Instead of checking current dnstap output file size and potentially
synchronously calling dns_dt_reopen() upon every call to dns_dt_send():
- call dns_dt_reopen() asynchronously by queuing an event to the task
specified at dnstap environment creation time,
- ensure no roll event is outstanding before checking dnstap output
file size and potentially queuing another roll event.
This causes dnstap output files to exceed their configured size limits,
but prevents any two threads from performing the roll simultaneously
(which causes crashes).
Instead of relying on the caller to set up task-exclusive mode, make
dns_dt_reopen() enforce task-exclusive mode itself, using the task
specified at dnstap environment creation time.
Implement a new variant of dns_dt_create() to enable a dnstap
environment structure to hold the task in the context of which
dns_dt_reopen() will be executed.
Remove a block of code which dates back to commit 8a2ab2b920, when
dns_zone_asyncload() did not yet check DNS_ZONEFLG_LOADPENDING.
Currently, no race in accessing DNS_ZONEFLG_LOADPENDING is possible any
more, because:
- dns_zone_asyncload() is still the only function which may queue
zone_asyncload(),
- dns_zone_asyncload() accesses DNS_ZONEFLG_LOADPENDING under a lock
(and potentially queues an event under the same lock),
- DNS_ZONEFLG_LOADPENDING is not cleared until the load actually
completes.
Thus, the rechecking code can be safely removed from zone_asyncload().
Note that this also brings zone_asyncload() to a state in which the
completion callback is always invoked. This is required to prevent
leaking memory in case something goes wrong in zone_asyncload() and a
zone table the zone belongs to is indefinitely left with a positive
reference count.
Code handling cancellation of asynchronous zone load events was likely
copied over from other functions when asynchronous zone loading was
first implemented in commit 8a2ab2b920. However, unlike those other
functions, asynchronous zone loading events currently have no way of
getting canceled once they get posted, which means the aforementioned
code is effectively dead. Remove it to prevent confusion.
zone_load() is not always synchronous, it may only initiate an
asynchronous load and return DNS_R_CONTINUE, which means zone loading
has not yet been completed. In such a case, zone_asyncload() must not
clear DNS_ZONEFLG_LOADPENDING immediately and leave that up to
zone_postload().
While this is not an issue in named, which only calls
dns_zone_asyncload() from task-exclusive mode, this function is exported
by libdns and thus may in theory be concurrently called for the same
zone by multiple threads. It also does not hurt to be consistent
locking-wise with other DNS_ZONEFLG_LOADPENDING accesses.
Tinderbox User
Ondřej Surý
Michał Kępień
Mark Andrews
Evan Hunt