On Tue, Sep 15, 2020 at 4:40 PM Andrew Cooper andrew.cooper3@citrix.com wrote:
On 16/09/2020 00:11, Andy Lutomirski wrote:
On Sep 15, 2020, at 2:24 PM, Nick Desaulniers ndesaulniers@google.com wrote:
On Tue, Sep 15, 2020 at 1:56 PM Andy Lutomirski luto@kernel.org wrote:
The old smap_save() code was:
pushf pop %0
with %0 defined by an "=rm" constraint. This is fine if the compiler picked the register option, but it was incorrect with an %rsp-relative memory operand.
It is incorrect because ... (I think mentioning the point about the red zone would be good, unless there were additional concerns?)
This isn’t a red zone issue — it’s a just-plain-wrong issue. The popf is storing the result in the wrong place in memory — it’s RSP-relative, but RSP is whatever the compiler thinks it should be minus 8, because the compiler doesn’t know that pushfq changed RSP.
It's worse than that. Even when stating that %rsp is modified in the asm, the generated code sequence is still buggy, for recent Clang and GCC.
It's clearly not safe to ever use memory operands with pushf/popf asm fragments.
Would this apply to native_save_fl() and native_restore_fl in arch/x86/include/asm/irqflags.h? It was like that two revisions ago, but it was changed (back) to "=rm" with a comment about it being safe.
This is something we should fix. Bill, James, and I are discussing this internally. Thank you for filing a bug; I owe you a beer just for that.
I’m looking forward to the day that beers can be exchanged in person again :)
+1 to that.
+100
-bw