mwrap 1.0.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,16 @@
1
+ # Copyright (C) 2018 mwrap hackers <mwrap-public@80x24.org>
2
+ # License: GPL-2.0+ <https://www.gnu.org/licenses/gpl-2.0.txt>
3
+ require 'rake/testtask'
4
+ begin
5
+ require 'rake/extensiontask'
6
+ Rake::ExtensionTask.new('mwrap')
7
+ rescue LoadError
8
+ warn 'rake-compiler not available, cross compiling disabled'
9
+ end
10
+
11
+ Rake::TestTask.new(:test)
12
+ task :test => :compile
13
+ task :default => :compile
14
+
15
+ c_files = File.readlines('MANIFEST').grep(%r{ext/.*\.[ch]$}).map!(&:chomp!)
16
+ task 'compile:mwrap' => c_files
@@ -0,0 +1,29 @@
1
+ #!/usr/bin/ruby
2
+ # frozen_string_literal: true
3
+ # Copyright (C) 2018 mwrap hackers <mwrap-public@80x24.org>
4
+ # License: GPL-2.0+ <https://www.gnu.org/licenses/gpl-2.0.txt>
5
+ require 'mwrap'
6
+ mwrap_so = $".grep(%r{/mwrap\.so\z})[0] or abort "mwrap.so not loaded"
7
+ cur = ENV['LD_PRELOAD']
8
+ if cur
9
+ cur = cur.split(/[:\s]+/)
10
+ if !cur.include?(mwrap_so)
11
+ # drop old versions
12
+ cur.delete_if { |path| path.end_with?('/mwrap.so') }
13
+ cur.unshift(mwrap_so)
14
+ ENV['LD_PRELOAD'] = cur.join(':')
15
+ end
16
+ else
17
+ ENV['LD_PRELOAD'] = mwrap_so
18
+ end
19
+
20
+ # work around close-on-exec by default behavior in Ruby:
21
+ opts = {}
22
+ if ENV['MWRAP'] =~ /dump_fd:(\d+)/
23
+ dump_fd = $1.to_i
24
+ if dump_fd > 2
25
+ dump_io = IO.new(dump_fd)
26
+ opts[dump_fd] = dump_io
27
+ end
28
+ end
29
+ exec *ARGV, opts
@@ -0,0 +1,13 @@
1
+ # frozen_string_literal: true
2
+ # Copyright (C) 2018 mwrap hackers <mwrap-public@80x24.org>
3
+ # License: GPL-2.0+ <https://www.gnu.org/licenses/gpl-2.0.txt>
4
+ require 'mkmf'
5
+
6
+ have_func 'mempcpy'
7
+ have_library 'urcu-cds' or abort 'userspace RCU not installed'
8
+ have_header 'urcu/rculfhash.h' or abort 'rculfhash.h not found'
9
+ have_library 'urcu-bp' or abort 'liburcu-bp not found'
10
+ have_library 'dl'
11
+ have_library 'c'
12
+ have_library 'execinfo' # FreeBSD
13
+ create_makefile 'mwrap'
@@ -0,0 +1,256 @@
1
+ #ifndef _JHASH_H
2
+ #define _JHASH_H
3
+
4
+ /*
5
+ * jhash.h
6
+ *
7
+ * Example hash function.
8
+ *
9
+ * Copyright 2009-2012 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
10
+ *
11
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
12
+ * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
13
+ *
14
+ * Permission is hereby granted to use or copy this program for any
15
+ * purpose, provided the above notices are retained on all copies.
16
+ * Permission to modify the code and to distribute modified code is
17
+ * granted, provided the above notices are retained, and a notice that
18
+ * the code was modified is included with the above copyright notice.
19
+ */
20
+
21
+ /*
22
+ * Hash function
23
+ * Source: http://burtleburtle.net/bob/c/lookup3.c
24
+ * Originally Public Domain
25
+ */
26
+
27
+ #define rot(x, k) (((x) << (k)) | ((x) >> (32 - (k))))
28
+
29
+ #define mix(a, b, c) \
30
+ do { \
31
+ a -= c; a ^= rot(c, 4); c += b; \
32
+ b -= a; b ^= rot(a, 6); a += c; \
33
+ c -= b; c ^= rot(b, 8); b += a; \
34
+ a -= c; a ^= rot(c, 16); c += b; \
35
+ b -= a; b ^= rot(a, 19); a += c; \
36
+ c -= b; c ^= rot(b, 4); b += a; \
37
+ } while (0)
38
+
39
+ #define final(a, b, c) \
40
+ { \
41
+ c ^= b; c -= rot(b, 14); \
42
+ a ^= c; a -= rot(c, 11); \
43
+ b ^= a; b -= rot(a, 25); \
44
+ c ^= b; c -= rot(b, 16); \
45
+ a ^= c; a -= rot(c, 4); \
46
+ b ^= a; b -= rot(a, 14); \
47
+ c ^= b; c -= rot(b, 24); \
48
+ }
49
+
50
+ #if (BYTE_ORDER == LITTLE_ENDIAN)
51
+ #define HASH_LITTLE_ENDIAN 1
52
+ #else
53
+ #define HASH_LITTLE_ENDIAN 0
54
+ #endif
55
+
56
+ /*
57
+ *
58
+ * hashlittle() -- hash a variable-length key into a 32-bit value
59
+ * k : the key (the unaligned variable-length array of bytes)
60
+ * length : the length of the key, counting by bytes
61
+ * initval : can be any 4-byte value
62
+ * Returns a 32-bit value. Every bit of the key affects every bit of
63
+ * the return value. Two keys differing by one or two bits will have
64
+ * totally different hash values.
65
+ *
66
+ * The best hash table sizes are powers of 2. There is no need to do
67
+ * mod a prime (mod is sooo slow!). If you need less than 32 bits,
68
+ * use a bitmask. For example, if you need only 10 bits, do
69
+ * h = (h & hashmask(10));
70
+ * In which case, the hash table should have hashsize(10) elements.
71
+ *
72
+ * If you are hashing n strings (uint8_t **)k, do it like this:
73
+ * for (i = 0, h = 0; i < n; ++i) h = hashlittle(k[i], len[i], h);
74
+ *
75
+ * By Bob Jenkins, 2006. bob_jenkins@burtleburtle.net. You may use this
76
+ * code any way you wish, private, educational, or commercial. It's free.
77
+ *
78
+ * Use for hash table lookup, or anything where one collision in 2^^32 is
79
+ * acceptable. Do NOT use for cryptographic purposes.
80
+ */
81
+ static
82
+ uint32_t hashlittle(const void *key, size_t length, uint32_t initval)
83
+ {
84
+ uint32_t a, b, c; /* internal state */
85
+ union {
86
+ const void *ptr;
87
+ size_t i;
88
+ } u;
89
+
90
+ /* Set up the internal state */
91
+ a = b = c = 0xdeadbeef + ((uint32_t)length) + initval;
92
+
93
+ u.ptr = key;
94
+ if (HASH_LITTLE_ENDIAN && ((u.i & 0x3) == 0)) {
95
+ const uint32_t *k = (const uint32_t *) key; /* read 32-bit chunks */
96
+
97
+ /*------ all but last block: aligned reads and affect 32 bits of (a,b,c) */
98
+ while (length > 12) {
99
+ a += k[0];
100
+ b += k[1];
101
+ c += k[2];
102
+ mix(a, b, c);
103
+ length -= 12;
104
+ k += 3;
105
+ }
106
+
107
+ /*----------------------------- handle the last (probably partial) block */
108
+ /*
109
+ * "k[2]&0xffffff" actually reads beyond the end of the string, but
110
+ * then masks off the part it's not allowed to read. Because the
111
+ * string is aligned, the masked-off tail is in the same word as the
112
+ * rest of the string. Every machine with memory protection I've seen
113
+ * does it on word boundaries, so is OK with this. But VALGRIND will
114
+ * still catch it and complain. The masking trick does make the hash
115
+ * noticably faster for short strings (like English words).
116
+ */
117
+ #ifndef VALGRIND
118
+
119
+ switch (length) {
120
+ case 12: c+=k[2]; b+=k[1]; a+=k[0]; break;
121
+ case 11: c+=k[2]&0xffffff; b+=k[1]; a+=k[0]; break;
122
+ case 10: c+=k[2]&0xffff; b+=k[1]; a+=k[0]; break;
123
+ case 9 : c+=k[2]&0xff; b+=k[1]; a+=k[0]; break;
124
+ case 8 : b+=k[1]; a+=k[0]; break;
125
+ case 7 : b+=k[1]&0xffffff; a+=k[0]; break;
126
+ case 6 : b+=k[1]&0xffff; a+=k[0]; break;
127
+ case 5 : b+=k[1]&0xff; a+=k[0]; break;
128
+ case 4 : a+=k[0]; break;
129
+ case 3 : a+=k[0]&0xffffff; break;
130
+ case 2 : a+=k[0]&0xffff; break;
131
+ case 1 : a+=k[0]&0xff; break;
132
+ case 0 : return c; /* zero length strings require no mixing */
133
+ }
134
+
135
+ #else /* make valgrind happy */
136
+ {
137
+ const uint8_t *k8;
138
+
139
+ k8 = (const uint8_t *) k;
140
+ switch (length) {
141
+ case 12: c+=k[2]; b+=k[1]; a+=k[0]; break;
142
+ case 11: c+=((uint32_t) k8[10])<<16; /* fall through */
143
+ case 10: c+=((uint32_t) k8[9])<<8; /* fall through */
144
+ case 9 : c+=k8[8]; /* fall through */
145
+ case 8 : b+=k[1]; a+=k[0]; break;
146
+ case 7 : b+=((uint32_t) k8[6])<<16; /* fall through */
147
+ case 6 : b+=((uint32_t) k8[5])<<8; /* fall through */
148
+ case 5 : b+=k8[4]; /* fall through */
149
+ case 4 : a+=k[0]; break;
150
+ case 3 : a+=((uint32_t) k8[2])<<16; /* fall through */
151
+ case 2 : a+=((uint32_t) k8[1])<<8; /* fall through */
152
+ case 1 : a+=k8[0]; break;
153
+ case 0 : return c;
154
+ }
155
+ }
156
+ #endif /* !valgrind */
157
+
158
+ } else if (HASH_LITTLE_ENDIAN && ((u.i & 0x1) == 0)) {
159
+ const uint16_t *k = (const uint16_t *) key; /* read 16-bit chunks */
160
+ const uint8_t *k8;
161
+
162
+ /*--------------- all but last block: aligned reads and different mixing */
163
+ while (length > 12)
164
+ {
165
+ a += k[0] + (((uint32_t) k[1])<<16);
166
+ b += k[2] + (((uint32_t) k[3])<<16);
167
+ c += k[4] + (((uint32_t) k[5])<<16);
168
+ mix(a, b, c);
169
+ length -= 12;
170
+ k += 6;
171
+ }
172
+
173
+ /*----------------------------- handle the last (probably partial) block */
174
+ k8 = (const uint8_t *) k;
175
+ switch(length)
176
+ {
177
+ case 12: c+=k[4]+(((uint32_t) k[5])<<16);
178
+ b+=k[2]+(((uint32_t) k[3])<<16);
179
+ a+=k[0]+(((uint32_t) k[1])<<16);
180
+ break;
181
+ case 11: c+=((uint32_t) k8[10])<<16; /* fall through */
182
+ case 10: c+=k[4];
183
+ b+=k[2]+(((uint32_t) k[3])<<16);
184
+ a+=k[0]+(((uint32_t) k[1])<<16);
185
+ break;
186
+ case 9 : c+=k8[8]; /* fall through */
187
+ case 8 : b+=k[2]+(((uint32_t) k[3])<<16);
188
+ a+=k[0]+(((uint32_t) k[1])<<16);
189
+ break;
190
+ case 7 : b+=((uint32_t) k8[6])<<16; /* fall through */
191
+ case 6 : b+=k[2];
192
+ a+=k[0]+(((uint32_t) k[1])<<16);
193
+ break;
194
+ case 5 : b+=k8[4]; /* fall through */
195
+ case 4 : a+=k[0]+(((uint32_t) k[1])<<16);
196
+ break;
197
+ case 3 : a+=((uint32_t) k8[2])<<16; /* fall through */
198
+ case 2 : a+=k[0];
199
+ break;
200
+ case 1 : a+=k8[0];
201
+ break;
202
+ case 0 : return c; /* zero length requires no mixing */
203
+ }
204
+
205
+ } else { /* need to read the key one byte at a time */
206
+ const uint8_t *k = (const uint8_t *)key;
207
+
208
+ /*--------------- all but the last block: affect some 32 bits of (a, b, c) */
209
+ while (length > 12) {
210
+ a += k[0];
211
+ a += ((uint32_t) k[1])<<8;
212
+ a += ((uint32_t) k[2])<<16;
213
+ a += ((uint32_t) k[3])<<24;
214
+ b += k[4];
215
+ b += ((uint32_t) k[5])<<8;
216
+ b += ((uint32_t) k[6])<<16;
217
+ b += ((uint32_t) k[7])<<24;
218
+ c += k[8];
219
+ c += ((uint32_t) k[9])<<8;
220
+ c += ((uint32_t) k[10])<<16;
221
+ c += ((uint32_t) k[11])<<24;
222
+ mix(a,b,c);
223
+ length -= 12;
224
+ k += 12;
225
+ }
226
+
227
+ /*-------------------------------- last block: affect all 32 bits of (c) */
228
+ switch (length) { /* all the case statements fall through */
229
+ case 12: c+=((uint32_t) k[11])<<24;
230
+ case 11: c+=((uint32_t) k[10])<<16;
231
+ case 10: c+=((uint32_t) k[9])<<8;
232
+ case 9 : c+=k[8];
233
+ case 8 : b+=((uint32_t) k[7])<<24;
234
+ case 7 : b+=((uint32_t) k[6])<<16;
235
+ case 6 : b+=((uint32_t) k[5])<<8;
236
+ case 5 : b+=k[4];
237
+ case 4 : a+=((uint32_t) k[3])<<24;
238
+ case 3 : a+=((uint32_t) k[2])<<16;
239
+ case 2 : a+=((uint32_t) k[1])<<8;
240
+ case 1 : a+=k[0];
241
+ break;
242
+ case 0 : return c;
243
+ }
244
+ }
245
+
246
+ final(a, b, c);
247
+ return c;
248
+ }
249
+
250
+ static inline
251
+ uint32_t jhash(const void *key, size_t length, uint32_t seed)
252
+ {
253
+ return hashlittle(key, length, seed);
254
+ }
255
+
256
+ #endif /* _JHASH_H */
@@ -0,0 +1,598 @@
1
+ /*
2
+ * Copyright (C) 2018 mwrap hackers <mwrap-public@80x24.org>
3
+ * License: GPL-2.0+ <https://www.gnu.org/licenses/gpl-2.0.txt>
4
+ */
5
+ #define _LGPL_SOURCE /* allows URCU to inline some stuff */
6
+ #include <ruby/ruby.h>
7
+ #include <ruby/thread.h>
8
+ #include <ruby/io.h>
9
+ #include <execinfo.h>
10
+ #include <stdio.h>
11
+ #include <stdlib.h>
12
+ #include <string.h>
13
+ #include <dlfcn.h>
14
+ #include <assert.h>
15
+ #include <errno.h>
16
+ #include <sys/types.h>
17
+ #include <sys/stat.h>
18
+ #include <fcntl.h>
19
+ #include <urcu-bp.h>
20
+ #include <urcu/rculfhash.h>
21
+ #include "jhash.h"
22
+
23
+ static ID id_uminus;
24
+ const char *rb_source_location_cstr(int *line); /* requires 2.6.0dev */
25
+ static int *(*has_gvl_p)(void);
26
+ #ifdef __FreeBSD__
27
+ void *__malloc(size_t);
28
+ void *__calloc(size_t, size_t);
29
+ void *__realloc(void *, size_t);
30
+ static void *(*real_malloc)(size_t) = __malloc;
31
+ static void *(*real_calloc)(size_t, size_t) = __calloc;
32
+ static void *(*real_realloc)(void *, size_t) = __realloc;
33
+ # define RETURN_IF_NOT_READY() do {} while (0) /* nothing */
34
+ #else
35
+ static int ready;
36
+ static void *(*real_malloc)(size_t);
37
+ static void *(*real_calloc)(size_t, size_t);
38
+ static void *(*real_realloc)(void *, size_t);
39
+
40
+ /*
41
+ * we need to fake an OOM condition while dlsym is running,
42
+ * as that calls calloc under glibc, but we don't have the
43
+ * symbol for the jemalloc calloc, yet
44
+ */
45
+ # define RETURN_IF_NOT_READY() do { \
46
+ if (!ready) { \
47
+ errno = ENOMEM; \
48
+ return NULL; \
49
+ } \
50
+ } while (0)
51
+
52
+ #endif /* !FreeBSD */
53
+
54
+ /*
55
+ * rb_source_location_cstr relies on GET_EC(), and it's possible
56
+ * to have a native thread but no EC during the early and late
57
+ * (teardown) phases of the Ruby process
58
+ */
59
+ static void **ec_loc;
60
+
61
+ static struct cds_lfht *totals;
62
+
63
+ static struct cds_lfht *
64
+ lfht_new(void)
65
+ {
66
+ return cds_lfht_new(16384, 1, 0, CDS_LFHT_AUTO_RESIZE, 0);
67
+ }
68
+
69
+ __attribute__((constructor)) static void resolve_malloc(void)
70
+ {
71
+ int err;
72
+
73
+ #ifndef __FreeBSD__
74
+ real_malloc = dlsym(RTLD_NEXT, "malloc");
75
+ real_calloc = dlsym(RTLD_NEXT, "calloc");
76
+ real_realloc = dlsym(RTLD_NEXT, "realloc");
77
+ if (!real_calloc || !real_malloc || !real_realloc) {
78
+ fprintf(stderr, "missing calloc/malloc/realloc %p %p %p\n",
79
+ real_calloc, real_malloc, real_realloc);
80
+ _exit(1);
81
+ }
82
+ ready = 1;
83
+ #endif
84
+
85
+ totals = lfht_new();
86
+ if (!totals)
87
+ fprintf(stderr, "failed to allocate totals table\n");
88
+
89
+ err = pthread_atfork(call_rcu_before_fork,
90
+ call_rcu_after_fork_parent,
91
+ call_rcu_after_fork_child);
92
+ if (err)
93
+ fprintf(stderr, "pthread_atfork failed: %s\n", strerror(err));
94
+
95
+ has_gvl_p = dlsym(RTLD_DEFAULT, "ruby_thread_has_gvl_p");
96
+
97
+ /*
98
+ * resolve dynamically so it doesn't break when LD_PRELOAD-ed
99
+ * into non-Ruby binaries
100
+ */
101
+ ec_loc = dlsym(RTLD_DEFAULT, "ruby_current_execution_context_ptr");
102
+ }
103
+
104
+ #ifndef HAVE_MEMPCPY
105
+ static void *
106
+ my_mempcpy(void *dest, const void *src, size_t n)
107
+ {
108
+ return (char *)memcpy(dest, src, n) + n;
109
+ }
110
+ #define mempcpy(dst,src,n) my_mempcpy(dst,src,n)
111
+ #endif
112
+
113
+ /* stolen from glibc: */
114
+ #define RETURN_ADDRESS(nr) \
115
+ (uintptr_t)(__builtin_extract_return_addr(__builtin_return_address(nr)))
116
+
117
+ static __thread size_t locating;
118
+
119
+ #define INT2STR_MAX (sizeof(int) == 4 ? 10 : 19)
120
+ static char *int2str(int num, char *dst, size_t * size)
121
+ {
122
+ if (num <= 9) {
123
+ *size -= 1;
124
+ *dst++ = (char)(num + '0');
125
+ return dst;
126
+ } else {
127
+ char buf[INT2STR_MAX];
128
+ char *end = buf + sizeof(buf);
129
+ char *p = end;
130
+ size_t adj;
131
+
132
+ do {
133
+ *size -= 1;
134
+ *--p = (char)((num % 10) + '0');
135
+ num /= 10;
136
+ } while (num && *size);
137
+
138
+ if (!num) {
139
+ adj = end - p;
140
+ return mempcpy(dst, p, adj);
141
+ }
142
+ }
143
+ return NULL;
144
+ }
145
+
146
+ static int has_ec_p(void)
147
+ {
148
+ return (ec_loc && *ec_loc);
149
+ }
150
+
151
+ struct src_loc {
152
+ struct rcu_head rcu_head;
153
+ size_t calls;
154
+ size_t total;
155
+ struct cds_lfht_node hnode;
156
+ uint32_t hval;
157
+ uint32_t capa;
158
+ char k[];
159
+ };
160
+
161
+ static int loc_is_addr(const struct src_loc *l)
162
+ {
163
+ return l->capa == 0;
164
+ }
165
+
166
+ static size_t loc_size(const struct src_loc *l)
167
+ {
168
+ return loc_is_addr(l) ? sizeof(uintptr_t) : l->capa;
169
+ }
170
+
171
+ static int loc_eq(struct cds_lfht_node *node, const void *key)
172
+ {
173
+ const struct src_loc *existing;
174
+ const struct src_loc *k = key;
175
+
176
+ existing = caa_container_of(node, struct src_loc, hnode);
177
+
178
+ return (k->hval == existing->hval &&
179
+ k->capa == existing->capa &&
180
+ memcmp(k->k, existing->k, loc_size(k)) == 0);
181
+ }
182
+
183
+ static void totals_add(struct src_loc *k)
184
+ {
185
+ struct cds_lfht_iter iter;
186
+ struct cds_lfht_node *cur;
187
+ struct src_loc *l;
188
+ struct cds_lfht *t;
189
+
190
+
191
+ again:
192
+ rcu_read_lock();
193
+ t = rcu_dereference(totals);
194
+ if (!t) goto out_unlock;
195
+ cds_lfht_lookup(t, k->hval, loc_eq, k, &iter);
196
+ cur = cds_lfht_iter_get_node(&iter);
197
+ if (cur) {
198
+ l = caa_container_of(cur, struct src_loc, hnode);
199
+ uatomic_add(&l->total, k->total);
200
+ uatomic_add(&l->calls, 1);
201
+ } else {
202
+ size_t n = loc_size(k);
203
+ l = malloc(sizeof(*l) + n);
204
+ if (!l) goto out_unlock;
205
+
206
+ memcpy(l, k, sizeof(*l) + n);
207
+ l->calls = 1;
208
+ cur = cds_lfht_add_unique(t, k->hval, loc_eq, l, &l->hnode);
209
+ if (cur != &l->hnode) { /* lost race */
210
+ rcu_read_unlock();
211
+ free(l);
212
+ goto again;
213
+ }
214
+ }
215
+ out_unlock:
216
+ rcu_read_unlock();
217
+ }
218
+
219
+ static void update_stats(size_t size, uintptr_t caller)
220
+ {
221
+ struct src_loc *k;
222
+ static const size_t xlen = sizeof(caller);
223
+ char *dst;
224
+
225
+ if (locating++) goto out; /* do not recurse into another *alloc */
226
+
227
+ if (has_gvl_p && has_gvl_p() && has_ec_p()) {
228
+ int line;
229
+ const char *ptr = rb_source_location_cstr(&line);
230
+ size_t len;
231
+ size_t int_size = INT2STR_MAX;
232
+
233
+ if (!ptr) goto unknown;
234
+
235
+ /* avoid vsnprintf or anything which could call malloc here: */
236
+ len = strlen(ptr);
237
+ k = alloca(sizeof(*k) + len + 1 + int_size + 1);
238
+ k->total = size;
239
+ dst = mempcpy(k->k, ptr, len);
240
+ *dst++ = ':';
241
+ dst = int2str(line, dst, &int_size);
242
+ if (dst) {
243
+ *dst = 0; /* terminate string */
244
+ k->capa = (uint32_t)(dst - k->k + 1);
245
+ k->hval = jhash(k->k, k->capa, 0xdeadbeef);
246
+ totals_add(k);
247
+ } else {
248
+ rb_bug("bad math making key from location %s:%d\n",
249
+ ptr, line);
250
+ }
251
+ } else {
252
+ unknown:
253
+ k = alloca(sizeof(*k) + xlen);
254
+ k->total = size;
255
+ memcpy(k->k, &caller, xlen);
256
+ k->capa = 0;
257
+ k->hval = jhash(k->k, xlen, 0xdeadbeef);
258
+ totals_add(k);
259
+ }
260
+ out:
261
+ --locating;
262
+ }
263
+
264
+ /*
265
+ * Do we care for *memalign? ruby/gc.c uses it in ways this lib
266
+ * doesn't care about, but maybe some gems use it, too.
267
+ */
268
+ void *malloc(size_t size)
269
+ {
270
+ RETURN_IF_NOT_READY();
271
+ update_stats(size, RETURN_ADDRESS(0));
272
+ return real_malloc(size);
273
+ }
274
+
275
+ void *calloc(size_t nmemb, size_t size)
276
+ {
277
+ RETURN_IF_NOT_READY();
278
+ /* ruby_xcalloc already does overflow checking */
279
+ update_stats(nmemb * size, RETURN_ADDRESS(0));
280
+ return real_calloc(nmemb, size);
281
+ }
282
+
283
+ void *realloc(void *ptr, size_t size)
284
+ {
285
+ RETURN_IF_NOT_READY();
286
+ update_stats(size, RETURN_ADDRESS(0));
287
+ return real_realloc(ptr, size);
288
+ }
289
+
290
+ struct dump_arg {
291
+ FILE *fp;
292
+ size_t min;
293
+ };
294
+
295
+ static void *dump_to_file(void *x)
296
+ {
297
+ struct dump_arg *a = x;
298
+ struct cds_lfht_iter iter;
299
+ struct src_loc *l;
300
+ struct cds_lfht *t;
301
+
302
+ ++locating;
303
+ rcu_read_lock();
304
+ t = rcu_dereference(totals);
305
+ if (!t)
306
+ goto out_unlock;
307
+ cds_lfht_for_each_entry(t, &iter, l, hnode) {
308
+ const void *p = l->k;
309
+ char **s = 0;
310
+ if (l->total <= a->min) continue;
311
+
312
+ if (loc_is_addr(l)) {
313
+ s = backtrace_symbols(p, 1);
314
+ p = s[0];
315
+ }
316
+ fprintf(a->fp, "%16zu %12zu %s\n",
317
+ l->total, l->calls, (const char *)p);
318
+ if (s) free(s);
319
+ }
320
+ out_unlock:
321
+ rcu_read_unlock();
322
+ --locating;
323
+ return 0;
324
+ }
325
+
326
+ /*
327
+ * call-seq:
328
+ *
329
+ * Mwrap.dump([[io] [, min]] -> nil
330
+ *
331
+ * Dumps the current totals to +io+ which must be an IO object
332
+ * (StringIO and similar are not supported). Total sizes smaller
333
+ * than or equal to +min+ are skipped.
334
+ *
335
+ * The output is space-delimited by 3 columns:
336
+ *
337
+ * total_size call_count location
338
+ */
339
+ static VALUE mwrap_dump(int argc, VALUE * argv, VALUE mod)
340
+ {
341
+ VALUE io, min;
342
+ struct dump_arg a;
343
+ rb_io_t *fptr;
344
+
345
+ rb_scan_args(argc, argv, "02", &io, &min);
346
+
347
+ if (NIL_P(io))
348
+ /* library may be linked w/o Ruby */
349
+ io = *((VALUE *)dlsym(RTLD_DEFAULT, "rb_stderr"));
350
+
351
+ a.min = NIL_P(min) ? 0 : NUM2SIZET(min);
352
+ io = rb_io_get_write_io(io);
353
+ GetOpenFile(io, fptr);
354
+ a.fp = rb_io_stdio_file(fptr);
355
+
356
+ rb_thread_call_without_gvl(dump_to_file, &a, 0, 0);
357
+ RB_GC_GUARD(io);
358
+ return Qnil;
359
+ }
360
+
361
+ static void
362
+ free_src_loc(struct rcu_head *head)
363
+ {
364
+ struct src_loc *l = caa_container_of(head, struct src_loc, rcu_head);
365
+ free(l);
366
+ }
367
+
368
+ static void *totals_clear(void *ign)
369
+ {
370
+ struct cds_lfht *new, *old;
371
+ struct cds_lfht_iter iter;
372
+ struct src_loc *l;
373
+
374
+ new = lfht_new();
375
+ rcu_read_lock();
376
+ old = rcu_dereference(totals);
377
+ rcu_assign_pointer(totals, new);
378
+ cds_lfht_for_each_entry(old, &iter, l, hnode) {
379
+ cds_lfht_del(old, &l->hnode);
380
+ call_rcu(&l->rcu_head, free_src_loc);
381
+ }
382
+ rcu_read_unlock();
383
+
384
+ synchronize_rcu(); /* ensure totals points to new */
385
+ cds_lfht_destroy(old, NULL);
386
+ return 0;
387
+ }
388
+
389
+ /*
390
+ * call-seq:
391
+ *
392
+ * Mwrap.clear -> nil
393
+ *
394
+ * Atomically replaces the totals table and destroys the old one.
395
+ * This resets all statistics. It is more expensive than `Mwrap.reset'
396
+ * as new allocations will need to be made to repopulate the new table.
397
+ */
398
+ static VALUE mwrap_clear(VALUE mod)
399
+ {
400
+ rb_thread_call_without_gvl(totals_clear, 0, 0, 0);
401
+ return Qnil;
402
+ }
403
+
404
+ static void *totals_reset(void *ign)
405
+ {
406
+ struct cds_lfht *t;
407
+ struct cds_lfht_iter iter;
408
+ struct src_loc *l;
409
+
410
+ rcu_read_lock();
411
+ t = rcu_dereference(totals);
412
+ cds_lfht_for_each_entry(t, &iter, l, hnode) {
413
+ uatomic_set(&l->total, 0);
414
+ uatomic_set(&l->calls, 0);
415
+ }
416
+ rcu_read_unlock();
417
+ return 0;
418
+ }
419
+
420
+ /*
421
+ * call-seq:
422
+ *
423
+ * Mwrap.reset -> nil
424
+ *
425
+ * Resets the the total tables by zero-ing all counters.
426
+ * This resets all statistics and is less costly than `Mwrap.clear'
427
+ * but is not an atomic operation.
428
+ */
429
+ static VALUE mwrap_reset(VALUE mod)
430
+ {
431
+ rb_thread_call_without_gvl(totals_reset, 0, 0, 0);
432
+ return Qnil;
433
+ }
434
+
435
+ static VALUE dump_ensure(VALUE ignored)
436
+ {
437
+ rcu_read_unlock();
438
+ --locating;
439
+ return Qfalse;
440
+ }
441
+
442
+ static VALUE dump_each_rcu(VALUE x)
443
+ {
444
+ struct dump_arg *a = (struct dump_arg *)x;
445
+ struct cds_lfht *t;
446
+ struct cds_lfht_iter iter;
447
+ struct src_loc *l;
448
+
449
+ t = rcu_dereference(totals);
450
+ cds_lfht_for_each_entry(t, &iter, l, hnode) {
451
+ VALUE v[3];
452
+ if (l->total <= a->min) continue;
453
+
454
+ if (loc_is_addr(l)) {
455
+ char **s = backtrace_symbols((void *)l->k, 1);
456
+ v[1] = rb_str_new_cstr(s[0]);
457
+ free(s);
458
+ }
459
+ else {
460
+ v[1] = rb_str_new(l->k, l->capa - 1);
461
+ }
462
+
463
+ /* deduplicate and try to free up some memory */
464
+ v[0] = rb_funcall(v[1], id_uminus, 0);
465
+ if (!OBJ_FROZEN_RAW(v[1]))
466
+ rb_str_resize(v[1], 0);
467
+
468
+ v[1] = SIZET2NUM(l->total);
469
+ v[2] = SIZET2NUM(l->calls);
470
+
471
+ rb_yield_values2(3, v);
472
+ assert(rcu_read_ongoing());
473
+ }
474
+ return Qnil;
475
+ }
476
+
477
+ /*
478
+ * call-seq:
479
+ *
480
+ * Mwrap.each([min]) { |location,total_bytes,call_count| ... }
481
+ *
482
+ * Yields each entry of the of the table to a caller-supplied block.
483
+ * +min+ may be specified to filter out lines with +total_bytes+
484
+ * equal-to-or-smaller-than the supplied minimum.
485
+ */
486
+ static VALUE mwrap_each(int argc, VALUE * argv, VALUE mod)
487
+ {
488
+ VALUE min;
489
+ struct dump_arg a;
490
+
491
+ rb_scan_args(argc, argv, "01", &min);
492
+ a.min = NIL_P(min) ? 0 : NUM2SIZET(min);
493
+
494
+ ++locating;
495
+ rcu_read_lock();
496
+
497
+ return rb_ensure(dump_each_rcu, (VALUE)&a, dump_ensure, 0);
498
+ }
499
+
500
+ /*
501
+ * Document-module: Mwrap
502
+ *
503
+ * require 'mwrap'
504
+ *
505
+ * Mwrap has a dual function as both a Ruby C extension and LD_PRELOAD
506
+ * wrapper. As a Ruby C extension, it exposes a limited Ruby API.
507
+ * To be effective at gathering status, mwrap must be loaded as a
508
+ * LD_PRELOAD (using the mwrap(1) executable makes it easy)
509
+ *
510
+ * ENVIRONMENT
511
+ *
512
+ * The "MWRAP" environment variable contains a comma-delimited list
513
+ * of key:value options for automatically dumping at program exit.
514
+ *
515
+ * * dump_fd: a writable FD to dump to
516
+ * * dump_path: a path to dump to, the file is opened in O_APPEND mode
517
+ * * dump_min: the minimum allocation size (total) to dump
518
+ *
519
+ * If both `dump_fd' and `dump_path' are specified, dump_path takes
520
+ * precedence.
521
+ */
522
+ void Init_mwrap(void)
523
+ {
524
+ VALUE mod = rb_define_module("Mwrap");
525
+ id_uminus = rb_intern("-@");
526
+
527
+ rb_define_singleton_method(mod, "dump", mwrap_dump, -1);
528
+ rb_define_singleton_method(mod, "clear", mwrap_clear, 0);
529
+ rb_define_singleton_method(mod, "reset", mwrap_reset, 0);
530
+ rb_define_singleton_method(mod, "each", mwrap_each, -1);
531
+ }
532
+
533
+ /* rb_cloexec_open isn't usable by non-Ruby processes */
534
+ #ifndef O_CLOEXEC
535
+ # define O_CLOEXEC 0
536
+ #endif
537
+
538
+ __attribute__ ((destructor))
539
+ static void mwrap_dump_destructor(void)
540
+ {
541
+ const char *opt = getenv("MWRAP");
542
+ const char *modes[] = { "a", "a+", "w", "w+", "r+" };
543
+ struct dump_arg a;
544
+ size_t i;
545
+ int dump_fd;
546
+ char *dump_path;
547
+
548
+ if (!opt)
549
+ return;
550
+
551
+ ++locating;
552
+ if ((dump_path = strstr(opt, "dump_path:")) &&
553
+ (dump_path += sizeof("dump_path")) &&
554
+ *dump_path) {
555
+ char *end = strchr(dump_path, ',');
556
+ if (end) {
557
+ char *tmp = alloca(end - dump_path + 1);
558
+ end = mempcpy(tmp, dump_path, end - dump_path);
559
+ *end = 0;
560
+ dump_path = tmp;
561
+ }
562
+ dump_fd = open(dump_path, O_CLOEXEC|O_WRONLY|O_APPEND|O_CREAT,
563
+ 0666);
564
+ if (dump_fd < 0) {
565
+ fprintf(stderr, "open %s failed: %s\n", dump_path,
566
+ strerror(errno));
567
+ goto out;
568
+ }
569
+ }
570
+ else if (!sscanf(opt, "dump_fd:%d", &dump_fd))
571
+ goto out;
572
+
573
+ if (!sscanf(opt, "dump_min:%zu", &a.min))
574
+ a.min = 0;
575
+
576
+ switch (dump_fd) {
577
+ case 0: goto out;
578
+ case 1: a.fp = stdout; break;
579
+ case 2: a.fp = stderr; break;
580
+ default:
581
+ if (dump_fd < 0)
582
+ goto out;
583
+ a.fp = 0;
584
+
585
+ for (i = 0; !a.fp && i < 5; i++)
586
+ a.fp = fdopen(dump_fd, modes[i]);
587
+
588
+ if (!a.fp) {
589
+ fprintf(stderr, "failed to open fd=%d: %s\n",
590
+ dump_fd, strerror(errno));
591
+ goto out;
592
+ }
593
+ /* we'll leak some memory here, but this is a destructor */
594
+ }
595
+ dump_to_file(&a);
596
+ out:
597
+ --locating;
598
+ }