|
amd_dbgapi_target_breakpoint::check_status
ROCgdb handles target events very slowly when running a test case like
this, where a breakpoint is preset on HipTest::vectorADD:
for (int i=0; i < numDevices; ++i) {
HIPCHECK(hipSetDevice(i));
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, stream[i],
static_cast<const int*>(A_d[i]), static_cast<const int*>(B_d[i]), C_d[i], N);
}
What happens is:
- A kernel is launched
- The internal runtime breakpoint is hit during the second
hipLaunchKernelGGL call, which causes
amd_dbgapi_target_breakpoint::check_status to be called
- Meanwhile, all waves of the kernel hit the breakpoint on vectorADD
- amd_dbgapi_target_breakpoint::check_status calls process_event_queue,
which pulls the thousand of breakpoint hit events from the kernel
- As part of handling the breakpoint hit events, we write the PC of the
waves that stopped to decrement it. Because the forward progress
requirement is not disabled, this causes a suspend/resume of the
queue each time, which is time-consuming.
The stack trace where this all happens is:
#32 0x00007ffff6b9abda in amd_dbgapi_write_register (wave_id=..., register_id=..., offset=0, value_size=8, value=0x7fffea9fdcc0) at /home/smarchi/src/amd-dbgapi/src/register.cpp:587
#33 0x00005555588c0bed in amd_dbgapi_target::store_registers (this=0x55555c7b1d20 <the_amd_dbgapi_target>, regcache=0x507000002240, regno=470) at /home/smarchi/src/wt/amd/gdb/amd-dbgapi-target.c:2504
#34 0x000055555a5186a1 in target_store_registers (regcache=0x507000002240, regno=470) at /home/smarchi/src/wt/amd/gdb/target.c:3973
#35 0x0000555559fab831 in regcache::raw_write (this=0x507000002240, regnum=470, src=...) at /home/smarchi/src/wt/amd/gdb/regcache.c:890
#36 0x0000555559fabd2b in regcache::cooked_write (this=0x507000002240, regnum=470, src=...) at /home/smarchi/src/wt/amd/gdb/regcache.c:915
#37 0x0000555559fc3ca5 in regcache::cooked_write<unsigned long, void> (this=0x507000002240, regnum=470, val=140737323456768) at /home/smarchi/src/wt/amd/gdb/regcache.c:850
#38 0x0000555559fab09a in regcache_cooked_write_unsigned (regcache=0x507000002240, regnum=470, val=140737323456768) at /home/smarchi/src/wt/amd/gdb/regcache.c:858
#39 0x0000555559fb0678 in regcache_write_pc (regcache=0x507000002240, pc=0x7ffff62bd900) at /home/smarchi/src/wt/amd/gdb/regcache.c:1460
#40 0x00005555588bb37d in process_one_event (event_id=..., event_kind=AMD_DBGAPI_EVENT_KIND_WAVE_STOP) at /home/smarchi/src/wt/amd/gdb/amd-dbgapi-target.c:1873
#41 0x00005555588bbf7b in process_event_queue (process_id=..., until_event_kind=AMD_DBGAPI_EVENT_KIND_BREAKPOINT_RESUME) at /home/smarchi/src/wt/amd/gdb/amd-dbgapi-target.c:2006
#42 0x00005555588b1aca in amd_dbgapi_target_breakpoint::check_status (this=0x511000140900, bs=0x50600014ed00) at /home/smarchi/src/wt/amd/gdb/amd-dbgapi-target.c:890
#43 0x0000555558c50080 in bpstat_stop_status (aspace=0x5070000061b0, bp_addr=0x7fffed0b9ab0, thread=0x518000026c80, ws=..., stop_chain=0x50600014ed00) at /home/smarchi/src/wt/amd/gdb/breakpoint.c:6126
#44 0x000055555984f4ff in handle_signal_stop (ecs=0x7fffeaa40ef0) at /home/smarchi/src/wt/amd/gdb/infrun.c:7169
#45 0x000055555984b889 in handle_inferior_event (ecs=0x7fffeaa40ef0) at /home/smarchi/src/wt/amd/gdb/infrun.c:6621
#46 0x000055555983eab6 in fetch_inferior_event () at /home/smarchi/src/wt/amd/gdb/infrun.c:4750
#47 0x00005555597caa5f in inferior_event_handler (event_type=INF_REG_EVENT) at /home/smarchi/src/wt/amd/gdb/inf-loop.c:42
#48 0x00005555588b838e in handle_target_event (client_data=0x0) at /home/smarchi/src/wt/amd/gdb/amd-dbgapi-target.c:1513
Fix that performance problem by disabling the forward progress
requirement in amd_dbgapi_target_breakpoint::check_status, before
calling process_event_queue, so that we can process all events
efficiently.
Since the same performance problem could theoritically happen any time
process_event_queue is called with forward progress requirement enabled,
add an assert to ensure that forward progress requirement is disabled
when process_event_queue is invoked. This makes it necessary to add a
require_forward_progress call to amd_dbgapi_finalize_core_attach. It
looks a bit strange, since core files don't have execution, but it
doesn't hurt.
Add a test that replicates this scenario. The test launches a kernel
that hits a breakpoint (with an always false condition) repeatedly.
Meanwhile, the host process loads an unloads a code object, causing
check_status to be called.
Bug: SWDEV-482511
Change-Id: Ida86340d679e6bd8462712953458c07ba3fd49ec
Approved-by: Lancelot Six <lancelot.six@amd.com>
|
