summaryrefslogtreecommitdiff
path: root/libraries
diff options
context:
space:
mode:
authorNiklas Hambüchen <mail@nh2.me>2017-02-11 19:24:49 -0500
committerBen Gamari <ben@smart-cactus.org>2017-02-11 19:58:35 -0500
commit6b4e46a1ad57f90726de7e5b7f59814aea5f03ef (patch)
tree8685dc16186d47440ff5c2a8becfe2a93dd3d899 /libraries
parent805db96544111bd548c9a32488a9c97996cc2b49 (diff)
downloadhaskell-6b4e46a1ad57f90726de7e5b7f59814aea5f03ef.tar.gz
bufWrite: Save extra syscall when data fills handle buffer completely.
The bug is that the check `if (size - w > count)` should be `if (size - w >= count)` instead (`>=` instead of `>`), because we can do the write all fine if it fits exactly. This allows us to do 1 instead of 2 write syscalls the case it fits. An example of when this matters is when an application writes output in chunks that are a fraction of the handle buffer size. For example, assume the buffer size is 8 KB, and the application writes four 2 KB chunks. Until now, this would result in 3 copies to the handle buffer, but the 4th one would not be allowed in by `size - w > count` (because `size - w == count` is the case), so we'd end up with a write syscall of only 6 KB data instead of 8 KB, thus creating more syscalls overall. Implementing this fix (switching to `size - w >= count`), we also have to flush the buffer if we fill it completely. If we made only the changes described so far, that would have the unintended side effect that writes of the size equal to the handle buffer size (`count == size`) suddenly also go to the handle buffer first: The data would first be copied to the handle buffer, and then immediately get flushed to the underlying FD. We don't want that extra `memcpy`, because it'd be unnecessary: The point of handle buffering is to coalesce smaller writes, and there are no smaller writes in this case. For example, if you specify 8 KB buffers (which menas you want your data to be written out in 8 KB blocks), and you get data that's already 8 KB in size, you can write that out as an 8 KB straight away, zero-copy fashion. For this reason, adding to the handle buffer now got an additional condition `count < size`. That way, writes equal to the buffer size go straight to the FD, as they did before this commit. Reviewers: simonmar, austin, hvr, bgamari Reviewed By: simonmar Subscribers: mpickering, thomie Differential Revision: https://phabricator.haskell.org/D3117
Diffstat (limited to 'libraries')
-rw-r--r--libraries/base/GHC/IO/Handle/Text.hs25
1 files changed, 21 insertions, 4 deletions
diff --git a/libraries/base/GHC/IO/Handle/Text.hs b/libraries/base/GHC/IO/Handle/Text.hs
index c53b4294b8..8d927384c8 100644
--- a/libraries/base/GHC/IO/Handle/Text.hs
+++ b/libraries/base/GHC/IO/Handle/Text.hs
@@ -736,13 +736,30 @@ bufWrite h_@Handle__{..} ptr count can_block =
old_buf@Buffer{ bufRaw=old_raw, bufR=w, bufSize=size }
<- readIORef haByteBuffer
- -- enough room in handle buffer?
- if (size - w > count)
- -- There's enough room in the buffer:
+ -- TODO: Possible optimisation:
+ -- If we know that `w + count > size`, we should write both the
+ -- handle buffer and the `ptr` in a single `writev()` syscall.
+
+ -- Need to buffer and enough room in handle buffer?
+ -- There's no need to buffer if the data to be written is larger than
+ -- the handle buffer (`count >= size`).
+ if (count < size && count <= size - w)
+ -- We need to buffer and there's enough room in the buffer:
-- just copy the data in and update bufR.
then do debugIO ("hPutBuf: copying to buffer, w=" ++ show w)
copyToRawBuffer old_raw w ptr count
- writeIORef haByteBuffer old_buf{ bufR = w + count }
+ let copied_buf = old_buf{ bufR = w + count }
+ -- If the write filled the buffer completely, we need to flush,
+ -- to maintain the "INVARIANTS on Buffers" from
+ -- GHC.IO.Buffer.checkBuffer: "a write buffer is never full".
+ if (count == size - w)
+ then do
+ debugIO "hPutBuf: flushing full buffer after writing"
+ flushed_buf <- Buffered.flushWriteBuffer haDevice copied_buf
+ -- TODO: we should do a non-blocking flush here
+ writeIORef haByteBuffer flushed_buf
+ else do
+ writeIORef haByteBuffer copied_buf
return count
-- else, we have to flush any existing handle buffer data