ZFS deadlock on db_mtx and dn_holds

[Simon Klinkert via illumos-zfs]

Hi,

what do you guys think about https://www.illumos.org/issues/5056 ? How should we fix it?

Simon

[Justin T. Gibbs via illumos-zfs]

Post by Simon Klinkert via illumos-zfs
Hi,
what do you guys think about https://www.illumos.org/issues/5056 ? How should we fix it?
Simon

There are a series of patches under review that change the handling of DMU buf user callbacks so that they occur without the dbuf mutex held.

This review has a full description of the proposed API changes:

https://reviews.csiden.org/r/65/

But the end result, assuming all of the review comments can be addressed, will look more like the patch here:

https://reviews.csiden.org/r/66/

Currently I’m working on a patch that splits out the removal of db->db_data aliases and the dbuf_user_ptr_ptr dbuf field, which is where the majority of the code churn comes from. This can be pushed before the larger DMU buf user API changes and should make the review of that portion easier.

—
Justin

[Simon Klinkert via illumos-zfs]

Excellent. Thanks for the update, Justin.

Simon

Post by Justin T. Gibbs via illumos-zfs

Post by Simon Klinkert via illumos-zfs
Hi,
what do you guys think about https://www.illumos.org/issues/5056 ? How should we fix it?
Simon

There are a series of patches under review that change the handling of DMU buf user callbacks so that they occur without the dbuf mutex held.
https://reviews.csiden.org/r/65/
https://reviews.csiden.org/r/66/
Currently I’m working on a patch that splits out the removal of db->db_data aliases and the dbuf_user_ptr_ptr dbuf field, which is where the majority of the code churn comes from. This can be pushed before the larger DMU buf user API changes and should make the review of that portion easier.
—
Justin