arm64: perf: Report the PC value in REGS_ABI_32 mode

commit 8dfe804a4031ca6ba3a3efb2048534249b64f3a5 upstream.

A 32-bit perf querying the registers of a compat task using REGS_ABI_32
will receive zeroes from w15, when it expects to find the PC.

Return the PC value for register dwarf register 15 when returning register
values for a compat task to perf.

Cc: <stable@vger.kernel.org>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Jiping Ma <jiping.ma2@windriver.com>
Link: https://lore.kernel.org/r/1589165527-188401-1-git-send-email-jiping.ma2@windriver.com
[will: Shuffled code and added a comment]
Signed-off-by: Will Deacon <will@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

diff --git a/arch/arm64/kernel/perf_regs.c b/arch/arm64/kernel/perf_regs.c
index 0bbac61..666b225 100644 (file)
--- a/arch/arm64/kernel/perf_regs.c
+++ b/arch/arm64/kernel/perf_regs.c
@@ -15,15 +15,34 @@ u64 perf_reg_value(struct pt_regs *regs, int idx)
                return 0;
 
        /*
-        * Compat (i.e. 32 bit) mode:
-        * - PC has been set in the pt_regs struct in kernel_entry,
-        * - Handle SP and LR here.
+        * Our handling of compat tasks (PERF_SAMPLE_REGS_ABI_32) is weird, but
+        * we're stuck with it for ABI compatability reasons.
+        *
+        * For a 32-bit consumer inspecting a 32-bit task, then it will look at
+        * the first 16 registers (see arch/arm/include/uapi/asm/perf_regs.h).
+        * These correspond directly to a prefix of the registers saved in our
+        * 'struct pt_regs', with the exception of the PC, so we copy that down
+        * (x15 corresponds to SP_hyp in the architecture).
+        *
+        * So far, so good.
+        *
+        * The oddity arises when a 64-bit consumer looks at a 32-bit task and
+        * asks for registers beyond PERF_REG_ARM_MAX. In this case, we return
+        * SP_usr, LR_usr and PC in the positions where the AArch64 SP, LR and
+        * PC registers would normally live. The initial idea was to allow a
+        * 64-bit unwinder to unwind a 32-bit task and, although it's not clear
+        * how well that works in practice, somebody might be relying on it.
+        *
+        * At the time we make a sample, we don't know whether the consumer is
+        * 32-bit or 64-bit, so we have to cater for both possibilities.
         */
        if (compat_user_mode(regs)) {
                if ((u32)idx == PERF_REG_ARM64_SP)
                        return regs->compat_sp;
                if ((u32)idx == PERF_REG_ARM64_LR)
                        return regs->compat_lr;
+               if (idx == 15)
+                       return regs->pc;
        }
 
        if ((u32)idx == PERF_REG_ARM64_SP)