| Age | Commit message (Collapse) | Author |
|---|
|
Based on a discussion with Niels Thykier who asked for Contents-all this
implements apt trying for all architecture dependent files to get a file
for the architecture all, which is treated internally now as an official
architecture which is always around (like native). This way arch:all
data can be shared instead of duplicated for each architecture requiring
the user to download the same information again and again.
There is one problem however: In Debian there is already a binary-all/
Packages file, but the binary-any files still include arch:all packages,
so that downloading this file now would be a waste of time, bandwidth
and diskspace. We therefore need a way to decide if it makes sense to
download the all file for Packages in Debian or not. The obvious answer
would be a special flag in the Release file indicating this, which would
need to default to 'no' and every reasonable repository would override
it to 'yes' in a few years time, but the flag would be there "forever".
Looking closer at a Release file we see the field "Architectures", which
doesn't include 'all' at the moment. With the idea outlined above that
'all' is a "proper" architecture now, we interpret this field as being
authoritative in declaring which architectures are supported by this
repository. If it says 'all', apt will try to get all, if not it will be
skipped. This gives us another interesting feature: If I configure a
source to download armel and mips, but it declares it supports only
armel apt will now print a notice saying as much. Previously this was a
very cryptic failure. If on the other hand the repository supports mips,
too, but for some reason doesn't ship mips packages at the moment, this
'missing' file is silently ignored (= that is the same as the repository
including an empty file).
The Architectures field isn't mandatory through, so if it isn't there,
we assume that every architecture is supported by this repository, which
skips the arch:all if not listed in the release file.
|
|
This allows running tests in parallel.
Git-Dch: Ignore
|
|
Our error reporting is historically grown into some kind of mess.
A while ago I implemented stacking for the global error which is used in
this commit now to wrap calls to functions which do not report (all)
errors via return, so that only failures in those calls cause a failure
to propergate down the chain rather than failing if anything
(potentially totally unrelated) has failed at some point in the past.
This way we can avoid stopping the entire acquire process just because a
single source produced an error for example. It also means that after
the acquire process the cache is generated – even if the acquire
process had failures – as we still have the old good data around we can and
should generate a cache for (again).
There are probably more instances of this hiding, but all these looked
like the easiest to work with and fix with reasonable (aka net-positive)
effects.
|
|
I never understood why there is an extra newline in those messages, so
now is as good time as any to drop them. Lets see if someone complains
with a good reason to keep it…
|
|
Again, consistency is the main sellingpoint here, but this way it is now
also easier to explain that some files move through different stages and
lines are printed for them hence multiple times: That is a bit hard to
believe if the number is changing all the time, but now that it keeps
consistent.
|
|
If we have a file on disk and the hashes are the same in the new Release
file and the old one we have on disk we know that if we ask the server
for the file, we will at best get an IMS hit – at worse the server
doesn't support this and sends us the (unchanged) file and we have to
run all our checks on it again for nothing. So, we can save ourselves
(and the servers) some unneeded requests if we figure this out on our
own.
|
|
Having every item having its own code to verify the file(s) it handles
is an errorprune process and easy to break, especially if items move
through various stages (download, uncompress, patching, …). With a giant
rework we centralize (most of) the verification to have a better
enforcement rate and (hopefully) less chance for bugs, but it breaks the
ABI bigtime in exchange – and as we break it anyway, it is broken even
harder.
It shouldn't effect most frontends as they don't deal with the acquire
system at all or implement their own items, but some do and will need to
be patched (might be an opportunity to use apt on-board material).
The theory is simple: Items implement methods to decide if hashes need to
be checked (in this stage) and to return the expected hashes for this
item (in this stage). The verification itself is done in worker message
passing which has the benefit that a hashsum error is now a proper error
for the acquire system rather than a Done() which is later revised to a
Failed().
|
|
If we e.g. fail on hash verification for Packages.xz its highly unlikely
that it will be any better with Packages.gz, so we just waste download
bandwidth and time. It also causes us always to fallback to the
uncompressed Packages file for which the error will finally be reported,
which in turn confuses users as the file usually doesn't exist on the
mirrors, so a bug in apt is suspected for even trying it…
|
|
Valid-Until protects us from long-living downgrade attacks, but not all
repositories have it and an attacker could still use older but still
valid files to downgrade us. While this makes it sounds like a security
improvement now, its a bit theoretical at best as an attacker with
capabilities to pull this off could just as well always keep us days
(but in the valid period) behind and always knows which state we have,
as we tell him with the If-Modified-Since header. This is also why this
is 'silently' ignored and treated as an IMSHit rather than screamed at
the user as this can at best be an annoyance for attackers.
An error here would 'regularily' be encountered by users by out-of-sync
mirrors serving a single run (e.g. load balancer) or in two consecutive
runs on the other hand, so it would just help teaching people ignore it.
That said, most of the code churn is caused by enforcing this additional
requirement. Crisscross from InRelease to Release.gpg is e.g. very
unlikely in practice, but if we would ignore it an attacker could
sidestep it this way.
|
|
Not all servers we are talking to support If-Modified-Since and some are
not even sending Last-Modified for us, so in an effort to detect such
hits we run a hashsum check on the 'old' compared to the 'new' file, we
got the hashes for the 'new' already for "free" from the methods anyway
and hence just need to calculated the old ones.
This allows us to detect hits even with unsupported servers, which in
turn means we benefit from all the new hit behavior also here.
|
|
If we get a IMSHit for the Transaction-Manager (= the InRelease file or
as its still supported fallback Release + Release.gpg combo) we can
assume that every file we would queue based on this manager, but already
have locally is current and hence would get an IMSHit, too. We therefore
save us and the server the trouble and skip the queuing in this case.
Beside speeding up repetative executions of 'apt-get update' this way we
also avoid hitting hashsum errors if the indexes are in fact already
updated, but the Release file isn't yet as it is the case on well
behaving mirrors as Release files is updated last.
The implementation is a bit harder than the theory makes it sound as we
still have to keep reverifying the Release files (e.g. to detect now expired
once to avoid an attacker being able to silently stale us) and have to
handle cases in which the Release file hits, but some indexes aren't
present (e.g. user added a new foreign architecture).
|
