'io_kiocb'. Error -22 > CPU: 0 UID: 0 PID: 1 Comm: swapper Not tainted 6.12.0-atari-03776-g7eaa1f99261a #1783 > Stack from 0102fe5c: > 0102fe5c 00514a2b 00514a2b ffffff00 00000001 0051f5ed 00425e78 00514a2b > 0041eb74 ffffffea 00000310 0051f5ed ffffffea ffffffea 00601f60 00000044 > 0102ff20 000e7a68 0051ab8e 004383b8 0051f5ed ffffffea 000000b8 00000007 > 01020c00 00000000 000e77f0 0041e5f0 005f67c0 0051f5ed 000000b6 0102fef4 > 00000310 0102fef4 00000000 00000016 005f676c 0060a34c 00000010 00000004 > 00000038 0000009a 01000000 000000b8 005f668e 0102e000 00001372 0102ff88 > Call Trace: [<00425e78>] dump_stack+0xc/0x10 > [<0041eb74>] panic+0xd8/0x26c > [<000e7a68>] __kmem_cache_create_args+0x278/0x2e8 > [<000e77f0>] __kmem_cache_create_args+0x0/0x2e8 > [<0041e5f0>] memset+0x0/0x8c > [<005f67c0>] io_uring_init+0x54/0xd2 > > The minimal alignment of an integral type may differ from its size, > hence is not safe to assume that an arbitrary freeptr_t (which is > basically an unsigned long) is always aligned to 4 or 8 bytes. > > As nothing seems to require the additional alignment, it is safe to fix > this by relaxing the check to the actual minimum alignment of freeptr_t. > > Fixes: aaa736b186239b7d ("io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache") > Fixes: d345bd2e9834e2da ("mm: add kmem_cache_create_rcu()") > Reported-by: Guenter Roeck <linux@roeck-us.net> > Closes: https://lore.kernel.org/37c588d4-2c32-4aad-a19e-642961f200d7@roeck-us.net > Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org> > --- Looks good to me, Reviewed-by: Christian Brauner <brauner@kernel.org>
On Thu, 21 Nov 2024, Geert Uytterhoeven wrote: > Linux has supported m68k since last century. Yeah I fondly remember the 80s where 68K systems were always out of reach for me to have. The dream system that I never could get my hands on. The creme de la creme du jour. I just had to be content with the 6800 and 6502 processors. Then IBM started the sick road down the 8088, 8086 that led from crap to more crap. Sigh. > Any new such assumptions are fixed quickly (at least in the kernel). > If you need a specific alignment, make sure to use __aligned and/or > appropriate padding in structures. > And yes, the compiler knows, and provides __alignof__. > > > How do you deal with torn reads/writes in such a scenario? Is this UP > > only? > > Linux does not support (rate) SMP m68k machines. Ah. Ok that explains it. Do we really need to maintain support for a platform that has been obsolete for decade and does not even support SMP?
On Thu, Nov 21, 2024 at 09:23:28AM -0800, Christoph Lameter (Ampere) wrote: > On Thu, 21 Nov 2024, Geert Uytterhoeven wrote: > > > Linux has supported m68k since last century. > > Yeah I fondly remember the 80s where 68K systems were always out of reach > for me to have. The dream system that I never could get my hands on. The > creme de la creme du jour. I just had to be content with the 6800 and > 6502 processors. Then IBM started the sick road down the 8088, 8086 > that led from crap to more crap. Sigh. > > > Any new such assumptions are fixed quickly (at least in the kernel). > > If you need a specific alignment, make sure to use __aligned and/or > > appropriate padding in structures. > > And yes, the compiler knows, and provides __alignof__. > > > > > How do you deal with torn reads/writes in such a scenario? Is this UP > > > only? > > > > Linux does not support (rate) SMP m68k machines. > > Ah. Ok that explains it. > > Do we really need to maintain support for a platform that has been > obsolete for decade and does not even support SMP? Since this keeps coming up, I think there is a much more important question to ask: Do we really need to continue supporting nommu machines ? Is anyone but me even boot testing those ? Guenter
On 11/21/24 11:30 AM, Guenter Roeck wrote: > On Thu, Nov 21, 2024 at 09:23:28AM -0800, Christoph Lameter (Ampere) wrote: >> On Thu, 21 Nov 2024, Geert Uytterhoeven wrote: >> >>> Linux has supported m68k since last century. >> >> Yeah I fondly remember the 80s where 68K systems were always out of reach >> for me to have. The dream system that I never could get my hands on. The >> creme de la creme du jour. I just had to be content with the 6800 and >> 6502 processors. Then IBM started the sick road down the 8088, 8086 >> that led from crap to more crap. Sigh. >> >>> Any new such assumptions are fixed quickly (at least in the kernel). >>> If you need a specific alignment, make sure to use __aligned and/or >>> appropriate padding in structures. >>> And yes, the compiler knows, and provides __alignof__. >>> >>>> How do you deal with torn reads/writes in such a scenario? Is this UP >>>> only? >>> >>> Linux does not support (rate) SMP m68k machines. >> >> Ah. Ok that explains it. >> >> Do we really need to maintain support for a platform that has been >> obsolete for decade and does not even support SMP? I asked that earlier in this thread too... > Since this keeps coming up, I think there is a much more important > question to ask: > > Do we really need to continue supporting nommu machines ? Is anyone > but me even boot testing those ? Getting rid of nommu would be nice for sure in terms of maintenance, it's one of those things that pop up as a build breaking thing because nobody is using/testing them. I'm all for axing relics from the codebase. Doesn't mean they can't be maintained out-of-tree, but that is where they belong imho.
On Thu, Nov 21, 2024 at 7:30 PM Guenter Roeck <linux@roeck-us.net> wrote: > On Thu, Nov 21, 2024 at 09:23:28AM -0800, Christoph Lameter (Ampere) wrote: > > On Thu, 21 Nov 2024, Geert Uytterhoeven wrote: > > > Linux has supported m68k since last century. > > > > Yeah I fondly remember the 80s where 68K systems were always out of reach > > for me to have. The dream system that I never could get my hands on. The > > creme de la creme du jour. I just had to be content with the 6800 and > > 6502 processors. Then IBM started the sick road down the 8088, 8086 > > that led from crap to more crap. Sigh. > > > > > Any new such assumptions are fixed quickly (at least in the kernel). > > > If you need a specific alignment, make sure to use __aligned and/or > > > appropriate padding in structures. > > > And yes, the compiler knows, and provides __alignof__. > > > > > > > How do you deal with torn reads/writes in such a scenario? Is this UP > > > > only? > > > > > > Linux does not support (rate) SMP m68k machines. s/rate/rare/ > > Ah. Ok that explains it. > > > > Do we really need to maintain support for a platform that has been > > obsolete for decade and does not even support SMP? > > Since this keeps coming up, I think there is a much more important > question to ask: > > Do we really need to continue supporting nommu machines ? Is anyone > but me even boot testing those ? Not all m68k platform are nommu. Gr{oetje,eeting}s, Geert
On Thu, Nov 21, 2024 at 07:50:33PM +0100, Geert Uytterhoeven wrote: > On Thu, Nov 21, 2024 at 7:30 PM Guenter Roeck <linux@roeck-us.net> wrote: > > On Thu, Nov 21, 2024 at 09:23:28AM -0800, Christoph Lameter (Ampere) wrote: > > > On Thu, 21 Nov 2024, Geert Uytterhoeven wrote: > > > > Linux has supported m68k since last century. > > > > > > Yeah I fondly remember the 80s where 68K systems were always out of reach > > > for me to have. The dream system that I never could get my hands on. The > > > creme de la creme du jour. I just had to be content with the 6800 and > > > 6502 processors. Then IBM started the sick road down the 8088, 8086 > > > that led from crap to more crap. Sigh. > > > > > > > Any new such assumptions are fixed quickly (at least in the kernel). > > > > If you need a specific alignment, make sure to use __aligned and/or > > > > appropriate padding in structures. > > > > And yes, the compiler knows, and provides __alignof__. > > > > > > > > > How do you deal with torn reads/writes in such a scenario? Is this UP > > > > > only? > > > > > > > > Linux does not support (rate) SMP m68k machines. > > s/rate/rare/ > > > > Ah. Ok that explains it. > > > > > > Do we really need to maintain support for a platform that has been > > > obsolete for decade and does not even support SMP? > > > > Since this keeps coming up, I think there is a much more important > > question to ask: > > > > Do we really need to continue supporting nommu machines ? Is anyone > > but me even boot testing those ? > > Not all m68k platform are nommu. > Yes, I wasn't trying to point to m68k, but to nommu in general. Guenter
On Thu, Nov 21, 2024 at 11:08:54AM -0800, Guenter Roeck wrote: > On Thu, Nov 21, 2024 at 07:50:33PM +0100, Geert Uytterhoeven wrote: > > On Thu, Nov 21, 2024 at 7:30 PM Guenter Roeck <linux@roeck-us.net> wrote: > > > On Thu, Nov 21, 2024 at 09:23:28AM -0800, Christoph Lameter (Ampere) wrote: > > > > On Thu, 21 Nov 2024, Geert Uytterhoeven wrote: > > > > > Linux has supported m68k since last century. > > > > > > > > Yeah I fondly remember the 80s where 68K systems were always out of reach > > > > for me to have. The dream system that I never could get my hands on. The > > > > creme de la creme du jour. I just had to be content with the 6800 and > > > > 6502 processors. Then IBM started the sick road down the 8088, 8086 > > > > that led from crap to more crap. Sigh. > > > > > > > > > Any new such assumptions are fixed quickly (at least in the kernel). > > > > > If you need a specific alignment, make sure to use __aligned and/or > > > > > appropriate padding in structures. > > > > > And yes, the compiler knows, and provides __alignof__. > > > > > > > > > > > How do you deal with torn reads/writes in such a scenario? Is this UP > > > > > > only? > > > > > > > > > > Linux does not support (rate) SMP m68k machines. > > > > s/rate/rare/ > > > > > > Ah. Ok that explains it. > > > > > > > > Do we really need to maintain support for a platform that has been > > > > obsolete for decade and does not even support SMP? > > > > > > Since this keeps coming up, I think there is a much more important > > > question to ask: > > > > > > Do we really need to continue supporting nommu machines ? Is anyone > > > but me even boot testing those ? > > > > Not all m68k platform are nommu. > > > Yes, I wasn't trying to point to m68k, but to nommu in general. > For some more context: I think it is highly unlikely that anyone is really using a recent version of Linux on a nommu machine. Maybe that was the case 10 or 20 years ago, but nowadays there are other operating systems which are much better suited than Linux for such systems. Yet, there is a _lot_ of nommu code in the kernel. In comparison, supporting m68k (mmu based) is a no brainer, plus there are actually people like Geert actively supporting it. If we are talking about dropping m68k support, we should really talk about dropping nommu support first to get some _real_ benefit. Guenter
On Wed, 2024-11-20 at 13:46 +0100, Geert Uytterhoeven wrote:
> On m68k, where the minimum alignment of unsigned long is 2 bytes:
Well, well, well, my old friend strikes again ;-).
These will always come up until we fix the alignment issue on m68k.
Adrian
On 22/11/24 04:30, Guenter Roeck wrote: > On Thu, Nov 21, 2024 at 09:23:28AM -0800, Christoph Lameter (Ampere) wrote: >> On Thu, 21 Nov 2024, Geert Uytterhoeven wrote: >> >>> Linux has supported m68k since last century. >> >> Yeah I fondly remember the 80s where 68K systems were always out of reach >> for me to have. The dream system that I never could get my hands on. The >> creme de la creme du jour. I just had to be content with the 6800 and >> 6502 processors. Then IBM started the sick road down the 8088, 8086 >> that led from crap to more crap. Sigh. >> >>> Any new such assumptions are fixed quickly (at least in the kernel). >>> If you need a specific alignment, make sure to use __aligned and/or >>> appropriate padding in structures. >>> And yes, the compiler knows, and provides __alignof__. >>> >>>> How do you deal with torn reads/writes in such a scenario? Is this UP >>>> only? >>> >>> Linux does not support (rate) SMP m68k machines. >> >> Ah. Ok that explains it. >> >> Do we really need to maintain support for a platform that has been >> obsolete for decade and does not even support SMP? > > Since this keeps coming up, I think there is a much more important > question to ask: > > Do we really need to continue supporting nommu machines ? Is anyone > but me even boot testing those ? Yes. Across many architectures. And yes on every release, and for m68k building and testing on every rc for nommu at a minimum. I rarely hit build or testing problems on nonmmu targets. At least every kernel release I build and test armnommu (including thumb2 on cortex), m68k, RISC-V and xtensa. They are all easy, qemu targets for them all. Thats just me. So I would guess there are others building and testing too. But what has that got to do with this thread, seems somewhat tangential to the discussions here so far... Regards Greg
On Thu, 21 Nov 2024, John Paul Adrian Glaubitz wrote: > On Wed, 2024-11-20 at 13:46 +0100, Geert Uytterhoeven wrote: > > On m68k, where the minimum alignment of unsigned long is 2 bytes: > > Well, well, well, my old friend strikes again ;-). > > These will always come up until we fix the alignment issue on m68k. > Hmmm. That patch you're replying too. Does it make the kernel source code better or worse?
On Fri, Nov 22, 2024 at 3:11 AM Finn Thain <fthain@linux-m68k.org> wrote: > On Thu, 21 Nov 2024, John Paul Adrian Glaubitz wrote: > > On Wed, 2024-11-20 at 13:46 +0100, Geert Uytterhoeven wrote: > > > On m68k, where the minimum alignment of unsigned long is 2 bytes: > > > > Well, well, well, my old friend strikes again ;-). > > > > These will always come up until we fix the alignment issue on m68k. > > Hmmm. That patch you're replying too. Does it make the kernel source code > better or worse? Touché ;-) The same can be said about commit d811ac148f0afd2f ("virtchnl: fix m68k build."): if you rely on a specific alignment, make sure to use __aligned__ and/or struct padding. Gr{oetje,eeting}s, Geert -- Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that. -- Linus Torvalds
On Fri, Nov 22, 2024 at 1:23 AM Greg Ungerer <gerg@kernel.org> wrote: > On 22/11/24 04:30, Guenter Roeck wrote: > > Do we really need to continue supporting nommu machines ? Is anyone > > but me even boot testing those ? > > Yes. Across many architectures. And yes on every release, and for m68k building > and testing on every rc for nommu at a minimum. > > I rarely hit build or testing problems on nonmmu targets. At least every kernel > release I build and test armnommu (including thumb2 on cortex), m68k, RISC-V and > xtensa. They are all easy, qemu targets for them all. Thats just me. So I would > guess there are others building and testing too. FTR, I do regular boot tests on K210 (SiPEED MAiX BiT RISC-V nommu). Getting harder, as 8 MiB of RAM is not much... Gr{oetje,eeting}s, Geert
On Thu, Nov 21, 2024 at 10:30 AM Guenter Roeck <linux@roeck-us.net> wrote: > Do we really need to continue supporting nommu machines ? Is anyone > but me even boot testing those ? I do rather regular boot tests on nommu xtensa (esp32, esp32-s3).
On Thu, Nov 21, 2024 at 11:22:58AM -0800, Guenter Roeck wrote: > On Thu, Nov 21, 2024 at 11:08:54AM -0800, Guenter Roeck wrote: > > On Thu, Nov 21, 2024 at 07:50:33PM +0100, Geert Uytterhoeven wrote: > > > On Thu, Nov 21, 2024 at 7:30 PM Guenter Roeck <linux@roeck-us.net> wrote: > > > > On Thu, Nov 21, 2024 at 09:23:28AM -0800, Christoph Lameter (Ampere) wrote: > > > > > On Thu, 21 Nov 2024, Geert Uytterhoeven wrote: > > > > > > Linux has supported m68k since last century. > > > > > > > > > > Yeah I fondly remember the 80s where 68K systems were always out of reach > > > > > for me to have. The dream system that I never could get my hands on. The > > > > > creme de la creme du jour. I just had to be content with the 6800 and > > > > > 6502 processors. Then IBM started the sick road down the 8088, 8086 > > > > > that led from crap to more crap. Sigh. > > > > > > > > > > > Any new such assumptions are fixed quickly (at least in the kernel). > > > > > > If you need a specific alignment, make sure to use __aligned and/or > > > > > > appropriate padding in structures. > > > > > > And yes, the compiler knows, and provides __alignof__. > > > > > > > > > > > > > How do you deal with torn reads/writes in such a scenario? Is this UP > > > > > > > only? > > > > > > > > > > > > Linux does not support (rate) SMP m68k machines. > > > > > > s/rate/rare/ > > > > > > > > Ah. Ok that explains it. > > > > > > > > > > Do we really need to maintain support for a platform that has been > > > > > obsolete for decade and does not even support SMP? > > > > > > > > Since this keeps coming up, I think there is a much more important > > > > question to ask: > > > > > > > > Do we really need to continue supporting nommu machines ? Is anyone > > > > but me even boot testing those ? > > > > > > Not all m68k platform are nommu. > > > > > Yes, I wasn't trying to point to m68k, but to nommu in general. > > > > For some more context: I think it is highly unlikely that anyone is really > using a recent version of Linux on a nommu machine. Maybe that was the case > 10 or 20 years ago, but nowadays there are other operating systems which are > much better suited than Linux for such systems. Yet, there is a _lot_ of > nommu code in the kernel. In comparison, supporting m68k (mmu based) is a no > brainer, plus there are actually people like Geert actively supporting it. > > If we are talking about dropping m68k support, we should really talk about > dropping nommu support first to get some _real_ benefit. > > Guenter > > I couldn't agree more re: nommu, it is the real source of maintenance issues at least for us in mm, and one I've personally run into many times. An aside, but note that there is a proposal to add nommu support to UML, which would entirely prevent our ability to eliminate it [0] (though it would make testing it easier! :) [0]:https://lore.kernel.org/all/cover.1731290567.git.thehajime@gmail.com/
On Fri, Nov 22, 2024 at 09:45:52AM +0000, Lorenzo Stoakes wrote: > On Thu, Nov 21, 2024 at 11:22:58AM -0800, Guenter Roeck wrote: > > On Thu, Nov 21, 2024 at 11:08:54AM -0800, Guenter Roeck wrote: > > > On Thu, Nov 21, 2024 at 07:50:33PM +0100, Geert Uytterhoeven wrote: > > > > On Thu, Nov 21, 2024 at 7:30 PM Guenter Roeck <linux@roeck-us.net> wrote: > > > > > On Thu, Nov 21, 2024 at 09:23:28AM -0800, Christoph Lameter (Ampere) wrote: > > > > > > On Thu, 21 Nov 2024, Geert Uytterhoeven wrote: > > > > > > > Linux has supported m68k since last century. > > > > > > > > > > > > Yeah I fondly remember the 80s where 68K systems were always out of reach > > > > > > for me to have. The dream system that I never could get my hands on. The > > > > > > creme de la creme du jour. I just had to be content with the 6800 and > > > > > > 6502 processors. Then IBM started the sick road down the 8088, 8086 > > > > > > that led from crap to more crap. Sigh. > > > > > > > > > > > > > Any new such assumptions are fixed quickly (at least in the kernel). > > > > > > > If you need a specific alignment, make sure to use __aligned and/or > > > > > > > appropriate padding in structures. > > > > > > > And yes, the compiler knows, and provides __alignof__. > > > > > > > > > > > > > > > How do you deal with torn reads/writes in such a scenario? Is this UP > > > > > > > > only? > > > > > > > > > > > > > > Linux does not support (rate) SMP m68k machines. > > > > > > > > s/rate/rare/ > > > > > > > > > > Ah. Ok that explains it. > > > > > > > > > > > > Do we really need to maintain support for a platform that has been > > > > > > obsolete for decade and does not even support SMP? > > > > > > > > > > Since this keeps coming up, I think there is a much more important > > > > > question to ask: > > > > > > > > > > Do we really need to continue supporting nommu machines ? Is anyone > > > > > but me even boot testing those ? > > > > > > > > Not all m68k platform are nommu. > > > > > > > Yes, I wasn't trying to point to m68k, but to nommu in general. > > > > > > > For some more context: I think it is highly unlikely that anyone is really > > using a recent version of Linux on a nommu machine. Maybe that was the case > > 10 or 20 years ago, but nowadays there are other operating systems which are > > much better suited than Linux for such systems. Yet, there is a _lot_ of > > nommu code in the kernel. In comparison, supporting m68k (mmu based) is a no > > brainer, plus there are actually people like Geert actively supporting it. > > > > If we are talking about dropping m68k support, we should really talk about > > dropping nommu support first to get some _real_ benefit. > > > > Guenter > > > > > > I couldn't agree more re: nommu, it is the real source of maintenance > issues at least for us in mm, and one I've personally run into many times. > > An aside, but note that there is a proposal to add nommu support to UML, > which would entirely prevent our ability to eliminate it [0] (though it > would make testing it easier! :) > > [0]:https://lore.kernel.org/all/cover.1731290567.git.thehajime@gmail.com/ To update, some interesting discussion in this thread suggests that indeed, there is an ongoing need for nommu regardless [1]. In which case this nommu uml series is rather helpful for testing :) [1]:https://lore.kernel.org/linux-mm/09060fcf-47e4-424f-9ab7-ee2f7919dbf5@lucifer.local/T/#m0cb0ace28f3905182369790ddc1b494d408587b9
On Fri, 22 Nov 2024, Lorenzo Stoakes wrote: > An aside, but note that there is a proposal to add nommu support to UML, > which would entirely prevent our ability to eliminate it [0] (though it > would make testing it easier! :) Ok back to the alignment check. The patch is ok and an improvement since it properly checks for the alignment of the scalar and does not assume that a scalar is naturally aligned. That may not be necessary but it is formally more correct. So Acked-by: Christoph Lameter <cl@linux.com>
diff --git a/mm/slab_common.c b/mm/slab_common.c index 893d320599151845..f2f201d865c108bd 100644 --- a/mm/slab_common.c +++ b/mm/slab_common.c @@ -230,7 +230,7 @@ static struct kmem_cache *create_cache(const char *name, if (args->use_freeptr_offset && (args->freeptr_offset >= object_size || !(flags & SLAB_TYPESAFE_BY_RCU) || - !IS_ALIGNED(args->freeptr_offset, sizeof(freeptr_t)))) + !IS_ALIGNED(args->freeptr_offset, __alignof(freeptr_t)))) goto out; err = -ENOMEM;
On m68k, where the minimum alignment of unsigned long is 2 bytes: Kernel panic - not syncing: __kmem_cache_create_args: Failed to create slab 'io_kiocb'. Error -22 CPU: 0 UID: 0 PID: 1 Comm: swapper Not tainted 6.12.0-atari-03776-g7eaa1f99261a #1783 Stack from 0102fe5c: 0102fe5c 00514a2b 00514a2b ffffff00 00000001 0051f5ed 00425e78 00514a2b 0041eb74 ffffffea 00000310 0051f5ed ffffffea ffffffea 00601f60 00000044 0102ff20 000e7a68 0051ab8e 004383b8 0051f5ed ffffffea 000000b8 00000007 01020c00 00000000 000e77f0 0041e5f0 005f67c0 0051f5ed 000000b6 0102fef4 00000310 0102fef4 00000000 00000016 005f676c 0060a34c 00000010 00000004 00000038 0000009a 01000000 000000b8 005f668e 0102e000 00001372 0102ff88 Call Trace: [<00425e78>] dump_stack+0xc/0x10 [<0041eb74>] panic+0xd8/0x26c [<000e7a68>] __kmem_cache_create_args+0x278/0x2e8 [<000e77f0>] __kmem_cache_create_args+0x0/0x2e8 [<0041e5f0>] memset+0x0/0x8c [<005f67c0>] io_uring_init+0x54/0xd2 The minimal alignment of an integral type may differ from its size, hence is not safe to assume that an arbitrary freeptr_t (which is basically an unsigned long) is always aligned to 4 or 8 bytes. As nothing seems to require the additional alignment, it is safe to fix this by relaxing the check to the actual minimum alignment of freeptr_t. Fixes: aaa736b186239b7d ("io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache") Fixes: d345bd2e9834e2da ("mm: add kmem_cache_create_rcu()") Reported-by: Guenter Roeck <linux@roeck-us.net> Closes: https://lore.kernel.org/37c588d4-2c32-4aad-a19e-642961f200d7@roeck-us.net Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org> --- mm/slab_common.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)