tmpfs: implement FIOSEEKDATA and FIOSEEKHOLE

[kib]

Summary

Also:
tmpfs: change return type of tmpfs_pages_check_avail() to bool
vn_bmap_seekhole: check that passed offset is non-negative

[markj]

sys/fs/tmpfs/tmpfs_vnops.c

1851 ↗

(On Diff #111870)

You could optimize for the case where no pages are swapped: check if the swap block trie is empty, and if so use vm_page_find_least(). With a bit more work, the swap pager could find the next block after a given pindex.

I remember that we needed similar optimizations to UFS to avoid pathological behaviour found by syzkaller. It would create a giant sparse file, then FIOSEEKDATA would spend eternity checking each block one by one in an unkillable loop.

1852 ↗

(On Diff #111870)

I'd avoid accessing vm_page fields like this, and instead write m != NULL && !vm_page_none_valid(m)). And maybe add vm_page_any_valid() so that we don't need a double negative.

[kib]

Make tmpfs_seek_data_locked() loop-less (I do not see a way to avoid walking the object page queue in tmpfs_seek_hole_locked()).
Add vm_page_any_valid() and use helpers in more places.

[markj]

In D37024#841021, @kib wrote:

(I do not see a way to avoid walking the object page queue in tmpfs_seek_hole_locked()).

Yes, for SEEKHOLE we need to check each index.

Add vm_page_any_valid() and use helpers in more places.

Thanks.

sys/fs/tmpfs/tmpfs_vnops.c

1841 ↗

(On Diff #111914)