* src/cuirass/database.scm (db-add-derivation): Catch 'sqlite-error and
handle SQLITE_CONSTRAINT_PRIMARYKEY.
(SQLITE_CONSTRAINT_UNIQUE): New variable.
* tests/database.scm ("database")["db-add-derivation"]: Add extra call to
'db-add-derivation'.
Before that, 'run-server' would force sequential processing of client
requests one after another.
* src/cuirass/http.scm (run-cuirass-server): Rewrite to use its own loop
instead of 'run-server'.
That makes it safe from SQL injection.
* src/cuirass/database.scm (db-get-builds): Rewrite to use question
marks in SQL queries and binding through '%sqlite-exec'.
* tests/database.scm ("database")["db-get-builds"]: Exercise 'WHERE'
clauses.
The new macro automatically takes care of inserting question marks in
the SQL queries, which in turn guarantees that there are always as many
question marks and arguments.
* src/cuirass/database.scm (sqlite-exec): Rename to...
(%sqlite-exec): ... this.
(sqlite-exec/bind, sqlite-exec): New macros.
(assq-refs): Remove.
(db-add-specification): Use the new 'sqlite-exec' form.
(db-get-specifications): Correctly deal with REV or TAG being #f.
(db-add-derivation, db-get-derivation, db-add-evaluation)
(db-add-build, db-update-build-status!, db-get-outputs)
(db-get-build, db-get-stamp, db-add-stamp): Adjust to the new
'sqlite-exec' form.
Previously any derivation not directly built by Cuirass would be
considered as failed because 'handle-build-event' wouldn't see any build
event. Here we just make sure the build status recorded in the database
corresponds to reality.
* src/cuirass/base.scm (update-build-statuses!): New procedure.
(spawn-builds): Call it after 'build-derivations&'.
* src/cuirass/database.scm (db-update-build-status!): And "AND status != ?"
in SQL queries.
* tests/database.scm ("database")["db-update-build-status!"]: Add call
to 'db-update-build-status!'.
This is purely "defensive programming" since the default value for
'print-build-trace' is currently #t.
* src/cuirass/base.scm (with-store): Pass #:print-build-trace to
'set-build-options'.
* src/cuirass/database.scm (%sqlite-exec): Remove.
(sqlite-exec): Turn back into a procedure. Use 'sqlite-bind'. Add
'normalize' procedure and use it.
(db-add-specification, db-add-derivation, db-get-derivation)
(db-add-evaluation, db-add-build, db-update-build-status!)
(db-get-build, db-get-stamp, db-add-stamp): Use question marks in SQL
queries.
* src/cuirass/base.scm (build-packages)[register]: Make #:log
non-false.
* tests/database.scm (make-dummy-job): Add #:job-name, #:system,
#:nix-name, and #:eval-id. This is necessary because 'sqlite-bind'
would now translate #f to a real NULL (before it would translate to the
string "#f"...), and would thus report violations of the non-NULL
constraint.
Co-authored-by: Ludovic Courtès <ludo@gnu.org>
Reported by Danny Milosavljevic.
* src/cuirass/http.scm (url-handler) <"jobsets"> Return the result of
'db-get-specifications' as-is, not just the car.
Works around <https://github.com/wingo/fibers/issues/19>.
The effect in practice would be that we'd usually not reach the
'close-pipe' call in 'evaluate', leaving zombie processes behind us,
never executing the continuation, and additionally spinning fast on a
sequence of epoll_wait/epoll_ctl calls.
* src/cuirass/base.scm <top level>: Monkey-patch (fibers internal).
We used to have 'build-derivations' write to the custom binary port
returned by 'build-event-output-port'. However, custom binary ports
constitute continuation barriers, thereby preventing fibers from being
suspended.
To make build log processing non-blocking, we therefore invert this
inversion of control and use a suspendable I/O procedure,
'read-line/non-blocking', when reading the build log.
* src/cuirass/base.scm (read-line/non-blocking, process-build-log)
(build-derivations&): New procedures.
(%newline, build-event-output-port): Remove.
(spawn-builds): Use 'build-derivations&' instead of 'build-derivations'
with 'build-event-output-port'.
This should placate 'guix-hydra-queued-builds', for instance, which
checks these values to choose between "Running" and "Scheduled".
* src/cuirass/http.scm (build->hydra-build): Provide the correct value
for #:finished and #:busy.
Fixes a bug whereby 'build-event-output-port' would sometimes read more
than COUNT from BV, which would usually result in a 'decoding-error'
exception from 'utf8->string'.
* src/cuirass/base.scm (build-event-output-port)[write!]: Use
'bytevector-range' to honor COUNT and OFFSET.
Previously these calls would effectively suspend execution of all the
fibers until they had completed.
* src/cuirass/base.scm (process-specs): Wrap 'fetch-repository' and
'compile' calls in 'non-blocking'.
* src/cuirass/logging.scm (call-with-time-logging): New procedure.
(with-time-logging): New macro.
* src/cuirass/http.scm (handle-builds-request): Use it.
* src/cuirass/base.scm (handle-build-event)[valid?]: New procedure.
Use it when handling 'build-started', 'build-succeeded', and
'build-failed' events.
* src/cuirass/database.scm (db-get-builds)[format-order-clause]: Add
'order' clauses.
* src/cuirass/http.scm (url-handler): Default to (order finish-time) for
/latestbuilds and (order submission-time) for /queue.
This works around a scalability issue in guix-daemon when passing a long
list of derivations to 'build-derivations'.
* src/cuirass/base.scm (spawn-builds): New procedure.
(restart-builds, build-packages): Use it.
Ludovic Courtès
Danny Milosavljevic