it's always legal to call curl_multi_remove_handle on a valid pair of
multi and easy handles. removing an easy handle that is not currently
attached to a multi handle is a no-op, and removing an easy handle of
a different multi handle is something we can't reasonably trigger. if
we *did* ever manage it would result in an error we'd ignore, and the
handles in question would not be changed at all. this is just simpler
Change-Id: I85ec62ff89385981ca49d243376b9c32586bd128
return it as a separate item in a pair instead. this will let us remove
enqueueDownload() in favor of returning metadata from download() itself
Change-Id: I74fad2ca15f920da1eefabc950c2baa2c360f2ba
it's just a uri and some headers now. those can be function arguments
with no loss of clarity. *actual* additional arguments, for example a
TLS context with additional certificates, could be added on a new and
improved FileTransfer class that carries not just a backend reference
but some real, visible context for its transfers. curl not being very
multi-threading-friendly when using multi handles will make sharing a
bit hard anyway once we drop the single global download worker thread
Change-Id: Id2112c95cbd118c6d920488f38d272d7da926460
we don't even need this outside of tests. maybe we should not do
automatic retries at this level at all and use retrying wrappers
instead? at some point we may have to do this, but not just yet.
Change-Id: If0088aa55215be81f1770c25b3bb1b5268c65cf8
never set explicitly, and transfers are never instantiated with one
current activity but submitted with a *different* current activity.
Change-Id: I1a3ec57c02013565aeb9e9398ea42d0c4279095e
it's only used by HttpBinaryCacheStore, and even there used in only on
place. this one place can set the header explicitly, which it now does
Change-Id: Id89228150669e25e7f59a3d6bd939e46059ce29e
it only sets the one field anyway (and the parent activity as a side
effect that does not depend on the exact location of the constructor
call). when FileTransferRequest goes away we would need this anyway.
Change-Id: I35cf2ed3533239181449a62cf34cd282b395e5db
this is fragile, slow as fuck, breaks constantly under high concurrency,
and completely unnecessary since ssh bypasses the stdio file descriptors
*anyway*. we do still check that we see ssh messages to ensure that none
of our subprocess handling messes with ssh's /dev/tty, but that's it now
Change-Id: Ib8e31e1999f813d07a27efc63a9d3454a9e4fcdd
a first little step into pushing the event loops up, up and away.
eventually we will want them to be instantiated only at the roots
of every thread (since kj binds loops to threads), but not today.
Change-Id: Ic97f1debba382a5a3f46daeaf2d6d434ee42569f
goals should be considered internal to the worker architecture due to
the tight coupling of the two, and we can finally do that. doing this
is also a prerequisite for turning Worker::run() into a real promise.
Change-Id: I7cf273d4a6fdb75b8d192fce1af07c6265ff6980
previously it was possible to fetchurl a dict server, or an ldap server,
or an imap server. this is a bit of a problem, both because rare schemes
may not be available on all systems, and because some schemes (e.g. scp)
are inherently insecure in potentially surprising ways we needn't allow.
Change-Id: I18fc567c6f58c3221b5ea8ce927f4da780057828
we already have a results type for the entire worker call, we may as
well put this bit of info in there. we do keep the assumption of the
old code that the field will only be read if some goals have failed;
fixing that is a very different mess, and not immediately necessary.
Change-Id: If3fc32649dcd88e1987cdd1758c6c5743e3b35ac
without this derivations do not show as completely processed in the
internal-json logs (or the newer multiline output). the former also
breaks external tools like nix-output-monitor which, like multiline
output, grow vertically until at least some goals are finally freed
Change-Id: I55758daf526ba29ae15fb82e0d88da8afb45bf5c
some promises capture `this`. we could also allocate a shared state,
but this thing doesn't really need to ever be moved anyway. so there.
Change-Id: I50b5c44684a8ab4e984b1323de21f97ace4a864a
not doing this can freeze slots until the goal that occupied them is
freed (rather than simply complete), and then can freeze the system.
fixes#549
Change-Id: I042df04222f8ffbaa18ef4a4eae6cbd6f89b679e
these two functions are now nearly trivial and much better inline into
makeGoalCommon. keeping them separate also separates information about
goal completion flows and how failure information ends up in `Worker`.
Change-Id: I6af86996e4a2346583371186595e3013c88fb082
we can use our newfound powers of Goal::work Is A Real Promise to remove
completed goals from continuation promises. apart from being much easier
to follow it's also a lot more efficient because we have the iterator to
the item we are trying to remove, skipping a linear search of the cache.
Change-Id: Ie0190d051c5f4b81304d98db478348b20c209df5
Goal::work() is a fully usable promise that does not rely on the worker
to report completion conditions. as such we no longer need the `notify`
field that enabled this interplay. we do have to clear goal caches when
destroying the worker though, otherwise goal promises may (incorrectly)
keep goals alive due to strong shared pointers created by childStarted.
Change-Id: Ie607209aafec064dbdf3464fe207d70ba9ee158a
derivation goals still hold a BuildResult member variable since parts of
these results of accumulated in different places, but the Goal class now
no longer has such a field. substitution goals don't need it at all, and
derivation goals should also be refactored to not drop their buildResult
Change-Id: Ic6d3d471cdbe790a6e09a43445e25bedec6ed446
the field is simply duplicated between the two, and now that we can
return WorkResults from Worker::run we no longer need both of them.
Change-Id: I82fc47d050b39b7bb7d1656445630d271f6c9830
this will be needed to move all interesting result fields out of Goal
proper and into WorkResult. once that is done we can treat goals as a
totally internal construct of the worker mechanism, which also allows
us to fully stop exposing unclear intermediate state to Worker users.
Change-Id: I98d7778a4b5b2590b7b070bdfc164a22a0ef7190
since we now propagate goal exceptions properly we no longer need to
check topGoals for a reason to abort early. any early abort reasons,
whether by exception or a clean top goal failure, can now be handled
by inspecting the goal result in the main loop. this greatly reduces
goal-to-goal interactions that do not happen at the main loop level.
since the underscore-free name is now available for use as variables
we'll migrate to that where we currently use `_topGoals` for locals.
Change-Id: I5727c5ea7799647c0a69ab76975b1a03a6558aa6
drop childException since it's no longer needed. also makes
waitForInput, childFinished, and childTerminated redundant.
Change-Id: I05d88ffd323c5b5c909ac21056162f69ffb0eb9f
there's no reason to have the worker set information on goals that the
goals themselves return from their entry point. doing this in the goal
`work()` function is much cleaner, and a prerequisite to removing more
implicit strong shared references to goals that are currently running.
Change-Id: Ibb3e953ab8482a6a21ce2ed659d5023a991e7923
this simplifies the worker loop, and lets us remove it entirely later.
note that ideally only one promise waiting for interrupts should exist
in the entire system. not one per event loop, one per *process*. extra
interrupt waiters make interrupt response nondeterministic and as such
aren't great for user experience. if anything wants to react to aborts
caused by explicit interruptions, or anything else, those things would
be much better served using RAII guards such as Finally (or KJ_DEFER).
Change-Id: I41d035ff40172d536e098153c7375b0972110d51
this was a triumph. i'm making a note here: huge success. it's hard to
overstate my satisfaction! i'm not even angry. i'm being so sincere ri
actually, no. we *are* angry. this was one dumbass odyssey. nobody has
asked for this. but not doing it would have locked us into old, broken
protocols forever or (possibly worse) forced us to write our own async
framework building on the old did-you-mean-continuations in Worker. if
we had done that we'd be locked into ever more, and ever more complex,
manual state management all over the place. this just could not stand.
Change-Id: I43a6de1035febff59d2eff83be9ad52af4659871
with async runtime scheduling we can no longer guarantee exact error
counts for builds that do not set keepGoing. the old behavior can be
recovered with a number of hacks that affect scheduling, but none of
those are very easy to follow now advisable. exact error counts will
like not be needed for almost all uses except tests, and *those* had
better check the actual messages rather than how many they got. more
messages can even help to avoid unnecessary rebuilds for most users.
Change-Id: I1c9aa7a401227dcaf2e19975b8cb83c5d4f85d64
due to event loop scheduling behavior it's possible for a derivation
goal to fully finish (having seen all paths it was asked to create),
but to not notify the worker of this in time to prevent another goal
asking the recently-finished goal for more outputs. if this happened
the finished goal would ignore the request for more outputs since it
considered itself fully done, and the delayed result reporting would
cause the requesting goal to assume its request had been honored. if
the requested goal had finished *properly* the worker would recreate
it instead of asking for more outputs, and this would succeed. it is
thus safe to always recreate goals once they are done, so we now do.
Change-Id: Ifedd69ca153372c623abe9a9b49cd1523588814f
we'll now loop to update displayed statistics, and use this loop to
limit the update rate to 50 times per second. we could have updated
much more frequently before this (once per iteration of `runImpl`),
much faster than would ever be useful in practice. aggressive stats
updates can even impede progress due to terminal or network delays.
Change-Id: Ifba755a2569f73c919b1fbb06a142c0951395d6d
Worker::run() is now entirely based on the kj event loop and promises,
so we need not handle awakeness of goals manually any more. every goal
can instead, once it has finished a partial work call, defer itself to
being called again in the next iteration of the loop. same end effect.
Change-Id: I320eee2fa60bcebaabd74d1323fa96d1402c1d15
notably we will check whether we want to do GC at all only once during
startup, and we'll only attempt GC every ten seconds rather than every
time a goal has finished a partial work call. this shouldn't cause any
problems in practice since relying on auto-gc is not deterministic and
stores in which builds can fill all remaining free space in merely ten
seconds are severely troubled even when gargage collection runs a lot.
Change-Id: I1175a56bf7f4e531f8be90157ad88750ff2ddec4
nothing needs to signal being still active but not actively pollable,
only that immediate polling for the next goal work phase is in order.
Change-Id: Ia43c1015e94ba4f5f6b9cb92943da608c4a01555
this was immensely inefficient on large caches, as can exist when many
derivations are buildable simultaneously. since we have smart pointers
to goals we can do cache maintenance in goal deleters instead, and use
the exact iterators instead of doing a linear search. this *does* rely
on goals being deleted to remove them from the cache, which isn't true
for toplevel goals. those would have previously been removed when done
in all cases, removing the cache entry when keep-going is set. this is
arguably incorrect since it might result in those goals being retried,
although that could only happen with dynamic derivations or the likes.
(luckily dynamic derivations not complete enough to allow this at all)
Change-Id: I8e750b868393588c33e4829333d370f2c509ce99
makeDerivationGoalCommon had the right idea, but it didn't quite go far
enough. let's do the rest and remove the remaining factory duplication.
Change-Id: I1fe32446bdfb501e81df56226fd962f85720725b
this was a debugging aid from day one that should not have any impact on
build semantics, and if it *does* have an impact on build semantics then
build semantics are seriously broken. keeping the order imposed by these
keys will be impossible once we let a real event loop schedule our jobs.
Change-Id: I5c313324e1f213ab6453d82f41ae5e59de809a5b
without circular references we do not need weak goal pointers except for
caches, which should not prevent goal destructors running. caches though
cannot create circular references even when they keep strong references.
if we removed goals from caches when their work() is fully finished, not
when their destructors are run, we could keep strong pointers in caches.
since we do not gain much from this we keep those pointers weak for now.
Change-Id: I1d4a6850ff5e264443c90eb4531da89f5e97a3a0
have DerivationGoal and its subclasses produce a wrapper promise for
their intermediate results instead, and return this wrapper promise.
Worker already handles promises that do not complete immediately, so
we do not have to duplicate this into an entire result type variant.
Change-Id: Iae8dbf63cfc742afda4d415922a29ac5a3f39348
the new event loop could very occasionally notice that a dependency of
some goal has failed, process the failure, cause the depending goal to
fail accordingly, and in the doing of the latter two steps let further
dependencies that previously have not been reported as failed do their
reporting anyway. in such cases a goal could fail with "1 dependencies
failed", but more than one dependency failure message was shown. we'll
now report the correct number of failed dependency goals in all cases.
Change-Id: I5aa95dcb2db4de4fd5fee8acbf5db833531d81a8
these can be unique rather than shared because shared_ptr has a
converting constructor. preparatory refactor for something else
and not necessary on its own, and the extra allocations we must
do for shared_ptr control blocks isn't usually relevant anyway.
Change-Id: I5391715545240c6ec8e83a031206edafdfc6462f
also gets rid of explicit strong references to dependencies of any goal,
and weak references to dependers as well. those are now only held within
promises representing goal completion and thus independent of the goal's
relation to each other. the weak references to dependers was only needed
for notifications, and that's much better handled entirely by kj itself.
Change-Id: I00d06df9090f8d6336ee4bb0c1313a7052fb016b
now that we have an event loop in the worker we can use it and its
magical execution suspending properties to replace the slot counts
we managed explicitly with semaphores and raii tokens. technically
this would not have needed an event loop base to be doable, but it
is a whole lot easier to wait for a token to be available if there
is a callback mechanism ready for use that doesn't require a whole
damn dedicated abstract method in Goal to work, and specific calls
to that dedicated method strewn all over the worker implementation
Change-Id: I1da7cf386d94e2bbf2dba9b53ff51dbce6a0cff7
with waitForAWhile turned into promised the core functionality of
waitForInput is now merely to let gc run every so often if needed
Change-Id: I68da342bbc1d67653901cf4502dabfa5bc947628
this simplifies waitForInput quite a lot, and at the same time makes
polling less thundering-herd-y. it even fixes early polling wakeups!
Change-Id: I6dfa62ce91729b8880342117d71af5ae33366414
this removes the rather janky did-you-mean-async poll loop we had so
far. sadly kj does not play well with pty file descriptors, so we do
have to add our own async input stream that does not eat pty EIO and
turns it into an exception. that's still a *lot* better than the old
code, and using a real even loop makes everything else easier later.
Change-Id: Idd7e0428c59758602cc530bcad224cd2fed4c15e
like kj::joinPromisesFailFast this allows waiting for the results of
multiple promises at once, but unlike it not all input promises must
be complete (or any of them failed) for results to become available.
Change-Id: I0e4a37e7bd90651d56b33d0bc5afbadc56cde70c
like a normal semaphore, but with awaitable acquire actions. this is
primarily intended as an intermediate concurrency limiting device in
the Worker code, but it may find other uses over time. we do not use
std::counting_semaphore as a base because the counter of that is not
inspectable as will be needed for Worker. we also do not need atomic
operations for cross-thread consistency since we don't have multiple
threads (thanks to kj event loops being confined to a single thread)
Change-Id: Ie2bcb107f3a2c0185138330f7cbba4cec6cbdd95
while gcc 12 and older miscompile our generators, gcc 13 and older
outright crash on kj coroutines. (newer gcc versions may fix this)
Change-Id: I19f12c8c147239680eb0fa5a84ef5c7de38c9263
using a proper event loop basis we no longer have to worry about most of
the intricacies of poll(), or platform-dependent replacements for it. we
may even be able to use the event loop and its promise system for all of
our scheduling in the future. we don't do any real async processing yet,
this is just preparation to separate the first such change from the huge
api design difference with the async framework we chose (kj from capnp):
kj::Promise, unlike std::future, doesn't return exceptions unmangled. it
instead wraps any non-kj exception into a kj exception, erasing all type
information and preserving mostly the what() string in the process. this
makes sense in the capnp rpc use case where unrestricted exception types
can't be transferred, and since it moves error handling styles closer to
a world we'd actually like there's no harm in doing it only here for now
Change-Id: I20f888de74d525fb2db36ca30ebba4bcfe9cc838
we're using boost::outcome rather than leaf or stl types because stl
types are not available everywhere and leaf does not provide its own
storage for error values, relying on thread-locals and the stack. if
we want to use promises we won't have a stack and would have to wrap
everything into leaf-specific allocating wrappers, so outcome it is.
Change-Id: I35111a1f9ed517e7f12a839e2162b1ba6a993f8f
it just makes sense to have it too, rather than just the pass/fail
information we keep so far. once we turn goals into something more
promise-shaped it'll also help detangle the current data flow mess
Change-Id: I915cf04d177cad849ea7a5833215d795326f1946
it doesn't have a purpose except cache priming, which is largely
irrelevant by default (since another code path already runs this
exact query). our store implementations do not benefit that much
from this either, and the more bursty load may indeed harm them.
Change-Id: I1cc12f8c21cede42524317736d5987f1e43fc9c9
updating statistics *immediately* when any counter changes declutters
things somewhat and makes useful status reports less dependent on the
current worker main loop. using callbacks will make it easier to move
the worker loop into kj entirely, using only promises for scheduling.
Change-Id: I695dfa83111b1ec09b1a54cff268f3c1d7743ed6
there's no reason to go through the event loop in these cases. returning
ContinueImmediately here is just a very convoluted way of jumping to the
state we've just set after unwinding one frame of the stack, which never
matters in the cases changed here because there are no live RAII guards.
Change-Id: I7c00948c22e3caf35e934c1a14ffd2d40efc5547
this is not ideal, but it's better than having this stuck in the worker
loop itself. setting ex on all failing goals is not problematic because
only toplevel goals can ever be observable, all the others are ignored.
notably only derivation goals ever set `ex`, substitution goals do not.
Change-Id: I02e2164487b2955df053fef3c8e774d557aa638a
this doesn't serve a great purpose yet except to confine construction of
goals to the stack frame of Worker::run() and its child frames. we don't
need this yet (and the goal constructors remain fully visible), but in a
future change that fully removes the current worker loop we'll need some
way of knowing which goals are top-level goals without passing the goals
themselves around. once that's possible we can remove visible goals as a
concept and rely on build result futures and a scheduler built upon them
Change-Id: Ia73cdeffcfb9ba1ce9d69b702dc0bc637a4c4ce6
whether goal errors are reported via the `ex` member or just printed to
the log depends on whether the goal is a toplevel goal or a dependency.
if goals are aware of this themselves we can move error printing out of
the worker loop, and since a running worker can only be used by running
goals it's totally sufficient to keep a `Worker::running` flag for this
Change-Id: I6b5cbe6eccee1afa5fde80653c4b968554ddd16f
the current test relies on derivation build order being deterministic,
which will not be a reasonable expectation for all that long any more.
Change-Id: I9be44a7725185f614a9a4c724045b8b1e6962c03
this makes WorkResult copyable, and just all around easier to deal with.
in the future we'll need this to let Goal::work() return a promise for a
WorkResult (or even just a Finished) that can be awaited by other goals.
Change-Id: Ic5a1ce04c5a0f8e683bd00a2ed2b77a2e28989c1
this should be done where we're actually trying to build something, not
in the main worker loop that shouldn't have to be aware of such details
Change-Id: I07276740c0e2e5591a8ce4828a4bfc705396527e
this can be a proper WorkResult now. childTerminated is unfortunately a
lot more stubborn and won't be made private for quite a while yet. once
we can get rid of the Worker poll loop that *should* be possible though
Change-Id: I2218df202da5cb84e852f6a37e4c20367495b617
we'll need this once we want to pass extra information out of accepting
replies, such as fd sets or possibly even async output reader promises.
Change-Id: I5e2f18cdb80b0d2faf3067703cc18bd263329b3f
don't keep fds open we're not using. currently this does not cause any
problems, but it does increase the size of our fd table needlessly and
in the future, when we have proper async processing, having builderOut
open in the daemon once the hook has been fully started is problematic
Change-Id: I6e7fb773b280b042873103638d3e04272ca1e4fc
this is useless to do on the face of it, but it'll make it easier to
convert the entire output handling to use async io and promises soon
Change-Id: I2d1eb62c4bbf8f57bd558b9599c08710a389b1a8
only DerivationGoal can set the hook to anything at all. it always sets
buildOutFD to something that is not related to fromHook in any way, and
mixing the two would have rather dire consequences for log consistency.
Change-Id: Ida86727fd1cd5e1ecd78f07f3bde330a346658a8
all derivation goals need a log fd of some description. let's save this
single fd in a dedicated pointer field for all subclasses so that later
we have just the one spot to change if we turn this into async promises
Change-Id: If223adf90909247363fb823d751cae34d25d0c0b
we don't need to expose information about how busy a Worker is if the
worker can instead tell its work items whether they are in a slot. in
the future we might use this to not start items waiting for a slot if
no slots are currently available, but that requires more preparation.
Change-Id: Ibe01ac536da7e6d6f80520164117c43e772f9bd9
this is only used to close non-stdio files in derivation sandboxes. we
may as well encode that in its name, drop the unnecessary integer set,
and use close_range to deal with the actual closing of files. not only
is this clearer, it also makes sandbox setup on linux fast by 1ms each
Change-Id: Id90e259a49c7bc896189e76bfbbf6ef2c0bcd3b2
implementing a build hook is pretty much impossible without either being
a nix, or blindly forwarding the important bits of all build requests to
some kind of nix. we've found no uses of build-hook in the wild, and the
build-hook protocol (apart from being entirely undocumented) is not able
to convey any kind of versioning information between hook and daemon. if
we want to upgrade this infrastructure (which we do), this must not stay
Change-Id: I1ec4976a35adf8105b8ca9240b7984f8b91e147e
this is a bit of a hack, but it's apparently the cleanest way of doing
this in the absence of any kind of priority/provenance information for
values of some given setting. we'll need this to deprecate build-hook.
Change-Id: I03644a9c3f17681c052ecdc610b4f1301266ab9e
sure, linux has been providing argv[0] by default for a while now. other
OSes may not be as forthcoming though, and relying on the OS to create a
world in which we can just make assumptions we could test for instead is
unnecessarily lazy. we *could* default argv0, but that's a little silly.
notably we abort instead of returning normally to avoid confusions where
a caller interprets our exit status like a Worker build results bitmask.
Change-Id: Id73f8cd0a630293b789c59a8c4b0c4a2b936b505
eldritch horrors