File Descriptor Disconnect Handling Broken

[bonefish]

The disconnect handling for file descriptors has several issues:

• Disconnected file_descriptors are leaked (put_fd() sets the ref_count to -1).
• There are race conditions, particularly in put_fd().
• Disconnected file descriptors can be dup()ed and copied on fork(), messing with the mangled ref_count.

[emitrax]

Could this be the cause of what I've experienced in bug #2381 ?

[bonefish]

Replying to emitrax:

Could this be the cause of what I've experienced in bug #2381 ?

#2381 has been closed as user error (leaking FDs). It doesn't have anything to do with this ticket.

[bonefish]

#4473 could be caused by this.

[bonefish]

#5123 could be caused by this.

[scottmc]

sliding this one out to R1/beta1.

[richienyhus]

Have there been any improvements in the interim?

Could this be moved to beta2 ?

[axeld]

Should be fixed with hrev49631. I'd feel better closing this bug if someone else could review the changes, though.

[ttcoder]

FWIW... We were seeing this with increasing frequency back when our reference hrev was 49137 (up to once a week towards the end!), but we didn't get a single instance of this since switching over to 49856.

[scottmc]

Sounds like this one might be fixed but should be one or two more reviews look it over and determine if we can close this one? @bonefish @mmlr can either of you weigh in on this one?

[axeld]

Seems fixed.