Repository navigation
Guilty Gear -Strive- (PPSA02257) #18025
Replies: 1 comment
Update: tutorial-load crash, causes ruled outSame build and host as above ( It is memory corruption, and timing changes where it showsWith the same input script (Cross at 90, 120, 150, 180, 210 s), the tutorial load fails about 15 s after the last press, but the symptom varies:
With At the Ruled out (measured)
Still openFile bytes match the disk when read (31,409 reads checked earlier), and the title decompresses them itself, so something changes data inside the process between the read and the parse. The next step is probably hardware write-watching on the buffer being parsed. I haven't done that. AI-assisted report. Raw logs and game data stay private. 🤖 Generated with Claude Code |
Uh oh!
There was an error while loading. Please reload this page.
Status: Menus
Guilty Gear -Strive-, PPSA02257, content version
01.016.000, my own dump.The title reaches the main menu at 60 fps. Choosing Yes on the first-launch tutorial prompt crashes it about 14 s later on
FAsyncLoadingThread.Build and host
main62653c8 with feat(Party): export the Party module's functions #15945 (Party) merged on top. The title doesn't importParty, so this is effectivelymain.Selected GPU name=NVIDIA GeForce RTX 3080), driver 610.88, Windows 11 Pro 10.0.26300, WinLibs GCC 15.2.0 r7, Release.tools/audit_dump.py --windows: 2926 unique imports, none absent, no stubs, 93 partial.PlayGo: this dump needs
APS5_PLAYGO_UNLISTED_CHUNKS=absentThe dump has no
playgo-chunk.datorplaygo-scenario.json. All data is inpakchunk0-ps5.pak(pluspakchunk0optional-ps5.pak).installed:scePlayGoGetLocusis called 245,290,326 times in about 30 s (APS5_TRACE_CALLS, executable converted with--trace-imports). The title then claims all of direct memory in halving steps (6912, 3456, 1728 … MiB) and stops with Unreal's ownLowLevelFatalError … Ran out of memory allocating 5565054976 bytes. WithAPS5_DIRECT_MEMORY_FROM_FLEXIBLEit claims 12 GiB instead, butUsedPhysical(9278521344) and the failing request are unchanged. My reading, not measured on a console: the title probes chunk IDs until it getsSCE_PLAYGO_ERROR_BAD_CHUNK_ID, andinstallednever returns it.absent: start-up is normal.How far it gets (with
absent)APS5_DUMP_FRAMES).FATAL: unhandled exception 0xc0000005 on thread … 'FAsyncLoadingThread'.The crash
Reproduced in 7 runs; the symptom varies but the place doesn't:
0x11613e7a4e,0x11613e694e(fromlibc.prx.guest.prx+0x20326),0x12e195f239(from title code);LowLevelFatalError … Ran out of memory allocating 548682072064 bytes, identical in 3 runs. That is0x7FC0000000, exactly the size of the range the title prints at start-up (Memory va range 1000000000 - 8fc0000000).Frame chain of the write (guest addresses):
0x1d64997←0x1e13871←0x1e0f265←0x1e137c3←0x428d1c6←0x1d02569←0x1d00e51←0x1d00931←0x1d05d99←0x1d0ac5f←0x1d0ba85←0x1ad672d←0x1a4427e. The fatal error goes through the same callers (0x1d64b94←0x1e1386d←0x1e0f261← …).0x1d64992is amemset(base + offset, 0, size)in a loop over 16-byte entries of a lazily built, per-type schema cached at+0xa8of the type object. It is steered by a byte header read from the serialized data (counts at+0,+2,+4,+6). That looks like Unreal's unversioned property serialization zeroing the properties the header marks as zero. A wrongoffset/sizethere means the serialized header and the in-memory schema disagree.Steps to reproduce: start with
APS5_PLAYGO_UNLISTED_CHUNKS=absent, then press Cross at about 90, 120, 150, 180 and 210 s (tools/pad_input.py). The fifth press answers Yes to the tutorial prompt; the crash follows about 14 s later.Ruled out (measured)
sceAmprAprCommandBufferReadFile→sceKernelAprSubmitCommandBufferAndGetResult→sceKernelAprWaitCommandBuffer, completion viasceAmprCommandBufferWriteKernelEventQueueOnCompletion). I checksummed every read (APS5_TRACE_APRplus a local FNV probe, since removed) against the bytes on disk: 31,409 reads, 18,758 of them past 4 GiB. All match except 3 large reads (2–4 MiB) whose destination changed between the copy and the checksum, all in the last seconds before the crash.sceKernelBatchMap(12,681 calls). A local probe, since removed, logged no failed batch. The faulting addresses were never covered by any traced map (APS5_TRACE_MEMORY).APS5_GPU_NO_WRITEBACK=1gives the same crash and the same 548682072064-byte request.sceKernelAddAmprEventkeeps up to 64 pending completions per event (kMaxPendingEvents). A local probe, since removed, saw at most 31 pending and nothing dropped up to the crash.Not found
Which AnyPS5 answer makes the header and the schema disagree. The next step is probably to watch the header bytes and the schema entries for the failing type around
0x1d648b5, and compare them with the serialized bytes at that pak offset.AI-assisted report. Raw logs, captures and game data stay private.
🤖 Generated with Claude Code
All reactions