Release laundered vnode pages to the head of the inactive queue

[markj]

Summary

The swap pager enqueues laundered pages near the head of the inactive queue to avoid another trip through LRU before reclamation. This change implements corresponding support for this behaviour to the vnode pager and makes use of it in UFS and ext2fs.

No changes to ZFS are required, due to the missing VM integration: zfs_putpages() copies the dirty pages for writing and returns VM_PAGER_OK in the non-error case, and vm_pageout_flush() takes care of moving the pages to the inactive queue. msdosfs requires a bit more work, so I've held off on that in this review.

[kib]

sys/ufs/ffs/ffs_vnops.c

461 ↗

(On Diff #22366)

Why do you need fs-specific variants of the function ? IMO it is cleaner to put the helper into vfs_bio.c.

Note that b_dep could be non-empty only for FFS.

[markj]

sys/ufs/ffs/ffs_vnops.c

461 ↗

(On Diff #22366)

No good reason, I'll move it.

[markj]

• Use a common function to release read bufs.

[kib]

This breaks LRU for buffers, but this is probably fine when pagedaemon wants to reclaim the pages anyway.

sys/ufs/ffs/ffs_vnops.c

808 ↗

(On Diff #22387)

If you add another boolean parameter to vfs_bio_rbrelse(), to indicate should brelse()/bqrelse() be called, you may avoid inlining this snippet there as well as in ext2 code.

[alc]

In D8589#178507, @kib wrote:

This breaks LRU for buffers, but this is probably fine when pagedaemon wants to reclaim the pages anyway.

Yes, I think so. In this case, the page(s) are known to be in the laundry queue and not wired into a buffer. In other words, we're creating a buffer just to perform the page out, so flagging that buffer as "no reuse" seems ok. Am I missing something?

[kib]

In D8589#178560, @alc wrote:

In D8589#178507, @kib wrote:

This breaks LRU for buffers, but this is probably fine when pagedaemon wants to reclaim the pages anyway.

Yes, I think so. In this case, the page(s) are known to be in the laundry queue and not wired into a buffer. In other words, we're creating a buffer just to perform the page out, so flagging that buffer as "no reuse" seems ok. Am I missing something?

This consideration was mostly what I mean. I am sure that laundered pages are returned to the inactive queue, which is confirmed by the description of the revision. Note that if buffer is waited for by other thread meantime, LRU is broken. I agree that this is acceptable.

[markj]

• Move more duplicated code to vfs_bio.c.

[markj]

sys/ufs/ffs/ffs_vnops.c

808 ↗

(On Diff #22387)

I made it a separate function since I couldn't come up with a name that made sense for both usages (name suggestions welcome). I also pushed IO_DIRECT handling there. IO_INVAL was left alone as it appears to be unused.

[markj]

• Fix a typo.

[kib]

sys/kern/vfs_bio.c

4423 ↗

(On Diff #22452)

b_io_dismiss() or vfs_bio_io_finish() ?

[markj]

sys/kern/vfs_bio.c

4423 ↗

(On Diff #22452)

b_io_dismiss() seems marginally better, but still doesn't feel quite right. I'll just go with that.