net/http: short writes with FileServer on macos

[a-palchikov]

Go version

go version go1.23.2 darwin/arm64

Output of go env in your module/workspace:

GO111MODULE=''
GOARCH='arm64'
GOBIN=''
GOCACHE='/user/Library/Caches/go-build'
GOENV='/user/Library/Application Support/go/env'
GOEXE=''
GOEXPERIMENT=''
GOFLAGS=''
GOHOSTARCH='arm64'
GOHOSTOS='darwin'
GOINSECURE=''
GOMODCACHE='/user/go/pkg/mod'
GONOPROXY=''
GONOSUMDB=''
GOOS='darwin'
GOPATH='/user/go'
GOPRIVATE=''
GOPROXY='https://proxy.golang.org,direct'
GOROOT='/user/dev/go/go'
GOSUMDB='sum.golang.org'
GOTMPDIR=''
GOTOOLCHAIN='auto'
GOTOOLDIR='/user/dev/go/go/pkg/tool/darwin_arm64'
GOVCS=''
GOVERSION='go1.23.2'
GODEBUG='httpmuxgo121=0'
GOTELEMETRY='local'
GOTELEMETRYDIR='/user/Library/Application Support/go/telemetry'
GCCGO='gccgo'
GOARM64='v8.0'
AR='ar'
CC='clang'
CXX='clang++'
CGO_ENABLED='1'
GOMOD='/dev/null'
GOWORK='/user/dev/path/to/go.work'
CGO_CFLAGS='-O2 -g'
CGO_CPPFLAGS=''
CGO_CXXFLAGS='-O2 -g'
CGO_FFLAGS='-O2 -g'
CGO_LDFLAGS='-O2 -g'
PKG_CONFIG='pkg-config'
GOGCCFLAGS='-fPIC -arch arm64 -pthread -fno-caret-diagnostics -Qunused-arguments -fmessage-length=0 -ffile-prefix-map=/var/folders/_y/<snip>/T/go-build<id>=/tmp/go-build -gno-record-gcc-switches -fno-common'

What did you do?

I have a small webserver that serves static files in a testing setup.

What did you see happen?

I'm on go 1.23.2 on macos 14.7 and I'm having issues with the stdlib http.FileServer.
After much of poking around, I found that with the BSD sendfile implementation in internal/poll/sendfile_bsd.go having
the maxSendfileSize set to 4MiB is causing all large file transfers to fail. For example, the following is the curl failure:

...
{ [512 bytes data]
* transfer closed with 361300436 bytes remaining to read
  1  350M    1 5567k    0     0  80.7M      0  0:00:04 --:--:--  0:00:04 81.1M
* Closing connection
curl: (18) transfer closed with 361300436 bytes remaining to read

so all requests are failing with short writes.

If I update the maxSendfileSize to maximum (adopting the change from this commit) I don't observe any failures.

Just setting the maxSendfileSize feels like a hack though (maybe I'm wrong) so I'd appreciate any hints on this behavior.

What did you expect to see?

The sendfile API should be able to send files in their entirety.

[gabyhelp]

Related Issues and Documentation

• net/http: FileServer on WSL fails with wget and curl for files over ~4MB #27128 (closed)
• net/http: ServeFile fails to serve files over 4GB on Windows #33193 (closed)
• net: sendFile can't handle os.fileWithoutWriteTo #66988 (closed)
• net/http: ServeFile adds garbage at end of file - Stat() gets wrong size from Mac OS #24423 (closed)
• net/http: FileServer no longer serves content for POST #59375 (closed)
• net/http: #62547 (closed)
• net/http: using io.Copy to copy a file to http.ResponseWriter does not use sendfile #69097

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[ianlancetaylor]

Can you show us code that we can use the replicate the problem? Thanks.

CC @golang/darwin

[ianlancetaylor]

CC @panjf2000

[a-palchikov]

There's nothing special to it, really:

m := http.NewServeMux()
m.Handle("/large-payload/", http.StripPrefix("/large-payload/", http.FileServer(http.Dir("path/to/large-payload"))))

it looks at a directory with a bunch of numeric files all 350MiB in size:

$ tree path/to/large-payload | head -n3
path/to/large-payload
├── 1
├── 10
...

and I simply hit the server with:

curl localhost:44444/large-payload/1 -o ./test

In an OK case, this results in:

  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  350M  100  350M    0     0  1199M      0 --:--:-- --:--:-- --:--:-- 1202M

In the error case (go 1.23.2):

  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
  2  350M    2 8031k    0     0   361M      0 --:--:-- --:--:-- --:--:--  373M
curl: (18) transfer closed with 358777300 bytes remaining to read

[panjf2000]

Can we get the error returned by your server-side? This 970b1c0 will ship in Go1.24.

[a-palchikov]

@panjf2000 I'll gladly comply if you can narrow the context. On high-level it writes short here:

go/src/net/http/fs.go

Line 424 in ed07b32

io.CopyN(w, sendContent, sendSize)

From what I observed, there's no error, e.g.:

go/src/net/sendfile_unix_alt.go

Line 77 in ed07b32

written, werr, handled = poll.SendFile(&c.pfd, int(fd), pos, remain)

returns with

written=4194304, werr=<nil>, handled=true

the number of bytes written fluctuates, e.g.:

written=12524536, werr=<nil>, handled=true

the main point is that it's a short write.

Please let me know if you need more details.

[jiahongli18]

Can we get the error returned by your server-side? This 970b1c0 will ship in Go1.24.

As @a-palchikov mentioned, we don't see any errors from the server itself. The observation from the client is a partially fetched file.

[prattmic]

cc @neild

[a-palchikov]

This seems to do the trick:

diff --git a/src/internal/poll/sendfile_bsd.go b/src/internal/poll/sendfile_bsd.go
index 669df94cc1..35338730aa 100644
--- a/src/internal/poll/sendfile_bsd.go
+++ b/src/internal/poll/sendfile_bsd.go
@@ -47,7 +47,7 @@ func SendFile(dstFD *FD, src int, pos, remain int64) (written int64, err error,
                // don't implement sendfile), and other errors.
                // We should end the loop when there is no error
                // returned from sendfile(2) or it is not a retryable error.
-               if err != syscall.EAGAIN {
+               if err != nil && err != syscall.EAGAIN {
                        break
                }
                if err = dstFD.pd.waitWrite(dstFD.isFile); err != nil {

The sequence of events is like following:

n= 1408044 , err= resource temporarily unavailable
n= 434176 , err= resource temporarily unavailable
n= 4194304

given the series of EAGAIN followed by no error, when it reaches the limit (4MiB), it bails out.

[ianlancetaylor]

@a-palchikov Thanks. Looks like we need something similar in the other sendfile implementations as well.

[jonjohnsonjr]

Just wanted to chime in that this exhibits worse behavior on the client side when uploading large files. I have a client that will hang indefinitely on macOS pretty regularly, which I believe is due to this issue, as it uses an os.File as the Body of a request.

This was a bit of a heisenbug for me because when I tried wrapping the os.File in another struct to observe what was going on, the problem disappeared. That was enough of a clue to lead me here (and also serve as a workaround until it's fixed).

Thanks for uncovering this!

[gopherbot]

Change https://go.dev/cl/622235 mentions this issue: internal/poll: keep copying after successful Sendfile return on BSD