Home | History | Annotate | Line # | Download | only in netinet
tcp_vtw.h revision 1.1
      1 /*
      2  * Copyright (c) 2011 The NetBSD Foundation, Inc.
      3  * All rights reserved.
      4  *
      5  * This code is derived from software contributed to The NetBSD Foundation
      6  * by Coyote Point Systems, Inc.
      7  *
      8  * Redistribution and use in source and binary forms, with or without
      9  * modification, are permitted provided that the following conditions
     10  * are met:
     11  * 1. Redistributions of source code must retain the above copyright
     12  *    notice, this list of conditions and the following disclaimer.
     13  * 2. Redistributions in binary form must reproduce the above copyright
     14  *    notice, this list of conditions and the following disclaimer in the
     15  *    documentation and/or other materials provided with the distribution.
     16  *
     17  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     18  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     19  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     20  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     21  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     22  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     23  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     24  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     25  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     26  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     27  * POSSIBILITY OF SUCH DAMAGE.
     28  */
     29 
     30 /*
     31  * Vestigial time-wait.
     32  *
     33  * This implementation uses cache-efficient techniques, which will
     34  * appear somewhat peculiar.  The main philosophy is to optimise the
     35  * amount of information available within a cache line.  Cache miss is
     36  * expensive.  So we employ ad-hoc techniques to pull a series of
     37  * linked-list follows into a cache line.  One cache line, multiple
     38  * linked-list equivalents.
     39  *
     40  * One such ad-hoc technique is fat pointers.  Additional degrees of
     41  * ad-hoqueness result from having to hand tune it for pointer size
     42  * and for cache line size.
     43  *
     44  * The 'fat pointer' approach aggregates, for x86_32, 15 linked-list
     45  * data structures into one cache line.  The additional 32 bits in the
     46  * cache line are used for linking fat pointers, and for
     47  * allocation/bookkeeping.
     48  *
     49  * The 15 32-bit tags encode the pointers to the linked list elements,
     50  * and also encode the results of a search comparison.
     51  *
     52  * First, some more assumptions/restrictions.
     53  *
     54  * All the fat pointers are from a contiguous allocation arena.  Thus,
     55  * we can refer to them by offset from a base, not as full pointers.
     56  *
     57  * All the linked list data elements are also from a contiguous
     58  * allocation arena, again so that we can refer to them as offset from
     59  * a base.
     60  *
     61  * In order to add a data element to a fat pointer, a key value is
     62  * computed, based on unique data within the data element.  It is the
     63  * linear searching of the linked lists of these elements based on
     64  * these unique data that are being optimised here.
     65  *
     66  * Lets call the function that computes the key k(e), where e is the
     67  * data element.  In this example, k(e) returns 32-bits.
     68  *
     69  * Consider a set E (say of order 15) of data elements.  Let K be
     70  * the set of the k(e) for e in E.
     71  *
     72  * Let O be the set of the offsets from the base of the data elements in E.
     73  *
     74  * For each x in K, for each matching o in O, let t be x ^ o.  These
     75  * are the tags. (More or less).
     76  *
     77  * In order to search all the data elements in E, we compute the
     78  * search key, and one at a time, XOR the key into the tags.  If any
     79  * result is a valid data element index, we have a possible match.  If
     80  * not, there is no match.
     81  *
     82  * The no-match cases mean we do not have to de-reference the pointer
     83  * to the data element in question.  We save cache miss penalty and
     84  * cache load decreases.  Only in the case of a valid looking data
     85  * element index, do we have to look closer.
     86  *
     87  * Thus, in the absence of false positives, 15 data elements can be
     88  * searched with one cache line fill, as opposed to 15 cache line
     89  * fills for the usual implementation.
     90  *
     91  * The vestigial time waits (vtw_t), the data elements in the above, are
     92  * searched by faddr, fport, laddr, lport.  The key is a function of
     93  * these values.
     94  *
     95  * We hash these keys into the traditional hash chains to reduce the
     96  * search time, and use fat pointers to reduce the cache impacts of
     97  * searching.
     98  *
     99  * The vtw_t are, per requirement, in a contiguous chunk.  Allocation
    100  * is done with a clock hand, and all vtw_t within one allocation
    101  * domain have the same lifetime, so they will always be sorted by
    102  * age.
    103  *
    104  * A vtw_t will be allocated, timestamped, and have a fixed future
    105  * expiration.  It will be added to a hash bucket implemented with fat
    106  * pointers, which means that a cache line will be allocated in the
    107  * hash bucket, placed at the head (more recent in time) and the vtw_t
    108  * will be added to this.  As more entries are added, the fat pointer
    109  * cache line will fill, requiring additional cache lines for fat
    110  * pointers to be allocated. These will be added at the head, and the
    111  * aged entries will hang down, tapeworm like.  As the vtw_t entries
    112  * expire, the corresponding slot in the fat pointer will be
    113  * reclaimed, and eventually the cache line will completely empty and
    114  * be re-cycled, if not at the head of the chain.
    115  *
    116  * At times, a time-wait timer is restarted.  This corresponds to
    117  * deleting the current entry and re-adding it.
    118  *
    119  * Most of the time, they are just placed here to die.
    120  */
    121 #ifndef _NETINET_TCP_VTW_H
    122 #define _NETINET_TCP_VTW_H
    123 
    124 #include <sys/types.h>
    125 #include <sys/socket.h>
    126 #include <sys/sysctl.h>
    127 #include <net/if.h>
    128 #include <net/route.h>
    129 #include <netinet/in.h>
    130 #include <netinet/in_systm.h>
    131 #include <netinet/ip.h>
    132 #include <netinet/in_pcb.h>
    133 #include <netinet/in_var.h>
    134 #include <netinet/ip_var.h>
    135 #include <netinet/in.h>
    136 #include <netinet/tcp.h>
    137 #include <netinet/tcp_timer.h>
    138 #include <netinet/tcp_var.h>
    139 #include <netinet6/in6.h>
    140 #include <netinet/ip6.h>
    141 #include <netinet6/ip6_var.h>
    142 #include <netinet6/in6_pcb.h>
    143 #include <netinet6/ip6_var.h>
    144 #include <netinet6/in6_var.h>
    145 #include <netinet/icmp6.h>
    146 #include <netinet6/nd6.h>
    147 
    148 #define	VTW_NCLASS	(1+3)		/* # different classes */
    149 
    150 /*
    151  * fat pointers, MI.
    152  */
    153 struct fatp_mi;
    154 
    155 typedef uint32_t fatp_word_t;
    156 
    157 typedef struct fatp_mi	fatp_t;
    158 
    159 /* Supported cacheline sizes: 32 64 128 bytes.  See fatp_key(),
    160  * fatp_slot_from_key(), fatp_xtra[].
    161  */
    162 #define	FATP_NTAGS	(CACHE_LINE_SIZE / sizeof(fatp_word_t) - 1)
    163 #define	FATP_NXT_WIDTH	(sizeof(fatp_word_t) * NBBY - FATP_NTAGS)
    164 
    165 #define	FATP_MAX	(1 << FATP_NXT_WIDTH)
    166 
    167 /* Worked example: ULP32 with 64-byte cacheline (32-bit x86):
    168  * 15 tags per cacheline.  At most 2^17 fat pointers per fatp_ctl_t.
    169  * The comments on the fatp_mi members, below, correspond to the worked
    170  * example.
    171  */
    172 struct fatp_mi {
    173 	fatp_word_t	inuse	: FATP_NTAGS;	/* (1+15)*4 == CL_SIZE */
    174 	fatp_word_t	nxt	: FATP_NXT_WIDTH;/* at most 2^17 fat pointers */
    175 	fatp_word_t	tag[FATP_NTAGS];	/* 15 tags per CL */
    176 };
    177 
    178 static inline int
    179 fatp_ntags(void)
    180 {
    181 	return FATP_NTAGS;
    182 }
    183 
    184 static inline int
    185 fatp_full(fatp_t *fp)
    186 {
    187 	fatp_t full;
    188 
    189 	full.inuse = ~0;
    190 
    191 	return (fp->inuse == full.inuse);
    192 }
    193 
    194 struct vtw_common;
    195 struct vtw_v4;
    196 struct vtw_v6;
    197 struct vtw_ctl;
    198 
    199 /*!\brief common to all vtw
    200  */
    201 typedef struct vtw_common {
    202 	struct timeval	expire;		/* date of birth+msl */
    203 	uint32_t	key;		/* hash key: full hash */
    204 	uint32_t	port_key;	/* hash key: local port hash */
    205 	uint32_t	rcv_nxt;
    206 	uint32_t	rcv_wnd;
    207 	uint32_t	snd_nxt;
    208 	uint32_t	snd_scale	: 8;	/* window scaling for send win */
    209 	uint32_t	msl_class	: 2;	/* TCP MSL class {0,1,2,3} */
    210 	uint32_t	reuse_port	: 1;
    211 	uint32_t	reuse_addr	: 1;
    212 	uint32_t	v6only		: 1;
    213 	uint32_t	hashed		: 1;	/* reachable via FATP */
    214 	uint32_t	uid;
    215 } vtw_t;
    216 
    217 /*!\brief vestigial timewait for IPv4
    218  */
    219 typedef struct vtw_v4 {
    220 	vtw_t		common;		/*  must be first */
    221 	uint16_t	lport;
    222 	uint16_t	fport;
    223 	uint32_t	laddr;
    224 	uint32_t	faddr;
    225 } vtw_v4_t;
    226 
    227 /*!\brief vestigial timewait for IPv6
    228  */
    229 typedef struct vtw_v6 {
    230 	vtw_t		common;		/* must be first */
    231 	uint16_t	lport;
    232 	uint16_t	fport;
    233 	struct in6_addr	laddr;
    234 	struct in6_addr	faddr;
    235 } vtw_v6_t;
    236 
    237 struct fatp_ctl;
    238 typedef struct vtw_ctl		vtw_ctl_t;
    239 typedef struct fatp_ctl		fatp_ctl_t;
    240 
    241 /*
    242  * The vestigeal time waits are kept in a contiguous chunk.
    243  * Allocation and free pointers run as clock hands thru this array.
    244  */
    245 struct vtw_ctl {
    246 	fatp_ctl_t	*fat;		/* collection of fatp to use	*/
    247 	vtw_ctl_t	*ctl;		/* <! controller's controller	*/
    248 	union {
    249 		vtw_t		*v;	/* common			*/
    250 		struct vtw_v4	*v4;	/* IPv4 resources		*/
    251 		struct vtw_v6	*v6;	/* IPv6 resources		*/
    252 	}		base,		/* base of vtw_t array		*/
    253 		/**/	lim,		/* extent of vtw_t array	*/
    254 		/**/	alloc,		/* allocation pointer		*/
    255 		/**/	oldest;		/* ^ to oldest			*/
    256 	uint32_t	nfree;		/* # free			*/
    257 	uint32_t	nalloc;		/* # allocated			*/
    258 	uint32_t	idx_mask;	/* mask capturing all index bits*/
    259 	uint32_t	is_v4	: 1;
    260 	uint32_t	is_v6	: 1;
    261 	uint32_t	idx_bits: 6;
    262 	uint32_t	clidx	: 3;	/* <! class index */
    263 };
    264 
    265 /*!\brief Collections of fat pointers.
    266  */
    267 struct fatp_ctl {
    268 	vtw_ctl_t	*vtw;		/* associated VTWs		*/
    269 	fatp_t		*base;		/* base of fatp_t array		*/
    270 	fatp_t		*lim;		/* extent of fatp_t array	*/
    271 	fatp_t		*free;		/* free list			*/
    272 	uint32_t	mask;		/* hash mask			*/
    273 	uint32_t	nfree;		/* # free			*/
    274 	uint32_t	nalloc;		/* # allocated			*/
    275 	fatp_t		**hash;		/* hash anchors			*/
    276 	fatp_t		**port;		/* port hash anchors		*/
    277 };
    278 
    279 /*!\brief stats
    280  */
    281 struct vtw_stats {
    282 	uint64_t	ins;		/* <! inserts */
    283 	uint64_t	del;		/* <! deleted */
    284 	uint64_t	kill;		/* <! assassination */
    285 	uint64_t	look[2];	/* <! lookup: full hash, port hash */
    286 	uint64_t	hit[2];		/* <! lookups that hit */
    287 	uint64_t	miss[2];	/* <! lookups that miss */
    288 	uint64_t	probe[2];	/* <! hits+miss */
    289 	uint64_t	losing[2];	/* <! misses requiring dereference */
    290 	uint64_t	max_chain[2];	/* <! max fatp chain traversed */
    291 	uint64_t	max_probe[2];	/* <! max probes in any one chain */
    292 	uint64_t	max_loss[2];	/* <! max losing probes in any one
    293 					 * chain
    294 					 */
    295 };
    296 
    297 typedef struct vtw_stats	vtw_stats_t;
    298 
    299 /*!\brief	follow fatp next 'pointer'
    300  */
    301 static inline fatp_t *
    302 fatp_next(fatp_ctl_t *fat, fatp_t *fp)
    303 {
    304 	return fp->nxt ? fat->base + fp->nxt-1 : 0;
    305 }
    306 
    307 /*!\brief determine a collection-relative fat pointer index.
    308  */
    309 static inline uint32_t
    310 fatp_index(fatp_ctl_t *fat, fatp_t *fp)
    311 {
    312 	return fp ? 1 + (fp - fat->base) : 0;
    313 }
    314 
    315 
    316 static inline uint32_t
    317 v4_tag(uint32_t faddr, uint32_t fport, uint32_t laddr, uint32_t lport)
    318 {
    319 	return (ntohl(faddr)   + ntohs(fport)
    320 		+ ntohl(laddr) + ntohs(lport));
    321 }
    322 
    323 static inline uint32_t
    324 v6_tag(const struct in6_addr *faddr, uint16_t fport,
    325        const struct in6_addr *laddr, uint16_t lport)
    326 {
    327 #ifdef IN6_HASH
    328 	return IN6_HASH(faddr, fport, laddr, lport);
    329 #else
    330 	return 0;
    331 #endif
    332 }
    333 
    334 static inline uint32_t
    335 v4_port_tag(uint16_t lport)
    336 {
    337 	uint32_t tag = lport ^ (lport << 11);
    338 
    339 	tag ^= tag << 3;
    340 	tag += tag >> 5;
    341 	tag ^= tag << 4;
    342 	tag += tag >> 17;
    343 	tag ^= tag << 25;
    344 	tag += tag >> 6;
    345 
    346 	return tag;
    347 }
    348 
    349 static inline uint32_t
    350 v6_port_tag(uint16_t lport)
    351 {
    352 	return v4_port_tag(lport);
    353 }
    354 
    355 struct tcpcb;
    356 struct tcphdr;
    357 
    358 int  vtw_add(int, struct tcpcb *);
    359 void vtw_del(vtw_ctl_t *, vtw_t *);
    360 int vtw_lookup_v4(const struct ip *ip, const struct tcphdr *th,
    361 		  uint32_t faddr, uint16_t fport,
    362 		  uint32_t laddr, uint16_t lport);
    363 #ifdef INET6
    364 struct ip6_hdr;
    365 struct in6_addr;
    366 
    367 int vtw_lookup_v6(const struct ip6_hdr *ip, const struct tcphdr *th,
    368 		  const struct in6_addr *faddr, uint16_t fport,
    369 		  const struct in6_addr *laddr, uint16_t lport);
    370 #endif
    371 
    372 typedef struct vestigial_inpcb {
    373 	union {
    374 		struct in_addr	v4;
    375 #ifdef INET6
    376 		struct in6_addr	v6;
    377 #endif
    378 	} faddr, laddr;
    379 	uint16_t		fport, lport;
    380 	uint32_t		valid		: 1;
    381 	uint32_t		v4		: 1;
    382 	uint32_t		reuse_addr	: 1;
    383 	uint32_t		reuse_port	: 1;
    384 	uint32_t		v6only		: 1;
    385 	uint32_t		more_tbd	: 1;
    386 	uint32_t		uid;
    387 	uint32_t		rcv_nxt;
    388 	uint32_t		rcv_wnd;
    389 	uint32_t		snd_nxt;
    390 	struct vtw_common	*vtw;
    391 	struct vtw_ctl		*ctl;
    392 } vestigial_inpcb_t;
    393 
    394 void vtw_restart(vestigial_inpcb_t*);
    395 int vtw_earlyinit(void);
    396 
    397 #ifdef VTW_DEBUG
    398 typedef struct sin_either {
    399 	uint8_t		sin_len;
    400 	uint8_t		sin_family;
    401 	uint16_t	sin_port;
    402 	union {
    403 		struct in_addr	v4;
    404 #ifdef INET6
    405 		struct in6_addr	v6;
    406 #endif
    407 	}		sin_addr;
    408 } sin_either_t;
    409 
    410 int vtw_debug_add(int af, sin_either_t *, sin_either_t *, int, int);
    411 
    412 typedef struct vtw_sysargs {
    413 	uint32_t	op;
    414 	sin_either_t	fa;
    415 	sin_either_t	la;
    416 } vtw_sysargs_t;
    417 
    418 #endif /* VTW_DEBUG */
    419 
    420 #endif /* _NETINET_TCP_VTW_H */
    421