From: Yanfei Xu <yanfei...@windriver.com> This change is for arm64 platform compat mode.
------[cut here: not apply patch]-------- Samples: 119K of event 'cycles' Event count (approx.): 114092698680 Children Self Command Shared Object Symbol ........ ........ ......... .................. 99.85% 0.00% perf-test [unknown] [.] 0x00000000ff80f4ec | ---0xff80f4ec | |--37.68%--__aeabi_idivmod | |--25.35%--calculate_meaning_of_life | |--20.06%--____aeabi_idivmod_from_thumb | |--7.25%--0 | | | |--1.76%--0xf543df9c | | __gettimeofday ------[cut here: apply patch]------------ Samples: 114K of event 'cycles' Event count (approx.): 109799966925 Children Self Command Shared Object Symbol ........ ........ ......... .................. 99.91% 0.00% perf-test perf-test [.] main | ---main | |--57.49%--pthread_create@@GLIBC_2.4 | | | --57.44%--0xf7cf2f88 | | | |--20.16%--__aeabi_idivmod | | | |--13.79%--calculate_meaning_of_life | | | |--10.89%--____aeabi_idivmod_from_thumb | | | |--8.99%--current_timestamp | | | | | --8.98%--__gettimeofday 1/1 [ Author: Fang Jia Email: fang....@windriver.com Subject: arm64/perf: fix backtrace for AAPCS with FP enabled Date: Fri, 28 Dec 2018 16:28:34 +0800 This change is for arm64 platform compat mode. The change for arm32 platform has been included in this commit "perf: fix backtrace for AAPCS with FP enabled". This change replaces code designed for the obsolete ARM APCS ABI, which causes failures of the perf backtrace logic unless the gcc option -mapcs-frame is used to build all binaries on the platform. This obsolete gcc option forces the compiler to include the stack pointer along with the frame pointer and link register in the stack frame for each funciton call. The current AAPCS ABI document, doesn't explicitly describe the frame structure when the gcc frame pointer option, -fno-omit-frame-pointer, is enabled. However, with this option enabled, examination of the emitted prologue instructions shows that 1) R11 is used as the frame pointer, 2) only the R11 and LR are saved onto the stack, not the stack pointer, 3) after this prologue setup, the frame pointer, R11 points to the saved location of LR on the stack. The use of unsigned int arithmetic in the commit is required since the gcc emitted pointer arithmetic uses 8-byte pointer sizes, which are incorrect addresses for the 4-byte stack address size. Signed-off-by: Fang Jia <fang....@windriver.com> Reviewed-by: Jiwei Sun <jiwei....@windriver.com> Signed-off-by: De Huo <de....@windriver.com> ] Signed-off-by: Yanfei Xu <yanfei...@windriver.com> --- ...-backtrace-for-AAPCS-with-FP-enabled.patch | 93 +++++++++++++++++++ patches/misc/misc.scc | 1 + 2 files changed, 94 insertions(+) create mode 100644 patches/misc/arm64-perf-fix-backtrace-for-AAPCS-with-FP-enabled.patch diff --git a/patches/misc/arm64-perf-fix-backtrace-for-AAPCS-with-FP-enabled.patch b/patches/misc/arm64-perf-fix-backtrace-for-AAPCS-with-FP-enabled.patch new file mode 100644 index 00000000..4c1e78cf --- /dev/null +++ b/patches/misc/arm64-perf-fix-backtrace-for-AAPCS-with-FP-enabled.patch @@ -0,0 +1,93 @@ +From cbbce37ccc6041d3ae3d3cb3b1918a61a39820a0 Mon Sep 17 00:00:00 2001 +From: Fang Jia <fang....@windriver.com> +Date: Fri, 28 Dec 2018 16:28:34 +0800 +Subject: [PATCH] arm64/perf: fix backtrace for AAPCS with FP enabled + +This change is for arm64 platform compat mode. +The change for arm32 platform has been included in this commit "perf: fix +backtrace for AAPCS with FP enabled". + +This change replaces code designed for the obsolete ARM APCS ABI, which +causes failures of the perf backtrace logic unless the gcc option +-mapcs-frame is used to build all binaries on the platform. This +obsolete gcc option forces the compiler to include the stack pointer +along with the frame pointer and link register in the stack frame +for each funciton call. The current AAPCS ABI document, doesn't +explicitly describe the frame structure when the gcc frame pointer +option, -fno-omit-frame-pointer, is enabled. However, with this option +enabled, examination of the emitted prologue instructions shows that +1) R11 is used as the frame pointer, +2) only the R11 and LR are saved onto the stack, not the stack pointer, +3) after this prologue setup, the frame pointer, R11 points to the +saved location of LR on the stack. + +The use of unsigned int arithmetic in the commit is required since +the gcc emitted pointer arithmetic uses 8-byte pointer sizes, which are +incorrect addresses for the 4-byte stack address size. + +Signed-off-by: Fang Jia <fang....@windriver.com> +Reviewed-by: Jiwei Sun <jiwei....@windriver.com> +Signed-off-by: De Huo <de....@windriver.com> +--- + arch/arm64/kernel/perf_callchain.c | 25 +++++++++++++++---------- + 1 file changed, 15 insertions(+), 10 deletions(-) + +diff --git a/arch/arm64/kernel/perf_callchain.c b/arch/arm64/kernel/perf_callchain.c +index 9d63514b9836..b9bbe0b04fc1 100644 +--- a/arch/arm64/kernel/perf_callchain.c ++++ b/arch/arm64/kernel/perf_callchain.c +@@ -54,16 +54,22 @@ user_backtrace(struct frame_tail __user *tail, + + #ifdef CONFIG_COMPAT + /* +- * The registers we're interested in are at the end of the variable +- * length saved register structure. The fp points at the end of this +- * structure so the address of this struct is: +- * (struct compat_frame_tail *)(xxx->fp)-1 ++ * The AAPCS ABI, the most current replacing the obsolete APCS ABI, ++ * does not specifically describe the stack frame with respect to the ++ * frame pointer. However, the examination of emitted prologue ++ * instructions for ARM implies that with -fno-omit-framepointer, ++ * register R11 is used as the frame pointer register and saved on the ++ * stack, with LR. + * +- * This code has been adapted from the ARM OProfile support. ++ * After the prolog, the FP points to the location of the saved LR and ++ * FP+4 points to the previous frames FP as shown below: ++ * Stack Hi Mem ++ * (Value of FP)+4 Saved FP for caller ++ * (Value of FP) LR set by caller ++ * Stack Lo Mem + */ + struct compat_frame_tail { + compat_uptr_t fp; /* a (struct compat_frame_tail *) in compat mode */ +- u32 sp; + u32 lr; + } __attribute__((packed)); + +@@ -91,11 +97,10 @@ compat_user_backtrace(struct compat_frame_tail __user *tail, + * Frame pointers should strictly progress back up the stack + * (towards higher addresses). + */ +- if (tail + 1 >= (struct compat_frame_tail __user *) +- compat_ptr(buftail.fp)) ++ if ((u32)tail + 4 >= buftail.fp) + return NULL; + +- return (struct compat_frame_tail __user *)compat_ptr(buftail.fp) - 1; ++ return (struct compat_frame_tail __user *)(buftail.fp - 4); + } + #endif /* CONFIG_COMPAT */ + +@@ -123,7 +128,7 @@ void perf_callchain_user(struct perf_callchain_entry_ctx *entry, + /* AARCH32 compat mode */ + struct compat_frame_tail __user *tail; + +- tail = (struct compat_frame_tail __user *)regs->compat_fp - 1; ++ tail = (struct compat_frame_tail __user *)(regs->compat_fp - 4); + + while ((entry->nr < entry->max_stack) && + tail && !((unsigned long)tail & 0x3)) +-- +2.17.1 + diff --git a/patches/misc/misc.scc b/patches/misc/misc.scc index dc091c36..910b4d36 100644 --- a/patches/misc/misc.scc +++ b/patches/misc/misc.scc @@ -8,3 +8,4 @@ patch uvesafb-provide-option-to-specify-timeout-for-task-c.patch patch linux-yocto-Handle-bin-awk-issues.patch patch modpost-srcversion-sometimes-incorrect.patch patch x86-boot-compressed-64-Define-__force_order-only-whe.patch +patch arm64-perf-fix-backtrace-for-AAPCS-with-FP-enabled.patch -- 2.18.2
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#8624): https://lists.yoctoproject.org/g/linux-yocto/message/8624 Mute This Topic: https://lists.yoctoproject.org/mt/73303101/21656 Group Owner: linux-yocto+ow...@lists.yoctoproject.org Unsubscribe: https://lists.yoctoproject.org/g/linux-yocto/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-