Change the test to be robust in slow environments and move to parallel.
The previous version of the test failed for me in parallel with just two
or four simultaneous versions running. This version passes 96
simultaneous versions running, but still fails as designed if the
request writes fail to prevent the request timeout from occurring.
PR-URL: https://github.com/nodejs/node/pull/30677
Reviewed-By: Gireesh Punathil <gpunathi@in.ibm.com>
Reviewed-By: James M Snell <jasnell@gmail.com>
test-http-server-consumed-timeout has code to that causes it to be
skipped on busy machines. Instead, use an exponential backoff for the
timeout if the machine is busy.
PR-URL: https://github.com/nodejs/node/pull/30677
Reviewed-By: Gireesh Punathil <gpunathi@in.ibm.com>
Reviewed-By: James M Snell <jasnell@gmail.com>
Replace setTimeout() with setImmediate() in
test-stream-writable-clear-buffer. The test still fails in Node.js 8.6.0
(if you comment out the `common` module, since it uses BigInts and those
aren't supported in 8.6.0).
PR-URL: https://github.com/nodejs/node/pull/30561
Reviewed-By: Denys Otrishko <shishugi@gmail.com>
Reviewed-By: Anna Henningsen <anna@addaleax.net>
This adds long-requested methods for asynchronously interacting and
iterating through directory entries by using `uv_fs_opendir`,
`uv_fs_readdir`, and `uv_fs_closedir`.
`fs.opendir()` and friends return an `fs.Dir`, which contains methods
for doing reads and cleanup. `fs.Dir` also has the async iterator
symbol exposed.
The `read()` method and friends only return `fs.Dirent`s for this API.
Having a entry type or doing a `stat` call is deemed to be necessary in
the majority of cases, so just returning dirents seems like the logical
choice for a new api.
Reading when there are no more entries returns `null` instead of a
dirent. However the async iterator hides that (and does automatic
cleanup).
The code lives in separate files from the rest of fs, this is done
partially to prevent over-pollution of those (already very large)
files, but also in the case of js allows loading into `fsPromises`.
Due to async_hooks, this introduces a new handle type of `DIRHANDLE`.
This PR does not attempt to make complete optimization of
this feature. Notable future improvements include:
- Moving promise work into C++ land like FileHandle.
- Possibly adding `readv()` to do multi-entry directory reads.
- Aliasing `fs.readdir` to `fs.scandir` and doing a deprecation.
Refs: https://github.com/nodejs/node-v0.x-archive/issues/388
Refs: https://github.com/nodejs/node/issues/583
Refs: https://github.com/libuv/libuv/pull/2057
PR-URL: https://github.com/nodejs/node/pull/29349
Reviewed-By: Anna Henningsen <anna@addaleax.net>
Reviewed-By: David Carlier <devnexen@gmail.com>
This improves dgram performance by avoiding unnecessary async
operations.
One issue with this commit is that it seems hard to actually create
conditions under which the fallback path to the async case is
actually taken, for all supported OS, so an internal CLI option
is used for testing that path.
Another caveat is that the lack of an async operation means
that there are slight timing differences (essentially `nextTick()`
rather than `setImmediate()` for the send callback).
PR-URL: https://github.com/nodejs/node/pull/29832
Reviewed-By: David Carlier <devnexen@gmail.com>
Reviewed-By: James M Snell <jasnell@gmail.com>
Reviewed-By: Ben Noordhuis <info@bnoordhuis.nl>
test-cluster-net-listen-ipv6only-none was using port `0` for an
IPv6-only operation and assuming that the operating system would supply
a port that was also available in IPv4. However, CI results seem to
indicate that a port can be supplied that is in use by IPv4 but
available to IPv6, resulting in the test failing. Use `common.PORT` to
avoid this issue.
Fixes: https://github.com/nodejs/node/issues/29679
PR-URL: https://github.com/nodejs/node/pull/29681
Reviewed-By: Sam Roberts <vieuxtech@gmail.com>
This simplifies the implementation of ELDHistogram a bit,
and more generally allows us to have weak JS references
associated with `HandleWrap`s.
PR-URL: https://github.com/nodejs/node/pull/29317
Reviewed-By: James M Snell <jasnell@gmail.com>
This test imposes a limit on the average bytes of space per chunk
for network traffic. However this number depends on VM
implementation details, and upcoming changes to V8's
array buffer management require a small bump to this
limit in this test.
PR-URL: https://github.com/nodejs/node/pull/28492
Reviewed-By: Anna Henningsen <anna@addaleax.net>
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
Reviewed-By: Rich Trott <rtrott@gmail.com>
Reviewed-By: Luigi Pinca <luigipinca@gmail.com>
Reviewed-By: Ruben Bridgewater <ruben@bridgewater.de>
Historically `error.errno` of system errors thrown by Node.js
can sometimes be the same as `err.code`, which are string
representations of the error numbers. This is useless and incorrect,
and results in an information loss for users since then they
will have to resort to something like
`process.binding('uv'[`UV_${errno}`])` to get to the numeric
error codes.
This patch corrects this behavior by always setting `error.errno`
to be negative numbers. For fabricated errors like `ENOTFOUND`,
`error.errno` is now undefined since there is no numeric equivalent
for them anyway. For c-ares errors, `error.errno` is now undefined
because the numeric representations (negated) can be in conflict
with libuv error codes - this is fine since numeric codes was
not available for c-ares errors anyway.
Users can use the public API `util.getSystemErrorName(errno)`
to retrieve string codes for these numbers.
PR-URL: https://github.com/nodejs/node/pull/28140
Reviewed-By: James M Snell <jasnell@gmail.com>
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
Reviewed-By: Rich Trott <rtrott@gmail.com>
Reviewed-By: Trivikram Kamat <trivikr.dev@gmail.com>
The tests still fail after being split into multiple files,
(2 out of 30 runs in roughly 48 hours) and the causes are missing
target frames in the samples. This patch moves them to sequential
to observe if the flakiness can be fixed when the tests are
run on a system with less load.
If the flake ever shows up again even after the tests are moved
to sequential, we should consider make the test conditions more
lenient - that is, we would only assert that there are *some* frames
in the generated CPU profile but do not look for the target
function there.
PR-URL: https://github.com/nodejs/node/pull/28210
Refs: https://github.com/nodejs/node/issues/27611
Reviewed-By: Rich Trott <rtrott@gmail.com>
Reviewed-By: Michael Dawson <michael_dawson@ca.ibm.com>
Reviewed-By: James M Snell <jasnell@gmail.com>
Lots of changes, but mostly just search/replace of
fixtures.readSync(...) to fixtures.readKey([new key]...)
Benchmarks modified to use fixtures.readKey(...):
benchmark/tls/throughput.js
benchmark/tls/tls-connect.js
benchmark/tls/secure-pair.js
Also be sure to review the change to L16 of
test/parallel/test-crypto-sign-verify.js
PR-URL: https://github.com/nodejs/node/pull/27962
Reviewed-By: Sam Roberts <vieuxtech@gmail.com>
Reviewed-By: Ujjwal Sharma <usharma1998@gmail.com>
Reviewed-By: Rich Trott <rtrott@gmail.com>
Lots of changes, but mostly just search/replace of
fixtures.readSync(...) to fixtures.readKey([new key]...)
Benchmarks modified to use fixtures.readKey(...):
benchmark/tls/throughput.js
benchmark/tls/tls-connect.js
benchmark/tls/secure-pair.js
Also be sure to review the change to L16 of
test/parallel/test-crypto-sign-verify.js
PR-URL: https://github.com/nodejs/node/pull/27962
Reviewed-By: Sam Roberts <vieuxtech@gmail.com>
Reviewed-By: Ujjwal Sharma <usharma1998@gmail.com>
Reviewed-By: Rich Trott <rtrott@gmail.com>
This patch implements --cpu-prof-interval to specify the sampling
interval of the CPU profiler started by --cpu-prof from the command
line. Also adjust the interval to 100 in test-cpu-prof.js to make
the test less flaky - it would fail if the time taken to finish
the workload is smaller than the sampling interval, which was
more likely on powerful machines when the interval was 1000.
PR-URL: https://github.com/nodejs/node/pull/27535
Reviewed-By: Jan Krems <jan.krems@gmail.com>
Reviewed-By: Anna Henningsen <anna@addaleax.net>
Reviewed-By: Franziska Hinkelmann <franziska.hinkelmann@gmail.com>
Reviewed-By: Rich Trott <rtrott@gmail.com>
Fix some issues introduced/not fixed via
https://github.com/nodejs/node/pull/25094:
* Init hook is not emitted for a reused HTTPParser
* HTTPParser was still used as resource in init hook
* type used in init hook was always HTTPINCOMINGMESSAGE even for client
requests
* some tests have not been adapted to new resource names
With this change the async hooks init event is emitted during a call
to Initialize() as the type and resource object is available at this
time. As a result Initialize() must be called now which could be seen
as breaking change even HTTPParser is not part of documented API.
It was needed to put the ClientRequest instance into a wrapper object
instead passing it directly as async resource otherwise
test-domain-multi fails. I think this is because adding an EventEmitter
to a Domain adds a property 'domain' and the presence of this changes
the context propagation in domains.
Besides that tests still refering to resource HTTPParser have been
updated/improved.
Fixes: https://github.com/nodejs/node/issues/27467
Fixes: https://github.com/nodejs/node/issues/26961
Refs: https://github.com/nodejs/node/pull/25094
PR-URL: https://github.com/nodejs/node/pull/27477
Reviewed-By: Anna Henningsen <anna@addaleax.net>
Reviewed-By: Matteo Collina <matteo.collina@gmail.com>
Reviewed-By: Rich Trott <rtrott@gmail.com>
Rich Trott