cuse(3): Allow shared memory allocations up to and including 2 GBytes.

[• hselasky]

Summary

Currently the cuse(3) mmap(2) offset is split into 128 banks of 16 Mbytes.
Allow cuse(3) to make allocations that span multiple banks at the expense
of any fragmentation issues that may arise. Typically mmap(2) buffers are
well below 16 Mbytes. This allows 8K video resolution to work using webcamd.

Some minor fixes are included aswell.

MFC after: 1 week
Sponsored by: NVIDIA Networking

Test Plan

V4L2 loopback with 8K resolution

[• hselasky]

Hopefull this works on CheriBSD aswell.

[• hselasky]

markj@ : Do you have time to have a quick look at this?

[markj]

lib/libcuse/cuse.3

128 ↗

(On Diff #108229)

limited by what?

lib/libcuse/cuse_lib.c

208

Don't you need to reserve all units in the run, not just the first one? What if two threads concurrently attempt an allocation, one of 32MB and the other of 16MB.

284

Shouldn't this zero the buffer before unmapping? Since we're mapping with MAP_SHARED, any time a segment is recycled, I believe the old contents will be preserved.

[• hselasky]

Update manual page as suggested by Mark.

[• hselasky]

lib/libcuse/cuse.3

128 ↗

(On Diff #108229)

It is limited by the mmap offset being 32-bit also for 64-bit platforms due to 32-bit user-space programs. I'll get that in.

lib/libcuse/cuse_lib.c

166

I might want to change the argument to unsigned to support the 2GByte case.

208

If the size is set and you allocate 32MB it will skip two entries because one entry can only hold 16MB.
Meaning any concurrent threads will see that window a busy or occupied, due to the "check if there are m free units ahead".

284

I want that it's the applications responsibility to clear memory before free. Also, if the application crashes, this needs to be done in the kernel and not here.

[• hselasky]

Make size argument for malloc unsigned. Really we want to use size_t, but because all allocations are limited to 32-bits it is kept 32-bit for now. Will need to introduce some symbol versioning to make this smooth.

[markj]

I can't spot any obvious problems with this diff.

lib/libcuse/cuse_lib.c

177

The cast shouldn't be necessary.

208

Ok, I see now.

284

It's moot anyway, I see now that CUSE_IOCTL_FREE_MEMORY will deallocate the pager for the unit. So the next allocation will get zeroed pages no matter what.

[• hselasky]

Remove not needed cast to unsigned long.

[• hselasky]

Thank you Mark for taking the time to review this.