This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
Re: [PATCH] Aarch64 SVE: Fix stack smashing when calling functions
- From: Alan Hayward <Alan dot Hayward at arm dot com>
- To: Joel Brobecker <brobecker at adacore dot com>
- Cc: "gdb-patches at sourceware dot org" <gdb-patches at sourceware dot org>, nd <nd at arm dot com>
- Date: Tue, 18 Sep 2018 09:08:39 +0000
- Subject: Re: [PATCH] Aarch64 SVE: Fix stack smashing when calling functions
- References: <20180917144540.78906-1-alan.hayward@arm.com> <20180917184205.GG19172@adacore.com>
> On 17 Sep 2018, at 19:42, Joel Brobecker <brobecker@adacore.com> wrote:
>
> Hi Alan,
>
>> Using "call" on a function that passes arguments via float registers can cause
>> gdb to overflow buffers.
>
>> Ensure enough memory is reserved to hold a full FP register.
>>
>> 2018-09-17 Alan Hayward <alan.hayward@arm.com>
>>
>> * aarch64-tdep.c (pass_in_v): Use register size.
>> (aarch64_extract_return_value): Likewise.
>> (aarch64_store_return_value): Likewise.
>
> Do we have a testcase already that demonstrates the problem?
> Otherwise, it would be nice to add one.
I should have mentioned that in the description. I can add:
“This fixes gdb.base/callfuncs.exp for Aarch64 SVE."
>
>> diff --git a/gdb/aarch64-tdep.c b/gdb/aarch64-tdep.c
>> index 6993e9061e..516eb138dc 100644
>> --- a/gdb/aarch64-tdep.c
>> +++ b/gdb/aarch64-tdep.c
>> @@ -1358,7 +1358,10 @@ pass_in_v (struct gdbarch *gdbarch,
>> if (info->nsrn < 8)
>> {
>> int regnum = AARCH64_V0_REGNUM + info->nsrn;
>> - gdb_byte reg[V_REGISTER_SIZE];
>> + /* Enough space for a full vector register. */
>> + gdb_byte reg[register_size (gdbarch, regnum)];
>> + gdb_static_assert (AARCH64_V0_REGNUM == AARCH64_SVE_Z0_REGNUM);
>> + gdb_assert (len <= sizeof (reg));
>
> Could you explain the relationship between making the buffer large
> enough, which is the purpose of this patch, and the assertion that
> AARCH64_V0_REGNUM == AARCH64_SVE_Z0_REGNUM?
>
> I don't see a problem with that assertion, but for archeology
> purposes, it is better to decorelate changes that are independent.
> It helps better document why we introduced changes.
Thinking about it, that assert should be removed.
I was reusing three lines from aarch64_pseudo_read_value_1, which
passed AARCH64_V0_REGNUM into register_size. There the assert made
sense. When I switched to use regnum I didn’t rethink the assert.
I’ll remove the three new instances of that assert from this patch.
Are you happy with those those changes?
Thanks,
Alan.