liblzma: Add a new flag LZMA_FAIL_FAST for threaded decoder.

In most cases if the input file is corrupt the application won't
care about the uncompressed content at all. With this new flag
the threaded decoder will return an error as soon as any thread
has detected an error; it won't wait to copy out the data before
the location of the error.

I don't plan to use this in xz to keep the behavior consistent
between single-threaded and multi-threaded modes.

src/liblzma/api/lzma/container.h

@@ -73,7 +73,7 @@ typedef struct {
 	 *
 	 * Decoder: Bitwise-or of zero or more of the decoder flags:
 	 * LZMA_TELL_NO_CHECK, LZMA_TELL_UNSUPPORTED_CHECK,
-	 * LZMA_TELL_ANY_CHECK, LZMA_CONCATENATED
+	 * LZMA_TELL_ANY_CHECK, LZMA_CONCATENATED, LZMA_FAIL_FAST
 	 */
 	uint32_t flags;
 
@@ -615,6 +615,29 @@ extern LZMA_API(lzma_ret) lzma_microlzma_encoder(
 #define LZMA_CONCATENATED               UINT32_C(0x08)
 
 
+/**
+ * This flag makes the threaded decoder report errors (like LZMA_DATA_ERROR)
+ * as soon as they are detected. This saves time when the application has no
+ * interest in a partially decompressed truncated or corrupt file. Note that
+ * due to timing randomness, if the same truncated or corrupt input is
+ * decompressed multiple times with this flag, a different amount of output
+ * may be produced by different runs, and even the error code might vary.
+ *
+ * Without this flag the threaded decoder will provide as much output as
+ * possible at first and then report the pending error. This default behavior
+ * matches the single-threaded decoder and provides repeatable behavior
+ * with truncated or corrupt input. There are a few special cases where the
+ * behavior can still differ like memory allocation failures (LZMA_MEM_ERROR).
+ *
+ * Single-threaded decoders currently ignore this flag.
+ *
+ * Support for this flag was added in liblzma 5.3.3alpha. Note that in older
+ * versions this flag isn't supported (LZMA_OPTIONS_ERROR) even by functions
+ * that ignore this flag in newer liblzma versions.
+ */
+#define LZMA_FAIL_FAST                  UINT32_C(0x20)
+
+
 /**
  * \brief       Initialize .xz Stream decoder
  *

src/liblzma/common/common.h

@@ -67,14 +67,15 @@
 #define LZMA_FILTER_RESERVED_START (LZMA_VLI_C(1) << 62)
 
 
-/// Supported flags that can be passed to lzma_stream_decoder()
-/// or lzma_auto_decoder().
+/// Supported flags that can be passed to lzma_stream_decoder(),
+/// lzma_auto_decoder(), or lzma_stream_decoder_mt().
 #define LZMA_SUPPORTED_FLAGS \
 	( LZMA_TELL_NO_CHECK \
 	| LZMA_TELL_UNSUPPORTED_CHECK \
 	| LZMA_TELL_ANY_CHECK \
 	| LZMA_IGNORE_CHECK \
-	| LZMA_CONCATENATED )
+	| LZMA_CONCATENATED \
+	| LZMA_FAIL_FAST )
 
 
 /// Largest valid lzma_action value as unsigned integer.

src/liblzma/common/stream_decoder_mt.c

@@ -300,6 +300,10 @@ struct lzma_stream_coder {
 	/// Stream Padding is a multiple of four bytes.
 	bool concatenated;
 
+	/// If true, we will return any errors immediately instead of first
+	/// producing all output before the location of the error.
+	bool fail_fast;
+
 
 	/// When decoding concatenated Streams, this is true as long as we
 	/// are decoding the first Stream. This is needed to avoid misleading
@@ -711,13 +715,12 @@ read_output_and_wait(struct lzma_stream_coder *coder,
 					coder->pending_error
 							= coder->thread_error;
 
-				// FIXME? Add a flag to do this conditionally?
-				// That way errors would get reported to the
-				// application without a delay.
-// 				if (coder->fast_errors) {
-// 					ret = coder->thread_error;
-// 					break;
-// 				}
+				// If LZMA_FAIL_FAST was used, report errors
+				// from worker threads immediately.
+				if (coder->fail_fast) {
+					ret = coder->thread_error;
+					break;
+				}
 			}
 
 			// Check if decoding of the next Block can be started.
@@ -1690,22 +1693,24 @@ stream_decode_mt(void *coder_ptr, const lzma_allocator *allocator,
 		break;
 
 	case SEQ_ERROR:
-		// Let the application get all data before the point where
-		// the error was detected. This matches the behavior of
-		// single-threaded use.
-		//
-		// FIXME? Some errors (LZMA_MEM_ERROR) don't get here,
-		// they are returned immediately. Thus in rare cases the
-		// output will be less than in single-threaded mode. But
-		// maybe this doesn't matter much in practice.
-		return_if_error(read_output_and_wait(coder, allocator,
-				out, out_pos, out_size,
-				NULL, true, &wait_abs, &has_blocked));
+		if (!coder->fail_fast) {
+			// Let the application get all data before the point
+			// where the error was detected. This matches the
+			// behavior of single-threaded use.
+			//
+			// FIXME? Some errors (LZMA_MEM_ERROR) don't get here,
+			// they are returned immediately. Thus in rare cases
+			// the output will be less than in the single-threaded
+			// mode. Maybe this doesn't matter much in practice.
+			return_if_error(read_output_and_wait(coder, allocator,
+					out, out_pos, out_size,
+					NULL, true, &wait_abs, &has_blocked));
 
-		// We get here only if the error happened in the main thread,
-		// for example, unsupported Block Header.
-		if (!lzma_outq_is_empty(&coder->outq))
-			return LZMA_OK;
+			// We get here only if the error happened in the main
+			// thread, for example, unsupported Block Header.
+			if (!lzma_outq_is_empty(&coder->outq))
+				return LZMA_OK;
+		}
 
 		return coder->pending_error;
 
@@ -1900,6 +1905,7 @@ stream_decoder_mt_init(lzma_next_coder *next, const lzma_allocator *allocator,
 	coder->tell_any_check = (options->flags & LZMA_TELL_ANY_CHECK) != 0;
 	coder->ignore_check = (options->flags & LZMA_IGNORE_CHECK) != 0;
 	coder->concatenated = (options->flags & LZMA_CONCATENATED) != 0;
+	coder->fail_fast = (options->flags & LZMA_FAIL_FAST) != 0;
 
 	coder->first_stream = true;
 	coder->out_was_filled = false;