Re: [RFC][gdb/symtab] Lazy expansion of full symbol table

[Tom de Vries <tdevries@suse.de>]

On 6/21/21 1:44 AM, Tom Tromey wrote:
> Tom> Tried the updated branch and ran into a race condition, fixed in
> Tom> attached patch.
> 
> How did you observe it?
> 

Download and extract
https://download.opensuse.org/tumbleweed/repo/debug/x86_64/glibc-debuginfo-2.33-8.1.x86_64.rpm
.

Then:
...
$ gdb -q -batch
~/reproduce/usr/lib/debug/usr/lib64/ld-2.33.so-2.33-8.1.x86_64.debug
Dwarf Error: DW_FORM_line_strp used without .debug_line_str section [in
module
/home/vries/reproduce/usr/lib/debug/usr/lib64/ld-2.33.so-2.33-8.1.x86_64.debug]
...

> Tom> +  static std::mutex mutex;
> Tom>    asection *sectp;
> Tom>    bfd *abfd;
> Tom>    gdb_byte *buf, *retbuf;
>  
> Tom> +  std::lock_guard<std::mutex> guard (mutex);
> Tom>    if (readin)
> Tom>      return;
> Tom>    buffer = NULL;
> 
> I don't think we want to do this, because BFD isn't thread-safe, and
> I've been avoiding touching it.  Instead I think we'll have to track
> down the problems spots and arrange for the required sections to be read
> in before any threaded work is started.
> 

Ack, I see.

---

I rebuild the updated branch and ran the tests and investigated this
dwarf error:
...
(gdb) file
/home/vries/gdb_versions/devel/build/gdb/testsuite/outputs/gdb.arch/amd64-disp-step-avx/amd64-disp-step-avx^M
Reading symbols from
/home/vries/gdb_versions/devel/build/gdb/testsuite/outputs/gdb.arch/amd64-disp-step-avx/amd64-disp-step-avx...^M
Dwarf Error: Could not find abbrev number 2391 in CU at offset 0x29 [in
module
/home/vries/gdb_versions/devel/build/gdb/testsuite/outputs/gdb.arch/amd64-disp-step-avx/amd64-disp-step-avx]^M
...
when using gcc-11.

I managed to fix that using this patch:
...
diff --git a/gdb/dwarf2/read.c b/gdb/dwarf2/read.c
index 265ebee1de8..f98acfa33ed 100644
--- a/gdb/dwarf2/read.c
+++ b/gdb/dwarf2/read.c
@@ -7204,7 +7204,8 @@ skip_one_die (const struct die_reader_specs
*reader, const gdb_b
yte *info_ptr,
       /* We only handle DW_FORM_ref4 here.  */
       const gdb_byte *sibling_data = info_ptr + abbrev->sibling_offset;
       unsigned int offset = read_4_bytes (abfd, sibling_data);
-      const gdb_byte *sibling_ptr = buffer + offset;
+      const gdb_byte *sibling_ptr
+       = buffer + to_underlying (cu->header.sect_off) + offset;
       if (sibling_ptr >= info_ptr && sibling_ptr < reader->buffer_end)
        return sibling_ptr;
       /* Fall through to the slow way.  */
...

The DW_FORM_ref4 offset is relative to the start of the CU, but buffer
points to the start of .debug_info.

Thanks,
- Tom