Adds `WalkBuilder::filter_entry` that takes a predicate to be applied to
all entries. If the predicate returns `false` on a given entry, that
entry and all children will be skipped.
Fixes#1555, Closes#1557
While Linux distributions (at least Arch Linux, RHEL, Debian) do not support
compressing files with compress(1), macOS & AIX do (the utility is part of
POSIX). Additionally, gzip is able to uncompress such compressed files and
provides an `uncompress` binary.
Closes#1547
This replaces the use of channels in the parallel directory traversal
with a simple stack. The primary motivation for this change is to reduce
peak memory usage. In particular, when using a channel (which is a
queue), we wind up visiting files in a breadth first fashion. Using a
stack switches us to a depth first traversal. While there are no real
intrinsic differences, depth first traversal generally tends to use less
memory because directory trees are more commonly wide than they are
deep.
In particular, the queue/stack size itself is not the only concern. In
one recent case documented in #1550, a user wanted to search all Rust
crates. The directory structure was shallow but extremely wide, with a
single directory containing all crates. This in turn results is in
descending into each of those directories and building a gitignore
matcher for each (since most crates have `.gitignore` files) before ever
searching a single file. This means that ripgrep has all such matchers
in memory simultaneously, which winds up using quite a bit of memory.
In a depth first traversal, peak memory usage is much lower because
gitignore matches are built and discarded more quickly. In the case of
searching all crates, the peak memory usage decrease is dramatic. On my
system, it shrinks by an order magnitude, from almost 1GB to 50MB. The
decline in peak memory usage is consistent across other use cases as
well, but is typically more modest. For example, searching the Linux
repo has a 50% decrease in peak memory usage and searching the Chromium
repo has a 25% decrease in peak memory usage.
Search times generally remain unchanged, although some ad hoc benchmarks
that I typically run have gotten a bit slower. As far as I can tell,
this appears to be result of scheduling changes. Namely, the depth first
traversal seems to result in searching some very large files towards the
end of the search, which reduces the effectiveness of parallelism and
makes the overall search take longer. This seems to suggest that a stack
isn't optimal. It would instead perhaps be better to prioritize
searching larger files first, but it's not quite clear how to do this
without introducing more overhead (getting the file size for each file
requires a stat call).
Fixes#1550
We should not assume that the commondir file actually exists. If it
doesn't, then just move on. This otherwise emits an error message when
searching normal submodules, which is not OK.
This regression was introduced in #1446.
Fixes#1520
We can just ask the channel whether any work has been loaded. Normally
querying a channel for its length is a strong predictor of bugs, but in
this case, we do it before we ever attempt a `recv`, so it should work.
Kudos to @zsugabubus for suggesting this!
It turns out that the previous version wasn't quite correct. Namely, it
was possible for the following sequence to occur:
1. Consider that all workers, except for one, are `waiting`.
2. The last remaining worker finds one more job to do and sends it on
the channel.
3. One of the previously `waiting` workers wakes up from the job that
the last running worker sent, but `self.resume()` has not been
called yet.
4. The last worker, from (2), calls `get_work` and sees that the
channel has nothing on it, so it executes `self.waiting() ==
1`. Since the worker in (3) hasn't called `self.resume()` yet,
`self.waiting() == 1` evaluates to true.
5. This sets off a chain reaction that stops all workers, despite that
fact that (3) got more work (which could itself spawn more work).
The end result is that the traversal may terminate while their are still
outstanding work items to process. This problem was observed through
spurious failures in CI. I was not actually able to reproduce the bug
locally.
We fix this by changing our strategy to detect termination using a
counter. Namely, we increment the counter just before sending new work
and decrement the counter just after finishing work. In this way, we
guarantee that the counter only ever reaches 0 once there is no more
work to process.
See #1337 for more discussion. Many thanks to @zsugabubus for helping me
work through this.
The top-level listing was just getting a bit too long for my taste. So
put all of the code in one directory and shrink the large top-level mess
to a small top-level mess.
NOTE: This commit only contains renames. The subsequent commit will
actually make ripgrep build again. We do it this way with the naive hope
that this will make it easier for git history to track the renames.
Sigh.
