blob: 385f3860bbb27a8f45b9d30b56d44b3fca93ff33 [file] [log] [blame]
/*
* HPACK header table management (RFC7541) - type definitions and prototypes
*
* Copyright (C) 2014-2017 Willy Tarreau <willy@haproxy.org>
* Copyright (C) 2017 HAProxy Technologies
*
* Permission is hereby granted, free of charge, to any person obtaining
* a copy of this software and associated documentation files (the
* "Software"), to deal in the Software without restriction, including
* without limitation the rights to use, copy, modify, merge, publish,
* distribute, sublicense, and/or sell copies of the Software, and to
* permit persons to whom the Software is furnished to do so, subject to
* the following conditions:
*
* The above copyright notice and this permission notice shall be
* included in all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
* OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
* HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
* WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
* OTHER DEALINGS IN THE SOFTWARE.
*/
#ifndef _COMMON_HPACK_TBL_H
#define _COMMON_HPACK_TBL_H
#include <stdint.h>
#include <stdlib.h>
#include <common/config.h>
#include <common/http-hdr.h>
#include <common/ist.h>
/* Dynamic Headers Table, usable for tables up to 4GB long and values of 64kB-1.
* The model can be improved by using offsets relative to the table entry's end
* or to the end of the area, or by moving the descriptors at the end of the
* table and the data at the beginning. This entry is 8 bytes long, which is 1/4
* of the bookkeeping planned by the HPACK spec. Thus it saves 24 bytes per
* header field, meaning that even with a single header, 24 extra bytes can be
* stored (ie one such descriptor). At 29.2 average bytes per header field as
* found in the hpack test case, that's slightly more than 1.5kB of space saved
* from a 4kB block, resulting in contiguous space almost always being
* available.
*
* Principle: the table is stored in a contiguous array containing both the
* descriptors and the contents. Descriptors are stored at the beginning of the
* array while contents are stored starting from the end. Most of the time there
* is enough room left in the table to insert a new header field, thanks to the
* savings on the descriptor size. Thus by inserting headers from the end it's
* possible to maximize the delay before a collision of DTEs and data. In order
* to always insert from the right, we need to keep a reference to the latest
* inserted element and look before it. The last inserted cell's address defines
* the lowest konwn address still in use, unless the area wraps in which case
* the available space lies between the end of the tail and the beginning of the
* head.
*
* In order to detect collisions between data blocks and DTEs, we also maintain
* an index to the lowest element facing the DTE table, called "front". This one
* is updated each time an element is inserted before it. Once the buffer wraps,
* this element doesn't have to be updated anymore until it is released, in
* which case the buffer doesn't wrap anymore and the front element becomes the
* head again.
*
* Various heuristics are possible concerning the opportunity to wrap the
* entries to limit the risk of collisions with the DTE, but experimentation
* shows that thanks to the important savings made on the descriptors, the
* likeliness of finding a large amount of free space at the end of the area is
* much higher than the risk of colliding, so in the end the most naive
* algorithms work pretty fine. Typical ratios of 1 collision per 2000 requests
* have been observed.
*
* The defragmentation should be rare ; a study on live data shows on average
* 29.2 bytes used per header field. This plus the 32 bytes overhead fix an
* average of 66.9 header fields per 4kB table. This brings a 1606 bytes saving
* using the current storage description, ensuring that oldest headers are
* linearly removed by the sender before fragmentation occurs. This means that
* for all smaller header fields there will not be any requirement to defragment
* the area and most of the time it will even be possible to copy the old values
* directly within the buffer after creating a new entry. On average within the
* available space there will be enough room to store 1606/(29.2+8)=43 extra
* header fields without switching to another place.
*
* The table header fits in the table itself, it only takes 16 bytes, so in the
* worst case (1 single header) it's possible to store 4096 - 16 - 8 = 4072
* data bytes, which is larger than the 4064 the protocol requires (4096 - 32).
*/
/* One dynamic table entry descriptor */
struct hpack_dte {
uint32_t addr; /* storage address, relative to the dte address */
uint16_t nlen; /* header name length */
uint16_t vlen; /* header value length */
};
/* Note: the table's head plus a struct hpack_dte must be smaller than or equal to 32
* bytes so that a single large header can always fit. Here that's 16 bytes for
* the header, plus 8 bytes per slot.
* Note that when <used> == 0, front, head, and wrap are undefined.
*/
struct hpack_dht {
uint32_t size; /* allocated table size in bytes */
uint32_t total; /* sum of nlen + vlen in bytes */
uint16_t front; /* slot number of the first node after the idx table */
uint16_t wrap; /* number of allocated slots, wraps here */
uint16_t head; /* last inserted slot number */
uint16_t used; /* number of slots in use */
struct hpack_dte dte[0]; /* dynamic table entries */
};
/* supported hpack encoding/decoding errors */
enum {
HPACK_ERR_NONE = 0, /* no error */
HPACK_ERR_ALLOC_FAIL, /* memory allocation error */
HPACK_ERR_UNKNOWN_OPCODE, /* invalid first byte */
HPACK_ERR_TRUNCATED, /* truncated stream */
HPACK_ERR_HUFFMAN, /* huffman decoding error */
HPACK_ERR_INVALID_PHDR, /* invalid pseudo header field name */
HPACK_ERR_MISPLACED_PHDR, /* pseudo header field after a regular header field */
HPACK_ERR_DUPLICATE_PHDR, /* duplicate pseudo header field */
HPACK_ERR_DHT_INSERT_FAIL, /* failed to insert into DHT */
HPACK_ERR_TOO_LARGE, /* decoded request/response is too large */
HPACK_ERR_MISSING_METHOD, /* :method is missing */
HPACK_ERR_MISSING_SCHEME, /* :scheme is missing */
HPACK_ERR_MISSING_PATH, /* :path is missing */
HPACK_ERR_MISSING_AUTHORITY, /* :authority is missing with CONNECT */
HPACK_ERR_SCHEME_NOT_ALLOWED, /* :scheme not allowed with CONNECT */
HPACK_ERR_PATH_NOT_ALLOWED, /* :path not allowed with CONNECT */
};
/* static header table as in RFC7541 Appendix A. [0] unused. */
#define HPACK_SHT_SIZE 62
extern const struct http_hdr hpack_sht[HPACK_SHT_SIZE];
extern int __hpack_dht_make_room(struct hpack_dht *dht, unsigned int needed);
extern int hpack_dht_insert(struct hpack_dht *dht, struct ist name, struct ist value);
/* return a pointer to the entry designated by index <idx> (starting at 1) or
* NULL if this index is not there.
*/
static inline const struct hpack_dte *hpack_get_dte(const struct hpack_dht *dht, uint16_t idx)
{
idx--;
if (idx >= dht->used)
return NULL;
if (idx <= dht->head)
idx = dht->head - idx;
else
idx = dht->head - idx + dht->wrap;
return &dht->dte[idx];
}
/* returns non-zero if <idx> is valid for table <dht> */
static inline int hpack_valid_idx(const struct hpack_dht *dht, uint32_t idx)
{
return idx < dht->used + HPACK_SHT_SIZE;
}
/* return a pointer to the header name for entry <dte>. */
static inline struct ist hpack_get_name(const struct hpack_dht *dht, const struct hpack_dte *dte)
{
struct ist ret = {
.ptr = (void *)dht + dte->addr,
.len = dte->nlen,
};
return ret;
}
/* return a pointer to the header value for entry <dte>. */
static inline struct ist hpack_get_value(const struct hpack_dht *dht, const struct hpack_dte *dte)
{
struct ist ret = {
.ptr = (void *)dht + dte->addr + dte->nlen,
.len = dte->vlen,
};
return ret;
}
/* takes an idx, returns the associated name */
static inline struct ist hpack_idx_to_name(const struct hpack_dht *dht, uint32_t idx)
{
const struct hpack_dte *dte;
if (idx < HPACK_SHT_SIZE)
return hpack_sht[idx].n;
dte = hpack_get_dte(dht, idx - HPACK_SHT_SIZE + 1);
if (!dte)
return ist("### ERR ###"); // error
return hpack_get_name(dht, dte);
}
/* takes an idx, returns the associated value */
static inline struct ist hpack_idx_to_value(const struct hpack_dht *dht, uint32_t idx)
{
const struct hpack_dte *dte;
if (idx < HPACK_SHT_SIZE)
return hpack_sht[idx].v;
dte = hpack_get_dte(dht, idx - HPACK_SHT_SIZE + 1);
if (!dte)
return ist("### ERR ###"); // error
return hpack_get_value(dht, dte);
}
/* Purges table dht until a header field of <needed> bytes fits according to
* the protocol (adding 32 bytes overhead). Returns non-zero on success, zero
* on failure (ie: table empty but still not sufficient).
*/
static inline int hpack_dht_make_room(struct hpack_dht *dht, unsigned int needed)
{
if (dht->used * 32 + dht->total + needed + 32 <= dht->size)
return 1;
else if (!dht->used)
return 0;
return __hpack_dht_make_room(dht, needed);
}
/* allocate a dynamic headers table of <size> bytes and return it initialized */
static inline void hpack_dht_init(struct hpack_dht *dht, uint32_t size)
{
dht->size = size;
dht->total = 0;
dht->used = 0;
}
/* allocate a dynamic headers table of <size> bytes and return it initialized */
static inline struct hpack_dht *hpack_dht_alloc(uint32_t size)
{
struct hpack_dht *dht;
dht = malloc(size);
if (!dht)
return dht;
hpack_dht_init(dht, size);
return dht;
}
/* free a dynamic headers table */
static inline void hpack_dht_free(struct hpack_dht *dht)
{
free(dht);
}
#endif /* _COMMON_HPACK_TBL_H */