fix: 畸形 DHT 拒绝 + 渐进式 restart 重同步(M23 r1 P1-1/P1-2)

P1-1(内存安全):mj_huff_build 构造 9 位快表前校验规范码空间
(code + bits[l] > (1<<l) → 表非法),填表时再夹一道 <512 边界;mj_dht 失败
即返回 0 → 后端拒绝 → 回退 stb。此前过订阅 DHT(如「长度 2 的码 100 个」)会把
fast[512] 下标推到 12799,28~220 字节的畸形 JPEG 就栈溢出(ASan 实测
stack-buffer-overflow,非 ASan 直接 SIGSEGV/SIGBUS)。
回归:码长 1..4 × 符号数 3..255 共 1012 组畸形 DHT 全部后端拒绝、无崩溃;
「有效 JPEG + 注入坏 DHT」走拒绝 + 回退(不花屏)。

P1-2(能力):三处扫描循环的区间重同步统一到 mj_boundary,**只在 total == rst
(编码器区间长度)时重同步**。原实现一见「over && marker∈D0..D7」就重同步,但熵码
位缓冲有最多 9 位前瞻,over 会在区间真正结束前 1~2 个 MCU 置位——提前重同步会丢掉
bitbuf 里本区间剩余块的有效比特,随即失步(渐进式 DC 扫描每块仅 ~1 bit,前瞻能跨过
整个 MCU)。这正是 Pillow restart_marker_blocks=1/2/3/4/5/8/16/64/200 的渐进式
100% 回退 stb(25.8MiB)的根因。另修 mj_restart 裸扫把 FF 00(数据字节 0xFF 的填充)
里的 00 误当 marker。
回归:9 个渐进式 + restart 夹具全部 BACKEND=1、峰值 1.86MiB(与无 restart 相同)、
与同编码无 restart 文件逐字节一致(mad=0.000);baseline + restart 仍逐字节一致。

附带(评审 P2 两项,改动极小且已验证):
- PNG 8bit gray/RGB 的 tRNS 透明接通(8bit 采样值在 tRNS 低字节),与 stb 逐点 alpha 一致;
- 拒绝 16bit 量化表(只属 12bit 精度;收下会让 F*qt 在 mj_accim 里 int32 溢出,UB)。
- meow_img_load_file_step 的 JPEG 走真缩解后端(PNG/BMP 暂留 stb:阅读器预算模型仍是
  M22「整幅物化」口径,待 M27 同步后放开)。

验证:ASan 构建全 PASS;test_meow_img_stream 111 项全 PASS;test_meow_img_budget /
test_meow_img_color 全 PASS;tuku run_tests.sh rc=0(zip 106 / cbz 296,含 membudget)、
run_webtoon_tests.sh rc=0;build_sdk.sh 零警告,MIPS32 rel2 目标码。
峰值不变(修复不增加分配):2000x3000 baseline/progressive 1.86MiB、PNG 1.33MiB。
体积:meow_jpeg.o 17,284B + meow_png_stream.o 8,456B = 25.1KB。
This commit is contained in:
AskaEth committed 2026-11-05 22:39:09 +08:00
1 parent e9c0749609
commit 11a4fcd730
6 files changed
+377 -54

No files matched your search

+254 -4
View File
@@ -180,6 +180,17 @@ static const char *g_gen_py =
" chunk(b'IDAT', zlib.compress(bytes(raw))) + chunk(b'IEND', b''))\n"
" open(path, 'wb').write(data)\n"
"adam7(os.path.join(out, 'adam7.png'), 1600, 2400)\n"
"for rb in (1, 2, 3, 4, 5, 8, 16, 64, 200):\n"
" big.save(os.path.join(out, 'prog_rst%d.jpg' % rb), 'JPEG', quality=60,\n"
" subsampling=2, progressive=True, restart_marker_blocks=rb)\n"
"g8 = Image.new('L', (640, 480), 0)\n"
"dg = ImageDraw.Draw(g8)\n"
"dg.rectangle([100, 100, 400, 300], fill=255)\n"
"g8.save(os.path.join(out, 'gray8_trns.png'), 'PNG', transparency=0)\n"
"r8 = Image.new('RGB', (640, 480), (255, 0, 255))\n"
"dr = ImageDraw.Draw(r8)\n"
"dr.rectangle([100, 100, 400, 300], fill=(10, 200, 30))\n"
"r8.save(os.path.join(out, 'rgb8_trns.png'), 'PNG', transparency=(255, 0, 255))\n"
"open(os.path.join(out, 'broken.jpg'), 'wb').write(b'\\xff\\xd8\\xff\\xe0' +\n"
" bytes((i * 37 + 11) & 0xFF for i in range(256)))\n"
"for n in sorted(os.listdir(out)):\n"
@@ -390,6 +401,75 @@ static size_t old_dim_peak(const char *path, int maxdim)
return peak;
}
static long alpha_zero_count(const meow_img *im)
{
int w = meow_img_w(im), h = meow_img_h(im);
const unsigned char *p = meow_img_pixels(im);
long n = 0;
if (!p) return 0;
for (long i = 0; i < (long)w * h; i++) if (p[i * 4 + 3] == 0) n++;
return n;
}
/* 目标图 vs 整幅参考图的 alpha 一致性:按 nearest 反向映射采样源像素,返回不一致像素数
* (-1 = 参数错误)。PNG 的采样公式与后端一致(tx*W/tw、ty*H/th),应逐点相等。 */
static long alpha_mismatch(const meow_img *got, const meow_img *ref)
{
int gw = meow_img_w(got), gh = meow_img_h(got);
int rw = meow_img_w(ref), rh = meow_img_h(ref);
const unsigned char *pg = meow_img_pixels(got), *pr = meow_img_pixels(ref);
long bad = 0;
if (!pg || !pr || gw <= 0 || gh <= 0) return -1;
for (int y = 0; y < gh; y++) {
int sy = (int)((long)y * rh / gh);
for (int x = 0; x < gw; x++) {
int sx = (int)((long)x * rw / gw);
if (pg[((size_t)y * gw + x) * 4 + 3] != pr[((size_t)sy * rw + sx) * 4 + 3])
bad++;
}
}
return bad;
}
/* ------------------------------------------------------------------ */
/* P1-1 回归:过订阅 Huffman 表(畸形 DHT) */
/* ------------------------------------------------------------------ */
/* 构造「SOI + DHT(码长 lc 有 nsym 个符号) + EOI」的最小 JPEG(≤ 280B)。 */
static size_t build_badhuff(unsigned char *buf, int lc, int nsym)
{
size_t p = 0;
int i;
int seglen = 2 + 1 + 16 + nsym; /* DHT 长度字段计入自身 2 字节 */
buf[p++] = 0xFF; buf[p++] = 0xD8; /* SOI */
buf[p++] = 0xFF; buf[p++] = 0xC4; /* DHT */
buf[p++] = (unsigned char)(seglen >> 8);
buf[p++] = (unsigned char)(seglen & 0xFF);
buf[p++] = 0x00; /* Tc=0, Th=0 */
for (i = 1; i <= 16; i++)
buf[p++] = (unsigned char)(i == lc ? nsym : 0);
for (i = 0; i < nsym; i++) buf[p++] = (unsigned char)i;
buf[p++] = 0xFF; buf[p++] = 0xD9; /* EOI */
return p;
}
/* 把同一段坏 DHT 插到真实 JPEG 的 SOI 之后(其余字节原样):验证「有效文件 + 畸形
* 码表」走的是「拒绝 + 回退」,而不是花屏/崩溃。 */
static unsigned char *inject_badhuff(const unsigned char *orig, size_t n,
int lc, int nsym, size_t *outn)
{
unsigned char seg[320];
size_t all = build_badhuff(seg, lc, nsym);
size_t seglen = all - 4; /* 去掉首 FFD8 与尾 FFD9 */
size_t total = n + seglen;
unsigned char *out = (unsigned char *)malloc(total);
if (!out) return NULL;
memcpy(out, orig, 2); /* SOI */
memcpy(out + 2, seg + 2, seglen); /* 坏 DHT 段 */
memcpy(out + 2 + seglen, orig + 2, n - 2); /* 原文件其余部分 */
*outn = total;
return out;
}
/* ================================================================== */
int main(int argc, char **argv)
{
@@ -681,22 +761,192 @@ int main(int argc, char **argv)
}
{
size_t base = peak_reset();
meow_img *got = meow_img_load_file_dim(prs, 720);
size_t pk = peak_since(base);
meow_img *ref = ref_thumb(prs, 720);
int tw = 0, th = 0;
unsigned char *px = NULL;
int handled = meow_img_backend_file(prs, 720, &tw, &th, &px);
meow_img *got;
size_t pk;
meow_img *ref;
double corr = 0.0;
double d = (got && ref) ? mean_abs_diff(got, ref, 64, &corr) : 1e9;
double d;
if (px) free(px);
base = peak_reset();
got = meow_img_load_file_dim(prs, 720);
pk = peak_since(base);
ref = ref_thumb(prs, 720);
d = (got && ref) ? mean_abs_diff(got, ref, 64, &corr) : 1e9;
peak_report("restart marker 2000x3000", pk);
CHECK(handled == 1, "baseline + restart 走后端(BACKEND=1,不回退 stb)");
CHECK(got && nonblank(got), "restart marker JPEG 解码成功");
CHECK(d < 24.0 && corr > 0.85, "restart marker JPEG 与 stb 结构一致(RST 重同步正确)");
CHECK(pk <= PEAK_2500KB, "restart marker JPEG 峰值 ≤ 2.5 MiB");
meow_img_free(got);
meow_img_free(ref);
}
/* baseline + restart 与同参数无 restart 文件:RST 只切分熵段,解码结果应逐字节相同 */
{
char *pbase = path_join("base2000.jpg");
meow_img *a = meow_img_load_file_dim(prs, 720);
meow_img *b = meow_img_load_file_dim(pbase, 720);
CHECK(a && b && meow_img_w(a) == meow_img_w(b) && meow_img_h(a) == meow_img_h(b) &&
memcmp(meow_img_pixels(a), meow_img_pixels(b),
(size_t)meow_img_w(a) * meow_img_h(a) * 4) == 0,
"baseline restart 与无 restart 结果逐字节一致(mad=0)");
meow_img_free(a);
meow_img_free(b);
free(pbase);
}
free(psw);
free(prs);
}
/* ---- S11 畸形 DHT(过订阅 Huffman 表):拒绝 + 回退,不崩 ---- */
printf("[S11] 畸形 DHT 矩阵(码长 1..4 × 符号数 3..255)\n");
{
unsigned char buf[320];
int cases = 0, rejected = 0;
int lc, nsym;
for (lc = 1; lc <= 4; lc++) {
for (nsym = 3; nsym <= 255; nsym++) {
size_t n = build_badhuff(buf, lc, nsym);
int tw = 0, th = 0;
unsigned char *px = NULL;
cases++;
/* 崩溃即测试进程死;返回 1 说明过订阅表被当成合法表收下(缺陷) */
if (meow_img_backend_mem(buf, n, 720, &tw, &th, &px) == 0) rejected++;
if (px) free(px);
}
}
{
char name[128];
snprintf(name, sizeof name, "过订阅 DHT 全部拒绝(%d/%d,无崩溃)", rejected, cases);
CHECK(rejected == cases, name);
}
/* 有效 JPEG + 注入坏 DHT:后端必须拒绝;公共入口不得崩(stb 兜底或返回 NULL) */
{
size_t flen = 0, ilen = 0;
uint8_t *real = read_whole(p_base, &flen);
CHECK(real != NULL, "读入 baseline JPEG 用于注入");
if (real) {
uint8_t *inj = inject_badhuff(real, flen, 2, 100, &ilen);
int tw = 0, th = 0;
unsigned char *px = NULL;
meow_img *im;
CHECK(inj && meow_img_backend_mem(inj, ilen, 720, &tw, &th, &px) == 0,
"有效 JPEG + 注入过订阅 DHT → 后端拒绝");
if (px) free(px);
im = inj ? meow_img_load_mem_dim(inj, (int)ilen, 720) : NULL;
CHECK(im == NULL || nonblank(im), "注入坏表后公共入口不崩、不花屏");
meow_img_free(im);
free(inj);
free(real);
}
}
}
/* ---- S12 渐进式 + restart marker(P1-2 回归) ---- */
printf("[S12] progressive + restart marker(BACKEND 必须为 1,峰值 ≤2.5MiB)\n");
{
static const int rb[] = { 1, 2, 3, 4, 5, 8, 16, 64, 200 };
meow_img *ref = meow_img_load_file_dim(p_prog, 720);
int all_backend = 1, all_peak = 1, all_equal = 1;
for (size_t i = 0; i < sizeof rb / sizeof rb[0]; i++) {
char fname[48];
char *p;
int tw = 0, th = 0;
unsigned char *px = NULL;
meow_img *im;
size_t base, pk;
snprintf(fname, sizeof fname, "prog_rst%d.jpg", rb[i]);
p = path_join(fname);
if (meow_img_backend_file(p, 720, &tw, &th, &px) != 1) all_backend = 0;
if (px) free(px);
base = peak_reset();
im = meow_img_load_file_dim(p, 720);
pk = peak_since(base);
if (!(im && nonblank(im) && pk <= PEAK_2500KB)) all_peak = 0;
if (!(im && ref && meow_img_w(im) == meow_img_w(ref) && meow_img_h(im) == meow_img_h(ref) &&
memcmp(meow_img_pixels(im), meow_img_pixels(ref),
(size_t)meow_img_w(ref) * meow_img_h(ref) * 4) == 0)) {
double c2 = 0.0;
double d2 = (im && ref) ? mean_abs_diff(im, ref, 64, &c2) : 1e9;
if (!(d2 < 0.1)) all_equal = 0; /* 允许极个别像素 ±1(编码器分表差异) */
printf(" RST=%-3d 与无 restart 平均差 %.4f\n", rb[i], d2);
}
/* 关键正确性:与本文件自己的 stb 整幅解码比对(不同编码器脚本下 restart 文件的
* 系数可能不同,所以拿「无 restart 文件」当基准只能作参考) */
{
meow_img *rr = ref_thumb(p, 720);
double c3 = 0.0;
double d3 = (im && rr) ? mean_abs_diff(im, rr, 64, &c3) : 1e9;
if (!(d3 < 24.0 && c3 > 0.85)) all_peak = 0;
printf(" RST=%-3d vs stb 平均差 %.2f 相关 %.3f\n", rb[i], d3, c3);
meow_img_free(rr);
}
printf(" RST=%-3d → %dx%d peak %.2f MiB\n", rb[i],
meow_img_w(im), meow_img_h(im), (double)pk / 1048576.0);
free(p);
meow_img_free(im);
}
CHECK(all_backend == 1, "9 个 progressive+restart 全部走后端(BACKEND=1)");
CHECK(all_peak == 1, "9 个 progressive+restart 全部解码成功且峰值 ≤2.5MiB");
CHECK(all_equal == 1, "与无 restart 的同编码 progressive 结果逐字节一致(mad=0)");
meow_img_free(ref);
}
/* ---- S13 8bit gray/RGB PNG 的 tRNS 透明(P2-1 回归) ---- */
printf("[S13] PNG 8bit tRNS(gray / RGB)\n");
{
struct { const char *tag; const char *name; } fx[] = {
{ "gray8 tRNS", "gray8_trns.png" },
{ "rgb8 tRNS", "rgb8_trns.png" },
};
for (size_t i = 0; i < sizeof fx / sizeof fx[0]; i++) {
char *p = path_join(fx[i].name);
int tw = 0, th = 0;
unsigned char *px = NULL;
int handled = meow_img_backend_file(p, 320, &tw, &th, &px);
meow_img *got;
meow_img *full;
char name[128];
if (px) free(px);
got = meow_img_load_file_dim(p, 320);
full = meow_img_load_file(p); /* stb 整幅(tRNS 处理正确) */
snprintf(name, sizeof name, "%s 走后端(BACKEND=1)", fx[i].tag);
CHECK(handled == 1, name);
snprintf(name, sizeof name, "%s 目标图有透明像素(%ld 个)", fx[i].tag,
alpha_zero_count(got));
CHECK(got && alpha_zero_count(got) > 0, name);
snprintf(name, sizeof name, "%s alpha 与 stb 整幅逐点一致(不一致 %ld)", fx[i].tag,
alpha_mismatch(got, full));
CHECK(got && full && alpha_mismatch(got, full) == 0, name);
meow_img_free(got);
meow_img_free(full);
free(p);
}
}
/* ---- S14 M22 显式档位入口 _step:JPEG 走真缩解(PNG/BMP 待 M27 同步预算模型) ---- */
printf("[S14] meow_img_load_file_step(M22 档位入口)JPEG 走后端\n");
{
size_t base = peak_reset();
meow_img *im = meow_img_load_file_step(p_base, 720, 8);
size_t pk = peak_since(base);
char name[128];
snprintf(name, sizeof name, "_step(720,8) → %dx%d(长边 ≤720,非空)",
meow_img_w(im), meow_img_h(im));
CHECK(im && nonblank(im) && meow_img_w(im) <= 720 && meow_img_h(im) <= 720, name);
peak_report("_step(720, 8)", pk);
CHECK(pk <= PEAK_2500KB, "_step 峰值 ≤2.5MiB(走后端,不整幅物化)");
meow_img_free(im);
base = peak_reset();
im = meow_img_load_file_step(p_base, 720, 1);
pk = peak_since(base);
CHECK(im && nonblank(im) && meow_img_w(im) <= 720, "_step(720,1) 成功且长边 ≤720");
CHECK(pk <= PEAK_2500KB, "_step(720,1) 峰值 ≤2.5MiB");
meow_img_free(im);
}
/* ---- S8 内存入口 ---- */
printf("[S8] 内存入口 meow_img_load_mem_dim\n");
{