RE: [PATCH v2 1/3] gdb: Make tagged pointer support configurable.

["Schimpe, Christina" <christina.schimpe@intel.com>]

> -----Original Message-----
> From: Luis Machado <luis.machado@arm.com>
> Sent: Monday, June 10, 2024 4:01 PM
> To: Schimpe, Christina <christina.schimpe@intel.com>; Willgerodt, Felix
> <felix.willgerodt@intel.com>; gdb-patches@sourceware.org
> Cc: eliz@gnu.org
> Subject: Re: [PATCH v2 1/3] gdb: Make tagged pointer support configurable.
> 
> On 6/3/24 15:13, Schimpe, Christina wrote:
> > Hi Luis,
> >
> >>>>> -/* AArch64 implementation of the remove_non_address_bits gdbarch
> >>>> hook.
> >>>>> Remove
> >>>>> -   non address bits from a pointer value.  */
> >>>>> -
> >>>>> -static CORE_ADDR
> >>>>> +CORE_ADDR
> >>>>>  aarch64_remove_non_address_bits (struct gdbarch *gdbarch,
> >> CORE_ADDR
> >>>>> pointer)
> >>>>
> >>>> Shouldn't there still be some sort of comment for this function in
> >>>> the c
> >> file?
> >>>> At least some "see header file"? Though it seems like no function
> >>>> in
> >>>> aarch64- tdep.h has any comment, and all comments are in
> >>>> aarch64-tdep.c. I would also be fine with that for consistency.
> >>>> Maybe Luis can comment how he prefers it.
> >>>>
> >>>> But I am fine with this patch. Let's see if someone else objects this split.
> >>>>
> >>>> Reviewed-By: Felix Willgerodt <felix.willgerodt@intel.com>
> >>>>
> >>>> Thanks,
> >>>> Felix
> >>>
> >>> Ok, let's wait for further comments here.
> >>
> >> Sorry, only spotted this now.
> >>
> >> The comment on the above function would be a repetition of the hook
> >> explanation, and the gdbarch hook explanation covers all the details
> >> we need. But if the comment is being removed due to it being in the
> >> header, we should point to the header instead.
> >>
> >> "See aarch64-tdep.h" should do.
> >
> > Ok, I will add that line.
> >
> >> There is something odd about this patch though. The aarch64 code is
> >> being changed to call aarch64_remove_non_address_bits directly, but 3
> >> new hooks are being set:
> >>
> >> +  set_gdbarch_remove_non_address_bits_watchpoint
> >> +    (gdbarch, aarch64_remove_non_address_bits);
> >> + set_gdbarch_remove_non_address_bits_breakpoint
> >> +    (gdbarch, aarch64_remove_non_address_bits);
> >> + set_gdbarch_remove_non_address_bits_memory
> >> +    (gdbarch, aarch64_remove_non_address_bits);
> >>
> >> But the above hooks (the gdbarch_remove_non_address_bits_memory at
> >> least) never get used in aarch64 code. Is there a reason for that?
> >
> > I thought it's wrong to call the gdbarch_remove_non_address_bits_memory
> hook in aarch64 code.
> > The reason is described here:
> > https://sourceware.org/pipermail/gdb-patches/2024-May/209420.html
> >
> > "Also I noticed that the gdbarch function
> 'gdbarch_remove_non_addr_bits_memory'
> > should not be used in several functions of aarch64* files (patch 1)
> > due to the following:
> >
> > The function description for 'gdbarch_remove_non_addr_bits_memory'
> states:
> >
> > "Given a pointer for the architecture, this hook removes all the
> > non-significant bits and sign-extends things as needed. It gets used
> > to remove non-address bits from any pointer used to access memory. "
> >
> > We don't know if the pointer passed to
> 'gdbarch_remove_non_addr_bits_memory'
> > by 'gdbarch_tagged_address_p' will be used to access memory.  That's
> > why it's wrong to call this function in 'gdbarch_tagged_address_p'.
> >
> > There are several similar wrong calls of
> 'gdbarch_remove_non_addr_bits_memory'
> > in the changes of patch 1.  I replaced all of them with
> > 'aarch64_remove_non_address_bits'."
> >
> 
> That's fine, but previously the hook was correct for what is was being used for
> in aarch64. This particular change means we no longer have a hook that
> aarch64 can use for the same purposes it used before. This part feels a bit odd.
> It feels like the memory-specific one would be appropriate, but it isn't clear.

Hm, I thought it is not a problem to remove the hook in the aarch64 files, as it's target
specific code. 

> Also, from the 1/3 patch submission:
> 
> "- In contrast to the implementation for ARM's TBI, the Linux kernel supports
>   tagged pointers for memory access."
> 
> Is there a negative impact if we feed the kernel an untagged pointer for a
> particular memory access? If not, it sounds like it might be simpler (from gdb's
> perspective) to let gdb always unmask things, no? Of course, there might be a
> performance impact of checking if a region is tagged.
> I'm just trying to understand if the additional hooks are really needed/useful.
> 
> Right now it looks as if we could make it work with the existing hook in place,
> instead of splitting it into 3 different hooks.

I think in case of amd64 there is a negative impact.  In contrast to ARM's TBI,
LAM can be enabled at runtime.  So we would have to parse the /proc/<pid>/status
file every time when the pointer is used to access memory. With the 3 hooks I
wanted to avoid this.

I tried to make the purpose more clear by adding the two sentences in the commit
message for v2:

"This way, one can make sure that addresses are only adjusted when necessary. 
In case of LAM, this avoids unnecessary parsing of the /proc/<pid>/status file to
get the untag mask."

So yes, it does work with the existing hook, but when it comes to performance, I
think 3 hooks would be better.

Christina
Intel Deutschland GmbH
Registered Address: Am Campeon 10, 85579 Neubiberg, Germany
Tel: +49 89 99 8853-0, www.intel.de
Managing Directors: Sean Fennelly, Jeffrey Schneiderman, Tiffany Doon Silva
Chairperson of the Supervisory Board: Nicole Lau
Registered Office: Munich
Commercial Register: Amtsgericht Muenchen HRB 186928