164288 - In the WINNT build incoming data may be lost after PR_IO_TIMEOUT_ERROR from PR_Recv

Status: NEW

(NSPR :: NSPR, defect)

Description

[Wan-Teh Chang]

In the WINNT build of NSPR, incoming data may
be lost after a PR_Recv call fails with the
PR_IO_TIMEOUT_ERROR or PR_PENDING_INTERRUPT_ERROR.
This problem won't be noticed if you immediately
close the socket after the PR_IO_TIMEOUT_ERROR
or PR_PENDING_INTERRUPT_ERROR, but may occur if
you call PR_NT_CancelIo and then try to receive
data from the socket again.

The incoming data will be lost with the following
sequence of events.

1. PR_Recv times out or is interrupted by another
thread.  The 'io_suspended' flag of the thread is
set to PR_TRUE.

2. Before the thread calls PR_Close or PR_NT_CancelIo
on the socket, the overlapped I/O completes and the
completion packet is posted to NSPR's I/O completion
port.  Note that incoming data have been placed in
the buffer passed to the PR_Recv call.

3. An NSPR "CPU idle" thread reads the completion
packet, sees that the thread is already running
(because its 'io_suspended' flag is true), and
goes on to read the next completion packet.
Nothing is done to the incoming data in the buffer
passed to the PR_Recv call; the data are lost.

A similar problem may exist for PR_Accept.

For PR_Send, the outgoing data may still be
sent even if PR_Send times out or is interrupted
by another thread.  This may be a problem in
some cases.

Comment 1

[Wan-Teh Chang]

The best way to fix this bug and the similar problems
with PR_Accept and PR_Send is to have these NSPR I/O
functions call NT's CancelIo function after a timeout or
interrupt.  If the overlapped I/O completes before we
can cancel it, these I/O functions will return with the
result of the completed overlapped I/O instead of the
timeout or interrupt error.

If we can implement this fix, PR_NT_CancelIo will no
longer be necessary.  It will become a no-op and be
deprecated.

Comment 2

[BugBot [:suhaib / :marco/ :calixte]]

The bug assignee didn't login in Bugzilla in the last 7 months, so the assignee is being reset.