The Pathless (PPSA01826)

[JJenkx]

Existing thread

• I searched this category for the title ID and no thread exists for this title on any platform.

Title

The Pathless (PPSA01826)

Content version tested

02.000.000

AnyPS5 commit

main 38d3c0858a with #8847 (12223bf208).

System tested

Linux 7.2.9 (Arch), Ryzen 9 7950X, Radeon RX 6650 XT, RADV Mesa 26.2.4, GCC, Release.

Furthest point

Boots (stops during engine start-up, before the first frame)

What is shown

• Without a PlayGo setting it stops 2 s in: scePlayGo: sce_sys/playgo-chunk.dat and sce_sys/playgo-scenario.json are missing. That's a property of this dump. With APS5_PLAYGO_UNLISTED_CHUNKS=installed it starts.
• Unreal Engine 4 then starts and stops itself 22 s in, before any frame:

Project file not found: ../../../Pathless/Pathless.uproject
PS5 Out of Memory:
    AvailablePhysical: 4863229952
    UsedPhysical:      9317646336
LowLevelFatalError [File:Unknown] [Line: 186]
Ran out of memory allocating 5565054976 bytes with alignment 65536

• Under GDB the process ends in the title's own libc.prx abort path (int $0x45 on the main thread). No AnyPS5 driver failure is logged.
• The physical total the engine sees (available + used) is 14,180,876,288 bytes. That's 300 MiB below AnyPS5's 13,824 MiB direct memory size. Severed Steel (PPSA05106) stops at the same total. Why the title has used 9.3 GB 22 s into start-up, and asks for 5.5 GB more, isn't investigated. The .uproject message alone may be harmless in a packaged build.

Current blocker

Unreal out-of-memory fatal error during start-up (5.5 GB allocation), cause not investigated.

How it was run

relinker --registry --rpath <build>/core/libs/libs eboot.elf app.elf

Modules in sce_module and prx/ are converted with the title. APS5_PLAYGO_UNLISTED_CHUNKS=installed is required for this dump.

AI-assisted testing and report: yes, Claude Code.

[JJenkx]

Follow-up on the out-of-memory stop, from APS5_TRACE_MEMORY=1 runs on the same build (40 s each):

• The Pathless: it allocates 32 KiB, then 300 MiB, then claims the rest of direct memory in halving blocks: 6912, 3456, 1728, 864, 432 and 108 MiB, then smaller ones. Each block is half the previous size, which fits starting from sceKernelGetDirectMemorySize() (13,824 MiB) and halving whenever a request is refused. It then releases the 300 MiB block. That leaves 13.5 GiB in separate blocks, the largest 6.75 GiB. The fatal request is a single 5,565,054,976-byte allocation, made while the engine is remapping large regions of its own pool. No file activity comes right before it.
• Control, Afterimage (PPSA14053, Unreal Engine 4, runs fine): the identical sequence, the same addresses and block sizes, and the same 300 MiB release.

So the halving claim is ordinary Unreal start-up on this build, and sceKernelAvailableDirectMemorySize reports the largest free range correctly. What differs is that The Pathless asks for one 5.5 GB block that its pool can't hold contiguously. That size isn't a file size: the largest pak, pakchunk0-ps5.pak, is 6,541,948,331 bytes. I couldn't trace where the request comes from without engine symbols. Both traces are in my private evidence and I can attach them if useful.

AI-assisted testing and report: yes, Claude Code.

[EagleBearerYS]

@JJenkx thanks for the traces. To tell whether AnyPS5's direct-memory accounting refuses something a console would allow, two things from the APS5_TRACE_MEMORY log of The Pathless (PPSA01826) would settle it:

• The fatal request: the lines around the 5,565,054,976-byte allocation. Which call makes it (allocate direct memory, map flexible memory, or reserve or map a virtual range), with its size, alignment, flags and memory type, and the error AnyPS5 returned.
• The start of the halving claim: whether there was a first request for the whole size, before the 6,912 MiB block, and what AnyPS5 returned for it, or whether the claim started at 6,912 MiB.

If Call of the Sea (PPSA01546) or Severed Steel (PPSA05106) show the same failing call, that would also tell us it's one path.

[JJenkx]

Cross-title note: Call of the Sea (PPSA01546), another Unreal Engine 4 title, stops in start-up on the identical request, Ran out of memory allocating 5565054976 bytes with alignment 65536, on current main 6a6a35fdf5: #9864. An identical 5,307.25 MiB size in two different titles points at an engine-side value, such as a pool size, rather than either title's data. It isn't a simple fraction of AnyPS5's 13,824 MiB direct memory size. Afterimage, also Unreal 4, makes no such request and runs.

AI-assisted: yes (Claude Code).

[JJenkx]

A third Unreal Engine 4 title stops on the identical request: Unbound: Worlds Apart (PPSA03274), Ran out of memory allocating 5565054976 bytes with alignment 65536 22 s into start-up on main 44a59edc3a: #10204. With The Pathless and Call of the Sea (#9864), that makes three titles requesting exactly 5,307.25 MiB, so the size comes from the engine, not from any one title.

AI-assisted: yes (Claude Code).

[QRMOE]

@JJenkx could you attach the APS5_TRACE_MEMORY lines around the 5,565,054,976-byte request in The Pathless, as @EagleBearerYS asked above? The request itself plus roughly the 50 lines before it would be enough: which call makes it (direct memory allocation, flexible map, or virtual range reserve/map), its size, alignment, flags and memory type, and the error AnyPS5 returns. The map and release calls during the remapping just before it would also show whether AnyPS5 still counts memory the engine has already released.

With that I can check the accounting in code and add a regression test for whatever differs.

AI-assisted: yes

[JJenkx]

An experiment on the shared 5,565,054,976-byte request (The Pathless, Call of the Sea #9864, Unbound #10204). I used Unbound (PPSA03274) on main 44a59edc3a, with private builds where only libkernel's DIRECT_MEMORY_SIZE (what sceKernelGetDirectMemorySize returns) was changed, and APS5_TRACE_MEMORY on:

reported direct memory	engine's first grab block	UsedPhysical at the failure	AvailablePhysical	5,307.25 MiB request
12,288 MiB	6,144 MiB	8,847.75 MiB	3,140 MiB	fails
13,824 MiB (main)	6,912 MiB	8,847.75 MiB	4,676 MiB	fails
16,384 MiB	8,192 MiB	11,343.6 MiB	4,740 MiB	fails

• The request is the same 5,565,054,976 bytes at every size, so it is a fixed engine value.
• The engine sizes its start-up grab from the reported size. It takes half first and keeps halving, in all three runs.
• Its pre-request usage grew by about 2.5 GB at 16,384 MiB, still leaving under 5.3 GB free. So reporting more direct memory would not make this request fit; something else must differ on the console.

From the earlier memory trace of The Pathless, the usage before the request is mostly the engine's own pools mapped from that grab: 2,807 MiB, 1,485 MiB, four of about 1,080 MiB (two physical ranges, each mapped at two addresses), 962 MiB, 785 MiB, and so on. Afterimage (PPSA14053), also Unreal Engine 4, makes the same start-up grab, but its largest mappings are about 64 MiB, and it runs. I don't have the console's value for sceKernelGetDirectMemorySize. AnyPS5's 13,824 MiB has had no recorded source since libkernel's first commit.

AI-assisted: yes (Claude Code).

[QRMOE]

Thanks, that rules out the reported size. Could you post the APS5_TRACE_MEMORY lines for two things from The Pathless run?

1. The two physical ranges of about 1,080 MiB that are each mapped at two addresses: every allocate, map, protect and unmap line for those ranges, in order, so it shows whether the engine maps them twice on purpose or a second mapping is left behind after something fails or moves.
2. The fatal 5,565,054,976-byte request: the call that makes it, its arguments, and the error AnyPS5 returns.

AI-assisted: yes