Message ID | 20210729131244.2595519-1-mpe@ellerman.id.au (mailing list archive) |
---|---|
State | Accepted |
Headers | show |
Series | powerpc/vdso: Don't use r30 to avoid breaking Go lang | expand |
Context | Check | Description |
---|---|---|
snowpatch_ozlabs/github-powerpc_selftests | success | Successfully ran 8 jobs. |
snowpatch_ozlabs/github-powerpc_sparse | success | Successfully ran 4 jobs. |
snowpatch_ozlabs/github-powerpc_ppctests | success | Successfully ran 8 jobs. |
snowpatch_ozlabs/github-powerpc_clang | success | Successfully ran 8 jobs. |
snowpatch_ozlabs/github-powerpc_kernel_qemu | success | Successfully ran 25 jobs. |
Dear Michael, Am 29.07.21 um 15:12 schrieb Michael Ellerman: > The Go runtime uses r30 for some special value called 'g'. It assumes > that value will remain unchanged even when calling VDSO functions. > Although r30 is non-volatile across function calls, the callee is free > to use it, as long as the callee saves the value and restores it before > returning. > > It used to be true by accident that the VDSO didn't use r30, because the > VDSO was hand-written asm. When we switched to building the VDSO from C > the compiler started using r30, at least in some builds, leading to > crashes in Go. eg: > > ~/go/src$ ./all.bash > Building Go cmd/dist using /usr/lib/go-1.16. (go1.16.2 linux/ppc64le) > Building Go toolchain1 using /usr/lib/go-1.16. > go build os/exec: /usr/lib/go-1.16/pkg/tool/linux_ppc64le/compile: signal: segmentation fault > go build reflect: /usr/lib/go-1.16/pkg/tool/linux_ppc64le/compile: signal: segmentation fault > go tool dist: FAILED: /usr/lib/go-1.16/bin/go install -gcflags=-l -tags=math_big_pure_go compiler_bootstrap bootstrap/cmd/...: exit status 1 > > There are patches in flight to fix Go[1], but until they are released > and widely deployed we can workaround it in the VDSO by avoiding use of Nit: work around is spelled with a space. > r30. > > Note this only works with GCC, clang does not support -ffixed-rN. Maybe the clang/LLVM build support folks (in CC) have an idea. > 1: https://go-review.googlesource.com/c/go/+/328110 > > Fixes: ab037dd87a2f ("powerpc/vdso: Switch VDSO to generic C implementation.") > Cc: stable@vger.kernel.org # v5.11+ > Reported-by: Paul Menzel <pmenzel@molgen.mpg.de> > Tested-by: Paul Menzel <pmenzel@molgen.mpg.de> > Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> > --- > arch/powerpc/kernel/vdso64/Makefile | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/arch/powerpc/kernel/vdso64/Makefile b/arch/powerpc/kernel/vdso64/Makefile > index 2813e3f98db6..3c5baaa6f1e7 100644 > --- a/arch/powerpc/kernel/vdso64/Makefile > +++ b/arch/powerpc/kernel/vdso64/Makefile > @@ -27,6 +27,13 @@ KASAN_SANITIZE := n > > ccflags-y := -shared -fno-common -fno-builtin -nostdlib \ > -Wl,-soname=linux-vdso64.so.1 -Wl,--hash-style=both > + > +# Go prior to 1.16.x assumes r30 is not clobbered by any VDSO code. That used to be true > +# by accident when the VDSO was hand-written asm code, but may not be now that the VDSO is > +# compiler generated. To avoid breaking Go tell GCC not to use r30. Impact on code > +# generation is minimal, it will just use r29 instead. > +ccflags-y += $(call cc-option, -ffixed-r30) > + > asflags-y := -D__VDSO64__ -s > > targets += vdso64.lds > The rest looks good. Kind regards, Paul
On Thu, Jul 29, 2021 at 6:42 AM Paul Menzel <pmenzel@molgen.mpg.de> wrote: > > Dear Michael, > > > Am 29.07.21 um 15:12 schrieb Michael Ellerman: > > The Go runtime uses r30 for some special value called 'g'. It assumes > > that value will remain unchanged even when calling VDSO functions. > > Although r30 is non-volatile across function calls, the callee is free > > to use it, as long as the callee saves the value and restores it before > > returning. > > > > It used to be true by accident that the VDSO didn't use r30, because the > > VDSO was hand-written asm. When we switched to building the VDSO from C > > the compiler started using r30, at least in some builds, leading to > > crashes in Go. eg: > > > > ~/go/src$ ./all.bash > > Building Go cmd/dist using /usr/lib/go-1.16. (go1.16.2 linux/ppc64le) > > Building Go toolchain1 using /usr/lib/go-1.16. > > go build os/exec: /usr/lib/go-1.16/pkg/tool/linux_ppc64le/compile: signal: segmentation fault > > go build reflect: /usr/lib/go-1.16/pkg/tool/linux_ppc64le/compile: signal: segmentation fault > > go tool dist: FAILED: /usr/lib/go-1.16/bin/go install -gcflags=-l -tags=math_big_pure_go compiler_bootstrap bootstrap/cmd/...: exit status 1 > > > > There are patches in flight to fix Go[1], but until they are released > > and widely deployed we can workaround it in the VDSO by avoiding use of > > Nit: work around is spelled with a space. > > > r30. > > > > Note this only works with GCC, clang does not support -ffixed-rN. > > Maybe the clang/LLVM build support folks (in CC) have an idea. Right, we've had issues with these in the past. Generally, we need to teach clang about which registers are valid for `N` so that it can diagnose invalid values ASAP. This has to be done on a per arch basis in LLVM to steal the register from the register allocator. For example, this was used previously for aarch64 (but removed from use in the kernel) and IIRC is used for m68k (which we're working to get builds online for). I've filed https://bugs.llvm.org/show_bug.cgi?id=51272. Thanks for the report. > > > 1: https://go-review.googlesource.com/c/go/+/328110 > > > > Fixes: ab037dd87a2f ("powerpc/vdso: Switch VDSO to generic C implementation.") > > Cc: stable@vger.kernel.org # v5.11+ > > Reported-by: Paul Menzel <pmenzel@molgen.mpg.de> > > Tested-by: Paul Menzel <pmenzel@molgen.mpg.de> > > Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> > > --- > > arch/powerpc/kernel/vdso64/Makefile | 7 +++++++ > > 1 file changed, 7 insertions(+) > > > > diff --git a/arch/powerpc/kernel/vdso64/Makefile b/arch/powerpc/kernel/vdso64/Makefile > > index 2813e3f98db6..3c5baaa6f1e7 100644 > > --- a/arch/powerpc/kernel/vdso64/Makefile > > +++ b/arch/powerpc/kernel/vdso64/Makefile > > @@ -27,6 +27,13 @@ KASAN_SANITIZE := n > > > > ccflags-y := -shared -fno-common -fno-builtin -nostdlib \ > > -Wl,-soname=linux-vdso64.so.1 -Wl,--hash-style=both > > + > > +# Go prior to 1.16.x assumes r30 is not clobbered by any VDSO code. That used to be true > > +# by accident when the VDSO was hand-written asm code, but may not be now that the VDSO is > > +# compiler generated. To avoid breaking Go tell GCC not to use r30. Impact on code > > +# generation is minimal, it will just use r29 instead. > > +ccflags-y += $(call cc-option, -ffixed-r30) > > + > > asflags-y := -D__VDSO64__ -s > > > > targets += vdso64.lds > > > > The rest looks good. > > > Kind regards, > > Paul
On Thu, Jul 29, 2021 at 09:25:43AM -0700, Nick Desaulniers wrote: > On Thu, Jul 29, 2021 at 6:42 AM Paul Menzel <pmenzel@molgen.mpg.de> wrote: > > Am 29.07.21 um 15:12 schrieb Michael Ellerman: > > > Note this only works with GCC, clang does not support -ffixed-rN. > > > > Maybe the clang/LLVM build support folks (in CC) have an idea. > > Right, we've had issues with these in the past. Generally, we need to > teach clang about which registers are valid for `N` so that it can > diagnose invalid values ASAP. This has to be done on a per arch basis > in LLVM to steal the register from the register allocator. For > example, this was used previously for aarch64 (but removed from use in > the kernel) and IIRC is used for m68k (which we're working to get > builds online for). In GCC, it is -ffixed-* (note: no "r"). The string is stripped of the standard prefix for the target (for Power, none), and possibly of one "%" or "#". If the string is a recognised register name (or alternative register name) for the target, that is used. If not, and it is a decimal number, then the internal GCC register of that number is used (these numbers can differ from one GCC release to another, and in fact we have changed the numbering for Power before -- but 0..31 have always been the GPRs, and 32..63 have always been the FPRs). The names for the Power registers are: GPRs: 0..31, or alternatively r0..r31 FPRs: 0..31, or alternatively fr0..fr31, or alternatively vs0..vs31 VRs: 0..31, or alternatively v0..v31, or alternatively vs32..vs63 CR fields: 0..7, or alternatively cr0..cr7 (There are more, but changing the default calling convention for those will not work anyway. Also, some non-Linux configurations use different names.) Segher
On Thu, 29 Jul 2021 23:12:44 +1000, Michael Ellerman wrote: > The Go runtime uses r30 for some special value called 'g'. It assumes > that value will remain unchanged even when calling VDSO functions. > Although r30 is non-volatile across function calls, the callee is free > to use it, as long as the callee saves the value and restores it before > returning. > > It used to be true by accident that the VDSO didn't use r30, because the > VDSO was hand-written asm. When we switched to building the VDSO from C > the compiler started using r30, at least in some builds, leading to > crashes in Go. eg: > > [...] Applied to powerpc/fixes. [1/1] powerpc/vdso: Don't use r30 to avoid breaking Go lang https://git.kernel.org/powerpc/c/a88603f4b92ecef9e2359e40bcb99ad399d85dd7 cheers
Michael Ellerman <mpe@ellerman.id.au> writes: > The Go runtime uses r30 for some special value called 'g'. It assumes > that value will remain unchanged even when calling VDSO functions. > Although r30 is non-volatile across function calls, the callee is free > to use it, as long as the callee saves the value and restores it before > returning. > > It used to be true by accident that the VDSO didn't use r30, because the > VDSO was hand-written asm. When we switched to building the VDSO from C > the compiler started using r30, at least in some builds, leading to > crashes in Go. eg: > > ~/go/src$ ./all.bash > Building Go cmd/dist using /usr/lib/go-1.16. (go1.16.2 linux/ppc64le) > Building Go toolchain1 using /usr/lib/go-1.16. > go build os/exec: /usr/lib/go-1.16/pkg/tool/linux_ppc64le/compile: signal: segmentation fault > go build reflect: /usr/lib/go-1.16/pkg/tool/linux_ppc64le/compile: signal: segmentation fault > go tool dist: FAILED: /usr/lib/go-1.16/bin/go install -gcflags=-l -tags=math_big_pure_go compiler_bootstrap bootstrap/cmd/...: exit status 1 > > There are patches in flight to fix Go[1], but until they are released > and widely deployed we can workaround it in the VDSO by avoiding use of > r30. > > Note this only works with GCC, clang does not support -ffixed-rN. > > 1: https://go-review.googlesource.com/c/go/+/328110 > > Fixes: ab037dd87a2f ("powerpc/vdso: Switch VDSO to generic C implementation.") > Cc: stable@vger.kernel.org # v5.11+ In practice, with GCC 10.3.0, that commit doesn't result in r30 being used by the compiler. It's commit 74205b3fc2ef ("powerpc/vdso: Add support for time namespaces"), which went into v5.13-rc1, which causes r30 to be used in __c_kernel_clock_gettime(): 00000000000006e0 <__c_kernel_clock_gettime>: 6e0: 0f 00 03 28 cmplwi r3,15 6e4: ec 00 81 41 bgt 7d0 <__c_kernel_clock_gettime+0xf0> 6e8: 01 00 20 39 li r9,1 6ec: 30 18 29 7d slw r9,r9,r3 6f0: 83 08 2a 71 andi. r10,r9,2179 6f4: fc 00 82 41 beq 7f0 <__c_kernel_clock_gettime+0x110> 6f8: e4 26 63 78 rldicr r3,r3,4,59 6fc: ff 7f 20 3d lis r9,32767 700: f0 ff c1 fb std r30,-16(r1) 704: f8 ff e1 fb std r31,-8(r1) 708: 14 1a c5 7c add r6,r5,r3 70c: ff ff 2b 61 ori r11,r9,65535 710: 00 00 05 81 lwz r8,0(r5) 714: 01 00 09 71 andi. r9,r8,1 718: 98 00 82 40 bne 7b0 <__c_kernel_clock_gettime+0xd0> 71c: ac 04 20 7c lwsync 720: a6 42 cc 7f mftb r30 cheers
diff --git a/arch/powerpc/kernel/vdso64/Makefile b/arch/powerpc/kernel/vdso64/Makefile index 2813e3f98db6..3c5baaa6f1e7 100644 --- a/arch/powerpc/kernel/vdso64/Makefile +++ b/arch/powerpc/kernel/vdso64/Makefile @@ -27,6 +27,13 @@ KASAN_SANITIZE := n ccflags-y := -shared -fno-common -fno-builtin -nostdlib \ -Wl,-soname=linux-vdso64.so.1 -Wl,--hash-style=both + +# Go prior to 1.16.x assumes r30 is not clobbered by any VDSO code. That used to be true +# by accident when the VDSO was hand-written asm code, but may not be now that the VDSO is +# compiler generated. To avoid breaking Go tell GCC not to use r30. Impact on code +# generation is minimal, it will just use r29 instead. +ccflags-y += $(call cc-option, -ffixed-r30) + asflags-y := -D__VDSO64__ -s targets += vdso64.lds