* a task that is marked woken but didn't actually wake before being cancelled will instead wake the next task in the queue
* a task that wakes but doesn't get a connection will put itself back in the queue instead of waiting until it times out with no way to be woken
* the idle reaper now won't run if there are tasks waiting for a connection, and also uses
the proper `SharedPool::release()` to return validated connections to the pool so waiting tasks get woken
closes#622, #1210
(hopefully for good this time)
Signed-off-by: Austin Bonander <austin@launchbadge.com>
While running an application a lot of warnings were printed about my
pgpass file, like:
Malformed line in pgpass file
This was due to the fact that my pgpass file contains whitespace and
comments to organize it in a better way.
This commit ensures we will ignore empty lines and lines that (barring
whitespace) start with a comment. This is in line with how PostgreSQL
treats these entries in the pgpass file:
- https://www.postgresql.org/docs/current/libpq-pgpass.html
- function passwordFromFile in src/interfaces/libpq/fe-connect.c
Unlike `Executor.fetch_optional()`, `Executor.fetch_many()` does not
have a single exit. The stream can be dropped at any time. To catch
this event, we create a `StatementResetter` structure inside the stream
loop and reset the statement when it is dropped.
A test case `it_resets_prepared_statement_after_fetch_many` is
similar to `it_resets_prepared_statement_after_fetch_one` which tests
`Executor.fetch_optional()`.
64-bit pointers can't be casted to i32 correctly.
It does not seem to cause problems on amd64 as long as both pointers
reside in the same page, but it's not correct generally.
The provided callback doesn't need to be `'static`. It should be enough that it doesn't borrow anything from the `Connection` itself, even if it does borrow data from elsewhere.
* fix `DecrementSizeGuard::drop()` only waking one `Waiter` regardless of whether that waiter was already woken
* fix connect-backoff loop giving up the size guard
* don't cut in line to open a new connection
* have tasks waiting on `acquire()` wake periodically to check if there's a connection in the queue
Signed-off-by: Austin Bonander <austin@launchbadge.com>
Fix commit updates the `postgres::connection::describe` module to add full support for domain types. Domain types were previously confused with their category which caused invalid oid resolution.
Fixeslaunchbadge/sqlx#110
It is *much* too easy to misread `x.min(y)` as "`x` should be *at least* `y`" when in fact it means the *exact* opposite, and same with `x.max(y)`. This has bitten us in the gluteus maximus a number of times both in SQLx and in private projects.
Signed-off-by: Austin Bonander <austin@launchbadge.com>
BREAKING CHANGE: some columns in `query!()` et. al. output will change from `T` to `Option<T>`
breakage should be minimal in practice as
these columns will need to have been manually
overridden anyway to avoid runtime errors
Signed-off-by: Austin Bonander <austin@launchbadge.com>
Io errors can describe a lot of things, when connecting to a database in general only
ECONNREFUSED can be hoped to be resolved by waiting. Other errors will require
user intervention
If the connection to the database is refused, it is not helpfull to hammer it with connection
attempts, instead we should wait a bit before connecting again. Here we start waiting 10ms,
doubling the wait time for every attempt before setteling out at 2s
If wait_for_conn is dropped before completing, release will call
wake on a waker that noone is listining on. This leads to a state where
waiting will gro indefinitly while all connections are idle.
To fix this we turn waiting into a queue of unique Weak pointers.
This way if wait_for_conn is dropped the pointer in waiting cannot
be upgraded, we can use this as a signal that we shoud wake the
next one instead.
