Re: [PATCH 03/14] io_uring: specify freeptr usage for SLAB_TYPESAFE_BY_RCU io_kiocb cache

[Vlastimil Babka <vbabka@suse.cz>]

On 11/20/24 09:19, Geert Uytterhoeven wrote:
> Hi Jens,
> 
> CC Christian (who added the check)
> CC Vlastimil (who suggested the check)
> 
> On Tue, Nov 19, 2024 at 11:30 PM Jens Axboe <axboe@kernel.dk> wrote:
>> On 11/19/24 2:46 PM, Guenter Roeck wrote:
>> > On 11/19/24 11:49, Jens Axboe wrote:
>> >> On 11/19/24 12:44 PM, Jens Axboe wrote:
>> >>>> On Tue, Nov 19, 2024 at 8:30?PM Jens Axboe <axboe@kernel.dk> wrote:
>> >>>>> On 11/19/24 12:25 PM, Geert Uytterhoeven wrote:
>> >>>>>> On Tue, Nov 19, 2024 at 8:10?PM Jens Axboe <axboe@kernel.dk> wrote:
>> >>>>>>> On 11/19/24 12:02 PM, Geert Uytterhoeven wrote:
>> >>>>>>>> On Tue, Nov 19, 2024 at 8:00?PM Jens Axboe <axboe@kernel.dk> wrote:
>> >>>>>>>>> On 11/19/24 10:49 AM, Geert Uytterhoeven wrote:
>> >>>>>>>>>> On Tue, Nov 19, 2024 at 5:21?PM Guenter Roeck <linux@roeck-us.net> wrote:
>> >>>>>>>>>>> On 11/19/24 08:02, Jens Axboe wrote:
>> >>>>>>>>>>>> On 11/19/24 8:36 AM, Guenter Roeck wrote:
>> >>>>>>>>>>>>> On Tue, Oct 29, 2024 at 09:16:32AM -0600, Jens Axboe wrote:
>> >>>>>>>>>>>>>> Doesn't matter right now as there's still some bytes left for it, but
>> >>>>>>>>>>>>>> let's prepare for the io_kiocb potentially growing and add a specific
>> >>>>>>>>>>>>>> freeptr offset for it.
>> >>>>>>>>>>>>>>
>> >>>>>>>>>>>>>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>> >>>>>>>>>>>>>
>> >>>>>>>>>>>>> This patch triggers:
>> >>>>>>>>>>>>>
>> >>>>>>>>>>>>> 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-mac-00971-g158f238aa69d #1
>> >>>>>>>>>>>>> Stack from 00c63e5c:
>> >>>>>>>>>>>>>           00c63e5c 00612c1c 00612c1c 00000300 00000001 005f3ce6 004b9044 00612c1c
>> >>>>>>>>>>>>>           004ae21e 00000310 000000b6 005f3ce6 005f3ce6 ffffffea ffffffea 00797244
>> >>>>>>>>>>>>>           00c63f20 000c6974 005ee588 004c9051 005f3ce6 ffffffea 000000a5 00c614a0
>> >>>>>>>>>>>>>           004a72c2 0002cb62 000c675e 004adb58 0076f28a 005f3ce6 000000b6 00c63ef4
>> >>>>>>>>>>>>>           00000310 00c63ef4 00000000 00000016 0076f23e 00c63f4c 00000010 00000004
>> >>>>>>>>>>>>>           00000038 0000009a 01000000 00000000 00000000 00000000 000020e0 0076f23e
>> >>>>>>>>>>>>> Call Trace: [<004b9044>] dump_stack+0xc/0x10
>> >>>>>>>>>>>>>    [<004ae21e>] panic+0xc4/0x252
>> >>>>>>>>>>>>>    [<000c6974>] __kmem_cache_create_args+0x216/0x26c
>> >>>>>>>>>>>>>    [<004a72c2>] strcpy+0x0/0x1c
>> >>>>>>>>>>>>>    [<0002cb62>] parse_args+0x0/0x1f2
>> >>>>>>>>>>>>>    [<000c675e>] __kmem_cache_create_args+0x0/0x26c
>> >>>>>>>>>>>>>    [<004adb58>] memset+0x0/0x8c
>> >>>>>>>>>>>>>    [<0076f28a>] io_uring_init+0x4c/0xca
>> >>>>>>>>>>>>>    [<0076f23e>] io_uring_init+0x0/0xca
>> >>>>>>>>>>>>>    [<000020e0>] do_one_initcall+0x32/0x192
>> >>>>>>>>>>>>>    [<0076f23e>] io_uring_init+0x0/0xca
>> >>>>>>>>>>>>>    [<0000211c>] do_one_initcall+0x6e/0x192
>> >>>>>>>>>>>>>    [<004a72c2>] strcpy+0x0/0x1c
>> >>>>>>>>>>>>>    [<0002cb62>] parse_args+0x0/0x1f2
>> >>>>>>>>>>>>>    [<000020ae>] do_one_initcall+0x0/0x192
>> >>>>>>>>>>>>>    [<0075c4e2>] kernel_init_freeable+0x1a0/0x1a4
>> >>>>>>>>>>>>>    [<0076f23e>] io_uring_init+0x0/0xca
>> >>>>>>>>>>>>>    [<004b911a>] kernel_init+0x0/0xec
>> >>>>>>>>>>>>>    [<004b912e>] kernel_init+0x14/0xec
>> >>>>>>>>>>>>>    [<004b911a>] kernel_init+0x0/0xec
>> >>>>>>>>>>>>>    [<0000252c>] ret_from_kernel_thread+0xc/0x14
>> >>>>>>>>>>>>>
>> >>>>>>>>>>>>> when trying to boot the m68k:q800 machine in qemu.
>> >>>>>>>>>>>>>
>> >>>>>>>>>>>>> An added debug message in create_cache() shows the reason:
>> >>>>>>>>>>>>>
>> >>>>>>>>>>>>> #### freeptr_offset=154 object_size=182 flags=0x310 aligned=0 sizeof(freeptr_t)=4
>> >>>>>>>>>>>>>
>> >>>>>>>>>>>>> freeptr_offset would need to be 4-byte aligned but that is not the
>> >>>>>>>>>>>>> case on m68k.
>> >>>>>>>>>>>>
>> >>>>>>>>>>>> Why is ->work 2-byte aligned to begin with on m68k?!
>> >>>>>>>>>>>
>> >>>>>>>>>>> My understanding is that m68k does not align pointers.
>> >>>>>>>>>>
>> >>>>>>>>>> The minimum alignment for multi-byte integral values on m68k is
>> >>>>>>>>>> 2 bytes.
>> >>>>>>>>>>
>> >>>>>>>>>> See also the comment at
>> >>>>>>>>>> https://elixir.bootlin.com/linux/v6.12/source/include/linux/maple_tree.h#L46
>> >>>>>>>>>
>> >>>>>>>>> Maybe it's time we put m68k to bed? :-)
>> >>>>>>>>>
>> >>>>>>>>> We can add a forced alignment ->work to be 4 bytes, won't change
>> >>>>>>>>> anything on anything remotely current. But does feel pretty hacky to
>> >>>>>>>>> need to align based on some ancient thing.
>> >>>>>>>>
>> >>>>>>>> Why does freeptr_offset need to be 4-byte aligned?
>> >>>>>>>
>> >>>>>>> Didn't check, but it's slab/slub complaining using a 2-byte aligned
>> >>>>>>> address for the free pointer offset. It's explicitly checking:
>> >>>>>>>
>> >>>>>>>          /* If a custom freelist pointer is requested make sure it's sane. */
>> >>>>>>>          err = -EINVAL;
>> >>>>>>>          if (args->use_freeptr_offset &&
>> >>>>>>>              (args->freeptr_offset >= object_size ||
>> >>>>>>>               !(flags & SLAB_TYPESAFE_BY_RCU) ||
>> >>>>>>>               !IS_ALIGNED(args->freeptr_offset, sizeof(freeptr_t))))
>                                                           ^^^^^^
> 
>> >>>>>>>                  goto out;
>> >>>>>>
>> >>>>>> It is not guaranteed that alignof(freeptr_t) >= sizeof(freeptr_t)
>> >>>>>> (free_ptr is sort of a long). If freeptr_offset must be a multiple of
>> >>>>>> 4 or 8 bytes,
>> >>>>>> the code that assigns it must make sure that is true.
>> >>>>>
>> >>>>> Right, this is what the email is about...
>> >>>>>
>> >>>>>> I guess this is the code in fs/file_table.c:
>> >>>>>>
>> >>>>>>      .freeptr_offset = offsetof(struct file, f_freeptr),
>> >>>>>>
>> >>>>>> which references:
>> >>>>>>
>> >>>>>>      include/linux/fs.h:           freeptr_t               f_freeptr;
>> >>>>>>
>> >>>>>> I guess the simplest solution is to add an __aligned(sizeof(freeptr_t))
>> >>>>>> (or __aligned(sizeof(long)) to the definition of freeptr_t:
>> >>>>>>
>> >>>>>>      include/linux/slab.h:typedef struct { unsigned long v; } freeptr_t;
>> >>>>>
>> >>>>> It's not, it's struct io_kiocb->work, as per the stack trace in this
>> >>>>> email.
>> >>>>
>> >>>> Sorry, I was falling out of thin air into this thread...
>> >>>>
>> >>>> linux-next/master:io_uring/io_uring.c:          .freeptr_offset =
>> >>>> offsetof(struct io_kiocb, work),
>> >>>> linux-next/master:io_uring/io_uring.c:          .use_freeptr_offset = true,
>> >>>>
>> >>>> Apparently io_kiocb.work is of type struct io_wq_work, not freeptr_t?
>> >>>> Isn't that a bit error-prone, as the slab core code expects a freeptr_t?
>> >>>
>> >>> It just needs the space, should not matter otherwise. But may as well
>> >>> just add the union and align the freeptr so it stop complaining on m68k.
>> >>
>> >> Ala the below, perhaps alignment takes care of itself then?
>> >
>> > No, that doesn't work (I tried), at least not on its own, because the pointer
>> > is still unaligned on m68k.
>>
>> Yeah we'll likely need to force it. The below should work, I pressume?
>> Feels pretty odd to have to align it to the size of it, when that should
>> naturally occur... Crusty legacy archs.
>>
>> diff --git a/include/linux/io_uring_types.h b/include/linux/io_uring_types.h
>> index 593c10a02144..8ed9c6923668 100644
>> --- a/include/linux/io_uring_types.h
>> +++ b/include/linux/io_uring_types.h
>> @@ -674,7 +674,11 @@ struct io_kiocb {
>>         struct io_kiocb                 *link;
>>         /* custom credentials, valid IFF REQ_F_CREDS is set */
>>         const struct cred               *creds;
>> -       struct io_wq_work               work;
>> +
>> +       union {
>> +               struct io_wq_work       work;
>> +               freeptr_t               freeptr __aligned(sizeof(freeptr_t));
> 
> I'd rather add the __aligned() to the definition of freeptr_t, so it
> applies to all (future) users.
> 
> But my main question stays: why is the slab code checking
> IS_ALIGNED(args->freeptr_offset, sizeof(freeptr_t)?

I believe it's to match how SLUB normally calculates the offset if no
explicit one is given, in calculate_sizes():

s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));

Yes there's a sizeof(void *) because freepointer used to be just that and we
forgot to update this place when freepointer_t was introduced (by Jann in
44f6a42d49350) for handling CONFIG_SLAB_FREELIST_HARDENED. In
get_freepointer() you can see how there's a cast to a pointer eventually.

Does m68k have different alignment for pointer and unsigned long or both are
2 bytes? Or any other arch, i.e. should get_freepointer be a union with
unsigned long and void * instead? (or it doesn't matter?)

> Perhaps that was just intended to be __alignof__ instead of sizeof()?

Would it do the right thing everywhere, given the explanation above?

Thanks,
Vlastimil

>> +       };
>>
>>         struct {
>>                 u64                     extra1;
>> diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
>> index 73af59863300..86ac7df2a601 100644
>> --- a/io_uring/io_uring.c
>> +++ b/io_uring/io_uring.c
>> @@ -3812,7 +3812,7 @@ static int __init io_uring_init(void)
>>         struct kmem_cache_args kmem_args = {
>>                 .useroffset = offsetof(struct io_kiocb, cmd.data),
>>                 .usersize = sizeof_field(struct io_kiocb, cmd.data),
>> -               .freeptr_offset = offsetof(struct io_kiocb, work),
>> +               .freeptr_offset = offsetof(struct io_kiocb, freeptr),
>>                 .use_freeptr_offset = true,
>>         };
> 
> Gr{oetje,eeting}s,
> 
>                         Geert
>