On Wed, Feb 26, 2025 at 12:58:21PM -0700, Jens Axboe wrote:

On 2/26/25 12:43 PM, Jens Axboe wrote:

quoted

Ignore this one, that was bogus. Trying to understand a crash in here.

Just to wrap this up, it's the combining of table and cache that's the
issue. The table can get torn down and freed while nodes are still
alive - which is fine, but not with this change, as there's then no
way to sanely put the cached entry.

Yes, the imu can outlive the table it came from, which means Pavel was
right to suggest I not tied the cache to the table. Thanks for pointing
this out.