[fix] Socket: Fix stale indexes and surface close/error reliably in _hx_std_socket_poll_events - #1252
Open
dimensionscape wants to merge 5 commits into
Open
[fix] Socket: Fix stale indexes and surface close/error reliably in _hx_std_socket_poll_events#1252dimensionscape wants to merge 5 commits into
dimensionscape wants to merge 5 commits into
Conversation
…ents` on error/empty sets Previously, `_hx_std_socket_poll_events` could exit early on `select()/poll()` errors without resetting the ready index arrays. This left stale values in `p->ridx/p->widx`, causing incorrect results when the socket set was empty or when sockets were removed. In Haxe this manifested as phantom entries (e.g. index "1" still marked ready after deregistration, or `poll([])` returning `[null]`). This change ensures `p->ridx[0]` and `p->widx[0]` are always initialized to `-1` at function entry, and remain consistent even if `select()/poll()` returns an error. With this fix, empty socket sets no longer yield spurious ready indexes, and poll() on an empty registry correctly returns `[]` without mutating the input array.
Windows: pass/handle exceptfds in select() and on SOCKET_ERROR surface current read fds for cleanup. POSIX: include POLLERR|POLLNVAL in readiness. Ensures peer FIN/RST/error is observed by default once-per-frame events(0) loops without scanning or keepalives as it should.
12 tasks
Contributor
Author
|
This is actually a kind of important fix for me personally as my framework is broken without it. If it's helpful, I can demonstrate a sample case that it resolves. @Simn But, for simplicity, I think we can agree that |
Member
|
Would it be feasible to add some tests? E.g. like here: Line 483 in 0b59d98 |
Contributor
Author
|
Yeah I will add tests for it. |
The existing testPoll only checks that polling an empty set returns; its timing assertions are commented out and it ends in Assert.pass(). It never reads readIndexes or writeIndexes, which is where both bugs in this branch lived. testPollReportsReadableSocket covers the first: a listening socket with a connection waiting to be accepted has to be reported readable. _hx_std_socket_poll_events could return without reporting it, leaving the caller to conclude nothing had arrived. Measured before and after the fix on Windows: readIndexes[0] was -1 with a connection pending, where select() on the same socket at the same moment saw it. testPollClearsIndexesBetweenCalls covers the second: the ready lists are the caller's only result, so they have to describe the poll that just ran rather than the one before it. Polling with nothing registered makes select() fail, which is the path that used to return early and leave the previous call's indexes in place; the test asserts both lists come back as -1. Both are bounded retry loops rather than single polls, because readiness is not required to appear on the first pass -- the assertion is that it appears at all.
Contributor
Author
|
Finally got around to adding a test for this. |
FDSIZE was sizeof(u_int) + n * sizeof(SOCKET). fd_array does not begin at sizeof(u_int): on 64-bit Windows SOCKET is eight bytes, so fd_set carries four bytes of padding after fd_count and the array starts at offset 8. Every memcpy of an fd_set therefore copied four bytes too few, truncating the first handle by half and leaving the remainder as whatever malloc had returned. select() was consequently handed something that was not a socket and failed with WSAENOTSOCK on every call. Confirmed by instrumenting the error branch: [poll] SOCKET_ERROR wsa=10038 rcount=0 wcount=1 The read path hid it. On select() failure the error branch marks every registered read socket ready, so poll appeared to work for reads while reporting readiness it had not measured -- an idle connection with nothing to read came back readable, and a caller woke to read nothing for as long as it stayed idle. Writes have no such fallback, so they simply never reported at all. Sizing the copy by offsetof(fd_set, fd_array) fixes all three: writes are reported, reads are reported because select actually said so, and an idle socket is correctly not ready. 32-bit was unaffected, where SOCKET is four bytes and no padding exists. Two tests come with it. testPollReportsWritableSocket covers the case that never worked. testPollDoesNotReportIdleSocketAsReadable covers the false positive, and is the one that distinguishes a working select from the error path pretending: it fails whenever poll answers from the fallback rather than from the call.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR contains two related fixes in
_hx_std_socket_poll_eventsto makepolling safer and more consistent:
Clear ridx/widx on entry and early return
Prevents stale values when polling with empty registries or when
select()/poll() errors out. Previously this could result in phantom
indexes or poll([]) returning [null].
Surface close/error conditions reliably
exceptfdsset to detect errorsand connection resets. On
SOCKET_ERROR, all current read fds aresurfaced so userland can clean up.
POLLERR | POLLNVALalongsidePOLLIN/HUPandPOLLOUT/HUPso teardown/error conditions are not missed.Together these changes ensure that:
userland can detect EOF and deregister them.