Re: [PATCH V3] riscv: asid: Fixup stale TLB entry cause application crash

[Gary Guo]

On Sat, 25 Feb 2023 23:24:40 -0500
Guo Ren <guoren@xxxxxxxxxx> wrote:

> On Sat, Feb 25, 2023 at 2:29 PM Sergey Matyukevich <geomatsi@xxxxxxxxx> wrote:
> >
> > On Fri, Feb 24, 2023 at 01:57:55AM +0800, Zong Li wrote:  
> > > Lad, Prabhakar <prabhakar.csengg@xxxxxxxxx> 於 2022年12月23日 週五 下午8:54寫道：  
> > > >
> > > > Hi Guo,
> > > >
> > > > Thank you for the patch.
> > > >
> > > > On Fri, Nov 11, 2022 at 8:00 AM <guoren@xxxxxxxxxx> wrote:  
> > > > >
> > > > > From: Guo Ren <guoren@xxxxxxxxxxxxxxxxx>
> > > > >
> > > > > After use_asid_allocator is enabled, the userspace application will
> > > > > crash by stale TLB entries. Because only using cpumask_clear_cpu without
> > > > > local_flush_tlb_all couldn't guarantee CPU's TLB entries were fresh.
> > > > > Then set_mm_asid would cause the user space application to get a stale
> > > > > value by stale TLB entry, but set_mm_noasid is okay.
> > > > >
> > > > > Here is the symptom of the bug:
> > > > > unhandled signal 11 code 0x1 (coredump)
> > > > >    0x0000003fd6d22524 <+4>:     auipc   s0,0x70
> > > > >    0x0000003fd6d22528 <+8>:     ld      s0,-148(s0) # 0x3fd6d92490  
> > > > > => 0x0000003fd6d2252c <+12>:    ld      a5,0(s0)  
> > > > > (gdb) i r s0
> > > > > s0          0x8082ed1cc3198b21       0x8082ed1cc3198b21
> > > > > (gdb) x /2x 0x3fd6d92490
> > > > > 0x3fd6d92490:   0xd80ac8a8      0x0000003f
> > > > > The core dump file shows that register s0 is wrong, but the value in
> > > > > memory is correct. Because 'ld s0, -148(s0)' used a stale mapping entry
> > > > > in TLB and got a wrong result from an incorrect physical address.
> > > > >
> > > > > When the task ran on CPU0, which loaded/speculative-loaded the value of
> > > > > address(0x3fd6d92490), then the first version of the mapping entry was
> > > > > PTWed into CPU0's TLB.
> > > > > When the task switched from CPU0 to CPU1 (No local_tlb_flush_all here by
> > > > > asid), it happened to write a value on the address (0x3fd6d92490). It
> > > > > caused do_page_fault -> wp_page_copy -> ptep_clear_flush ->
> > > > > ptep_get_and_clear & flush_tlb_page.
> > > > > The flush_tlb_page used mm_cpumask(mm) to determine which CPUs need TLB
> > > > > flush, but CPU0 had cleared the CPU0's mm_cpumask in the previous
> > > > > switch_mm. So we only flushed the CPU1 TLB and set the second version
> > > > > mapping of the PTE. When the task switched from CPU1 to CPU0 again, CPU0
> > > > > still used a stale TLB mapping entry which contained a wrong target
> > > > > physical address. It raised a bug when the task happened to read that
> > > > > value.
> > > > >
> > > > >    CPU0                               CPU1
> > > > >    - switch 'task' in
> > > > >    - read addr (Fill stale mapping
> > > > >      entry into TLB)
> > > > >    - switch 'task' out (no tlb_flush)
> > > > >                                       - switch 'task' in (no tlb_flush)
> > > > >                                       - write addr cause pagefault
> > > > >                                         do_page_fault() (change to
> > > > >                                         new addr mapping)
> > > > >                                           wp_page_copy()
> > > > >                                             ptep_clear_flush()
> > > > >                                               ptep_get_and_clear()
> > > > >                                               & flush_tlb_page()
> > > > >                                         write new value into addr
> > > > >                                       - switch 'task' out (no tlb_flush)
> > > > >    - switch 'task' in (no tlb_flush)
> > > > >    - read addr again (Use stale
> > > > >      mapping entry in TLB)
> > > > >      get wrong value from old phyical
> > > > >      addr, BUG!
> > > > >
> > > > > The solution is to keep all CPUs' footmarks of cpumask(mm) in switch_mm,
> > > > > which could guarantee to invalidate all stale TLB entries during TLB
> > > > > flush.
> > > > >
> > > > > Fixes: 65d4b9c53017 ("RISC-V: Implement ASID allocator")
> > > > > Signed-off-by: Guo Ren <guoren@xxxxxxxxxxxxxxxxx>
> > > > > Signed-off-by: Guo Ren <guoren@xxxxxxxxxx>
> > > > > Cc: Anup Patel <apatel@xxxxxxxxxxxxxxxx>
> > > > > Cc: Palmer Dabbelt <palmer@xxxxxxxxxxxx>
> > > > > ---
> > > > > Changes in v3:
> > > > >  - Move set/clear cpumask(mm) into set_mm (Make code more pretty
> > > > >    with Andrew's advice)
> > > > >  - Optimize comment description
> > > > >
> > > > > Changes in v2:
> > > > >  - Fixup nommu compile problem (Thx Conor, Also Reported-by: kernel
> > > > >    test robot <lkp@xxxxxxxxx>)
> > > > >  - Keep cpumask_clear_cpu for noasid
> > > > > ---
> > > > >  arch/riscv/mm/context.c | 30 ++++++++++++++++++++----------
> > > > >  1 file changed, 20 insertions(+), 10 deletions(-)
> > > > >  
> > > > As reported on the patch [0] I was seeing consistent failures on the
> > > > RZ/Five SoC while running bonnie++ utility. After applying this patch
> > > > on top of Palmer's for-next branch (eb67d239f3aa) I am no longer
> > > > seeing this issue.
> > > >
> > > > Tested-by: Lad Prabhakar <prabhakar.mahadev-lad.rj@xxxxxxxxxxxxxx>
> > > >
> > > > [0] https://patchwork.kernel.org/project/linux-riscv/patch/20220829205219.283543-1-geomatsi@xxxxxxxxx/
> > > >  
> > >
> > > Hi all,
> > > I got the same situation (i.e. unhandle signal 11) on our internal
> > > multi-core system, I tried the patch[0] & [1], but it still doesn't
> > > work, I guess there are still some potential problems. After applying
> > > this patch, the situation disappeared, I took some time to look at
> > > other arches' implementations, such as arc, they don't clear the
> > > mm_cpumask due to the similar issue. I can't say which approach might
> > > be better, but I'd like to point out that this patch works to me.
> > > Thanks.
> > >
> > > Tested-by: Zong Li <zong.li@xxxxxxxxxx>
> > >
> > > [0] https://lore.kernel.org/linux-riscv/20220829205219.283543-1-geomatsi@xxxxxxxxx/
> > > [1] https://lore.kernel.org/linux-riscv/20230129211818.686557-1-geomatsi@xxxxxxxxx/  
> >
> > Thanks for the report! By the way, could you please share some
> > information about the reproducing workload ?
> >
> > Initial idea was to reduce the number of TLB flushes by deferring (and
> > possibly avoiding) some of them. But we have already bug reports from
> > two different vendors, so apparently something is overlooked here.
> > Lets switch to 'aggrregating' mm_cpumask approach suggested by Guo Ren.
> >
> > @Guo Ren, do you mind if I re-send your v3 patch together with the
> > remaining reverts of my changes ?  
> Okay, thx for taking care. Let's make it work around first and then improve it.
> 
> Actually, the current riscv asid is from arm64 with hardware broadcast
> requirements. Maybe we need to consider x86 per-cpu asid pool way.

It should be noted that the spec expects supervisor software to
use a consistent meaning of non-zero ASIDs across different harts.

Also, a vendor could implement efficient hardware broadcasting ASID
invalidation with a custom instruction and expose it via SBI.

Best,
Gary