Re: [PATCH 2/4] perf report: Load PE files from debug cache only

[Nicholas Fraser <nfraser@codeweavers.com>]

Sorry, I should have been more clear in the commit message. The use case
you outlined still works even with this patch.

dso__load_bfd_symbols() is called in a loop from dso__load() for a variety
of paths. These are generated by the various DSO_BINARY_TYPEs in the
binary_type_symtab list at the top of util/symbol.c. In each case the
debugfile passed to dso__load_bfd_symbols() is the path to try.

One of those iterations (the first one I believe) passes the original path
as the debugfile. If the file still exists at the original path, this is
the one that ends up being used in case the debugcache was deleted or the
PE file doesn't have a build-id.

A later iteration (BUILD_ID_CACHE) passes debugfile as the file in the
debugcache if it has a build-id. Even if the file was previously loaded at
its original path, (if I understand correctly) this load will override it
so the debugcache file ends up being used.

Nick


On 2021-02-12 7:28 a.m., Arnaldo Carvalho de Melo wrote:
> Em Wed, Feb 10, 2021 at 02:17:38PM -0500, Nicholas Fraser escreveu:
>> dso__load_bfd_symbols() attempts to load a DSO at its original path,
>> then closes it and loads the file in the debug cache. This is incorrect.
>> It should ignore the original file and work with only the debug cache.
>> The original file may have changed or may not even exist, for example if
>> the debug cache has been transferred to another machine via "perf
>> archive".
>>
>> This fix makes it only load the file in the debug cache.
> 
> Well this improves your current use case and only affects PE files, so I
> am applying, but consider a slightly different workflow:
> 
>  1. perf record ./foo.exe
>  2. perf report     # works, finds the file in the ~/.debug cache, as stored
>                     # by 'perf record'
>  3. rm -rf ~/.debug # I need more space
>  4. perf report     # Fails, as it looks only in the ~/.debug cache, that
>                     # was nuked
> 
> 
> So at 4 it should look at the original pathname, and hope for the best.
> 
> All this is moot if we have something like a build-id in PE files,
> where we can look in any order since we'll verify the unique ID to see
> if it is the one we need, right?
> 
> - Arnaldo
>  
>> Signed-off-by: Nicholas Fraser <nfraser@codeweavers.com>
>> ---
>>  tools/perf/util/symbol.c | 8 +-------
>>  1 file changed, 1 insertion(+), 7 deletions(-)
>>
>> diff --git a/tools/perf/util/symbol.c b/tools/perf/util/symbol.c
>> index 64a039cbba1b..aa9ae875b995 100644
>> --- a/tools/perf/util/symbol.c
>> +++ b/tools/perf/util/symbol.c
>> @@ -1569,7 +1569,7 @@ int dso__load_bfd_symbols(struct dso *dso, const char *debugfile)
>>  	u_int i;
>>  	u64 start, len;
>>  
>> -	abfd = bfd_openr(dso->long_name, NULL);
>> +	abfd = bfd_openr(debugfile, NULL);
>>  	if (!abfd)
>>  		return -1;
>>  
>> @@ -1586,12 +1586,6 @@ int dso__load_bfd_symbols(struct dso *dso, const char *debugfile)
>>  	if (section)
>>  		dso->text_offset = section->vma - section->filepos;
>>  
>> -	bfd_close(abfd);
>> -
>> -	abfd = bfd_openr(debugfile, NULL);
>> -	if (!abfd)
>> -		return -1;
>> -
>>  	if (!bfd_check_format(abfd, bfd_object)) {
>>  		pr_debug2("%s: cannot read %s bfd file.\n", __func__,
>>  			  debugfile);
>> -- 
>> 2.30.0
>>
>