Re: [PATCH] Update auxv cache when there is no auxv cached data

[John Baldwin <jhb@FreeBSD.org>]

On 7/19/22 7:45 AM, Luis Machado via Gdb-patches wrote:
> While adding support for MTE corefiles and running the MTE corefile tests,
> I noticed a strange situation where loading the symbol file + core file
> through the command line has a different behavior compared to firing up
> GDB, loading the symbol file with the "file" command and then loading the
> core file with the "core" command.
> 
> I tracked this down to gdb/auxv.c:get_auxv_inferior_data returning a valid
> info struct, but with no auxv data.
> 
> We've been doing the auxv caching for a while now, but sometime between
> enabling auxv data caching and now, we turned the auxv data into an optional
> value.

The commit to use gdb::optional<> was this one in 2018:

commit 9018be22e022e6db2ba07c4e407c7244022bc69a
Author: Simon Marchi <simon.marchi@ericsson.com>
Date:   Sat Apr 7 13:19:12 2018 -0400

     Make target_read_alloc & al return vectors
     
     This patch started by changing target_read_alloc_1 to return a
     byte_vector, to avoid manual memory management (in target_read_alloc_1
     and in the callers).  To communicate failures to the callers, it
     actually returns a gdb::optional<gdb::byte_vector>.
     
     Adjusting target_read_stralloc was a bit more tricky, since it wants to
     return a buffer of char, and not gdb_byte.  Since you can't just cast a
     gdb::byte_vector into a gdb::def_vector<char>, I made
     target_read_alloc_1 templated, so both versions (that return vectors of
     gdb_byte and char) are generated.  Since target_read_stralloc now
     returns a gdb::char_vector instead of a gdb::unique_xmalloc_ptr<char>, a
     few callers need to be adjusted.

But looking at it's diff, it was always caching negative results even
before this change.  This is the relevant hunk of that commit:

@@ -358,9 +356,8 @@ get_auxv_inferior_data (struct target_ops *ops)
    info = (struct auxv_info *) inferior_data (inf, auxv_inferior_data);
    if (info == NULL)
      {
-      info = XCNEW (struct auxv_info);
-      info->length = target_read_alloc (ops, TARGET_OBJECT_AUXV,
-					NULL, &info->data);
+      info = new auxv_info;
+      info->data = target_read_alloc (ops, TARGET_OBJECT_AUXV, NULL);
        set_inferior_data (inf, auxv_inferior_data, info);
      }
  
So even before it seems that we would have cached a "negative" result.

I believe your change will mean that we will keep asking over and over
for the auxv data if a target doesn't support it and that it will only
support positive caching.  I wonder if a recent change means that in
your test case something is trying to fetch the AUXV vector before the
target can read it, e.g. trying to fetch AT_HWCAP* when determining the
target description for a core file or the like?  It seems like for
the core target in corelow.c it should be fine to fetch AUXV data in
the gdbarch_core_read_description hook though.  Did you try setting a
breakpoint for when this function was called to see when it is being
cached?

I wonder if the issue is that parsing the symbol file without a core
is trying to fetch hwcap actually.  Presumably a breakpoint would let
you see that case?  Maybe the fetch of hwcap needs to be guarded by
something like "target_has_execution" or the like?

-- 
John Baldwin