Skip to contentSkip to conversation

exec: flush CPU TB cache when breakpoint address translation fails

RFC In review Max Filippov 4 participants 5 messages Nov 26, 2019 – Dec 1, 2019

lore
Share & tools
lore
#0 lore raw

RFC 1/1 exec: flush CPU TB cache when breakpoint address translation fails

When a breakpoint is inserted at location for which there's currently no virtual to physical translation no action is taken on CPU TB cache. If a TB for that virtual address already exists but is not visible ATM the breakpoint won't be hit next time an instruction at that address will be executed.

Flush entire CPU TB cache when a breakpoint is set for a virtual address that cannot be translated to physical address.

This change fixes the following workflow:
- linux user application is running
- a breakpoint is inserted from QEMU gdbstub for a user address that is not currently present in the target CPU TLB
- an instruction at that address is executed, but the external debugger doesn't get control.

---
 exec.c | 2 ++
 1 file changed, 2 insertions(+)
exec.c +2 -0
index ffdb5185353b..918945f8097e 100644
@@ -1024,6 +1024,8 @@ static void breakpoint_invalidate(CPUState *cpu, target_ulong pc)
         /* Locks grabbed by tb_invalidate_phys_addr */
         tb_invalidate_phys_addr(cpu->cpu_ases[asidx].as,
                                 phys | (pc & ~TARGET_PAGE_MASK), attrs);
+    } else {
+        tb_flush(cpu);
     }
 }
 #endif
Signature · 1 line
2.20.1
#1 lore raw
Quoted 9 lines · Max Filippov

When a breakpoint is inserted at location for which there's currently no virtual to physical translation no action is taken on CPU TB cache. If a TB for that virtual address already exists but is not visible ATM the breakpoint won't be hit next time an instruction at that address will be executed.

Flush entire CPU TB cache when a breakpoint is set for a virtual address that cannot be translated to physical address.

Quoted 6 lines

This change fixes the following workflow:
- linux user application is running
- a breakpoint is inserted from QEMU gdbstub for a user address that is not currently present in the target CPU TLB
- an instruction at that address is executed, but the external debugger doesn't get control.

That's quite heavyweight, but perhaps since it's debugging we might as well _always_ just flush the TB cache?

In any case, it deserves a comment in the code.

Paolo

#2 lore raw
In reply to Max Filippov
Quoted 5 lines · Max Filippov

When a breakpoint is inserted at location for which there's currently no virtual to physical translation no action is taken on CPU TB cache. If a TB for that virtual address already exists but is not visible ATM the breakpoint won't be hit next time an instruction at that address will be executed.

So the userspace has run once but is currently paged out?

Quoted 29 lines

Flush entire CPU TB cache when a breakpoint is set for a virtual address that cannot be translated to physical address.

This change fixes the following workflow:
- linux user application is running
- a breakpoint is inserted from QEMU gdbstub for a user address that is not currently present in the target CPU TLB
- an instruction at that address is executed, but the external debugger doesn't get control.

---
 exec.c | 2 ++
 1 file changed, 2 insertions(+)
exec.c +2 -0
index ffdb5185353b..918945f8097e 100644
@@ -1024,6 +1024,8 @@ static void breakpoint_invalidate(CPUState *cpu, target_ulong pc)
         /* Locks grabbed by tb_invalidate_phys_addr */
         tb_invalidate_phys_addr(cpu->cpu_ases[asidx].as,
                                 phys | (pc & ~TARGET_PAGE_MASK), attrs);
+    } else {
+        tb_flush(cpu);
     }
 }
 #endif
Signature · 1 line
Alex Bennée
#3 lore raw
Quoted 9 lines · Alex Bennée
Quoted 5 lines · Max Filippov

When a breakpoint is inserted at location for which there's currently no virtual to physical translation no action is taken on CPU TB cache. If a TB for that virtual address already exists but is not visible ATM the breakpoint won't be hit next time an instruction at that address will be executed.

So the userspace has run once but is currently paged out?

Yes, but not necessarily paged out, just not in the CPU TLB. Or it has run to completion and when you start it next time it gets loaded to the same physical pages.

Signature · 2 lines
Thanks.
-- Max
#4 lore raw
In reply to Max Filippov
Quoted 9 lines · Max Filippov

When a breakpoint is inserted at location for which there's currently no virtual to physical translation no action is taken on CPU TB cache. If a TB for that virtual address already exists but is not visible ATM the breakpoint won't be hit next time an instruction at that address will be executed.

Flush entire CPU TB cache when a breakpoint is set for a virtual address that cannot be translated to physical address.

Quoted 8 lines

This change fixes the following workflow:
- linux user application is running
- a breakpoint is inserted from QEMU gdbstub for a user address that is not currently present in the target CPU TLB
- an instruction at that address is executed, but the external debugger doesn't get control.

As a short-term fix this works. I'm willing to apply this to 5.0 with cc:
stable so that as a bug fix this gets pulled back to 4.2.1. Paolo's request for a comment is en pointe, so I'll wait for a version 2.

However, while running in system mode with a decent sized amount of guest ram, I see 1-2M tbs live at once. We don't really want to throw all of those away.

I think it would be better to put the tbs into some sort of data structure that indexes the tbs by their virtual address. I have no strong feelings about what sort of data structure would be best, just something better than a linear search of all of the tbs. :-)

r~