powerpc/64/signal: balance return predictor stack in signal trampoline
Nicholas Piggin
reviewerPATCH v2 1/1 powerpc/64/signal: balance return predictor stack in signal trampoline
Returning from an interrupt or syscall to a signal handler currently begins execution directly at the handler's entry point, with LR set to the address of the sigreturn trampoline. When the signal handler function returns, it runs the trampoline. It looks like this:
# interrupt at user address xyz
# kernel stuff... signal is raised
rfid
# void handler(int sig)
addis 2,12,.TOC.-.LCF0@ha
addi 2,2,.TOC.-.LCF0@l
mflr 0
std 0,16(1)
stdu 1,-96(1)
# handler stuff
ld 0,16(1)
mtlr 0
blr
# __kernel_sigtramp_rt64
addi r1,r1,__SIGNAL_FRAMESIZE
li r0,__NR_rt_sigreturn
sc
# kernel executes rt_sigreturn
rfid
# back to user address xyzNote the blr with no matching bl. This can corrupt the return predictor.
Solve this by instead resuming execution at the signal trampoline which then calls the signal handler. qtrace-tools link_stack checker confirms the entire user/kernel/vdso cycle is balanced after this patch, whereas it's not upstream.
Alan confirms the dwarf unwind info still looks good. gdb still recognises the signal frame and can step into parent frames if it break inside a signal handler.
Performance is pretty noisy, not a very significant change on a POWER9 here, but branch misses are consistently a lot lower on a microbenchmark:
Performance counter stats for './signal':
13,085.72 msec task-clock # 1.000 CPUs utilized
45,024,760,101 cycles # 3.441 GHz
65,102,895,542 instructions # 1.45 insn per cycle
11,271,673,787 branches # 861.372 M/sec
59,468,979 branch-misses # 0.53% of all branches 12,989.09 msec task-clock # 1.000 CPUs utilized
44,692,719,559 cycles # 3.441 GHz
65,109,984,964 instructions # 1.46 insn per cycle
11,282,136,057 branches # 868.585 M/sec
39,786,942 branch-misses # 0.35% of all branches---
Since v1:
- Fix for the legacy on-stack trampoline path (TRAMP_TRACEBACK was not updated to 4).
- Tested legacy case (must also disable no-exec stack to test this).
arch/powerpc/include/asm/ppc-opcode.h | 1 + arch/powerpc/kernel/signal_64.c | 22 ++++++++++++---------- arch/powerpc/kernel/vdso64/sigtramp.S | 13 +++++-------- 3 files changed, 18 insertions(+), 18 deletions(-)
arch/powerpc/include/asm/ppc-opcode.h +1 -0
@@ -329,6 +329,7 @@#define PPC_INST_BLR 0x4e800020#define PPC_INST_BLRL 0x4e800021#define PPC_INST_BCTR 0x4e800420 +#define PPC_INST_BCTRL 0x4e800421#define PPC_INST_MULLD 0x7c0001d2#define PPC_INST_MULLW 0x7c0001d6#define PPC_INST_MULHWU 0x7c000016
arch/powerpc/kernel/signal_64.c +12 -10
@@ -40,8 +40,8 @@#define GP_REGS_SIZE min(sizeof(elf_gregset_t), sizeof(struct pt_regs))#define FP_REGS_SIZE sizeof(elf_fpregset_t) -#define TRAMP_TRACEBACK 3 -#define TRAMP_SIZE 6 +#define TRAMP_TRACEBACK 4 +#define TRAMP_SIZE 7/* * When we have signals to deliver, we set up on the user stack,
@@ -603,13 +603,15 @@ static long setup_trampoline(unsigned int syscall, unsigned int __user *tramp)int i;long err =0 ; +/* bctrl # call the handler */ + err |=__put_user (PPC_INST_BCTRL , &tramp[0 ]);/* addi r1, r1, __SIGNAL_FRAMESIZE # Pop the dummy stackframe */ err |=__put_user (PPC_INST_ADDI |__PPC_RT (R1 ) |__PPC_RA (R1 ) | - (__SIGNAL_FRAMESIZE &0xffff ), &tramp[0 ]); + (__SIGNAL_FRAMESIZE &0xffff ), &tramp[1 ]);/* li r0, __NR_[rt_]sigreturn| */ - err |=__put_user (PPC_INST_ADDI | (syscall &0xffff ), &tramp[1 ]); + err |=__put_user (PPC_INST_ADDI | (syscall &0xffff ), &tramp[2 ]);/* sc */ - err |=__put_user (PPC_INST_SC , &tramp[2 ]); + err |=__put_user (PPC_INST_SC , &tramp[3 ]);/* Minimal traceback info */ for (i=TRAMP_TRACEBACK ; i <TRAMP_SIZE ;i++)
@@ -867,12 +869,12 @@ int handle_rt_signal64(struct ksignal *ksig, sigset_t *set,/* Set up to return from userspace. */ if (vdso64_rt_sigtramp && tsk->mm->context.vdso_base) { - regs->link = tsk->mm->context.vdso_base + vdso64_rt_sigtramp; + regs->nip = tsk->mm->context.vdso_base + vdso64_rt_sigtramp; }else { err |=setup_trampoline (__NR_rt_sigreturn, &frame->tramp[0 ]);if (err) goto badframe; - regs->link = (unsigned long ) &frame->tramp[0 ]; + regs->nip = (unsigned long ) &frame->tramp[0 ]; }/* Allocate a dummy caller frame for the signal handler. */
@@ -881,8 +883,8 @@ int handle_rt_signal64(struct ksignal *ksig, sigset_t *set,/* Set up "regs" so we "return" to the signal handler. */ if (is_elf2_task ()) { - regs->nip = (unsigned long ) ksig->ka.sa.sa_handler; - regs->gpr[12 ] = regs->nip; + regs->ctr = (unsigned long ) ksig->ka.sa.sa_handler; + regs->gpr[12 ] = regs->ctr; }else {/* Handler is *really* a pointer to the function descriptor for * the signal routine. The first entry in the function
@@ -892,7 +894,7 @@ int handle_rt_signal64(struct ksignal *ksig, sigset_t *set,func_descr_t __user *funct_desc_ptr = (func_descr_t __user *) ksig->ka.sa.sa_handler; - err |=get_user (regs->nip, &funct_desc_ptr->entry); + err |=get_user (regs->ctr, &funct_desc_ptr->entry); err |=get_user (regs->gpr[2 ], &funct_desc_ptr->toc); }
arch/powerpc/kernel/vdso64/sigtramp.S +5 -8
@@ -6,6 +6,7 @@ * Copyright (C) 2004 Benjamin Herrenschmuidt ([email protected]), IBM Corp. * Copyright (C) 2004 Alan Modra ([email protected])), IBM Corp. */ +#include <asm/cache.h> /* IFETCH_ALIGN_BYTES */ #include <asm/processor.h> #include <asm/ppc_asm.h> #include <asm/unistd.h>
@@ -14,21 +15,17 @@ .text -/* The nop here is a hack. The dwarf2 unwind routines subtract 1 from - the return address to get an address in the middle of the presumed - call instruction. Since we don't have a call here, we artificially - extend the range covered by the unwind info by padding before the - real start. */ - nop .balign 8 + .balign IFETCH_ALIGN_BYTES V_FUNCTION_BEGIN(__kernel_sigtramp_rt64) -.Lsigrt_start = . - 4 +.Lsigrt_start: + bctrl /* call the handler */ addi r1, r1, __SIGNAL_FRAMESIZE li r0,__NR_rt_sigreturn sc .Lsigrt_end: V_FUNCTION_END(__kernel_sigtramp_rt64) -/* The ".balign 8" above and the following zeros mimic the old stack +/* The .balign 8 above and the following zeros mimic the old stack trampoline layout. The last magic value is the ucontext pointer, chosen in such a way that older libgcc unwind code returns a zero for a sigcontext pointer. */
Signature · 1 line
2.23.0
Michael Ellerman
AppliedQuoted 26 lines · Nicholas Piggin
Returning from an interrupt or syscall to a signal handler currently begins execution directly at the handler's entry point, with LR set to the address of the sigreturn trampoline. When the signal handler function returns, it runs the trampoline. It looks like this:
# interrupt at user address xyz # kernel stuff... signal is raised rfid # void handler(int sig) addis 2,12,.TOC.-.LCF0@ha addi 2,2,.TOC.-.LCF0@l mflr 0 std 0,16(1) stdu 1,-96(1) # handler stuff ld 0,16(1) mtlr 0 blr # __kernel_sigtramp_rt64 addi r1,r1,__SIGNAL_FRAMESIZE li r0,__NR_rt_sigreturn sc # kernel executes rt_sigreturn rfid # back to user address xyz
Quoted 1 line
[...]
Applied to powerpc/next.
[1/1] powerpc/64/signal: Balance return predictor stack in signal trampoline
https://git.kernel.org/powerpc/c/0138ba5783ae0dcc799ad401a1e8ac8333790df9
cheers
* Nicholas Piggin:
Quoted 28 lines
arch/powerpc/kernel/vdso64/sigtramp.S +4
-7@@ -6,6 +6,7 @@ * Copyright (C) 2004 Benjamin Herrenschmuidt ([email protected]), IBM Corp. * Copyright (C) 2004 Alan Modra ([email protected])), IBM Corp. */ +#include <asm/cache.h> /* IFETCH_ALIGN_BYTES */ #include <asm/processor.h> #include <asm/ppc_asm.h> #include <asm/unistd.h>@@ -14,21 +15,17 @@ .text -/* The nop here is a hack. The dwarf2 unwind routines subtract 1 from - the return address to get an address in the middle of the presumed - call instruction. Since we don't have a call here, we artificially - extend the range covered by the unwind info by padding before the - real start. */ - nop .balign 8 + .balign IFETCH_ALIGN_BYTES V_FUNCTION_BEGIN(__kernel_sigtramp_rt64) -.Lsigrt_start = . - 4 +.Lsigrt_start: + bctrl /* call the handler */
arch/powerpc/kernel/vdso64/sigtramp.S line 31
Quoted 10 lines
arch/powerpc/kernel/vdso64/sigtramp.S +1
-1addi r1, r1, __SIGNAL_FRAMESIZE li r0,__NR_rt_sigreturn sc .Lsigrt_end: V_FUNCTION_END(__kernel_sigtramp_rt64) -/* The ".balign 8" above and the following zeros mimic the old stack +/* The .balign 8 above and the following zeros mimic the old stack trampoline layout. The last magic value is the ucontext pointer, chosen in such a way that older libgcc unwind code returns a zero for a sigcontext pointer. */
As far as I understand it, this breaks cancellation handling on musl and future glibc because it is necessary to look at the signal delivery location to see if a system call sequence has result in an action, and that location is no longer in user code after this change.
We have a glibc test in preparation of our change, and it started failing:
Linux 5.10 breaks sigcontext_get_pc on powerpc64
<https://sourceware.org/bugzilla/show_bug.cgi?id=27223>
Isn't it possible to avoid the return predictor desynchronization by adding the appropriate hint?
Thanks,
Florian
Signature · 3 lines
Red Hat GmbH, https://de.redhat.com/ , Registered seat: Grasbrunn, Commercial register: Amtsgericht Muenchen, HRB 153243, Managing Directors: Charles Cachera, Brian Klemm, Laurie Krebs, Michael O'Neill
Quoted 46 lines · Florian Weimer
* Nicholas Piggin:
Quoted 28 lines
arch/powerpc/kernel/vdso64/sigtramp.S +4
-7@@ -6,6 +6,7 @@ * Copyright (C) 2004 Benjamin Herrenschmuidt ([email protected]), IBM Corp. * Copyright (C) 2004 Alan Modra ([email protected])), IBM Corp. */ +#include <asm/cache.h> /* IFETCH_ALIGN_BYTES */ #include <asm/processor.h> #include <asm/ppc_asm.h> #include <asm/unistd.h>@@ -14,21 +15,17 @@ .text -/* The nop here is a hack. The dwarf2 unwind routines subtract 1 from - the return address to get an address in the middle of the presumed - call instruction. Since we don't have a call here, we artificially - extend the range covered by the unwind info by padding before the - real start. */ - nop .balign 8 + .balign IFETCH_ALIGN_BYTES V_FUNCTION_BEGIN(__kernel_sigtramp_rt64) -.Lsigrt_start = . - 4 +.Lsigrt_start: + bctrl /* call the handler */arch/powerpc/kernel/vdso64/sigtramp.S line 31
Quoted 10 lines
arch/powerpc/kernel/vdso64/sigtramp.S +1
-1addi r1, r1, __SIGNAL_FRAMESIZE li r0,__NR_rt_sigreturn sc .Lsigrt_end: V_FUNCTION_END(__kernel_sigtramp_rt64) -/* The ".balign 8" above and the following zeros mimic the old stack +/* The .balign 8 above and the following zeros mimic the old stack trampoline layout. The last magic value is the ucontext pointer, chosen in such a way that older libgcc unwind code returns a zero for a sigcontext pointer. */As far as I understand it, this breaks cancellation handling on musl and future glibc because it is necessary to look at the signal delivery location to see if a system call sequence has result in an action, and that location is no longer in user code after this change.
Quoted 8 lines
We have a glibc test in preparation of our change, and it started failing:
Linux 5.10 breaks
sigcontext_get_pcon powerpc64
<https://sourceware.org/bugzilla/show_bug.cgi?id=27223>Isn't it possible to avoid the return predictor desynchronization by adding the appropriate hint?
Maybe I'm missing something but I don't see how this would break musl; we just inspect the PC in the mcontext, which I don't see any changes to and which should still point to the next instruction of the interrupted context. I don't have a test environment though so I'll have to wait for feedback from ppc users to be sure. Are there any further details on how it's breaking glibc?
Rich
Quoted 6 lines · Rich Felker
Maybe I'm missing something but I don't see how this would break musl; we just inspect the PC in the mcontext, which I don't see any changes to and which should still point to the next instruction of the interrupted context. I don't have a test environment though so I'll have to wait for feedback from ppc users to be sure. Are there any further details on how it's breaking glibc?
For glibc, backtrace() compares the return-address from each stack frame to the value of __kernel_sigtramp_rt64 to identify the frame with the mcontext information, but now the return-address is not the start of the routine, but the middle of it, so it fails to catch this special frame.
o/
Raoni Fassina
Quoted 8 lines · Raoni Fassina Firmino
Quoted 6 lines · Rich Felker
Maybe I'm missing something but I don't see how this would break musl; we just inspect the PC in the mcontext, which I don't see any changes to and which should still point to the next instruction of the interrupted context. I don't have a test environment though so I'll have to wait for feedback from ppc users to be sure. Are there any further details on how it's breaking glibc?
Quoted 4 lines
For glibc,
backtrace()compares the return-address from each stack frame to the value of__kernel_sigtramp_rt64to identify the frame with the mcontext information, but now the return-address is not the start of the routine, but the middle of it, so it fails to catch this special frame.
Is there a reason it's backtracing rather than just looking at the interrupted context (pointed to by the third argument to the signal handler)?
Rich
Quoted 9 lines · Rich Felker
Quoted 8 lines · Raoni Fassina Firmino
Quoted 6 lines · Rich Felker
Maybe I'm missing something but I don't see how this would break musl; we just inspect the PC in the mcontext, which I don't see any changes to and which should still point to the next instruction of the interrupted context. I don't have a test environment though so I'll have to wait for feedback from ppc users to be sure. Are there any further details on how it's breaking glibc?
Quoted 8 lines
Quoted 4 lines
For glibc,
backtrace()compares the return-address from each stack frame to the value of__kernel_sigtramp_rt64to identify the frame with the mcontext information, but now the return-address is not the start of the routine, but the middle of it, so it fails to catch this special frame.Is there a reason it's backtracing rather than just looking at the interrupted context (pointed to by the third argument to the signal handler)?
The regression is exposed in the backtrace() routine. More precisely, when the backtrace() is called from inside a signal handling. What I described is the way backtrace() uses to identify this special situation. What is failling in glibc is the test for this.
o/
Raoni Fassina
Quoted 46 lines · AL glibc-alpha
* Nicholas Piggin:
Quoted 28 lines
arch/powerpc/kernel/vdso64/sigtramp.S +4
-7@@ -6,6 +6,7 @@ * Copyright (C) 2004 Benjamin Herrenschmuidt ([email protected]), IBM Corp. * Copyright (C) 2004 Alan Modra ([email protected])), IBM Corp. */ +#include <asm/cache.h> /* IFETCH_ALIGN_BYTES */ #include <asm/processor.h> #include <asm/ppc_asm.h> #include <asm/unistd.h>@@ -14,21 +15,17 @@ .text -/* The nop here is a hack. The dwarf2 unwind routines subtract 1 from - the return address to get an address in the middle of the presumed - call instruction. Since we don't have a call here, we artificially - extend the range covered by the unwind info by padding before the - real start. */ - nop .balign 8 + .balign IFETCH_ALIGN_BYTES V_FUNCTION_BEGIN(__kernel_sigtramp_rt64) -.Lsigrt_start = . - 4 +.Lsigrt_start: + bctrl /* call the handler */arch/powerpc/kernel/vdso64/sigtramp.S line 31
Quoted 10 lines
arch/powerpc/kernel/vdso64/sigtramp.S +1
-1addi r1, r1, __SIGNAL_FRAMESIZE li r0,__NR_rt_sigreturn sc .Lsigrt_end: V_FUNCTION_END(__kernel_sigtramp_rt64) -/* The ".balign 8" above and the following zeros mimic the old stack +/* The .balign 8 above and the following zeros mimic the old stack trampoline layout. The last magic value is the ucontext pointer, chosen in such a way that older libgcc unwind code returns a zero for a sigcontext pointer. */As far as I understand it, this breaks cancellation handling on musl and future glibc because it is necessary to look at the signal delivery location to see if a system call sequence has result in an action, and that location is no longer in user code after this change.
Quoted 8 lines
We have a glibc test in preparation of our change, and it started failing:
Linux 5.10 breaks
sigcontext_get_pcon powerpc64
<https://sourceware.org/bugzilla/show_bug.cgi?id=27223>Isn't it possible to avoid the return predictor desynchronization by adding the appropriate hint?
I also caught this regression, I believe it was introduced in the kernel 5.9.
I don't know enough to comment on Florian suggestion, but I am working on some possible fixes:
On the kernel side we can keep __kernel_sigtramp_rt64 in the original place after bctrl and add a new symbol so the kernel can jump to the right place before bctrl. This would ensure backward compatibility.
On the other side, this change exposed how fragile backtrace() is to any changes in the trampoline code, which the libc has no control over in this case. So maybe there is something that can be improved in how backtrace decides that the return-address is to the trampoline. My fist option is to test a range, after __kernel_sigtramp_rt64 so see if the address is inside the routine. This would be better if we can know the size of the function, I know that the vdso.so has this info in the elf but I don't know if it is exposed to the glibc.
As Nicholas mentioned in his patch, GDB seems to keep working just fine, it is seems that GDB uses some heuristics to match the surround code of the return address to identify that it is the trampoline code. So maybe other option is to do something similar.
o/
Raoni Fassina