1/*
2 * Copyright © 2007,2008,2009,2010 Red Hat, Inc.
3 * Copyright © 2012,2018 Google, Inc.
4 *
5 * This is part of HarfBuzz, a text shaping library.
6 *
7 * Permission is hereby granted, without written agreement and without
8 * license or royalty fees, to use, copy, modify, and distribute this
9 * software and its documentation for any purpose, provided that the
10 * above copyright notice and the following two paragraphs appear in
11 * all copies of this software.
12 *
13 * IN NO EVENT SHALL THE COPYRIGHT HOLDER BE LIABLE TO ANY PARTY FOR
14 * DIRECT, INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES
15 * ARISING OUT OF THE USE OF THIS SOFTWARE AND ITS DOCUMENTATION, EVEN
16 * IF THE COPYRIGHT HOLDER HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH
17 * DAMAGE.
18 *
19 * THE COPYRIGHT HOLDER SPECIFICALLY DISCLAIMS ANY WARRANTIES, INCLUDING,
20 * BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
21 * FITNESS FOR A PARTICULAR PURPOSE. THE SOFTWARE PROVIDED HEREUNDER IS
22 * ON AN "AS IS" BASIS, AND THE COPYRIGHT HOLDER HAS NO OBLIGATION TO
23 * PROVIDE MAINTENANCE, SUPPORT, UPDATES, ENHANCEMENTS, OR MODIFICATIONS.
24 *
25 * Red Hat Author(s): Behdad Esfahbod
26 * Google Author(s): Behdad Esfahbod
27 */
28
29#ifndef HB_SANITIZE_HH
30#define HB_SANITIZE_HH
31
32#include "hb.hh"
33#include "hb-blob.hh"
34#include "hb-dispatch.hh"
35
36
37/*
38 * Sanitize
39 *
40 *
41 * === Introduction ===
42 *
43 * The sanitize machinery is at the core of our zero-cost font loading. We
44 * mmap() font file into memory and create a blob out of it. Font subtables
45 * are returned as a readonly sub-blob of the main font blob. These table
46 * blobs are then sanitized before use, to ensure invalid memory access does
47 * not happen. The toplevel sanitize API use is like, eg. to load the 'head'
48 * table:
49 *
50 * hb_blob_t *head_blob = hb_sanitize_context_t ().reference_table<OT::head> (face);
51 *
52 * The blob then can be converted to a head table struct with:
53 *
54 * const head *head_table = head_blob->as<head> ();
55 *
56 * What the reference_table does is, to call hb_face_reference_table() to load
57 * the table blob, sanitize it and return either the sanitized blob, or empty
58 * blob if sanitization failed. The blob->as() function returns the null
59 * object of its template type argument if the blob is empty. Otherwise, it
60 * just casts the blob contents to the desired type.
61 *
62 * Sanitizing a blob of data with a type T works as follows (with minor
63 * simplification):
64 *
65 * - Cast blob content to T*, call sanitize() method of it,
66 * - If sanitize succeeded, return blob.
67 * - Otherwise, if blob is not writable, try making it writable,
68 * or copy if cannot be made writable in-place,
69 * - Call sanitize() again. Return blob if sanitize succeeded.
70 * - Return empty blob otherwise.
71 *
72 *
73 * === The sanitize() contract ===
74 *
75 * The sanitize() method of each object type shall return true if it's safe to
76 * call other methods of the object, and false otherwise.
77 *
78 * Note that what sanitize() checks for might align with what the specification
79 * describes as valid table data, but does not have to be. In particular, we
80 * do NOT want to be pedantic and concern ourselves with validity checks that
81 * are irrelevant to our use of the table. On the contrary, we want to be
82 * lenient with error handling and accept invalid data to the extent that it
83 * does not impose extra burden on us.
84 *
85 * Based on the sanitize contract, one can see that what we check for depends
86 * on how we use the data in other table methods. Ie. if other table methods
87 * assume that offsets do NOT point out of the table data block, then that's
88 * something sanitize() must check for (GSUB/GPOS/GDEF/etc work this way). On
89 * the other hand, if other methods do such checks themselves, then sanitize()
90 * does not have to bother with them (glyf/local work this way). The choice
91 * depends on the table structure and sanitize() performance. For example, to
92 * check glyf/loca offsets in sanitize() would cost O(num-glyphs). We try hard
93 * to avoid such costs during font loading. By postponing such checks to the
94 * actual glyph loading, we reduce the sanitize cost to O(1) and total runtime
95 * cost to O(used-glyphs). As such, this is preferred.
96 *
97 * The same argument can be made re GSUB/GPOS/GDEF, but there, the table
98 * structure is so complicated that by checking all offsets at sanitize() time,
99 * we make the code much simpler in other methods, as offsets and referenced
100 * objects do not need to be validated at each use site.
101 */
102
103/* This limits sanitizing time on really broken fonts. */
104#ifndef HB_SANITIZE_MAX_EDITS
105#define HB_SANITIZE_MAX_EDITS 32
106#endif
107#ifndef HB_SANITIZE_MAX_OPS_FACTOR
108#define HB_SANITIZE_MAX_OPS_FACTOR 8
109#endif
110#ifndef HB_SANITIZE_MAX_OPS_MIN
111#define HB_SANITIZE_MAX_OPS_MIN 16384
112#endif
113#ifndef HB_SANITIZE_MAX_OPS_MAX
114#define HB_SANITIZE_MAX_OPS_MAX 0x3FFFFFFF
115#endif
116#ifndef HB_SANITIZE_MAX_SUTABLES
117#define HB_SANITIZE_MAX_SUTABLES 0x4000
118#endif
119
120struct hb_sanitize_context_t :
121 hb_dispatch_context_t<hb_sanitize_context_t, bool, HB_DEBUG_SANITIZE>
122{
123 hb_sanitize_context_t () :
124 start (nullptr), end (nullptr),
125 max_ops (0), max_subtables (0),
126 writable (false), edit_count (0),
127 blob (nullptr),
128 num_glyphs (65536),
129 num_glyphs_set (false) {}
130
131 const char *get_name () { return "SANITIZE"; }
132 template <typename T, typename F>
133 bool may_dispatch (const T *obj HB_UNUSED, const F *format)
134 { return format->sanitize (this); }
135 static return_t default_return_value () { return true; }
136 static return_t no_dispatch_return_value () { return false; }
137 bool stop_sublookup_iteration (const return_t r) const { return !r; }
138
139 bool visit_subtables (unsigned count)
140 {
141 max_subtables += count;
142 return max_subtables < HB_SANITIZE_MAX_SUTABLES;
143 }
144
145 private:
146 template <typename T, typename ...Ts> auto
147 _dispatch (const T &obj, hb_priority<1>, Ts&&... ds) HB_AUTO_RETURN
148 ( obj.sanitize (this, hb_forward<Ts> (ds)...) )
149 template <typename T, typename ...Ts> auto
150 _dispatch (const T &obj, hb_priority<0>, Ts&&... ds) HB_AUTO_RETURN
151 ( obj.dispatch (this, hb_forward<Ts> (ds)...) )
152 public:
153 template <typename T, typename ...Ts> auto
154 dispatch (const T &obj, Ts&&... ds) HB_AUTO_RETURN
155 ( _dispatch (obj, hb_prioritize, hb_forward<Ts> (ds)...) )
156
157
158 void init (hb_blob_t *b)
159 {
160 this->blob = hb_blob_reference (b);
161 this->writable = false;
162 }
163
164 void set_num_glyphs (unsigned int num_glyphs_)
165 {
166 num_glyphs = num_glyphs_;
167 num_glyphs_set = true;
168 }
169 unsigned int get_num_glyphs () { return num_glyphs; }
170
171 void set_max_ops (int max_ops_) { max_ops = max_ops_; }
172
173 template <typename T>
174 void set_object (const T *obj)
175 {
176 reset_object ();
177
178 if (!obj) return;
179
180 const char *obj_start = (const char *) obj;
181 if (unlikely (obj_start < this->start || this->end <= obj_start))
182 this->start = this->end = nullptr;
183 else
184 {
185 this->start = obj_start;
186 this->end = obj_start + hb_min (size_t (this->end - obj_start), obj->get_size ());
187 }
188 }
189
190 void reset_object ()
191 {
192 this->start = this->blob->data;
193 this->end = this->start + this->blob->length;
194 assert (this->start <= this->end); /* Must not overflow. */
195 }
196
197 void start_processing ()
198 {
199 reset_object ();
200 if (unlikely (hb_unsigned_mul_overflows (this->end - this->start, HB_SANITIZE_MAX_OPS_FACTOR)))
201 this->max_ops = HB_SANITIZE_MAX_OPS_MAX;
202 else
203 this->max_ops = hb_clamp ((unsigned) (this->end - this->start) * HB_SANITIZE_MAX_OPS_FACTOR,
204 (unsigned) HB_SANITIZE_MAX_OPS_MIN,
205 (unsigned) HB_SANITIZE_MAX_OPS_MAX);
206 this->edit_count = 0;
207 this->debug_depth = 0;
208
209 DEBUG_MSG_LEVEL (SANITIZE, start, 0, +1,
210 "start [%p..%p] (%lu bytes)",
211 this->start, this->end,
212 (unsigned long) (this->end - this->start));
213 }
214
215 void end_processing ()
216 {
217 DEBUG_MSG_LEVEL (SANITIZE, this->start, 0, -1,
218 "end [%p..%p] %u edit requests",
219 this->start, this->end, this->edit_count);
220
221 hb_blob_destroy (this->blob);
222 this->blob = nullptr;
223 this->start = this->end = nullptr;
224 }
225
226 unsigned get_edit_count () { return edit_count; }
227
228 bool check_range (const void *base,
229 unsigned int len) const
230 {
231 const char *p = (const char *) base;
232 bool ok = !len ||
233 (this->start <= p &&
234 p <= this->end &&
235 (unsigned int) (this->end - p) >= len &&
236 this->max_ops-- > 0);
237
238 DEBUG_MSG_LEVEL (SANITIZE, p, this->debug_depth+1, 0,
239 "check_range [%p..%p]"
240 " (%d bytes) in [%p..%p] -> %s",
241 p, p + len, len,
242 this->start, this->end,
243 ok ? "OK" : "OUT-OF-RANGE");
244
245 return likely (ok);
246 }
247
248 template <typename T>
249 bool check_range (const T *base,
250 unsigned int a,
251 unsigned int b) const
252 {
253 return !hb_unsigned_mul_overflows (a, b) &&
254 this->check_range (base, a * b);
255 }
256
257 template <typename T>
258 bool check_range (const T *base,
259 unsigned int a,
260 unsigned int b,
261 unsigned int c) const
262 {
263 return !hb_unsigned_mul_overflows (a, b) &&
264 this->check_range (base, a * b, c);
265 }
266
267 template <typename T>
268 bool check_array (const T *base, unsigned int len) const
269 {
270 return this->check_range (base, len, hb_static_size (T));
271 }
272
273 template <typename T>
274 bool check_array (const T *base,
275 unsigned int a,
276 unsigned int b) const
277 {
278 return this->check_range (base, a, b, hb_static_size (T));
279 }
280
281 template <typename Type>
282 bool check_struct (const Type *obj) const
283 { return likely (this->check_range (obj, obj->min_size)); }
284
285 bool may_edit (const void *base, unsigned int len)
286 {
287 if (this->edit_count >= HB_SANITIZE_MAX_EDITS)
288 return false;
289
290 const char *p = (const char *) base;
291 this->edit_count++;
292
293 DEBUG_MSG_LEVEL (SANITIZE, p, this->debug_depth+1, 0,
294 "may_edit(%u) [%p..%p] (%d bytes) in [%p..%p] -> %s",
295 this->edit_count,
296 p, p + len, len,
297 this->start, this->end,
298 this->writable ? "GRANTED" : "DENIED");
299
300 return this->writable;
301 }
302
303 template <typename Type, typename ValueType>
304 bool try_set (const Type *obj, const ValueType &v)
305 {
306 if (this->may_edit (obj, hb_static_size (Type)))
307 {
308 * const_cast<Type *> (obj) = v;
309 return true;
310 }
311 return false;
312 }
313
314 template <typename Type>
315 hb_blob_t *sanitize_blob (hb_blob_t *blob)
316 {
317 bool sane;
318
319 init (blob);
320
321 retry:
322 DEBUG_MSG_FUNC (SANITIZE, start, "start");
323
324 start_processing ();
325
326 if (unlikely (!start))
327 {
328 end_processing ();
329 return blob;
330 }
331
332 Type *t = reinterpret_cast<Type *> (const_cast<char *> (start));
333
334 sane = t->sanitize (this);
335 if (sane)
336 {
337 if (edit_count)
338 {
339 DEBUG_MSG_FUNC (SANITIZE, start, "passed first round with %d edits; going for second round", edit_count);
340
341 /* sanitize again to ensure no toe-stepping */
342 edit_count = 0;
343 sane = t->sanitize (this);
344 if (edit_count) {
345 DEBUG_MSG_FUNC (SANITIZE, start, "requested %d edits in second round; FAILLING", edit_count);
346 sane = false;
347 }
348 }
349 }
350 else
351 {
352 if (edit_count && !writable) {
353 start = hb_blob_get_data_writable (blob, nullptr);
354 end = start + blob->length;
355
356 if (start)
357 {
358 writable = true;
359 /* ok, we made it writable by relocating. try again */
360 DEBUG_MSG_FUNC (SANITIZE, start, "retry");
361 goto retry;
362 }
363 }
364 }
365
366 end_processing ();
367
368 DEBUG_MSG_FUNC (SANITIZE, start, sane ? "PASSED" : "FAILED");
369 if (sane)
370 {
371 hb_blob_make_immutable (blob);
372 return blob;
373 }
374 else
375 {
376 hb_blob_destroy (blob);
377 return hb_blob_get_empty ();
378 }
379 }
380
381 template <typename Type>
382 hb_blob_t *reference_table (const hb_face_t *face, hb_tag_t tableTag = Type::tableTag)
383 {
384 if (!num_glyphs_set)
385 set_num_glyphs (hb_face_get_glyph_count (face));
386 return sanitize_blob<Type> (hb_face_reference_table (face, tableTag));
387 }
388
389 const char *start, *end;
390 mutable int max_ops, max_subtables;
391 private:
392 bool writable;
393 unsigned int edit_count;
394 hb_blob_t *blob;
395 unsigned int num_glyphs;
396 bool num_glyphs_set;
397};
398
399struct hb_sanitize_with_object_t
400{
401 template <typename T>
402 hb_sanitize_with_object_t (hb_sanitize_context_t *c, const T& obj) : c (c)
403 { c->set_object (obj); }
404 ~hb_sanitize_with_object_t ()
405 { c->reset_object (); }
406
407 private:
408 hb_sanitize_context_t *c;
409};
410
411
412#endif /* HB_SANITIZE_HH */
413