Re: [RFC PATCH 3/5] arm64: aarch64-insn: Add barrier encodings

From: Julien Thierry
Date: Wed Feb 03 2021 - 03:48:54 EST




On 2/2/21 12:15 PM, Mark Rutland wrote:
On Wed, Jan 20, 2021 at 06:17:43PM +0100, Julien Thierry wrote:
Create necessary functions to encode/decode aarch64 data/instruction
barriers.

Signed-off-by: Julien Thierry <jthierry@xxxxxxxxxx>
---
arch/arm64/include/asm/aarch64-insn.h | 9 +++++++++
1 file changed, 9 insertions(+)

diff --git a/arch/arm64/include/asm/aarch64-insn.h b/arch/arm64/include/asm/aarch64-insn.h
index 200bee726172..d0fee47bbe6e 100644
--- a/arch/arm64/include/asm/aarch64-insn.h
+++ b/arch/arm64/include/asm/aarch64-insn.h
@@ -379,6 +379,9 @@ __AARCH64_INSN_FUNCS(eret_auth, 0xFFFFFBFF, 0xD69F0BFF)
__AARCH64_INSN_FUNCS(mrs, 0xFFF00000, 0xD5300000)
__AARCH64_INSN_FUNCS(msr_imm, 0xFFF8F01F, 0xD500401F)
__AARCH64_INSN_FUNCS(msr_reg, 0xFFF00000, 0xD5100000)
+__AARCH64_INSN_FUNCS(dmb, 0xFFFFF0FF, 0xD50330BF)
+__AARCH64_INSN_FUNCS(dsb, 0xFFFFF0FF, 0xD503309F)
+__AARCH64_INSN_FUNCS(isb, 0xFFFFF0FF, 0xD50330DF)

These match the encodings in ARM DDI 0487G.a, with a couple of caveats
for DSB.

Per section C6.2.82 on page C6-1000, when CRm != 0x00, the instruction
isn't considered a DSB. I believe per the "barriers" decode table on
page C4-289 that here "0x00" is actually a binary string and 'x' is a
"don't care" -- I've raised a ticket to get the documentation clarified.
I suspect we need to write a function to handle that.


Ah, I did miss that part. Thanks for pointing it out (and for clarifying it's probably not hexa, but the binary string makes sense since it's a 4 bits field)

There's also a secondary encoding for DSB with FEAT_XS, which we don't
currently use but might want to add.


Ah, yes, had to pick up a newer version of the Arm ARM! I'll add it.

#undef __AARCH64_INSN_FUNCS
@@ -390,6 +393,12 @@ static inline bool aarch64_insn_is_adr_adrp(u32 insn)
return aarch64_insn_is_adr(insn) || aarch64_insn_is_adrp(insn);
}
+static inline bool aarch64_insn_is_barrier(u32 insn)
+{
+ return aarch64_insn_is_dmb(insn) || aarch64_insn_is_dsb(insn) ||
+ aarch64_insn_is_isb(insn);
+}

I assume this is meant to match the barriers instruction class, as per
the table on page C4-289? That also contains CLREX, SB, SSBB, and PSSBB,
and it might be worth adding them at the same time.


Yes, I have to admit I only added the ones that objtool saw and complained about "unreachable instruction" (mostly barriers after ret/eret). I'll add them as well.

Thanks,

--
Julien Thierry