tcg: Always pass the full write size to notdirty_write()
cover tcg: Always pass the full write size to notdirty_write()
Hi,
this is the fix for an issue I complained about a few days back on the IRC. Unfortunately my reproducer [1] does not work anymore, so I'm sending this separately and as an RFC.
The user-visible effect was that:
- If a TB writes to itself;
- The address of the write is before the TB start;
- The newly written instruction causes an exception;
- Then single-stepping the instruction that does a write gets you to an interrupt handler and not to the newly written "bad" instruction.
You can still see that something is wrong with the softmmu reproducer in the debugger though. In helper_vstl(), probe_write_access() does not lead to cpu_loop_exit_noexc() - which I think it should, but cpu_stq_be_data_ra() still does, masking the issue.
Best regards,
Ilya
[1] https://lists.gnu.org/archive/html/qemu-devel/2023-08/msg01081.html
Ilya Leoshkevich (1):
tcg: Always pass the full write size to notdirty_write()
accel/tcg/cputlb.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
Signature · 1 line
2.41.0
RFC PATCH 1/1 tcg: Always pass the full write size to notdirty_write()
One of notdirty_write()'s responsibilities is detecting self-modifying code. Some functions pass the full size of a write to it, some pass 1. When a write to a code section begins before a TB start, but then overlaps the TB, the paths that pass 1 don't flush a TB and don't return to the translator loop.
This may be masked, one example being HELPER(vstl). There, probe_write_access() ultimately calls notdirty_write() with a size of 1 and misses self-modifying code. However, cpu_stq_be_data_ra() ultimately calls mmu_watch_or_dirty(), which in turn calls notdirty_write() with the full size.
It's still worth improving this, because there may still be user-visible adverse effects in other helpers.
--- accel/tcg/cputlb.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
accel/tcg/cputlb.c +4 -4
@@ -1582,7 +1582,7 @@ int probe_access_full(CPUArchState *env, vaddr addr, int size,/* Handle clean RAM pages. */ if (unlikely (flags &TLB_NOTDIRTY )) { -notdirty_write (env_cpu (env), addr,1 , *pfull, retaddr); +notdirty_write (env_cpu (env), addr, size, *pfull, retaddr); flags &= ~TLB_NOTDIRTY ; }
@@ -1605,7 +1605,7 @@ int probe_access_full_mmu(CPUArchState *env, vaddr addr, int size,/* Handle clean RAM pages. */ if (unlikely (flags &TLB_NOTDIRTY )) { -notdirty_write (env_cpu (env), addr,1 , *pfull,0 ); +notdirty_write (env_cpu (env), addr, size, *pfull,0 ); flags &= ~TLB_NOTDIRTY ; }
@@ -1626,7 +1626,7 @@ int probe_access_flags(CPUArchState *env, vaddr addr, int size,/* Handle clean RAM pages. */ if (unlikely (flags &TLB_NOTDIRTY )) { -notdirty_write (env_cpu (env), addr,1 , full, retaddr); +notdirty_write (env_cpu (env), addr, size, full, retaddr); flags &= ~TLB_NOTDIRTY ; }
@@ -1661,7 +1661,7 @@ void *probe_access(CPUArchState *env, vaddr addr, int size,/* Handle clean RAM pages. */ if (flags &TLB_NOTDIRTY ) { -notdirty_write (env_cpu (env), addr,1 , full, retaddr); +notdirty_write (env_cpu (env), addr, size, full, retaddr); } }
Signature · 1 line
2.41.0
Quoted 12 lines · Ilya Leoshkevich
One of notdirty_write()'s responsibilities is detecting self-modifying code. Some functions pass the full size of a write to it, some pass 1. When a write to a code section begins before a TB start, but then overlaps the TB, the paths that pass 1 don't flush a TB and don't return to the translator loop.
This may be masked, one example being HELPER(vstl). There,
probe_write_access()ultimately callsnotdirty_write()with a size of 1 and misses self-modifying code. However,cpu_stq_be_data_ra()ultimately callsmmu_watch_or_dirty(), which in turn callsnotdirty_write()with the full size.
Quoted 4 lines
It's still worth improving this, because there may still be user-visible adverse effects in other helpers.
Signed-off-by Ilya Leoshkevich <[email protected]>
IIRC there are some uses of probe_access_* that set size == 0. Should we adjust addr+size to cover the whole page for that case? That seems to be the intent, anyway.
r~
Quoted 44 lines
--- accel/tcg/cputlb.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)accel/tcg/cputlb.c +4
-4@@ -1582,7 +1582,7 @@ int probe_access_full(CPUArchState *env, vaddr addr, int size,/* Handle clean RAM pages. */ if (unlikely (flags &TLB_NOTDIRTY )) { -notdirty_write (env_cpu (env), addr,1 , *pfull, retaddr); +notdirty_write (env_cpu (env), addr, size, *pfull, retaddr); flags &= ~TLB_NOTDIRTY ; }@@ -1605,7 +1605,7 @@ int probe_access_full_mmu(CPUArchState *env, vaddr addr, int size,/* Handle clean RAM pages. */ if (unlikely (flags &TLB_NOTDIRTY )) { -notdirty_write (env_cpu (env), addr,1 , *pfull,0 ); +notdirty_write (env_cpu (env), addr, size, *pfull,0 ); flags &= ~TLB_NOTDIRTY ; }@@ -1626,7 +1626,7 @@ int probe_access_flags(CPUArchState *env, vaddr addr, int size,/* Handle clean RAM pages. */ if (unlikely (flags &TLB_NOTDIRTY )) { -notdirty_write (env_cpu (env), addr,1 , full, retaddr); +notdirty_write (env_cpu (env), addr, size, full, retaddr); flags &= ~TLB_NOTDIRTY ; }@@ -1661,7 +1661,7 @@ void *probe_access(CPUArchState *env, vaddr addr, int size,/* Handle clean RAM pages. */ if (flags &TLB_NOTDIRTY ) { -notdirty_write (env_cpu (env), addr,1 , full, retaddr); +notdirty_write (env_cpu (env), addr, size, full, retaddr); } }
Quoted 16 lines · Richard Henderson
Quoted 15 lines · Ilya Leoshkevich
One of notdirty_write()'s responsibilities is detecting self- modifying
code. Some functions pass the full size of a write to it, some pass 1.
When a write to a code section begins before a TB start, but then overlaps the TB, the paths that pass 1 don't flush a TB and don't return to the translator loop.This may be masked, one example being HELPER(vstl). There,
probe_write_access()ultimately callsnotdirty_write()with a size of
1 and misses self-modifying code. However,cpu_stq_be_data_ra()ultimately callsmmu_watch_or_dirty(), which in turn callsnotdirty_write()with the full size.
Quoted 8 lines
Quoted 4 lines
It's still worth improving this, because there may still be user-visible adverse effects in other helpers.
Signed-off-by Ilya Leoshkevich <[email protected]>IIRC there are some uses of probe_access_* that set size == 0. Should we adjust addr+size to cover the whole page for that case? That seems to be the intent, anyway.
There is a comment that says we shouldn't do watchpoint/smc detection in this case:
/* Per the interface, size == 0 merely faults the access. */
if (size == 0) {
return NULL;
}Come to think of it, qemu-user works this way too: SMC is detected on the actual access, not the probe:
helper_vstl()
cpu_stq_be_data_ra()
...
stq_he_p()
<signal handler called>
host_signal_handler()
handle_sigsegv_accerr_write()
page_unprotect()
tb_invalidate_phys_page_unwind()
cpu_loop_exit_noexc()So all this is probably fine, I now think it's better to leave the code as is, especially given that I cannot reproduce the original problem anymore.
Quoted 27 lines · Ilya Leoshkevich
Quoted 3 lines · Richard Henderson
IIRC there are some uses of probe_access_* that set size == 0. Should we adjust addr+size to cover the whole page for that case? That seems to be the intent, anyway.
There is a comment that says we shouldn't do watchpoint/smc detection in this case:
/* Per the interface, size == 0 merely faults the access. */ if (size == 0) { return NULL; }Come to think of it, qemu-user works this way too: SMC is detected on the actual access, not the probe:
helper_vstl() cpu_stq_be_data_ra() ... stq_he_p() <signal handler called> host_signal_handler() handle_sigsegv_accerr_write() page_unprotect() tb_invalidate_phys_page_unwind() cpu_loop_exit_noexc()
Quoted 3 lines
So all this is probably fine, I now think it's better to leave the code as is, especially given that I cannot reproduce the original problem anymore.
Ok then.
r~