Skip to contentSkip to conversation

avoid infinite loop in x86_64 interrupt return

PATCH Andrea Arcangeli 3 participants 7 messages May 4, 2005

lore
Share & tools
lore
#0 lore raw

Hello,

A few minutes ago I've got an unkillable task in R state with vanilla 2.6.12-rc3 on x86_64, luckily system was still up with the other cpu (on the desktop, so I had no kgdb environment set). Debugging revelaed rdi corrupt when entering retint_signal (not set to $_TIF_WORK_MASK as expected). This lead the rdx&rdi to return 0x20000 -> infinite loop. Precisely rdi is set to ffff810030923f58 instead of $_TIF_WORK_MASK. So it was the combination of ...2xxxx as rsp with TIF_IA32 in the task flags. After noticing the rdi screwup the bug was quite clear: rdi was set to pt_regs instead of $_TIF_WORK_MASK. Of course rsp is set to ffff810030923f58 too (which also means do_notify_resume didn't clobber rdi even if it could).

The below should fix the problem, I've no idea how to reproduce the problem but it works on basic testing. The task looping was java (32bit, that's where the 0x20000 come from), but it wasn't me starting java, it must have been some random website (java was hanging around with 100% system time for half an hour once I noticed it).

arch/x86_64/kernel/entry.S renamed from arch/x86_64/kernel/entry.S.orig +1 -0
@@ -489,6 +489,7 @@ retint_signal:
 	movq %rsp,%rdi		# &pt_regs
 	call do_notify_resume
 	RESTORE_REST
+	movl $_TIF_WORK_MASK,%edi
 	cli
 	GET_THREAD_INFO(%rcx)	
 	jmp retint_check
#1 lore raw

Hi,

Quoted 25 lines · Andrea Arcangeli

Hello,

A few minutes ago I've got an unkillable task in R state with vanilla 2.6.12-rc3 on x86_64, luckily system was still up with the other cpu (on the desktop, so I had no kgdb environment set). Debugging revelaed rdi corrupt when entering retint_signal (not set to $_TIF_WORK_MASK as expected). This lead the rdx&rdi to return 0x20000 -> infinite loop. Precisely rdi is set to ffff810030923f58 instead of $_TIF_WORK_MASK. So it was the combination of ...2xxxx as rsp with TIF_IA32 in the task flags. After noticing the rdi screwup the bug was quite clear: rdi was set to pt_regs instead of $_TIF_WORK_MASK. Of course rsp is set to ffff810030923f58 too (which also means do_notify_resume didn't clobber rdi even if it could).

The below should fix the problem, I've no idea how to reproduce the problem but it works on basic testing. The task looping was java (32bit, that's where the 0x20000 come from), but it wasn't me starting java, it must have been some random website (java was hanging around with 100% system time for half an hour once I noticed it).

arch/x86_64/kernel/entry.S retint_signal:

Quoted 8 lines
arch/x86_64/kernel/entry.S renamed from arch/x86_64/kernel/entry.S.orig +1 -1
@@ -489,6 +489,7 @@ retint_signal:
 	movq %rsp,%rdi		# &pt_regs
 	call do_notify_resume
 	RESTORE_REST
+	movl $_TIF_WORK_MASK,%edi
 	cli
 	GET_THREAD_INFO(%rcx)	
 	jmp retint_check
-

You also need to add two missing clis. Please have a look at the attached patch from Andi.

Greets,
Rafael

Signature · 3 lines
- Would you tell me, please, which way I ought to go from here?
- That depends a good deal on where you want to get to.
		-- Lewis Carroll "Alice's Adventures in Wonderland"
arch/x86_64/kernel/entry.S renamed from arch/x86_64/kernel/entry.S-o +4 -1
diff -u linux-2.6.12rc3/arch/x86_64/kernel/entry.S-o linux-2.6.12rc3/arch/x86_64/kernel/entry.S
@@ -296,6 +296,7 @@
 	call syscall_trace_leave
 	popq %rdi
 	andl $~(_TIF_SYSCALL_TRACE|_TIF_SYSCALL_AUDIT|_TIF_SINGLESTEP),%edi
+	cli	
 	jmp int_restore_rest
 	
 int_signal:
@@ -307,6 +308,7 @@
 1:	movl $_TIF_NEED_RESCHED,%edi	
 int_restore_rest:
 	RESTORE_REST
+	cli	
 	jmp int_with_check
 	CFI_ENDPROC
 		
@@ -490,7 +492,8 @@
 	call do_notify_resume
 	RESTORE_REST
 	cli
-	GET_THREAD_INFO(%rcx)	
+	GET_THREAD_INFO(%rcx)
+	movl $_TIF_WORK_MASK,%edi	
 	jmp retint_check
 
 #ifdef CONFIG_PREEMPT
#5 lore raw
In reply to Andrea Arcangeli

It is already fixed in mainline. Actually I think it was a merging problem I had actually fixed it before the last merge in my tree, but some patches got lost :/

-Andi

#2 lore raw
In reply to Rafael J. Wysocki
Quoted 2 lines · Rafael J. Wysocki

You also need to add two missing clis. Please have a look at the attached patch from Andi.

Those two clis seems unrelated, so I don't see why I should mix them in the same patch. I couldn't trigger the other problems, only the one with rdi corruption.

Note that those clis seems superflous, cli is only needed before swapgs, so adding cli before swapgs in retint_swapgs sounds a better idea than to keep irq off for a longer time for no apparent good reason. But I've no real idea why those cli are needed so I guess I must be missing something. there's no commentary attached to your patch that can exlain why the cli are needed _way_ before calling swapgs.

#6 lore raw
In reply to Andi Kleen
Quoted 4 lines · Andi Kleen

It is already fixed in mainline. Actually I think it was a merging problem I had actually fixed it before the last merge in my tree, but some patches got lost :/

Ok thanks. I'm still not in sync with git so I didn't notice, I looked into mercurial instead of git infact.

#3 lore raw
In reply to Andrea Arcangeli
Quoted 7 lines · Andrea Arcangeli
Quoted 2 lines · Rafael J. Wysocki

You also need to add two missing clis. Please have a look at the attached patch from Andi.

Those two clis seems unrelated, so I don't see why I should mix them in the same patch. I couldn't trigger the other problems, only the one with rdi corruption.

THere was a second patch which essentially got the line you posted together with the missing clis.

I originally fixed it in a slightly different way (in the version that got lost), but this one was equivalent.

Quoted 7 lines

Note that those clis seems superflous, cli is only needed before swapgs, so adding cli before swapgs in retint_swapgs sounds a better idea than to keep irq off for a longer time for no apparent good reason. But I've no real idea why those cli are needed so I guess I must be missing something. there's no commentary attached to your patch that can exlain why the cli are needed _way_ before calling swapgs.

To avoid losing schedule events and signals. Between checking for them and returning to user space interrupts need to be off. When they are reenabled everything needs to be rechecked.

-Andi

#4 lore raw
Quoted 3 lines · Andi Kleen

To avoid losing schedule events and signals. Between checking for them and returning to user space interrupts need to be off. When they are reenabled everything needs to be rechecked.

All right good point, the latency issues, they'd wait the timeslice to expire instead of being processed immediatly. That couldn't be noticed without some measurement but it's sure worth fixing agreed, thanks!