Re: [PATCH v2] drm/xe: Add bounds check for num_binds to prevent memory exhaustion

[Ramesh Adhikari <adhikari.resume@gmail.com>]

Hi Matthew and Thomas,

I apologize if you're receiving this message multiple times - I've tried 
sending replies twice before but they don't appear in the mailing list 
archives or on Patchwork, so I'm not sure if they reached you. I also 
attempted to send a v3 patch which similarly didn't appear. If you did 
receive my previous emails, I sincerely apologize for the duplicate messages.

When I was tracing through the code, I found that vm_bind_ioctl_ops_create
allocates about 160 bytes per bind (drm_gpuva_ops + xe_vma_op) in the loop,
and those allocations use GFP_KERNEL without __GFP_ACCOUNT. That's separate
from the main arrays you already have protected with __GFP_ACCOUNT.

At 2048 binds that's only 320KB unaccounted, which is why I thought it was
safe to start conservative. But you're right - at 64k binds it would still
only be about 10MB unaccounted, which is probably fine and won't force
unnecessary fallbacks.

Should I send a v4 with 64k? Or do you think the loop allocations I found
need a different approach?

Thanks,
Ramesh