[PATCH 3/3] malloc: Remove currently broken memory tagging

[Adhemerval Zanella Netto]

On 08/05/26 13:29, Florian Weimer wrote:
> * Yury Khrustalev:
> 
>> Have you run it on a system that has HWCAP2_MTE? One thing that doesn't
>> work is that calling free(ptr) results in a SIGSEGV if ptr is tagged and
>> memory it points to is also tagged. But of course for this to happen the
>> system should have HWCAP2_MTE, and Glibc (or your test) should do prctl
>> syscall with PR_SET_TAGGED_ADDR_CTRL and then malloc should use mmap with
>> PROT_MTE.
> 
> The HWCAP2 bit was set, and the prctl call succeeded.  It was definitely
> was not easy to get this point, though.
> 
>> If you manage to enable memory tagging in malloc correctly, a lot of
>> malloc tests from the Glibc testsuite fail (due to similar segfaults
>> resulting from tag mismatch).
> 
> I went back to my old notes, and found this (regarding test failures in
> particular):
> 
> | There's some issue with the static startup code, and ldconfig from our
> | RPM builds crashes (seems compiler flags dependent).  The glibc debug
> | malloc is completely broken because it calls mem2chunk on null
> | pointers, for free(NULL) and realloc(NULL, size).  That's not valid in
> | MTE mode.  Rebooting with MTE forced on in glibc (for dynamically
> | linked binaries) resulted in a crash in lvm.  But it was not as bad as
> | I expected: podman and Go seem to be working.

Was podman podmad built with gccgo or golang? If the latter it does not
give us much information about memory tagging status since afaik golang
binaries do not use libc malloc.

> 
> I think this was with glibc-2.39-derived code, so a bit different from
> the malloc in the current tree.
> 
> Thanks,
> Florian
>