189 struct hlist_head { 190 struct hlist_node *first; 191 }; 192 193 struct hlist_node { 194 struct hlist_node *next, **pprev; 195 };
hlist_head表示哈希表的头结点，是hash数组的一个元素。
哈希表中每一个entry(hlist_head)所对应的都是一个链表（hlist)，该链表的结点由hlist_node表示。
hlist_head结构体只有一个域，即first。 first指针指向该hlist链表的第一个节点。
hlist_node结构体有两个域，next 和pprev。 next指针很容易理解，它指向下个hlist_node结点，倘若该节点是链表的最后一个节点，next指向NULL。
pprev是一个二级指针， 它指向前一个节点的next指针。为什么我们需要这样一个指针呢？它的好处是什么？
在回答这个问题之前，我们先研究另一个问题：为什么散列表的实现需要两个不同的数据结构？
散列表的目的是为了方便快速的查找，所以散列表通常是一个比较大的数组，否则“冲突”的概率会非常大， 这样也就失去了散列表的意义。如何做到既能维护一张大表，又能不使用过多的内存呢？就只能从数据结构上下功夫了。所以对于散列表的每个entry，它的结构体中只存放一个指针，解决了占用空间的问题。
现在又出现了另一个问题：数据结构不一致。显然，如果hlist_node采用传统的next,prev指针， 对于第一个节点和后面其他节点的处理会不一致。这样并不优雅，而且效率上也有损失。
hlist_node巧妙地将pprev指向上一个节点的next指针的地址，由于hlist_head和hlist_node指向的下一个节点的指针类型相同，这样就解决了通用性！
以IP分片时hlist使用为例，说明一下hlist的使用。
31 /** 32 * struct inet_frag_queue - fragment queue 33 * 34 * @lock: spinlock protecting the queue 35 * @timer: queue expiration timer 36 * @list: hash bucket list 37 * @refcnt: reference count of the queue 38 * @fragments: received fragments head 39 * @fragments_tail: received fragments tail 40 * @stamp: timestamp of the last received fragment 41 * @len: total length of the original datagram 42 * @meat: length of received fragments so far 43 * @flags: fragment queue flags 44 * @max_size: maximum received fragment size 45 * @net: namespace that this frag belongs to 46 * @list_evictor: list of queues to forcefully evict (e.g. due to low memory) 47 */ 48 struct inet_frag_queue { 49 spinlock_t lock; 50 struct timer_list timer; 51 struct hlist_node list; //hash节点 52 atomic_t refcnt; 53 struct sk_buff *fragments; 54 struct sk_buff *fragments_tail; 55 ktime_t stamp; 56 int len; 57 int meat; 58 __u8 flags; 59 u16 max_size; 60 struct netns_frags *net; 61 struct hlist_node list_evictor; 62 }; 63 64 #define INETFRAGS_HASHSZ 1024 //hash数组的大小 65 66 /* averaged: 67 * max_depth = default ipfrag_high_thresh / INETFRAGS_HASHSZ / 68 * rounded up (SKB_TRUELEN(0) + sizeof(struct ipq or 69 * struct frag_queue)) 70 */ 71 #define INETFRAGS_MAXDEPTH 128 72 73 struct inet_frag_bucket { 74 struct hlist_head chain; //hash冲突链表 75 spinlock_t chain_lock; 76 }; 77 78 struct inet_frags { 79 struct inet_frag_bucket hash[INETFRAGS_HASHSZ]; //hash数组 80 81 struct work_struct frags_work; 82 unsigned int next_bucket; 83 unsigned long last_rebuild_jiffies; 84 bool rebuild; 85 86 /* The first call to hashfn is responsible to initialize 87 * rnd. This is best done with net_get_random_once. 88 * 89 * rnd_seqlock is used to let hash insertion detect 90 * when it needs to re-lookup the hash chain to use. 91 */ 92 u32 rnd; 93 seqlock_t rnd_seqlock; 94 int qsize; 95 96 unsigned int (*hashfn)(const struct inet_frag_queue *); 97 bool (*match)(const struct inet_frag_queue *q, 98 const void *arg); 99 void (*constructor)(struct inet_frag_queue *q, 100 const void *arg); 101 void (*destructor)(struct inet_frag_queue *); 102 void (*skb_free)(struct sk_buff *); 103 void (*frag_expire)(unsigned long data); 104 struct kmem_cache *frags_cachep; 105 const char *frags_cache_name; 106 };
1 /* 2 * inet fragments management 3 * 4 * This program is free software; you can redistribute it and/or 5 * modify it under the terms of the GNU General Public License 6 * as published by the Free Software Foundation; either version 7 * 2 of the License, or (at your option) any later version. 8 * 9 * Authors: Pavel Emelyanov <xemul@openvz.org> 10 * Started as consolidation of ipv4/ip_fragment.c, 11 * ipv6/reassembly. and ipv6 nf conntrack reassembly 12 */ 13 14 #include <linux/list.h> 15 #include <linux/spinlock.h> 16 #include <linux/module.h> 17 #include <linux/timer.h> 18 #include <linux/mm.h> 19 #include <linux/random.h> 20 #include <linux/skbuff.h> 21 #include <linux/rtnetlink.h> 22 #include <linux/slab.h> 23 24 #include <net/sock.h> 25 #include <net/inet_frag.h> 26 #include <net/inet_ecn.h> 27 28 #define INETFRAGS_EVICT_BUCKETS 128 29 #define INETFRAGS_EVICT_MAX 512 30 31 /* don't rebuild inetfrag table with new secret more often than this */ 32 #define INETFRAGS_MIN_REBUILD_INTERVAL (5 * HZ) 33 34 /* Given the OR values of all fragments, apply RFC 3168 5.3 requirements 35 * Value : 0xff if frame should be dropped. 36 * 0 or INET_ECN_CE value, to be ORed in to final iph->tos field 37 */ 38 const u8 ip_frag_ecn_table[16] = { 39 /* at least one fragment had CE, and others ECT_0 or ECT_1 */ 40 [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = INET_ECN_CE, 41 [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = INET_ECN_CE, 42 [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = INET_ECN_CE, 43 44 /* invalid combinations : drop frame */ 45 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE] = 0xff, 46 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0] = 0xff, 47 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_1] = 0xff, 48 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff, 49 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = 0xff, 50 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = 0xff, 51 [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff, 52 }; 53 EXPORT_SYMBOL(ip_frag_ecn_table); 54 55 static unsigned int //hash函数 56 inet_frag_hashfn(const struct inet_frags *f, const struct inet_frag_queue *q) 57 { 58 return f->hashfn(q) & (INETFRAGS_HASHSZ - 1); 59 } 60 61 static bool inet_frag_may_rebuild(struct inet_frags *f) 62 { 63 return time_after(jiffies, 64 f->last_rebuild_jiffies + INETFRAGS_MIN_REBUILD_INTERVAL); 65 } 66 67 static void inet_frag_secret_rebuild(struct inet_frags *f) 68 { 69 int i; 70 71 write_seqlock_bh(&f->rnd_seqlock); 72 73 if (!inet_frag_may_rebuild(f)) 74 goto out; 75 76 get_random_bytes(&f->rnd, sizeof(u32)); 77 78 for (i = 0; i < INETFRAGS_HASHSZ; i++) { 79 struct inet_frag_bucket *hb; 80 struct inet_frag_queue *q; 81 struct hlist_node *n; 82 83 hb = &f->hash[i]; 84 spin_lock(&hb->chain_lock); 85 86 hlist_for_each_entry_safe(q, n, &hb->chain, list) {//安全遍历所有hash冲突链表 87 unsigned int hval = inet_frag_hashfn(f, q); //计算hash值 88 89 if (hval != i) { 90 struct inet_frag_bucket *hb_dest; 91 92 hlist_del(&q->list); 93 94 /* Relink to new hash chain. */ 95 hb_dest = &f->hash[hval]; //hash冲突链表 96 97 /* This is the only place where we take 98 * another chain_lock while already holding 99 * one. As this will not run concurrently, 100 * we cannot deadlock on hb_dest lock below, if its 101 * already locked it will be released soon since 102 * other caller cannot be waiting for hb lock 103 * that we've taken above. 104 */ 105 spin_lock_nested(&hb_dest->chain_lock, 106 SINGLE_DEPTH_NESTING);
//把节点list加入到hash冲突链表chain里面 107 hlist_add_head(&q->list, &hb_dest->chain); 108 spin_unlock(&hb_dest->chain_lock); 109 } 110 } 111 spin_unlock(&hb->chain_lock); 112 } 113 114 f->rebuild = false; 115 f->last_rebuild_jiffies = jiffies; 116 out: 117 write_sequnlock_bh(&f->rnd_seqlock); 118 } 119 120 static bool inet_fragq_should_evict(const struct inet_frag_queue *q) 121 { 122 return q->net->low_thresh == 0 || 123 frag_mem_limit(q->net) >= q->net->low_thresh; 124 } 125 126 static unsigned int 127 inet_evict_bucket(struct inet_frags *f, struct inet_frag_bucket *hb) 128 { 129 struct inet_frag_queue *fq; 130 struct hlist_node *n; 131 unsigned int evicted = 0; 132 HLIST_HEAD(expired); 133 134 spin_lock(&hb->chain_lock); 135 136 hlist_for_each_entry_safe(fq, n, &hb->chain, list) { 137 if (!inet_fragq_should_evict(fq)) 138 continue; 139 140 if (!del_timer(&fq->timer)) 141 continue; 142 143 hlist_add_head(&fq->list_evictor, &expired); 144 ++evicted; 145 } 146 147 spin_unlock(&hb->chain_lock); 148 149 hlist_for_each_entry_safe(fq, n, &expired, list_evictor) 150 f->frag_expire((unsigned long) fq); 151 152 return evicted; 153 } 154 155 static void inet_frag_worker(struct work_struct *work) 156 { 157 unsigned int budget = INETFRAGS_EVICT_BUCKETS; 158 unsigned int i, evicted = 0; 159 struct inet_frags *f; 160 161 f = container_of(work, struct inet_frags, frags_work); 162 163 BUILD_BUG_ON(INETFRAGS_EVICT_BUCKETS >= INETFRAGS_HASHSZ); 164 165 local_bh_disable(); 166 167 for (i = ACCESS_ONCE(f->next_bucket); budget; --budget) { 168 evicted += inet_evict_bucket(f, &f->hash[i]); 169 i = (i + 1) & (INETFRAGS_HASHSZ - 1); 170 if (evicted > INETFRAGS_EVICT_MAX) 171 break; 172 } 173 174 f->next_bucket = i; 175 176 local_bh_enable(); 177 178 if (f->rebuild && inet_frag_may_rebuild(f)) 179 inet_frag_secret_rebuild(f); 180 } 181 182 static void inet_frag_schedule_worker(struct inet_frags *f) 183 { 184 if (unlikely(!work_pending(&f->frags_work))) 185 schedule_work(&f->frags_work); 186 } 187 188 int inet_frags_init(struct inet_frags *f) 189 { 190 int i; 191 192 INIT_WORK(&f->frags_work, inet_frag_worker); 193 194 for (i = 0; i < INETFRAGS_HASHSZ; i++) { //初始化hash数组中所有的元素 195 struct inet_frag_bucket *hb = &f->hash[i]; 196 197 spin_lock_init(&hb->chain_lock); //初始化一个hash冲突链表 198 INIT_HLIST_HEAD(&hb->chain); 199 } 200 201 seqlock_init(&f->rnd_seqlock); 202 f->last_rebuild_jiffies = 0; 203 f->frags_cachep = kmem_cache_create(f->frags_cache_name, f->qsize, 0, 0, 204 NULL); 205 if (!f->frags_cachep) 206 return -ENOMEM; 207 208 return 0; 209 } 210 EXPORT_SYMBOL(inet_frags_init); 211 212 void inet_frags_init_net(struct netns_frags *nf) 213 { 214 init_frag_mem_limit(nf); 215 } 216 EXPORT_SYMBOL(inet_frags_init_net); 217 218 void inet_frags_fini(struct inet_frags *f) 219 { 220 cancel_work_sync(&f->frags_work); 221 kmem_cache_destroy(f->frags_cachep); 222 } 223 EXPORT_SYMBOL(inet_frags_fini); 224 225 void inet_frags_exit_net(struct netns_frags *nf, struct inet_frags *f) 226 { 227 unsigned int seq; 228 int i; 229 230 nf->low_thresh = 0; 231 232 evict_again: 233 local_bh_disable(); 234 seq = read_seqbegin(&f->rnd_seqlock); 235 236 for (i = 0; i < INETFRAGS_HASHSZ ; i++) 237 inet_evict_bucket(f, &f->hash[i]); 238 239 local_bh_enable(); 240 cond_resched(); 241 242 if (read_seqretry(&f->rnd_seqlock, seq) || 243 percpu_counter_sum(&nf->mem)) 244 goto evict_again; 245 246 percpu_counter_destroy(&nf->mem); 247 } 248 EXPORT_SYMBOL(inet_frags_exit_net); 249 250 static struct inet_frag_bucket * 251 get_frag_bucket_locked(struct inet_frag_queue *fq, struct inet_frags *f) 252 __acquires(hb->chain_lock) 253 { 254 struct inet_frag_bucket *hb; 255 unsigned int seq, hash; 256 257 restart: 258 seq = read_seqbegin(&f->rnd_seqlock); 259 260 hash = inet_frag_hashfn(f, fq); 261 hb = &f->hash[hash]; 262 263 spin_lock(&hb->chain_lock); 264 if (read_seqretry(&f->rnd_seqlock, seq)) { 265 spin_unlock(&hb->chain_lock); 266 goto restart; 267 } 268 269 return hb; 270 } 271 272 static inline void fq_unlink(struct inet_frag_queue *fq, struct inet_frags *f) 273 { 274 struct inet_frag_bucket *hb; 275 276 hb = get_frag_bucket_locked(fq, f); 277 hlist_del(&fq->list); 278 fq->flags |= INET_FRAG_COMPLETE; 279 spin_unlock(&hb->chain_lock); 280 } 281 282 void inet_frag_kill(struct inet_frag_queue *fq, struct inet_frags *f) 283 { 284 if (del_timer(&fq->timer)) 285 atomic_dec(&fq->refcnt); 286 287 if (!(fq->flags & INET_FRAG_COMPLETE)) { 288 fq_unlink(fq, f); 289 atomic_dec(&fq->refcnt); 290 } 291 } 292 EXPORT_SYMBOL(inet_frag_kill); 293 294 static inline void frag_kfree_skb(struct netns_frags *nf, struct inet_frags *f, 295 struct sk_buff *skb) 296 { 297 if (f->skb_free) 298 f->skb_free(skb); 299 kfree_skb(skb); 300 } 301 302 void inet_frag_destroy(struct inet_frag_queue *q, struct inet_frags *f) 303 { 304 struct sk_buff *fp; 305 struct netns_frags *nf; 306 unsigned int sum, sum_truesize = 0; 307 308 WARN_ON(!(q->flags & INET_FRAG_COMPLETE)); 309 WARN_ON(del_timer(&q->timer) != 0); 310 311 /* Release all fragment data. */ 312 fp = q->fragments; 313 nf = q->net; 314 while (fp) { 315 struct sk_buff *xp = fp->next; 316 317 sum_truesize += fp->truesize; 318 frag_kfree_skb(nf, f, fp); 319 fp = xp; 320 } 321 sum = sum_truesize + f->qsize; 322 323 if (f->destructor) 324 f->destructor(q); 325 kmem_cache_free(f->frags_cachep, q); 326 327 sub_frag_mem_limit(nf, sum); 328 } 329 EXPORT_SYMBOL(inet_frag_destroy); 330 331 static struct inet_frag_queue *inet_frag_intern(struct netns_frags *nf, 332 struct inet_frag_queue *qp_in, 333 struct inet_frags *f, 334 void *arg) 335 { 336 struct inet_frag_bucket *hb = get_frag_bucket_locked(qp_in, f); 337 struct inet_frag_queue *qp; 338 339 #ifdef CONFIG_SMP 340 /* With SMP race we have to recheck hash table, because 341 * such entry could have been created on other cpu before 342 * we acquired hash bucket lock. 343 */ 344 hlist_for_each_entry(qp, &hb->chain, list) { 345 if (qp->net == nf && f->match(qp, arg)) { 346 atomic_inc(&qp->refcnt); 347 spin_unlock(&hb->chain_lock); 348 qp_in->flags |= INET_FRAG_COMPLETE; 349 inet_frag_put(qp_in, f); 350 return qp; 351 } 352 } 353 #endif 354 qp = qp_in; 355 if (!mod_timer(&qp->timer, jiffies + nf->timeout)) 356 atomic_inc(&qp->refcnt); 357 358 atomic_inc(&qp->refcnt); 359 hlist_add_head(&qp->list, &hb->chain); 360 361 spin_unlock(&hb->chain_lock); 362 363 return qp; 364 } 365 366 static struct inet_frag_queue *inet_frag_alloc(struct netns_frags *nf, 367 struct inet_frags *f, 368 void *arg) 369 { 370 struct inet_frag_queue *q; 371 372 if (frag_mem_limit(nf) > nf->high_thresh) { 373 inet_frag_schedule_worker(f); 374 return NULL; 375 } 376 377 q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC); 378 if (!q) 379 return NULL; 380 381 q->net = nf; 382 f->constructor(q, arg); 383 add_frag_mem_limit(nf, f->qsize); 384 385 setup_timer(&q->timer, f->frag_expire, (unsigned long)q); 386 spin_lock_init(&q->lock); 387 atomic_set(&q->refcnt, 1); 388 389 return q; 390 } 391 392 static struct inet_frag_queue *inet_frag_create(struct netns_frags *nf, 393 struct inet_frags *f, 394 void *arg) 395 { 396 struct inet_frag_queue *q; 397 398 q = inet_frag_alloc(nf, f, arg); 399 if (!q) 400 return NULL; 401 402 return inet_frag_intern(nf, q, f, arg); 403 } 404 405 struct inet_frag_queue *inet_frag_find(struct netns_frags *nf, 406 struct inet_frags *f, void *key, 407 unsigned int hash) 408 { 409 struct inet_frag_bucket *hb; 410 struct inet_frag_queue *q; 411 int depth = 0; 412 413 if (frag_mem_limit(nf) > nf->low_thresh) 414 inet_frag_schedule_worker(f); 415 416 hash &= (INETFRAGS_HASHSZ - 1); 417 hb = &f->hash[hash]; 418 419 spin_lock(&hb->chain_lock); 420 hlist_for_each_entry(q, &hb->chain, list) { 421 if (q->net == nf && f->match(q, key)) { 422 atomic_inc(&q->refcnt); 423 spin_unlock(&hb->chain_lock); 424 return q; 425 } 426 depth++; 427 } 428 spin_unlock(&hb->chain_lock); 429 430 if (depth <= INETFRAGS_MAXDEPTH) 431 return inet_frag_create(nf, f, key); 432 433 if (inet_frag_may_rebuild(f)) { 434 if (!f->rebuild) 435 f->rebuild = true; 436 inet_frag_schedule_worker(f); 437 } 438 439 return ERR_PTR(-ENOBUFS); 440 } 441 EXPORT_SYMBOL(inet_frag_find); 442 443 void inet_frag_maybe_warn_overflow(struct inet_frag_queue *q, 444 const char *prefix) 445 { 446 static const char msg[] = "inet_frag_find: Fragment hash bucket" 447 " list length grew over limit " __stringify(INETFRAGS_MAXDEPTH) 448 ". Dropping fragment.\n"; 449 450 if (PTR_ERR(q) == -ENOBUFS) 451 net_dbg_ratelimited("%s%s", prefix, msg); 452 } 453 EXPORT_SYMBOL(inet_frag_maybe_warn_overflow); 454
