On Fri, Oct 27, 2023 at 11:22:42AM -0400, Neal Gompa wrote: > On Fri, Oct 27, 2023 at 6:10 AM Heiko Carstens <hca@xxxxxxxxxxxxx> wrote: > > > > On Fri, Oct 27, 2023 at 10:00:53AM +0200, Daan De Meyer wrote: > > > > > > If the kernel gets support for s390x user space unwinding using the backchain, > > > we'll propose to enable -mbackchain in the default compilation flags for Fedora > > > so that s390x on Fedora will have the same profiling experience as x86-64, arm64 > > > and ppc64. For now we'll keep the status quo since compiling with the backchain > > > doesn't provide any benefit until the kernel unwinder can unwind user > > > space stacks > > > using it. > > > > > > Thanks for clarifying the current state of user space stack unwinding on s390x! > > > > We will implement the missing pieces and let you know when things are > > supposed to work. > > Do you think we could have an initial patch set for implementing the > missing pieces in time for the Linux 6.8 merge window? Then we can > look at enabling this for s390x as a Fedora Linux 40 Change. This will be very likely the case. Actually the plan is to go with the patch below. I gave it some testing with Fedora 38 and replaced (only) glibc with a variant that was compiled with -mbackchain. It seems to work, including the not avoidable problems described below. So I guess we will try to bring this upstream with the current merge window. >From 3ba1e255c8af61f172a8b31de2b76d39c7d4c0c7 Mon Sep 17 00:00:00 2001 From: Heiko Carstens <hca@xxxxxxxxxxxxx> Date: Mon, 30 Oct 2023 11:34:07 +0100 Subject: [PATCH] s390/perf: implement perf_callchain_user() Daan De Meyer and Neal Gompa reported that s390 does not support perf user stack unwinding. This was never implemented since this requires user space to be compiled with the -mbackchain compile option, which until now no distribution did. However this is going to change with Fedora. Therefore provide a perf_callchain_user() implementation. Note that due to the way s390 sets up stack frames the provided call chains can contain invalid values. This is especially true for the first stack frame, where it is not possible to tell if the return address has been written to the stack already or not. Reported-by: Daan De Meyer <daan.j.demeyer@xxxxxxxxx> Reported-by: Neal Gompa <ngompa@xxxxxxxxxxxxxxxxx> Closes: https://lore.kernel.org/all/CAO8sHcn3+_qrnvp0580aK7jN0Wion5F7KYeBAa4MnCY4mqABPA@xxxxxxxxxxxxxx/ Signed-off-by: Heiko Carstens <hca@xxxxxxxxxxxxx> --- arch/s390/kernel/perf_event.c | 47 +++++++++++++++++++++++++++++++++++ 1 file changed, 47 insertions(+) diff --git a/arch/s390/kernel/perf_event.c b/arch/s390/kernel/perf_event.c index c27321cb0969..fdbe7e1d9eea 100644 --- a/arch/s390/kernel/perf_event.c +++ b/arch/s390/kernel/perf_event.c @@ -212,6 +212,53 @@ void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, } } +struct stack_frame_user { + unsigned long back_chain; + unsigned long empty1[5]; + unsigned long gprs[10]; + unsigned long empty2[4]; +}; + +void perf_callchain_user(struct perf_callchain_entry_ctx *entry, + struct pt_regs *regs) +{ + struct stack_frame_user __user *sf; + unsigned long ip, sp; + bool first = true; + + if (perf_guest_state()) + return; + if (is_compat_task()) + return; + perf_callchain_store(entry, instruction_pointer(regs)); + sf = (void *)user_stack_pointer(regs); + pagefault_disable(); + while (entry->nr < entry->max_stack) { + if (__get_user(sp, &sf->back_chain)) + break; + if (__get_user(ip, &sf->gprs[8])) + break; + if (ip & 0x1) { + /* + * If the instruction address is invalid, and this + * is the first stack frame, assume r14 has not + * been written to the stack yet. Otherwise exit. + */ + if (first && !(regs->gprs[14] & 0x1)) + ip = regs->gprs[14]; + else + break; + } + perf_callchain_store(entry, ip); + /* Sanity check: ABI requires SP to be aligned 8 bytes. */ + if (!sp || sp & 0x7) + break; + sf = (void __user *)sp; + first = false; + } + pagefault_enable(); +} + /* Perf definitions for PMU event attributes in sysfs */ ssize_t cpumf_events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page) -- 2.39.2
