[newlib-cygwin] Cygwin: FIFO: update commentary

[newlib-cygwin] Cygwin: FIFO: update commentary

[Ken Brown]

https://sourceware.org/git/gitweb.cgi?p=newlib-cygwin.git;h=98dfadec3a07d5be5d1796d1e27e51892dd9daf0

commit 98dfadec3a07d5be5d1796d1e27e51892dd9daf0
Author: Ken Brown <kbrown@cornell.edu>
Date:   Tue Apr 14 09:45:44 2020 -0400

    Cygwin: FIFO: update commentary
    
    The beginning of fhandler_fifo.cc contains a long comment giving an
    overview of the FIFO implementation.  This is now updated to describe
    the support for multiple readers.

Diff:
---
 winsup/cygwin/fhandler_fifo.cc | 58 +++++++++++++++++++++++++-----------------
 1 file changed, 35 insertions(+), 23 deletions(-)

diff --git a/winsup/cygwin/fhandler_fifo.cc b/winsup/cygwin/fhandler_fifo.cc
index 4c7d51edb..ec7c5d427 100644
--- a/winsup/cygwin/fhandler_fifo.cc
+++ b/winsup/cygwin/fhandler_fifo.cc
@@ -26,29 +26,41 @@
 /*
    Overview:
 
-     Currently a FIFO can be opened once for reading and multiple
-     times for writing.  Any attempt to open the FIFO a second time
-     for reading fails with EACCES (from STATUS_ACCESS_DENIED).
-
-     When a FIFO is opened for reading,
-     fhandler_fifo::create_pipe_instance is called to create the first
-     instance of a Windows named pipe server (Windows terminology).  A
-     "fifo_reader" thread is also started; it waits for pipe clients
-     (Windows terminology again) to connect.  This happens every time
-     a process opens the FIFO for writing.
-
-     The fifo_reader thread creates new instances of the pipe server
-     as needed, so that there is always an instance available for a
-     writer to connect to.
-
-     The reader maintains a list of "fifo_client_handlers", one for
-     each pipe instance.  A fifo_client_handler manages the connection
-     between the pipe instance and a writer connected to that pipe
-     instance.
-
-     TODO: Allow a FIFO to be opened multiple times for reading.
-     Maybe this could be done by using shared memory, so that all
-     readers could have access to the same list of writers.
+     FIFOs are implemented via Windows named pipes.  The server end of
+     the pipe corresponds to an fhandler_fifo open for reading (a.k.a,
+     a "reader"), and the client end corresponds to an fhandler_fifo
+     open for writing (a.k.a, a "writer").
+
+     The server can have multiple instances.  The reader (assuming for
+     the moment that there is only one) creates a pipe instance for
+     each writer that opens.  The reader maintains a list of
+     "fifo_client_handler" structures, one for each writer.  A
+     fifo_client_handler contains the handle for the pipe server
+     instance and information about the state of the connection with
+     the writer.  Each writer holds the pipe instance's client handle.
+
+     The reader runs a "fifo_reader_thread" that creates new pipe
+     instances as needed and listens for client connections.
+
+     If there are multiple readers open, only one of them, called the
+     "owner", maintains the fifo_client_handler list.  The owner is
+     therefore the only reader that can read at any given time.  If a
+     different reader wants to read, it has to take ownership and
+     duplicate the fifo_client_handler list.
+
+     A reader that is not an owner also runs a fifo_reader_thread,
+     which is mostly idle.  The thread wakes up if that reader might
+     need to take ownership.
+
+     There is a block of shared memory, accessible to all readers,
+     that contains information needed for the owner change process.
+     It also contains some locks to prevent races and deadlocks
+     between the various threads.
+
+     At this writing, I know of only one application (Midnight
+     Commander when running under tcsh) that *explicitly* opens two
+     readers of a FIFO.  But many applications will have multiple
+     readers open via dup/fork/exec.
 */

[newlib-cygwin] Cygwin: FIFO: update commentary

[Ken Brown]

https://sourceware.org/git/gitweb.cgi?p=newlib-cygwin.git;h=ac371ee1baa27441d21b71fefbc817e9d821411e

commit ac371ee1baa27441d21b71fefbc817e9d821411e
Author: Ken Brown <kbrown@cornell.edu>
Date:   Sat Jul 11 15:43:44 2020 -0400

    Cygwin: FIFO: update commentary

Diff:
---
 winsup/cygwin/fhandler_fifo.cc | 21 +++++++++++++++++----
 1 file changed, 17 insertions(+), 4 deletions(-)

diff --git a/winsup/cygwin/fhandler_fifo.cc b/winsup/cygwin/fhandler_fifo.cc
index 30486304f..e9d0187d4 100644
--- a/winsup/cygwin/fhandler_fifo.cc
+++ b/winsup/cygwin/fhandler_fifo.cc
@@ -52,10 +52,23 @@
      which is mostly idle.  The thread wakes up if that reader might
      need to take ownership.
 
-     There is a block of shared memory, accessible to all readers,
-     that contains information needed for the owner change process.
-     It also contains some locks to prevent races and deadlocks
-     between the various threads.
+     There is a block of named shared memory, accessible to all
+     fhandlers for a given FIFO.  It keeps track of the number of open
+     readers and writers; it contains information needed for the owner
+     change process; and it contains some locks to prevent races and
+     deadlocks between the various threads.
+
+     The shared memory is created by the first reader to open the
+     FIFO.  It is opened by subsequent readers and by all writers.  It
+     is destroyed by Windows when the last handle to it is closed.
+
+     If a handle to it somehow remains open after all processes
+     holding file descriptors to the FIFO have closed, the shared
+     memory can persist and be reused with stale data by the next
+     process that opens the FIFO.  So far I've seen this happen only
+     as a result of a bug in the code, but there are some debug_printf
+     statements in fhandler_fifo::open to help detect this if it
+     happens again.
 
      At this writing, I know of only one application (Midnight
      Commander when running under tcsh) that *explicitly* opens two