Re: [PATCH] x86_64: uninline TASK_SIZE

[Ingo Molnar <mingo@kernel.org>]

* Andy Lutomirski <luto@kernel.org> wrote:

> > Saving 2KB on a defconfig is quite a lot.
> 
> Saving 2kB of text by adding 8 bytes to thread_info seems rather
> dubious to me.  You only need 256 tasks before you lose.  My
> not-particularly-loaded laptop has 865 tasks right now.

I was suggesting current->task_size or thread_info->task_size as a way to 
100% avoid the function call overhead. Worth a tiny amount of RAM - even 
with 1 million tasks it's only 4MB of RAM. ;-)

Some TASK_SIZE users are prominent syscalls: mmap(), 

> As a general principle, the mere existence of TIF_ADDR32 is a bug. The 
> value of that flag is *wrong* under the 32-bit variant of CRIU. How 
> about instead making some more progress toward getting rid of dubious 
> TASK_SIZE users?  I'm working on a little series to get rid of most of 
> them.  Meanwhile: it sure looks like a large fraction of the users are 
> confused as to whether TASK_SIZE is the highest user address or the 
> lowest non-user address.

I really like that, replacing TASK_SIZE with *nothing* would be even 
faster.

In fact instead of just reducing its usage I'd suggest removing TASK_SIZE 
altogether and renaming TASK_SIZE_MAX back to TASK_SIZE, or something 
like that - the confusion from the deceptively macro-ish naming of 
TASK_SIZE is real.

The original commit of making TASK_SIZE dynamic on the task's compat flag 
was done in:

  84929801e14d: [PATCH] x86_64: TASK_SIZE fixes for compatibility mode processes

Here's the justification given:

    Appended patch will setup compatibility mode TASK_SIZE properly.  This will
    fix atleast three known bugs that can be encountered while running
    compatibility mode apps.
    
    a) A malicious 32bit app can have an elf section at 0xffffe000.  During
       exec of this app, we will have a memory leak as insert_vm_struct() is
       not checking for return value in syscall32_setup_pages() and thus not
       freeing the vma allocated for the vsyscall page.  And instead of exec
       failing (as it has addresses > TASK_SIZE), we were allowing it to
       succeed previously.
    
    b) With a 32bit app, hugetlb_get_unmapped_area/arch_get_unmapped_area
       may return addresses beyond 32bits, ultimately causing corruption
       because of wrap-around and resulting in SEGFAULT, instead of returning
       ENOMEM.
    
    c) 32bit app doing this below mmap will now fail.
    
      mmap((void *)(0xFFFFE000UL), 0x10000UL, PROT_READ|PROT_WRITE,
            MAP_FIXED|MAP_PRIVATE|MAP_ANON, 0, 0);

I believe a) is addressed by getting rid of the vsyscall page - but it 
might also not be a current problem because the vsyscall page has its own 
gate-vma now.

b) shouldn't be an issue if the mmap allocator correctly treats the 
compat bit - this doesn't require generic TASK_SIZE variations though, as 
the mmap allocator is already specific to arch/x86/.

c) is a variant of a) I believe, which should be fixed by now.

I just looked through some of the current TASK_SIZE users, and *ALL* of 
them seem dubious to me, with the exception of the mmap allocators. In 
fact some of them seem to be active bugs:

kernel/:

 - PR_SET_MM_MAP, PR_SET_MM_MAP_SIZE, prctl_set_mm() et al. Ugh, what a 
   nasty prctl()! But besides that, the TASK_SIZE restriction to the ABI 
   looks questionable: if we offer this CRIU functionality then why 
   should it be impossible for a 32-bit CRIU task to switch to 64-bit?

 - kernel/events/core.c: TASK_SIZE_MAX should be a fine filter here, in 
   fact it's probably *wrong* to restrict the profiling data here just 
   because the task happens to be in 32-bit compat mode.

 - kernel/rseq.c: is this TASK_SIZE restriction even required, wouldn't 
   TASK_SIZE_MAX be sufficient?

mm/:

 - GUP's get_get_area() et al looks really weird - why do we special-case 
   vsyscalls:

      - Can we get rid of the vsyscall page in modern kernels?

      - I don't think anyone runs those ancient glibc versions with a 
        fresh kernel anymore - can we start generating a WARN()ing 
        perhaps to see whether there's any complaints?

      - Or at least pretend it doesn't exist in terms of a GUP target page?

 - mm/kasan/generic_report.c:get_wild_bug_type() - this can use 
   TASK_SIZE_MAX just fine IMHO.

 - mm/mempolicy.c:mpol_shared_policy_init() - unsure, but I suspect we 
   can just create the pseudo-vma with a TASK_SIZE_MAX vm_end just fine.

 - mm/mlock.c:mlockall() - I believe it could be considered an outright 
   *bug* if there any pages outside the 32-bit area and don't get mlocked 
   by mlockall, just because this is a compat task. Especially with the 
   CRIU prctl() having 64-bit vmas outside the 32-bit mappings is a real 
   possibility, right? I.e. TASK_SIZE_MAX would be the right solution 
   here.

To turn the argument around: beyond the memory allocators, which includes 
the mmap and huge-mmap variants plus the SysV shmem allocator, can we 
list all the places that absolutely *rely* on TASK_SIZE being TIF_ADDR32 
restricted on compat tasks? I couldn't find any.

So I concur 100% that most TASK_SIZE uses are questionable. In fact think 
84929801e14d was a mistake, and we should effectively revert it 
carefully, by:

 - First by moving almost all TASK_SIZE users over to TASK_SIZE_MAX, 
   analyzing and justifying the impact case by case.

 - Then making the mmap allocators compat compatible (ha) without relying 
   on TASK_SIZE.

 - Renaming TASK_SIZE back to TASK_SIZE_MAX and getting rid of the 
   TASK_SIZE and TASK_SIZE_MAX differentiation.

Or am I missing some complication?

Thanks,

	Ingo