summaryrefslogtreecommitdiffstats
path: root/generic/tclIO.c
diff options
context:
space:
mode:
authordgp <dgp@users.sourceforge.net>2014-10-28 20:10:32 (GMT)
committerdgp <dgp@users.sourceforge.net>2014-10-28 20:10:32 (GMT)
commit7ab8a9b2efc85997d0a6f576c20a69d272bd98ff (patch)
treef028b842e93ec6e43d37de0ddf424781dd34940d /generic/tclIO.c
parentf015a62dd6b90d27db2849f96b3a9af050a291dc (diff)
downloadtcl-7ab8a9b2efc85997d0a6f576c20a69d272bd98ff.zip
tcl-7ab8a9b2efc85997d0a6f576c20a69d272bd98ff.tar.gz
tcl-7ab8a9b2efc85997d0a6f576c20a69d272bd98ff.tar.bz2
Work in progress restoring ability to [read] after [eof] and get non-empty
strings back in those cases where the channel has them to offer. Also working through all the implications of this possibility on Tcl's more exotic channel features, like stacking.
Diffstat (limited to 'generic/tclIO.c')
-rw-r--r--generic/tclIO.c88
1 files changed, 67 insertions, 21 deletions
diff --git a/generic/tclIO.c b/generic/tclIO.c
index 0122ec9..c55c118 100644
--- a/generic/tclIO.c
+++ b/generic/tclIO.c
@@ -4414,6 +4414,17 @@ TclGetsObjBinary(
if (bufPtr == NULL) {
goto restore;
}
+ } else {
+ /*
+ * There's something already in the buffer. If
+ * CHANNEL_STICKY_EOF is set we know that something begins
+ * with the eofchar. Otherwise, if CHANNEL_EOF is set, we
+ * know some earlier inputproc call returned zero bytes when
+ * we were trying to get more bytes to put in the buffer.
+ * which means..... ???? Place to probe with tests.
+ */
+ assert (GotFlag(statePtr, CHANNEL_STICKY_EOF)
+ || !GotFlag(statePtr, CHANNEL_EOF) );
}
dst = (unsigned char *) RemovePoint(bufPtr);
@@ -4694,6 +4705,9 @@ FilterInputBytes(
gsPtr->rawRead = 0;
return -1;
}
+ } else {
+ assert( GotFlag(statePtr, CHANNEL_STICKY_EOF)
+ || !GotFlag(statePtr, CHANNEL_EOF) );
}
/*
@@ -5018,6 +5032,7 @@ Tcl_ReadRaw(
/* State info for channel */
int copied = 0;
+ assert(bytesToRead > 0);
if (CheckChannelErrors(statePtr, TCL_READABLE | CHANNEL_RAW_MODE) != 0) {
return -1;
}
@@ -5049,8 +5064,19 @@ Tcl_ReadRaw(
}
}
- /* Go to the driver if more data needed. */
+ /*
+ * Go to the driver only if we got nothing from pushback.
+ * Have to do it this way to avoid EOF mis-timings when we
+ * consider the ability that EOF may not be a permanent
+ * condition in the driver, and in that case we have to
+ * synchronize.
+ */
+
+ if (copied) {
+ return copied;
+ }
+ /* This test not needed. */
if (bytesToRead > 0) {
int nread = ChanRead(chanPtr, readBuf, bytesToRead);
@@ -5073,12 +5099,10 @@ Tcl_ReadRaw(
if (!GotFlag(statePtr, CHANNEL_BLOCKED) || copied == 0) {
copied = -1;
}
- } else if (copied > 0) {
+ } else {
/*
- * nread == 0. Driver is at EOF, but if copied>0 bytes
- * from pushback, then we should not signal it yet.
+ * nread == 0. Driver is at EOF. Let that state filter up.
*/
- ResetFlag(statePtr, CHANNEL_EOF);
}
}
return copied;
@@ -6120,18 +6144,39 @@ GetInput(
}
/*
- * For a channel at EOF do not bother allocating buffers; there's
- * nothing more to read. Avoid calling the driver inputproc in
- * case some of them do not react well to additional calls after
- * they've reported an eof state..
- * TODO: Candidate for a can't happen panic.
+ * Strangely named "STICKY_EOF" really means we've seen the
+ * eofchar for this channel, and nothing since has reset it.
+ * (changed the eofchar, [seek]ed to a new offset, etc.) So,
+ * we know we're still poised to read that eofchar again, and
+ * there's no need to actually do it.
*/
- if (GotFlag(statePtr, CHANNEL_EOF)) {
+ if (GotFlag(statePtr, CHANNEL_STICKY_EOF)) {
+ assert(statePtr->inEofChar);
+ assert(statePtr->inQueueHead);
+ assert(RemovePoint(statePtr->inQueueHead)[0] == statePtr->inEofChar);
return 0;
}
/*
+ * WARNING: There was once a comment here claiming that it was
+ * a bad idea to make another call to the inputproc of a channel
+ * driver when EOF has already been detected on the channel. Through
+ * much of Tcl's history, this warning was then completely negated
+ * by having all (most?) read paths clear the EOF setting before
+ * reaching here. So we had a guard that was never triggered.
+ *
+ * Don't be tempted to restore the guard. Even if EOF is set on
+ * the channel, continue through and call the inputproc again. This
+ * is the way to enable the ability to [read] again beyond the EOF,
+ * which seems a strange thing to do, but for which use cases exist
+ * [Tcl Bug 5adc350683] and which may even be essential for channels
+ * representing things like ttys or other devices where the stream
+ * might take the logical form of a series of 'files' separated by
+ * an EOF condition.
+ */
+
+ /*
* First check for more buffers in the pushback area of the topmost
* channel in the stack and use them. They can be the result of a
* transformation which went away without reading all the information
@@ -8848,16 +8893,6 @@ DoRead(
ChannelBuffer *bufPtr = statePtr->inQueueHead;
/*
- * When there's no buffered data to read, and we're at EOF,
- * escape to the caller.
- */
-
- if (GotFlag(statePtr, CHANNEL_EOF)
- && (bufPtr == NULL || IsBufferEmpty(bufPtr))) {
- break;
- }
-
- /*
* Don't read more data if we have what we need.
*/
@@ -8969,12 +9004,23 @@ DoRead(
statePtr->inQueueTail = NULL;
}
RecycleBuffer(statePtr, bufPtr, 0);
+ bufPtr = statePtr->inQueueHead;
}
if (GotFlag(statePtr, CHANNEL_NONBLOCKING|CHANNEL_BLOCKED)
== (CHANNEL_NONBLOCKING|CHANNEL_BLOCKED)) {
break;
}
+
+ /*
+ * When there's no buffered data to read, and we're at EOF,
+ * escape to the caller.
+ */
+
+ if (GotFlag(statePtr, CHANNEL_EOF)
+ && (bufPtr == NULL || IsBufferEmpty(bufPtr))) {
+ break;
+ }
}
if (bytesToRead == 0) {
ResetFlag(statePtr, CHANNEL_BLOCKED);