Quoting Gwan-gyeong Mun (2026-11-13 21:09:04)
> Hi Mika,
> the description of the "ttm_bo_access()" API states, "Should not be 
> directly exported to user space via a peak/poke interface."
> 
> When the newly added "ttm_bo_access_locked()" is used in "eudebug", it 
> is ultimately exposed to user space via functions such as "pread" and 
> "pwrite". @Mattew Brost,I would like to inquire whether this violates 
> the considerations made when the "ttm_bo_access()" API was initially 
> designed.

This was already discussed during earlier series review.

Here, the application itself doesn't get a peek/poke interface, it's
only the debugger process. The debugger process is also given heavy
hardware flushing operations that are needed to flush out eg.
instruction caches.

This interface is pretty much mandatory to facilitate memory access
for debugger on our HW as the RCS/CCS/BCS engines may be stalled due
to a later page-fault in the queue while dealing with first page-fault.

In that scenarion, submitting workloads to the engines to facilitate
single word modifications to memory would simply be blocked.

It'd also fall off performance cliff for single-stepping as GDB at
least on the Intel GPU instruction set is very heavy on sparsely
accessing a lot of memory during single-stepping. We already had to
separate memory access and memory flushing to different operations
to make that more user-friendly for interactive single-stepping.

Regards, Joonas

Reply via email to