- Missing check for TIF_WORK if trace/audit enabled
PATCH 1/1 - Missing check for TIF_WORK if trace/audit enabled
It appears that if trace/audit is enabled, the kernel fails to check for pending work before returning to user mode.
---
Tony / Ken - check this carefully. This is not easy code to understand :-)
I noticed the problem running a signal test that sent SIGUSR2 signals to itself. Signals were being dropped. I isolated the problem to an uninitialzed p6:
ia64_trace_syscall -> work_pending_syscall_end -> work_processed_syscall
work_processed_syscall assumes p6 indicates if the kernel is returning to user mode. Nothing set p6 for this particular path.
Index: linux/arch/ia64/kernel/entry.S
=================================--- linux.orig/arch/ia64/kernel/entry.S 2006-02-15 12:32:49.836316366 -0600
+++ linux/arch/ia64/kernel/entry.S 2006-02-15 15:09:49.983620290 -0600
@@ -569,6 +569,7 @@ GLOBAL_ENTRY(ia64_trace_syscall)
.mem.offset 0,0; st8.spill [r2]=r8 // store return value in slot for r8
.mem.offset 8,0; st8.spill [r3]=r10 // clear error indication in slot for r10
br.call.sptk.many rp=syscall_trace_leave // give parent a chance to catch return value
+(pUStk) cmp.eq.unc p6,p0=r0,r0 // p6 <- pUStk
.ret3: br.cond.sptk .work_pending_syscall_end
strace_error:
Jack Steiner wrote on Wednesday, February 15, 2006 1:29 PM
Quoted 8 lines
It appears that if trace/audit is enabled, the kernel fails to check for pending work before returning to user mode.
Tony / Ken - check this carefully. This is not easy code to understand :-)
I noticed the problem running a signal test that sent SIGUSR2 signals to itself. Signals were being dropped. I isolated the problem to an uninitialzed p6:
Jack, I agree with this fix. The only request I have is to move the .ret3 label up before cmp inst, so cmp and br end up in the same instruction bundle instead of two right now.
Quoted 11 lines
Index:
linux/arch/ia64/kernel/entry.S
=================================> ---linux.orig/arch/ia64/kernel/entry.S2006-02-15 12:32:49.836316366 -0600
+++linux/arch/ia64/kernel/entry.S2006-02-15 15:09:49.983620290 -0600(quoted diff) +1
-0@@ -569,6 +569,7 @@ GLOBAL_ENTRY(ia64_trace_syscall) .mem.offset 0,0; st8.spill [r2]=r8 // store return value in slot for r8 .mem.offset 8,0; st8.spill [r3]=r10 // clear error indication in slot for r10 br.call.sptk.many rp=syscall_trace_leave // give parent a chance to catch return value +(pUStk) cmp.eq.unc p6,p0=r0,r0 // p6 <- pUStk .ret3: br.cond.sptk .work_pending_syscall_end strace_error:
It appears that if auditing is enabled, the kernel fails to check for pending signals before returning to user mode.
---
Tony - I think this should go into 2.6.16
Second version. Moved the ".ret" label as suggested by Ken.
I noticed the problem running a signal test that sent SIGUSR2 signals to itself. Signals were being dropped. I isolated the problem to an uninitialzed p6:
ia64_trace_syscall -> work_pending_syscall_end -> work_processed_syscall
work_processed_syscall assumes p6 indicates if the kernel is returning to user mode. Nothing set p6 for this particular path.
Index: linux/arch/ia64/kernel/entry.S
=================================--- linux.orig/arch/ia64/kernel/entry.S 2006-02-15 12:32:49.836316366 -0600
+++ linux/arch/ia64/kernel/entry.S 2006-02-15 19:35:24.539592133 -0600
@@ -569,7 +569,9 @@ GLOBAL_ENTRY(ia64_trace_syscall)
.mem.offset 0,0; st8.spill [r2]=r8 // store return value in slot for r8
.mem.offset 8,0; st8.spill [r3]=r10 // clear error indication in slot for r10
br.call.sptk.many rp=syscall_trace_leave // give parent a chance to catch return value
-.ret3: br.cond.sptk .work_pending_syscall_end
+.ret3:
+(pUStk) cmp.eq.unc p6,p0=r0,r0 // p6 <- pUStk
+ br.cond.sptk .work_pending_syscall_end
strace_error:
ld8 r3=[r2] // load pt_regs.r8