Changeset: 835cc241a8d0 for MonetDB
URL: https://dev.monetdb.org/hg/MonetDB?cmd=changeset;node=835cc241a8d0
Added Files:
common/stream/pump.h
Modified Files:
common/stream/xz_stream.c
Branch: makelibstreamgreatagain
Log Message:
Introduce generic_pump_{in,out}, use in xz_stream
diffs (truncated from 475 to 300 lines):
diff --git a/common/stream/pump.h b/common/stream/pump.h
new file mode 100644
--- /dev/null
+++ b/common/stream/pump.h
@@ -0,0 +1,216 @@
+/*
+ * This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/.
+ *
+ * Copyright 1997 - July 2008 CWI, August 2008 - 2020 MonetDB B.V.
+ */
+
+/* streams working on a lzma/xz-compressed disk file */
+
+#include "monetdb_config.h"
+
+/* The logic of zlib's inflate and various other compression libraries
+ * is always very similar. At the heart there is a work function which is
+ * called with an input buffer, an output buffer and some internal state.
+ * The function tries to move bits out of the input buffer through the
+ * internal state to the output buffer until the input buffer is empty
+ * or the output buffer is full.
+ *
+ * It has an action parameter which controls flushing: sometimes, for example
+ * when closing the file, you want to flush the internal state out instead
+ * of hanging onto it waiting for more compression opportunities. If that
+ * parameter is set, the protocol is that you keep making room in the output
+ * buffer and keep offering *exactly the same* input buffer until the
+ * worker indicates all content has been flushed.
+ *
+ * All compression libraries we use have such an API but of course the
+ * parameter types etc are always different.
+ * This function encapsulates the logic of copying a whole buffer
+ * out to a stream in a type-agnostic way.
+ */
+
+/* Helper functions used to access the state in a generic way */
+
+typedef enum {
+ PUMP_OK,
+ PUMP_END,
+ PUMP_ERROR,
+} pump_result;
+
+typedef enum {
+ PUMP_WORK,
+ PUMP_FLUSH_DATA,
+ PUMP_FLUSH_ALL,
+ PUMP_FINISH,
+} pump_action;
+
+typedef pump_result (*pump_worker)(void *state, pump_action action);
+typedef ssize_t (*pump_io)(void *state, char *data, size_t len);
+
+typedef struct {
+ char *start;
+ size_t count;
+} pump_buffer;
+
+typedef struct {
+ char **start;
+ size_t *count;
+} pump_buffer_location;
+
+// These helper functions help make sure we don't accidentally
+// write `buf.start == 0` where we meant `*buf.start == 0`.
+
+static inline char *
+start(pump_buffer_location b)
+{
+ return *b.start;
+}
+
+static inline size_t
+count(pump_buffer_location b)
+{
+ return *b.count;
+}
+
+static inline void
+set_count(pump_buffer_location b, size_t count)
+{
+ *b.count = count;
+}
+
+static inline void
+set_start(pump_buffer_location b, char *start)
+{
+ *b.start = start;
+}
+
+static inline void
+reset_buffer(pump_buffer_location b, pump_buffer buf)
+{
+ *b.start = buf.start;
+ *b.count = buf.count;
+}
+
+static inline pump_result
+generic_pump_out(
+ void *state,
+ pump_action action,
+ pump_buffer buffer,
+ pump_buffer_location window_in,
+ pump_buffer_location window_out,
+ pump_worker worker,
+ pump_io ship_out)
+{
+ while (1) {
+ // Make sure there is room in the output buffer
+ if (count(window_out) == 0) {
+ size_t amount = start(window_out) - buffer.start;
+ ssize_t nwritten = ship_out(state, buffer.start,
amount);
+ if (nwritten != (ssize_t)amount)
+ return PUMP_ERROR;
+ reset_buffer(window_out, buffer);
+ }
+
+ // Try to make progress
+ pump_result ret = worker(state, action);
+ if (ret == PUMP_ERROR)
+ return PUMP_ERROR;
+
+ // There was no error but if input is still available, we
definitely
+ // need another round
+ if (count(window_in) > 0)
+ continue;
+
+ // Though the input data has been consumed, some may still
linger
+ // in the internal state.
+ if (action == PUMP_WORK) {
+ // Let it linger, we'll combine it with the next batch
+ assert(ret == PUMP_OK); // worker would never PUMP_END
+ return PUMP_OK;
+ }
+
+ // We are flushing or finishing or whatever.
+ // We may need to keep iterating to flush the internal state.
+ // Is there any internal state left?
+ if (ret == PUMP_OK)
+ // yes, there is
+ continue;
+
+ // All internal state has been drained.
+ // Now drain the output buffer
+ assert(ret == PUMP_END);
+ size_t amount = start(window_out) - buffer.start;
+ if (amount > 0) {
+ ssize_t nwritten = ship_out(state, buffer.start,
amount);
+ if (nwritten != (ssize_t)amount)
+ return PUMP_ERROR;
+ }
+ reset_buffer(window_out, buffer);
+ return PUMP_END;
+ }
+}
+
+/* Similar to generic_pump_out, but for reading.
+*
+* In every iteration, fill the input buffer if empty, and let the worker try to
+* make progress. Stop iterating if the output buffer is full, when ship_in
fails
+* or when it does not fill the whole buffer. The latter case is important when
+* we're reading for example from a socket. It's more important to return the
+* data that has come in so far than to wait until the output buffer is
+* completely full.
+*
+* Returns PUMP_END if the input is exhausted and no data lingers in the
internal
+* state, PUMP_ERROR when the input has a failure or PUMP_OK otherwise.
+ */
+static inline pump_result
+generic_pump_in(
+ void *state,
+ pump_buffer buffer,
+ pump_buffer_location window_in,
+ pump_buffer_location window_out,
+ pump_worker worker,
+ pump_io ship_in)
+{
+ char *orig_out = start(window_out); // nice for debugging
+ (void)orig_out;
+ while (1) {
+ if (count(window_out) == 0)
+ // Output buffer is sufficiently full.
+ return PUMP_OK;
+
+ // Handle input, if possible and necessary
+ if (start(window_in) != NULL && count(window_in) == 0) {
+ // start != NULL means we haven't encountered EOF yet
+ ssize_t nread = ship_in(state, buffer.start,
buffer.count);
+ if (nread < 0)
+ // Error. Return directly, discarding any data
lingering
+ // in the internal state.
+ return PUMP_ERROR;
+ if (nread == 0)
+ // Set to NULL so we'll remember next time.
+ // Maybe there is some data in the internal
state we don't
+ // return immediately.
+ set_start(window_in, NULL);
+ else
+ // All good
+ set_start(window_in, buffer.start);
+ set_count(window_in, (ssize_t)nread);
+ }
+
+ // Try to make some progress
+ pump_action action = (start(window_in) != NULL) ? PUMP_WORK :
PUMP_FINISH;
+ assert(count(window_out) > 0);
+ assert(count(window_in) > 0 || action == PUMP_FINISH);
+ pump_result ret = worker(state, action);
+ if (ret == PUMP_ERROR)
+ return PUMP_ERROR;
+
+ if (ret == PUMP_END)
+ // If you say so
+ return PUMP_END;
+
+ // If we get here we made some progress so we're ready for a
new iteration.
+ }
+}
+
diff --git a/common/stream/xz_stream.c b/common/stream/xz_stream.c
--- a/common/stream/xz_stream.c
+++ b/common/stream/xz_stream.c
@@ -11,6 +11,7 @@
#include "monetdb_config.h"
#include "stream.h"
#include "stream_internal.h"
+#include "pump.h"
#ifdef HAVE_LIBLZMA
@@ -21,6 +22,7 @@ typedef struct xz_state {
uint8_t buf[XZBUFSIZ];
} xz_state;
+
/* Keep calling lzma_code until the whole input buffer has been consumed
* and all necessary output has been written.
*
@@ -34,132 +36,95 @@ typedef struct xz_state {
*
* Returns > 0 on succes, 0 on error.
*/
-static int
-pump_out(stream *s, lzma_action action)
+static pump_result
+pumper(void *state, pump_action action)
{
- xz_state *xz = (xz_state *) s->stream_data.p;
-
- while (1) {
- // Make sure there is room in the output buffer
- if (xz->strm.avail_out == 0) {
- size_t nwritten = fwrite(xz->buf, 1, XZBUFSIZ, xz->fp);
- if (nwritten != XZBUFSIZ) {
- return 0;
- }
- xz->strm.next_out = xz->buf;
- xz->strm.avail_out = XZBUFSIZ;
- }
-
- lzma_ret ret = lzma_code(&xz->strm, action);
- if (ret != LZMA_OK && ret != LZMA_STREAM_END) {
- // Some kind of error.
- return 0;
- }
+ stream *s = (stream*) state;
+ xz_state *xz = (xz_state*) s->stream_data.p;
- if (xz->strm.avail_in > 0) {
- // Definitely not done yet. Flush the buffer and encode
- // some more.
- continue;
- }
-
- // Whether we are already done or not depends on the mode.
- if (action == LZMA_RUN) {
- assert(ret == LZMA_OK);
- // More input will follow so we can leave the output
buffer
- // for later.
- return 1;
- }
+ lzma_action a;
+ switch (action) {
+ case PUMP_WORK:
+ a = LZMA_RUN;
+ break;
+ case PUMP_FLUSH_DATA:
+ a = LZMA_SYNC_FLUSH;
+ break;
+ case PUMP_FLUSH_ALL:
+ a = LZMA_FULL_FLUSH;
+ break;
+ case PUMP_FINISH:
+ a = LZMA_FINISH;
+ break;
+ }
- // We need to flush all data out of the encoder and out of our
_______________________________________________
checkin-list mailing list
checkin-list@monetdb.org
https://www.monetdb.org/mailman/listinfo/checkin-list
