Connection establishment now happens only in one place in each IMAP loop.
Now all connection establishment happens in one place
and is limited by the ratelimit.
Backoff was removed from fake_idle
as it does not establish connections anymore.
If connection fails, fake_idle will return an error.
We then drop the connection and get back to the beginning of IMAP
loop.
Backoff may be still nice to have to delay retries
in case of constant connection failures
so we don't immediately hit ratelimit if the network is unusable
and returns immediate error on each connection attempt
(e.g. ICMP network unreachable error),
but adding backoff for connection failures is out of scope for this change.
ratelimit can be exhausted quickly if the network is not available,
i.e. if every connection attempt returns "network unreachable" error.
When the network becomes available, we want to retry connecting
as soon as maybe_network is called without waiting for ratelimiter.
Before group avatar was sent as an attachment. Let's do the same as with user avatar and send group
avatar as base64. Receiver code uses the same functions for user and chat avatars, so base64 avatars
are supported for most receivers already.
It was broken completely and before "fix: apply Autocrypt headers if timestamp is unchanged" that
didn't show up because the message from the second Bob's device never had "Date" greater than one
from the message sent before from the first device.
In this case connection failure
may be a connection timeout (currently 1 minute),
so it does not make sense to fake idle for another minute immediately after.
However, failure may be immediate if the port is closed
and the server refuses connection every time.
To prevent busy loop in this case
we apply ratelimit to connection attempts rather than login attempts.
This partially reverts ccec26ffa7
Even if 1:1 chat with alice is protected,
we should send vc-request unencrypted.
This happens if Alice changed the key
and QR-code Bob scans contains fingerprint
that is different from the verified fingerprint.
Sending vc-request encrypted to the old key
does not help because Alice is not able
to decrypt it in this case.
Use sync messages for that as it is done for e.g. Config::Displayname. Maybe we need to remove
self-status synchronisation via usual messages then, but let's think of it a bit.
Use sync messages for that as it is done for e.g. Config::Displayname. Maybe we need to remove
avatar synchronisation via usual messages then, but let's think of it a bit.
Encryption subkey is incorrectly referred to as public key
in variable names.
This is incorrect because generated encryption key
is secret too just as the signing primary key.
Generated OpenPGP secret key consists of primary signing key
and encryption subkey.
Then OpenPGP public key consisting of
the primary signing public key
and encryption public key is generated.
Keypair consists of the secret OpenPGP key and public OpenPGP key,
each of them has a primary key and subkey inside.
The limit is better enforced by webxdc distributors,
e.g. xdc store bots or actually email providers
to allow for experimentation with large frameworks
or porting existing apps and testing them
before reducing their size.
Besides that, the comment on WEBXDC_SENDING_LIMIT was outdated,
it was not updated when the limit was increased to 640 kB.
Even if `vc-request-with-auth` is received with a delay, the protection message must have the sort
timestamp equal to the Sent timestamp of `vc-request-with-auth`, otherwise subsequent chat messages
would also have greater sort timestamps and while it doesn't affect the chat itself (because Sent
timestamps are shown to a user), it affects the chat position in the chatlist because chats there
are sorted by sort timestamps of the last messages, so the user sees chats sorted out of
order. That's what happened in #5088 where a user restores the backup made before setting up a
verified chat with their contact and fetches new messages, including `vc-request-with-auth` and also
messages from other chats, after that.
Add a new crate `deltachat_time` with a fake `struct SystemTimeTools` for mocking
`SystemTime::now()` for test purposes. One still needs to use `std::time::SystemTime` as a struct
representing a system time. I think such a minimalistic approach is ok -- even if somebody uses the
original `SystemTime::now()` instead of the mock by mistake, that could break only tests but not the
program itself. The worst thing that can happen is that tests using `SystemTime::shift()` and
checking messages timestamps f.e. wouldn't catch the corresponding bugs, but now we don't have such
tests at all which is much worse.
PR #5099 removed some columns in the database that were actually in use.
usually, to not worsen UX unnecessarily
(releases take time - in between, "Add Second Device", "Backup" etc.
would fail), we try to avoid such schema changes (checking for
db-version would avoid import etc. but would still worse UX),
see discussion at #2294.
these are the errors, the user will be confronted with otherwise:
<img width=400
src=https://github.com/deltachat/deltachat-core-rust/assets/9800740/e3f0fd6e-a7a9-43f6-9023-0ae003985425>
it is not great to maintain the old columns, but well :)
as no official releases with newer cores are rolled out yet, i think, it
is fine to change the "107" migration
and not copy things a second time in a newer migration.
(this issue happens to me during testing, and is probably also the issue
reported by @lk108 for ubuntu-touch)
The system clock may be adjusted and even go back, so caching system time in code sections where
it's not supposed to change may even protect from races/bugs.
If a time value doesn't need to be sent to another host, saved to the db or otherwise used across
program restarts, a monotonically nondecreasing clock (`Instant`) should be used. But as `Instant`
may use `libc::clock_gettime(CLOCK_MONOTONIC)`, e.g. on Android, and does not advance while being in
deep sleep mode, get rid of `Instant` in favor of using `SystemTime`, but add `tools::Time` as an
alias for it with the appropriate comment so that it's clear why `Instant` isn't used in those
places and to protect from unwanted usages of `Instant` in the future. Also this can help to switch
to another clock impl if we find any.
Restart the IO scheduler if needed to make the new config value effective (for `MvboxMove,
OnlyFetchMvbox, SentboxWatch` currently). Also add `set_config_internal()` which doesn't affect
running the IO scheduler. The reason is that `Scheduler::start()` itself calls `set_config()`,
although not for the mentioned keys, but still, and also Rust complains about recursive async calls.
