Gather page output into one buffer

1 file changed, 45 insertions, 2 deletions

diff --git a/source/html.c b/source/html.cindex c3ab696..dbb0a57 100644--- a/source/html.c+++ b/source/html.c@@ -9,6 +9,9 @@ #include "cgit.h" #include "html.h" #include "url.h"+#define HTML_WRITE_BUFSIZE (64 * 1024)+static char html_buf[HTML_WRITE_BUFSIZE];+static size_t html_buflen;  /* Percent-encoding of each character, except: a-zA-Z0-9!$()*,./:;@- */ static const char* url_escape_table[256] = {@@ -78,12 +81,52 @@ char *cgit_fmtalloc(const char *format, ...) 	return strbuf_detach(&sb, NULL); } -void html_raw(const char *data, size_t size)+/*+ * Page output is collected here and written out in whole buffers. A page is+ * built from a great many small fragments, and writing each one cost a syscall+ * apiece: a tree listing spent more time entering the kernel than generating+ * anything.+ *+ * cgit does not own stdout by itself, so the buffer has to be emptied before+ * anything else writes there. Those points are a filter taking over stdout,+ * the header block that precedes output produced by git itself, the cache+ * slot being measured, and process exit.+ */+static void html_write(const char *data, size_t size) {-	if (write(STDOUT_FILENO, data, size) != size)+	// A blob, snapshot or patch reaches this with a size well past what one+	// write can move onto a pipe, so a short write is ordinary rather than+	// an error and has to be resumed instead of reported.+	if (write_in_full(STDOUT_FILENO, data, size) < 0) 		die_errno("write error on html output"); } +void html_flush(void)+{+	size_t len = html_buflen;++	if (!len)+		return;+	// Clear the length first: html_write can die, and the error page it+	// produces would otherwise try to flush the same bytes again.+	html_buflen = 0;+	html_write(html_buf, len);+}++void html_raw(const char *data, size_t size)+{+	if (size >= HTML_WRITE_BUFSIZE) {+		// Nothing is gained by copying a blob through the buffer.+		html_flush();+		html_write(data, size);+		return;+	}+	if (html_buflen + size > HTML_WRITE_BUFSIZE)+		html_flush();+	memcpy(html_buf + html_buflen, data, size);+	html_buflen += size;+}+ void html(const char *txt) { 	html_raw(txt, strlen(txt));
diff --git a/source/html.c b/source/html.c
index c3ab696..dbb0a57 100644
--- a/source/html.c
+++ b/source/html.c
@@ -9,6 +9,9 @@
 #include "cgit.h"
 #include "html.h"
 #include "url.h"
+#define HTML_WRITE_BUFSIZE (64 * 1024)
+static char html_buf[HTML_WRITE_BUFSIZE];
+static size_t html_buflen;
 
 /* Percent-encoding of each character, except: a-zA-Z0-9!$()*,./:;@- */
 static const char* url_escape_table[256] = {
@@ -78,12 +81,52 @@ char *cgit_fmtalloc(const char *format, ...)
 	return strbuf_detach(&sb, NULL);
 }
 
-void html_raw(const char *data, size_t size)
+/*
+ * Page output is collected here and written out in whole buffers. A page is
+ * built from a great many small fragments, and writing each one cost a syscall
+ * apiece: a tree listing spent more time entering the kernel than generating
+ * anything.
+ *
+ * cgit does not own stdout by itself, so the buffer has to be emptied before
+ * anything else writes there. Those points are a filter taking over stdout,
+ * the header block that precedes output produced by git itself, the cache
+ * slot being measured, and process exit.
+ */
+static void html_write(const char *data, size_t size)
 {
-	if (write(STDOUT_FILENO, data, size) != size)
+	// A blob, snapshot or patch reaches this with a size well past what one
+	// write can move onto a pipe, so a short write is ordinary rather than
+	// an error and has to be resumed instead of reported.
+	if (write_in_full(STDOUT_FILENO, data, size) < 0)
 		die_errno("write error on html output");
 }
 
+void html_flush(void)
+{
+	size_t len = html_buflen;
+
+	if (!len)
+		return;
+	// Clear the length first: html_write can die, and the error page it
+	// produces would otherwise try to flush the same bytes again.
+	html_buflen = 0;
+	html_write(html_buf, len);
+}
+
+void html_raw(const char *data, size_t size)
+{
+	if (size >= HTML_WRITE_BUFSIZE) {
+		// Nothing is gained by copying a blob through the buffer.
+		html_flush();
+		html_write(data, size);
+		return;
+	}
+	if (html_buflen + size > HTML_WRITE_BUFSIZE)
+		html_flush();
+	memcpy(html_buf + html_buflen, data, size);
+	html_buflen += size;
+}
+
 void html(const char *txt)
 {
 	html_raw(txt, strlen(txt));