Re: [PATCH] gdb/amdgpu: Fix accessing GPU global memory

[Simon Marchi <simark@simark.ca>]

On 2026-09-07 18:15, Pedro Alves wrote:
> Now that multi solib provider support is in, the gdb.rocm/ tests
> should work on Windows, in theory.  Turns out they still don't.  GDB
> is now able to discover GPU code objects, but still fails to stop at
> GPU breakpoints.
> 
> The reason is that accessing GPU global memory is not working on
> Windows.  It fails with STATUS_ERROR_INVALID_LANE_ID, like so:
> 
>   [amd-dbgapi-lib] amd_dbgapi_code_object_get_info (code_object_id=code_object_1 <"memory://5268#offset=0x90e9850&size=47824">, query=CODE_OBJECT_INFO_URI_NAME, value_size=
> 8, value=0x73739feee0) {
>   [amd-dbgapi-lib]    callback: allocate_memory (42) {
>   [amd-dbgapi-lib]    callback: } allocate_memory = 0x1f82fdcc110 (took 0.011000 ms)
>   [amd-dbgapi-lib] } amd_dbgapi_code_object_get_info = STATUS_SUCCESS, *value="memory://5268#offset=0x90e9850&size=47824"@000001f82fdcc110 (took 0.633000 ms)
>   [amd-dbgapi-lib] amd_dbgapi_read_memory (process_id=process_1, wave_id=WAVE_NONE, lane_id=0, address_space_id=address_space_1 <global>, segment_address=0x90e9850, value_s
> ize=47824@00000073739fde78, value=0x1f8309246b0) {
>   [amd-dbgapi-lib] } amd_dbgapi_read_memory = STATUS_ERROR_INVALID_LANE_ID (took 0.024000 ms)
> 
> We get STATUS_ERROR_INVALID_LANE_ID because we passed down
> AMD_DBGAPI_WAVE_NONE, and in that case dbgapi enforces that the caller
> pass down AMD_DBGAPI_LANE_NONE for lane.
> 
> From dbgapi/src/memory.cpp:
> 
> xfer_memory (amd_dbgapi_process_id_t process_id, amd_dbgapi_wave_id_t wave_id,
> ...
>   if (wave != nullptr)
>     {
>       ...
>     }
>   else if (wave_id != AMD_DBGAPI_WAVE_NONE)
>     {
>       THROW (AMD_DBGAPI_STATUS_ERROR_INVALID_WAVE_ID);
>     }
>   else if (lane_id != AMD_DBGAPI_LANE_NONE)
>     {
>       /* wave_id is WAVE_NONE and lane_id != LANE_NONE.  */
>       THROW (AMD_DBGAPI_STATUS_ERROR_INVALID_LANE_ID);          <<<<<<<<<< HERE
>     }
> ...
> 
> Actually, all the GPU global accesses currently fail on Linux too!
> The logs show the exact same STATUS_ERROR_INVALID_LANE_ID all over the
> place.  It just so happens to be papered over by accident there --
> when accessing memory via amd-dbgapi-target fails, the core of GDB
> will try the target beneath in the target stack (the process stratum
> layer), and that (linux-nat) will work, because unlike on Windows, on
> Linux we have unified memory so /proc/pid/mem successfully accesses
> the GPU memory.
> 
> Fix this by passing AMD_DBGAPI_LANE_NONE when wave_id is
> AMD_DBGAPI_WAVE_NONE.  We don't have proper lane support yet, so when
> wave_id is a valid wave, continue accessing lane 0.
> 
> Tested on x86_64-pc-linux-gnu and x86_64-pc-windows-msvc (with pending
> testsuite patches).
> 
> gdb.rocm/ tests now work on Windows, and most pass cleanly.
> 
> Change-Id: I7dd53bc0a6ffdfaca71e70772562cc7ac1362f05
> ---
>  gdb/amd-dbgapi-target.c | 8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/gdb/amd-dbgapi-target.c b/gdb/amd-dbgapi-target.c
> index 45ad8f98a92..93207a8fdd2 100644
> --- a/gdb/amd-dbgapi-target.c
> +++ b/gdb/amd-dbgapi-target.c
> @@ -769,16 +769,20 @@ amd_dbgapi_target::xfer_partial (enum target_object object, const char *annex,
>    amd_dbgapi_wave_id_t wave_id = (ptid_is_gpu (inferior_ptid)
>  				  ? get_amd_dbgapi_wave_id (inferior_ptid)
>  				  : AMD_DBGAPI_WAVE_NONE);
> +  /* dbgapi requires LANE_NONE when wave is WAVE_NONE.  */
> +  amd_dbgapi_lane_id_t lane_id = (wave_id != AMD_DBGAPI_WAVE_NONE
> +				  ? 0
> +				  : AMD_DBGAPI_LANE_NONE);
>  
>    size_t len = requested_len;
>    amd_dbgapi_status_t status;
>  
>    if (readbuf != nullptr)
> -    status = amd_dbgapi_read_memory (process_id, wave_id, 0,
> +    status = amd_dbgapi_read_memory (process_id, wave_id, lane_id,
>  				     AMD_DBGAPI_ADDRESS_SPACE_GLOBAL,
>  				     offset, &len, readbuf);
>    else
> -    status = amd_dbgapi_write_memory (process_id, wave_id, 0,
> +    status = amd_dbgapi_write_memory (process_id, wave_id, lane_id,
>  				      AMD_DBGAPI_ADDRESS_SPACE_GLOBAL,
>  				      offset, &len, writebuf);

I don't know if you want to wait for Lancelot's approval, but fwiw it
LGTM, and trivial enough.

Approved-By: Simon Marchi <simon.marchi@efficios.com>

Simon