Page MenuHomeFreeBSD

geom: add kqfilter support for geom dev
ClosedPublic

Authored by rew on Dec 13 2021, 2:18 AM.
Tags
None
Referenced Files
Unknown Object (File)
Sun, Apr 14, 12:28 PM
Unknown Object (File)
Sat, Mar 30, 10:57 PM
Unknown Object (File)
Feb 15 2024, 9:35 AM
Unknown Object (File)
Jan 2 2024, 12:44 AM
Unknown Object (File)
Dec 27 2023, 11:49 PM
Unknown Object (File)
Dec 20 2023, 6:04 AM
Unknown Object (File)
Dec 11 2023, 9:59 AM
Unknown Object (File)
Dec 2 2023, 8:40 AM
Subscribers

Details

Summary

Implement support for the EVFILT_VNODE filter to trigger the NOTE_ATTRIB
event.

This will allow bhyve to notify a guest when an underlying zvol is
resized from the host. This filter will catch events for zvols that are
being managed by geom (i.e., volmode=full|geom).

Diff Detail

Repository
rG FreeBSD src repository
Lint
Lint Not Applicable
Unit
Tests Not Applicable

Event Timeline

rew requested review of this revision.Dec 13 2021, 2:18 AM
jhb added inline comments.
sys/geom/geom_dev.c
946

Similar to the zvol request, perhaps you want to reject requests with any flags other than NOTE_ATTRIB?

don't add the knote to knlist if the event is not supported.

I think NOTE_REVOKE is also quite useful for geom devs.

sys/geom/geom_dev.c
327

I suspect this should be NOTE_EXTEND and not NOTE_ATTRIB?

In D33402#759444, @kib wrote:

I think NOTE_REVOKE is also quite useful for geom devs.

Noted, I will follow-up with separate reviews that implement other NOTE_* events, including NOTE_REVOKE.

I noticed geom access counters get leaked when a geom device is revoked - and then I stumbled on commit https://reviews.freebsd.org/rS286237. When a geom device is being revoked, can the access counters be zero'ed out so they don't get leaked when revoked?

sys/geom/geom_dev.c
327

I used NOTE_ATTRIB for consistency because that is the event that is triggered when the size of regular file changes.

In my mind, NOTE_EXTEND was to communicate something being extended (size increase) as opposed to shrinking (size decrease). Maybe I misunderstand that.

If it's acceptable to use NOTE_EXTEND when the size of a geom is extended/shrunk, then I'm good with using NOTE_EXTEND - let me know what you think.

In D33402#763772, @rew wrote:

I noticed geom access counters get leaked when a geom device is revoked - and then I stumbled on commit https://reviews.freebsd.org/rS286237. When a geom device is being revoked, can the access counters be zero'ed out so they don't get leaked when revoked?

I really do not know. An answer to this question can be obtained only by the experiment, try to do that and see what else get broken. From very high point of view, I doubt that it is possible, because there are other ways to keep a handle on geom providers than through devfs node. So despite devfs node is reclaimed, you might still have reference to provider and have associated rights (g_vfs_open() is about this, in fact).

sys/geom/geom_dev.c
327

Look for instance at ftruncate(2) that both extends and shrinks.

sys/geom/geom_dev.c
327

That'll trigger the NOTE_ATTRIB event when extending or shrinking.

For a regular file, NOTE_EXTEND will be triggered after a VOP_WRITE() if the size is larger than before.

sys/geom/geom_dev.c
327

I must admit that this is quite weird and not matches the actual machinery behind the implementation. For instance, on UFS the last byte of file must be always backed by the block, i.e. hole can exist only in between start and EOF, so UFS does allocate and zero-out extending block (but not through VOP_WRITE()).

This revision is now accepted and ready to land.Jan 9 2022, 5:58 AM
sys/geom/geom_dev.c
327

It's kind of complicated. It should perhaps be NOTE_EXTEND if it grows (but not shrinks!), but NOTE_ATTRIB always? The problem is that EVFILT_VNODE is a bit "filesystem like UFS"-specific. That is, a VOP_WRITE() that grows a file raises NOTE_EXTEND, but using truncate(2) or ftruncate(2) to resize a file uses VOP_SETATTR which raises NOTE_ATTRIB. Right now bhyve only lists for NOTE_ATTRIB, though I've suggested to Rob that perhaps it wants to listen to both NOTE_ATTRIB and NOTE_EXTEND.

(Somehow my earlier comment didn't post)

This revision was automatically updated to reflect the committed changes.