xref: /freebsd/sys/kern/subr_sfbuf.c (revision 031beb4e239bfce798af17f5fe8dba8bcaf13d99)
1 /*-
2  * Copyright (c) 2014 Gleb Smirnoff <glebius@FreeBSD.org>
3  * Copyright (c) 2003, 2005 Alan L. Cox <alc@cs.rice.edu>
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 #include <sys/param.h>
30 #include <sys/kernel.h>
31 #include <sys/lock.h>
32 #include <sys/malloc.h>
33 #include <sys/mutex.h>
34 #include <sys/proc.h>
35 #include <sys/sf_buf.h>
36 #include <sys/smp.h>
37 #include <sys/sysctl.h>
38 
39 #include <vm/vm.h>
40 #include <vm/vm_extern.h>
41 #include <vm/vm_page.h>
42 
43 #ifndef NSFBUFS
44 #define	NSFBUFS		(512 + maxusers * 16)
45 #endif
46 
47 static int nsfbufs;
48 static int nsfbufspeak;
49 static int nsfbufsused;
50 
51 SYSCTL_INT(_kern_ipc, OID_AUTO, nsfbufs, CTLFLAG_RDTUN, &nsfbufs, 0,
52     "Maximum number of sendfile(2) sf_bufs available");
53 SYSCTL_INT(_kern_ipc, OID_AUTO, nsfbufspeak, CTLFLAG_RD, &nsfbufspeak, 0,
54     "Number of sendfile(2) sf_bufs at peak usage");
55 SYSCTL_INT(_kern_ipc, OID_AUTO, nsfbufsused, CTLFLAG_RD, &nsfbufsused, 0,
56     "Number of sendfile(2) sf_bufs in use");
57 
58 static void	sf_buf_init(void *arg);
59 SYSINIT(sock_sf, SI_SUB_MBUF, SI_ORDER_ANY, sf_buf_init, NULL);
60 
61 LIST_HEAD(sf_head, sf_buf);
62 
63 /*
64  * A hash table of active sendfile(2) buffers
65  */
66 static struct sf_head *sf_buf_active;
67 static u_long sf_buf_hashmask;
68 
69 #define	SF_BUF_HASH(m)	(((m) - vm_page_array) & sf_buf_hashmask)
70 
71 static TAILQ_HEAD(, sf_buf) sf_buf_freelist;
72 static u_int	sf_buf_alloc_want;
73 
74 /*
75  * A lock used to synchronize access to the hash table and free list
76  */
77 static struct mtx sf_buf_lock;
78 
79 /*
80  * Allocate a pool of sf_bufs (sendfile(2) or "super-fast" if you prefer. :-))
81  */
82 static void
83 sf_buf_init(void *arg)
84 {
85 	struct sf_buf *sf_bufs;
86 	vm_offset_t sf_base;
87 	int i;
88 
89 	if (PMAP_HAS_DMAP)
90 		return;
91 
92 	nsfbufs = NSFBUFS;
93 	TUNABLE_INT_FETCH("kern.ipc.nsfbufs", &nsfbufs);
94 
95 	sf_buf_active = hashinit(nsfbufs, M_TEMP, &sf_buf_hashmask);
96 	TAILQ_INIT(&sf_buf_freelist);
97 	sf_base = kva_alloc(nsfbufs * PAGE_SIZE);
98 	sf_bufs = malloc(nsfbufs * sizeof(struct sf_buf), M_TEMP,
99 	    M_WAITOK | M_ZERO);
100 	for (i = 0; i < nsfbufs; i++) {
101 		sf_bufs[i].kva = sf_base + i * PAGE_SIZE;
102 		TAILQ_INSERT_TAIL(&sf_buf_freelist, &sf_bufs[i], free_entry);
103 	}
104 	sf_buf_alloc_want = 0;
105 	mtx_init(&sf_buf_lock, "sf_buf", NULL, MTX_DEF);
106 }
107 
108 /*
109  * Get an sf_buf from the freelist.  May block if none are available.
110  */
111 struct sf_buf *
112 sf_buf_alloc(struct vm_page *m, int flags)
113 {
114 	struct sf_head *hash_list;
115 	struct sf_buf *sf;
116 	int error;
117 
118 	if (PMAP_HAS_DMAP)
119 		return ((struct sf_buf *)m);
120 
121 	KASSERT(curthread->td_pinned > 0 || (flags & SFB_CPUPRIVATE) == 0,
122 	    ("sf_buf_alloc(SFB_CPUPRIVATE): curthread not pinned"));
123 	hash_list = &sf_buf_active[SF_BUF_HASH(m)];
124 	mtx_lock(&sf_buf_lock);
125 	LIST_FOREACH(sf, hash_list, list_entry) {
126 		if (sf->m == m) {
127 			sf->ref_count++;
128 			if (sf->ref_count == 1) {
129 				TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
130 				nsfbufsused++;
131 				nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
132 			}
133 #if defined(SMP) && defined(SFBUF_CPUSET)
134 			sf_buf_shootdown(sf, flags);
135 #endif
136 			goto done;
137 		}
138 	}
139 	while ((sf = TAILQ_FIRST(&sf_buf_freelist)) == NULL) {
140 		if (flags & SFB_NOWAIT)
141 			goto done;
142 		sf_buf_alloc_want++;
143 		SFSTAT_INC(sf_allocwait);
144 		error = msleep(&sf_buf_freelist, &sf_buf_lock,
145 		    (flags & SFB_CATCH) ? PCATCH | PVM : PVM, "sfbufa", 0);
146 		sf_buf_alloc_want--;
147 
148 		/*
149 		 * If we got a signal, don't risk going back to sleep.
150 		 */
151 		if (error)
152 			goto done;
153 	}
154 	TAILQ_REMOVE(&sf_buf_freelist, sf, free_entry);
155 	if (sf->m != NULL)
156 		LIST_REMOVE(sf, list_entry);
157 	LIST_INSERT_HEAD(hash_list, sf, list_entry);
158 	sf->ref_count = 1;
159 	sf->m = m;
160 	nsfbufsused++;
161 	nsfbufspeak = imax(nsfbufspeak, nsfbufsused);
162 	sf_buf_map(sf, flags);
163 done:
164 	mtx_unlock(&sf_buf_lock);
165 	return (sf);
166 }
167 
168 /*
169  * Remove a reference from the given sf_buf, adding it to the free
170  * list when its reference count reaches zero.  A freed sf_buf still,
171  * however, retains its virtual-to-physical mapping until it is
172  * recycled or reactivated by sf_buf_alloc(9).
173  */
174 void
175 sf_buf_free(struct sf_buf *sf)
176 {
177 
178 	if (PMAP_HAS_DMAP)
179 		return;
180 
181 	mtx_lock(&sf_buf_lock);
182 	sf->ref_count--;
183 	if (sf->ref_count == 0) {
184 		TAILQ_INSERT_TAIL(&sf_buf_freelist, sf, free_entry);
185 		nsfbufsused--;
186 		if (sf_buf_unmap(sf)) {
187 			sf->m = NULL;
188 			LIST_REMOVE(sf, list_entry);
189 		}
190 		if (sf_buf_alloc_want > 0)
191 			wakeup(&sf_buf_freelist);
192 	}
193 	mtx_unlock(&sf_buf_lock);
194 }
195 
196 void
197 sf_buf_ref(struct sf_buf *sf)
198 {
199 
200 	if (PMAP_HAS_DMAP)
201 		return;
202 
203 	mtx_lock(&sf_buf_lock);
204 	KASSERT(sf->ref_count > 0, ("%s: sf %p not allocated", __func__, sf));
205 	sf->ref_count++;
206 	mtx_unlock(&sf_buf_lock);
207 }
208 
209 #ifdef SFBUF_PROCESS_PAGE
210 /*
211  * Run callback function on sf_buf that holds a certain page.
212  */
213 boolean_t
214 sf_buf_process_page(vm_page_t m, void (*cb)(struct sf_buf *))
215 {
216 	struct sf_head *hash_list;
217 	struct sf_buf *sf;
218 
219 	hash_list = &sf_buf_active[SF_BUF_HASH(m)];
220 	mtx_lock(&sf_buf_lock);
221 	LIST_FOREACH(sf, hash_list, list_entry) {
222 		if (sf->m == m) {
223 			cb(sf);
224 			mtx_unlock(&sf_buf_lock);
225 			return (TRUE);
226 		}
227 	}
228 	mtx_unlock(&sf_buf_lock);
229 	return (FALSE);
230 }
231 #endif	/* SFBUF_PROCESS_PAGE */
232