Imrove memory usage during huffman decoding

- Do not allocate new memory during hpack decoding
  if some memory is already available. (It can be
  available because we allocate 150 % memory for
  each header).
- Move BUFFER_GET macros to hpack.c
- Some codestyle fixes according checkpatch.pl

Author: EvgeniiMekhanik

fw/hpack.c

@@ -30,6 +30,31 @@
 
 #include "hpack_tbl.h"
 
+/**
+ * There are to cases when we allocate space for header:
+ * - it->rspace is equal to zero. In this case we allocate
+ *   `len` bytes;
+ * - it->rspace is not equal to zero. This means that we
+ *   already have buffer with rspace bytes available.
+ *   Save delta = len - it->rspace bytes. Later when
+ *   `it->rspace` will be exceeded we use it to
+ *   allocate new chunk;
+ */
+#define	BUFFER_GET(len, it)						\
+do {									\
+	BUG_ON(!(len));							\
+	if (!(it)->rspace) {						\
+		(it)->rspace = len;					\
+		(it)->pos = tfw_pool_alloc_not_align((it)->pool, len);	\
+		T_DBG3("%s: get buffer, len=%lu, it->pos=[%p],"		\
+		       " it->pos=%lu\n", __func__, (unsigned long)len,	\
+		       (it)->pos, (unsigned long)(it)->pos);		\
+	} else {							\
+		(it)->to_alloc = len > (it)->rspace ?			\
+			len - (it)->rspace : 0;				\
+	}								\
+} while (0)
+
 #define HP_HDR_NAME(name)						\
 	(&(TfwStr){							\
 		.chunks = &(TfwStr){					\
@@ -278,25 +303,32 @@ do {								\
 	       last - src);					\
 } while (0)
 
-#define	BUFFER_HDR_INIT(length, it)				\
+#define	BUFFER_HDR_INIT(it)					\
 do {								\
 	(it)->hdr.data = (it)->pos;				\
-	(it)->hdr.len = length;					\
+	(it)->hdr.len = 0;					\
 	(it)->next = &(it)->hdr;				\
 } while (0)
 
 #define	BUFFER_NAME_OPEN(length)				\
 do {								\
 	WARN_ON_ONCE(!TFW_STR_EMPTY(&it->hdr));			\
 	if (state & HPACK_FLAGS_HUFFMAN_NAME) {			\
+		/*						\
+		 * Allocate extra 50% bytes. Huffman usually	\
+		 * gives compression benefit about 20 - 60 %.	\
+		 * Since we use extra allocated bytes for the	\
+		 * next header, we can allocate extra 50% bytes	\
+		 * without fear of losing a lot of memory.	\
+		 */						\
 		unsigned long len = length + (length >> 1);	\
 								\
 		BUFFER_GET(len, it);				\
 		if (!it->pos) {					\
 			r = -ENOMEM;				\
 			goto out;				\
 		}						\
-		BUFFER_HDR_INIT(len, it);			\
+		BUFFER_HDR_INIT(it);				\
 	}							\
 } while (0)
 
@@ -308,6 +340,13 @@ do {								\
 		? it->parsed_hdr->nchunks			\
 		: 1;						\
 	if (state & HPACK_FLAGS_HUFFMAN_VALUE) {		\
+		/*						\
+		 * Allocate extra 50% bytes. Huffman usually	\
+		 * gives compression benefit about 20 - 60 %.	\
+		 * Since we use extra allocated bytes for the	\
+		 * next header, we can allocate extra 50% bytes	\
+		 * without fear of losing a lot of memory.	\
+		 */						\
 		unsigned long len = length + (length >> 1);	\
 								\
 		BUFFER_GET(len, it);				\
@@ -317,13 +356,13 @@ do {								\
 		}						\
 		if (!TFW_STR_EMPTY(&it->hdr)) {			\
 			it->next = tfw_hpack_exp_hdr(req->pool,	\
-						     len, it);	\
+						     0, it);	\
 			if (!it->next) {			\
 				r = -ENOMEM;			\
 				goto out;			\
 			}					\
 		} else {					\
-			BUFFER_HDR_INIT(len, it);		\
+			BUFFER_HDR_INIT(it);			\
 		}						\
 	}							\
 } while (0)
@@ -459,38 +498,61 @@ tfw_hpack_exp_hdr(TfwPool *__restrict pool, unsigned long len,
 static inline int
 tfw_hpack_huffman_write(char sym, TfwHttpReq *__restrict req)
 {
-	bool np;
 	TfwMsgParseIter *it = &req->pit;
+	unsigned long to_alloc;
+	bool np;
+
+#define ADJUST_HDR_LEN(it)						\
+do {									\
+	TfwStr *hdr = &it->hdr;						\
+	TfwStr *last = TFW_STR_LAST(hdr);				\
+									\
+	T_DBG3("%s: add to hdr, hdr->len=%lu, last->len=%lu,"		\
+	       " last->data=%.*s\n", __func__, hdr->len, last->len,	\
+	       (int)last->len, last->data);				\
+									\
+	++hdr->len;							\
+	if (!TFW_STR_PLAIN(hdr))					\
+		++last->len;						\
+} while (0)
 
 	if (it->rspace) {
 		--it->rspace;
 		*it->pos++ = sym;
+		ADJUST_HDR_LEN(it);
 		return 0;
 	}
 
-	if (!(it->pos = tfw_pool_alloc_not_align_np(it->pool, 1, &np)))
-		return -ENOMEM;
+	to_alloc = it->to_alloc ? it->to_alloc : 1;
+	if (to_alloc > 1)
+		it->rspace = to_alloc - 1;
 
-	*it->pos = sym;
+	it->pos = tfw_pool_alloc_not_align_np(it->pool, to_alloc, &np);
+	if (!it->pos)
+		return -ENOMEM;
 
 	T_DBG3("%s: it->rspace=%lu, sym=%c, np=%d\n", __func__,
 	       it->rspace, sym, np);
 
-	if (!np) {
-		TfwStr *hdr = &it->hdr;
-		TfwStr *last = TFW_STR_LAST(hdr);
-
-		T_DBG3("%s: add to hdr, hdr->len=%lu, last->len=%lu,"
-		       " last->data=%.*s\n", __func__, hdr->len, last->len,
-		       (int)last->len, last->data);
-
-		++hdr->len;
-		if (!TFW_STR_PLAIN(hdr))
-			++last->len;
+	/*
+	 * If the new page was allocated or it->pos points to
+	 * the buffer, which was allocated to store previously
+	 * decoded header (it->to_alloc != 0) we should expand
+	 * header.
+	 */
+	if (!np && !it->to_alloc) {
+		*it->pos++ = sym;
+		ADJUST_HDR_LEN(it);
 		return 0;
 	}
 
-	return tfw_hpack_exp_hdr(req->pool, 1, it) ? 0 : -ENOMEM;
+	if (!tfw_hpack_exp_hdr(req->pool, 1, it))
+		return -ENOMEM;
+
+	*it->pos++ = sym;
+	it->to_alloc = 0;
+
+	return 0;
 }
 
 static int
@@ -501,28 +563,14 @@ huffman_decode_tail(TfwHPack *__restrict hp, TfwHttpReq *__restrict req,
 	unsigned int i;
 	int r;
 
-#define ADJUST_EXTRA_RSPACE(req)		\
-do {						\
-	TfwMsgParseIter *it = &req->pit;	\
-	TfwStr *hdr = &it->hdr;			\
-	TfwStr *last = TFW_STR_LAST(hdr);	\
-						\
-	hdr->len -= it->rspace;			\
-	if (!TFW_STR_PLAIN(hdr))		\
-		last->len -= it->rspace;	\
-	it->rspace = 0;				\
-} while (0)
-
 	for (;;) {
 		int shift;
 
 		if (hp->curr == -HT_NBITS) {
-			if (likely(offset == 0)) {
-				ADJUST_EXTRA_RSPACE(req);
+			if (likely(offset == 0))
 				return T_OK;
-			} else {
+			else
 				return T_COMPRESSION;
-			}
 		}
 
 		i = (hp->hctx << -hp->curr) & HT_NMASK;
@@ -549,10 +597,8 @@ do {						\
 			 */
 			if (likely(offset == 0)) {
 				if ((i ^ (HT_EOS_HIGH >> 1)) <
-				    (1U << -hp->curr)) {
-					ADJUST_EXTRA_RSPACE(req);
+				    (1U << -hp->curr))
 					return T_OK;
-				}
 			}
 			/*
 			 * The first condition here equivalent to the
@@ -576,10 +622,8 @@ do {						\
 		}
 	}
 	if (likely(offset == 0)) {
-		if ((i ^ (HT_EOS_HIGH >> 1)) < (1U << -hp->curr)) {
-			ADJUST_EXTRA_RSPACE(req);
+		if ((i ^ (HT_EOS_HIGH >> 1)) < (1U << -hp->curr))
 			return T_OK;
-		}
 	}
 	return T_COMPRESSION;
 

fw/hpack.h

@@ -296,17 +296,6 @@ typedef struct {
 	TfwStr			hdr_data;
 } TfwDecodeCacheIter;
 
-#define	BUFFER_GET(len, it)					\
-do {								\
-	BUG_ON(!(len));						\
-	WARN_ON_ONCE((it)->rspace);				\
-	(it)->rspace = len;					\
-	(it)->pos = tfw_pool_alloc_not_align((it)->pool, len);	\
-	T_DBG3("%s: get buffer, len=%lu, it->pos=[%p],"		\
-	       " it->pos=%lu\n", __func__, (unsigned long)len,	\
-	       (it)->pos, (unsigned long)(it)->pos);		\
-} while (0)
-
 void write_int(unsigned long index, unsigned short max, unsigned short mask,
 	       TfwHPackInt *__restrict res_idx);
 int tfw_hpack_init(TfwHPack *__restrict hp, unsigned int htbl_sz);

fw/msg.h

@@ -67,30 +67,33 @@ typedef struct {
  * @pool	- allocation pool for target buffer of decoded headers;
  * @parsed_hdr	- pointer to the message header which is currently processed;
  * @hdrs_len	- accumulated length of message's decoded and parsed headers;
+ * @rspace	- space remained in the allocated chunk;
+ * @pos		- pointer to the currently allocated chunk of decoded headers'
+ *		  buffer;
  * @hdrs_cnt	- count of all headers from message headers block;
  * @__off	- offset for iterator reinitializing before next processing
  *		  stage;
  * @hdr		- descriptor of currently decoded header in target buffer;
- * @pos		- pointer to the currently allocated chunk of decoded headers'
- *		  buffer;
- * @rspace	- space remained in the allocated chunk;
  * @next	- pointer to the decoded header part (name/value) to be
  *		- parsed next;
  * @nm_len	- length of the decoded header's name;
+ * @to_alloc	- count of bytes which, should be allocated, after rspace
+ *		  will be exceeded;
  * @nm_num	- chunks number of the decoded header's name;
  * @tag		- tag of currently processed decoded header.
  */
 typedef struct {
 	TfwPool		*pool;
 	TfwStr		*parsed_hdr;
 	unsigned long	hdrs_len;
+	unsigned long	rspace;
+	char		*pos;
 	unsigned int	hdrs_cnt;
 	char		__off[0];
 	TfwStr		hdr;
-	char		*pos;
-	unsigned long	rspace;
 	TfwStr		*next;
 	unsigned long	nm_len;
+	unsigned long	to_alloc;
 	unsigned int	nm_num;
 	unsigned int	tag;
 } TfwMsgParseIter;