Crimson Gem Saga:

[Vyguard]

For what ever reason the game seems to crash at totally random times. Sometimes you can get past it, sometimes you can't. I've run into an instance that I can't seem to get past. The ppsspp crashes, and I just get the nice message for windows to attempt to find a solution. Are there any fixes for this, or any settings that need to be changed? I'm running it on the ppsspp64 all the other games run perfectly fine. I'm also running default settings for the program.

[unknownbrackets]

This was reported here:
http://forums.ppsspp.org/showthread.php?tid=1243&pid=106156#pid106156

As a regression from 879baec. We weren't sure why, though - this just suggested it was somehow timing related.

@LunaMoo created a cheat workaround, but of course that's not really a full solution.

Looking into it further, the problem seems to be io timing related (noting for #7647). By increasing the seek delay from 0.1ms to 0.3ms, for example, I'm also able to avoid the problem. More tests may be needed, but changing the seek latency is probably the "correct" fix here.

For more technical detail:

1. The game creates a thread, th_load_voice.
2. th_load_voice locks a mutex and seeks in a file.
3. user_main calls memset 5 times on small amounts of bytes (about 0x900 total.)
4. th_load_voice awakes from sleep, unlocks the mutex, then locks it again.
5. A memory read gets a null pointer and crashes.

I haven't found the exact point where things diverge, but making memset take longer basically means the 5 memsets are not done before th_load_voice wakes up. Ultimately, the game seems to be depending on a race between the user_main thread and th_load_voice. Making memset take longer fixed the winner of the race, but changing seek latency has the same effect (the other way) and seems more correct.

-[Unknown]

[unknownbrackets]

It seems even 0.101ms latency instead of 0.1ms will work. It's very possible that 0.1ms is just the wrong value and slightly too low, but I need to do more tests to try to verify. Maybe we should change it for now, although I'm not sure if it fixes this issue everywhere.

The change is (all occurrences in sceIo.cpp):

return hleDelayResult(result, "io seek", 100);

To:

return hleDelayResult(result, "io seek", 101);

-[Unknown]

[y-salnikov]

101 does nothing but 200 is work for me. Don't know what i'm doing though.

[unknownbrackets]

Turns out it was actually too high - it works consistently for me without the too-large delay.

-[Unknown]