On 6/9/2022 11:03 AM, Enzo Matsumiya wrote:
On 06/09, Tom Talpey wrote:
On 6/8/2022 5:54 PM, Enzo Matsumiya wrote:
Hi,
These 2 patches are a simple way to fix the DNS issue that
currently exists in cifs, where the upcall to key.dns_resolver will
always return 0 for the record TTL, hence, making the resolve worker
always use the default value SMB_DNS_RESOLVE_INTERVAL_DEFAULT
(currently 600 seconds).
This also makes the new setting `dns_interval' user-configurable via
procfs (/proc/fs/cifs/dns_interval).
One disadvantage here is that the interval is applied to all hosts
resolution. This is still how it works today, because we're always using
the default value anyway, but should someday this be fixed, then we can
go back to rely on the keys infrastructure to cache each hostname with
its own separate TTL.
Curious, why did you choose a procfs global approach? Wouldn't it be
more appropriate to make this a mount option, so it would be applied
selectively per-server?
I tried to stick to the current behaviour, while also fixing the issue
of getting a TTL=0 from key.dns_resolver upcall.
A mount option looks indeed better initially, and that was also my
initial approach to this. But in a, e.g. multi-domain forest, with
multiple DFS targets, the per-mount setting starts to look irrelevant
again as well. So I took the "per-client" approach instead.
Ok, I guess. It seems like kicking the can down the road, a little.
I agree that rearchitecting the DNS cache might be extreme, but many
distros consider procfs to be a user API, and they'll require it to
be supported basically forever. That would be unfortunate...
Tom.
