Merge branch 'busybox' into merge

6 files changed, 395 insertions, 63 deletions

diff --git a/libbb/Kbuild.src b/libbb/Kbuild.src
index d7ab1b129..1f2ed36d9 100644
--- a/libbb/Kbuild.src
+++ b/libbb/Kbuild.src
@@ -152,6 +152,7 @@ lib-$(CONFIG_VLOCK) += pw_encrypt.o correct_password.o
 lib-$(CONFIG_SU) += pw_encrypt.o correct_password.o
 lib-$(CONFIG_LOGIN) += pw_encrypt.o correct_password.o
 lib-$(CONFIG_FEATURE_HTTPD_AUTH_MD5) += pw_encrypt.o
+lib-$(CONFIG_FEATURE_FTP_AUTHENTICATION) += pw_encrypt.o
 
 lib-$(CONFIG_DF) += find_mount_point.o
 lib-$(CONFIG_MKFS_MINIX) += find_mount_point.o
diff --git a/libbb/appletlib.c b/libbb/appletlib.c
index 1d0a8f711..2c85062b4 100644
--- a/libbb/appletlib.c
+++ b/libbb/appletlib.c
@@ -184,8 +184,7 @@ void lbb_prepare(const char *applet
 #endif
         applet_name = applet;
 
-        /* Set locale for everybody except 'init' */
-        if (ENABLE_LOCALE_SUPPORT && getpid() != 1)
+        if (ENABLE_LOCALE_SUPPORT)
                 setlocale(LC_ALL, "");
 
 #if ENABLE_PLATFORM_MINGW32
@@ -759,15 +758,25 @@ void FAST_FUNC run_applet_no_and_exit(int applet_no, char **argv)
 
         /* Reinit some shared global data */
         xfunc_error_retval = EXIT_FAILURE;
-
         applet_name = APPLET_NAME(applet_no);
-        if (argc == 2 && strcmp(argv[1], "--help") == 0) {
-                /* Special case. POSIX says "test --help"
-                 * should be no different from e.g. "test --foo".  */
-//TODO: just compare applet_no with APPLET_NO_test
-                if (!ENABLE_TEST || strcmp(applet_name, "test") != 0) {
-                        /* If you want "foo --help" to return 0: */
-                        xfunc_error_retval = 0;
+
+#if defined APPLET_NO_test
+        /* Special case. POSIX says "test --help"
+         * should be no different from e.g. "test --foo".
+         * Thus for "test", we skip --help check.
+         */
+        if (applet_no != APPLET_NO_test)
+#endif
+        {
+                if (argc == 2 && strcmp(argv[1], "--help") == 0) {
+#if defined APPLET_NO_false
+                        /* Someone insisted that "false --help" must exit 1. Sigh */
+                        if (applet_no != APPLET_NO_false)
+#endif
+                        {
+                                /* Make "foo --help" exit with 0: */
+                                xfunc_error_retval = 0;
+                        }
                         bb_show_usage();
                 }
         }
diff --git a/libbb/bb_askpass.c b/libbb/bb_askpass.c
index 77c1bcd95..1927ba9e9 100644
--- a/libbb/bb_askpass.c
+++ b/libbb/bb_askpass.c
@@ -30,9 +30,12 @@ char* FAST_FUNC bb_ask(const int fd, int timeout, const char *prompt)
         struct sigaction sa, oldsa;
         struct termios tio, oldtio;
 
+        tcflush(fd, TCIFLUSH);
+        /* Was buggy: was printing prompt *before* flushing input,
+         * which was upsetting "expect" based scripts of some users.
+         */
         fputs(prompt, stdout);
         fflush_all();
-        tcflush(fd, TCIFLUSH);
 
         tcgetattr(fd, &oldtio);
         tio = oldtio;
diff --git a/libbb/correct_password.c b/libbb/correct_password.c
index acadf3914..513c93028 100644
--- a/libbb/correct_password.c
+++ b/libbb/correct_password.c
@@ -30,6 +30,63 @@
 
 #include "libbb.h"
 
+#define SHADOW_BUFSIZE 256
+
+/* Retrieve encrypted password string for pw.
+ * If pw == NULL, return a string which fails password check against any
+ * password.
+ */
+#if !ENABLE_FEATURE_SHADOWPASSWDS
+#define get_passwd(pw, buffer) get_passwd(pw)
+#endif
+static const char *get_passwd(const struct passwd *pw, char buffer[SHADOW_BUFSIZE])
+{
+        const char *pass;
+
+        if (!pw)
+                return "aa"; /* "aa" will never match */
+
+        pass = pw->pw_passwd;
+#if ENABLE_FEATURE_SHADOWPASSWDS
+        /* Using _r function to avoid pulling in static buffers */
+        if ((pass[0] == 'x' || pass[0] == '*') && !pass[1]) {
+                struct spwd spw;
+                int r;
+                /* getspnam_r may return 0 yet set result to NULL.
+                 * At least glibc 2.4 does this. Be extra paranoid here. */
+                struct spwd *result = NULL;
+                r = getspnam_r(pw->pw_name, &spw, buffer, SHADOW_BUFSIZE, &result);
+                pass = (r || !result) ? "aa" : result->sp_pwdp;
+        }
+#endif
+        return pass;
+}
+
+/*
+ * Return 1 if PW has an empty password.
+ * Return 1 if the user gives the correct password for entry PW,
+ * 0 if not.
+ * NULL pw means "just fake it for login with bad username"
+ */
+int FAST_FUNC check_password(const struct passwd *pw, const char *plaintext)
+{
+        IF_FEATURE_SHADOWPASSWDS(char buffer[SHADOW_BUFSIZE];)
+        char *encrypted;
+        const char *pw_pass;
+        int r;
+
+        pw_pass = get_passwd(pw, buffer);
+        if (!pw_pass[0]) { /* empty password field? */
+                return 1;
+        }
+
+        encrypted = pw_encrypt(plaintext, /*salt:*/ pw_pass, 1);
+        r = (strcmp(encrypted, pw_pass) == 0);
+        free(encrypted);
+        return r;
+}
+
+
 /* Ask the user for a password.
  * Return 1 without asking if PW has an empty password.
  * Return -1 on EOF, error while reading input, or timeout.
@@ -41,42 +98,23 @@
 int FAST_FUNC ask_and_check_password_extended(const struct passwd *pw,
                 int timeout, const char *prompt)
 {
-        char *unencrypted, *encrypted;
-        const char *correct;
+        IF_FEATURE_SHADOWPASSWDS(char buffer[SHADOW_BUFSIZE];)
+        char *plaintext;
+        const char *pw_pass;
         int r;
-        /* fake salt. crypt() can choke otherwise. */
-        correct = "aa";
-        if (!pw) {
-                /* "aa" will never match */
-                goto fake_it;
-        }
-        correct = pw->pw_passwd;
-#if ENABLE_FEATURE_SHADOWPASSWDS
-        /* Using _r function to avoid pulling in static buffers */
-        if ((correct[0] == 'x' || correct[0] == '*') && !correct[1]) {
-                struct spwd spw;
-                char buffer[256];
-                /* getspnam_r may return 0 yet set result to NULL.
-                 * At least glibc 2.4 does this. Be extra paranoid here. */
-                struct spwd *result = NULL;
-                r = getspnam_r(pw->pw_name, &spw, buffer, sizeof(buffer), &result);
-                correct = (r || !result) ? "aa" : result->sp_pwdp;
-        }
-#endif
 
-        if (!correct[0]) /* empty password field? */
+        pw_pass = get_passwd(pw, buffer);
+        if (!pw_pass[0]) /* empty password field? */
                 return 1;
 
- fake_it:
-        unencrypted = bb_ask(STDIN_FILENO, timeout, prompt);
-        if (!unencrypted) {
+        plaintext = bb_ask(STDIN_FILENO, timeout, prompt);
+        if (!plaintext) {
                 /* EOF (such as ^D) or error (such as ^C) or timeout */
                 return -1;
         }
-        encrypted = pw_encrypt(unencrypted, correct, 1);
-        r = (strcmp(encrypted, correct) == 0);
-        free(encrypted);
-        nuke_str(unencrypted);
+
+        r = check_password(pw, plaintext);
+        nuke_str(plaintext);
         return r;
 }
 
diff --git a/libbb/hash_md5_sha.c b/libbb/hash_md5_sha.c
index 3f743ac75..1f63ccdee 100644
--- a/libbb/hash_md5_sha.c
+++ b/libbb/hash_md5_sha.c
@@ -926,10 +926,81 @@ void FAST_FUNC sha512_end(sha512_ctx_t *ctx, void *resbuf)
 # define SHA3_SMALL CONFIG_SHA3_SMALL
 #endif
 
+#define OPTIMIZE_SHA3_FOR_32 0
+/*
+ * SHA3 can be optimized for 32-bit CPUs with bit-slicing:
+ * every 64-bit word of state[] can be split into two 32-bit words
+ * by even/odd bits. In this form, all rotations of sha3 round
+ * are 32-bit - and there are lots of them.
+ * However, it requires either splitting/combining state words
+ * before/after sha3 round (code does this now)
+ * or shuffling bits before xor'ing them into state and in sha3_end.
+ * Without shuffling, bit-slicing results in -130 bytes of code
+ * and marginal speedup (but of course it gives wrong result).
+ * With shuffling it works, but +260 code bytes, and slower.
+ * Disabled for now:
+ */
+#if 0 /* LONG_MAX == 0x7fffffff */
+# undef OPTIMIZE_SHA3_FOR_32
+# define OPTIMIZE_SHA3_FOR_32 1
+#endif
+
 enum {
         SHA3_IBLK_BYTES = 72, /* 576 bits / 8 */
 };
 
+#if OPTIMIZE_SHA3_FOR_32
+/* This splits every 64-bit word into a pair of 32-bit words,
+ * even bits go into first word, odd bits go to second one.
+ * The conversion is done in-place.
+ */
+static void split_halves(uint64_t *state)
+{
+        /* Credit: Henry S. Warren, Hacker's Delight, Addison-Wesley, 2002 */
+        uint32_t *s32 = (uint32_t*)state;
+        uint32_t t, x0, x1;
+        int i;
+        for (i = 24; i >= 0; --i) {
+                x0 = s32[0];
+                t = (x0 ^ (x0 >> 1)) & 0x22222222; x0 = x0 ^ t ^ (t << 1);
+                t = (x0 ^ (x0 >> 2)) & 0x0C0C0C0C; x0 = x0 ^ t ^ (t << 2);
+                t = (x0 ^ (x0 >> 4)) & 0x00F000F0; x0 = x0 ^ t ^ (t << 4);
+                t = (x0 ^ (x0 >> 8)) & 0x0000FF00; x0 = x0 ^ t ^ (t << 8);
+                x1 = s32[1];
+                t = (x1 ^ (x1 >> 1)) & 0x22222222; x1 = x1 ^ t ^ (t << 1);
+                t = (x1 ^ (x1 >> 2)) & 0x0C0C0C0C; x1 = x1 ^ t ^ (t << 2);
+                t = (x1 ^ (x1 >> 4)) & 0x00F000F0; x1 = x1 ^ t ^ (t << 4);
+                t = (x1 ^ (x1 >> 8)) & 0x0000FF00; x1 = x1 ^ t ^ (t << 8);
+                *s32++ = (x0 & 0x0000FFFF) | (x1 << 16);
+                *s32++ = (x0 >> 16) | (x1 & 0xFFFF0000);
+        }
+}
+/* The reverse operation */
+static void combine_halves(uint64_t *state)
+{
+        uint32_t *s32 = (uint32_t*)state;
+        uint32_t t, x0, x1;
+        int i;
+        for (i = 24; i >= 0; --i) {
+                x0 = s32[0];
+                x1 = s32[1];
+                t = (x0 & 0x0000FFFF) | (x1 << 16);
+                x1 = (x0 >> 16) | (x1 & 0xFFFF0000);
+                x0 = t;
+                t = (x0 ^ (x0 >> 8)) & 0x0000FF00; x0 = x0 ^ t ^ (t << 8);
+                t = (x0 ^ (x0 >> 4)) & 0x00F000F0; x0 = x0 ^ t ^ (t << 4);
+                t = (x0 ^ (x0 >> 2)) & 0x0C0C0C0C; x0 = x0 ^ t ^ (t << 2);
+                t = (x0 ^ (x0 >> 1)) & 0x22222222; x0 = x0 ^ t ^ (t << 1);
+                *s32++ = x0;
+                t = (x1 ^ (x1 >> 8)) & 0x0000FF00; x1 = x1 ^ t ^ (t << 8);
+                t = (x1 ^ (x1 >> 4)) & 0x00F000F0; x1 = x1 ^ t ^ (t << 4);
+                t = (x1 ^ (x1 >> 2)) & 0x0C0C0C0C; x1 = x1 ^ t ^ (t << 2);
+                t = (x1 ^ (x1 >> 1)) & 0x22222222; x1 = x1 ^ t ^ (t << 1);
+                *s32++ = x1;
+        }
+}
+#endif
+
 /*
  * In the crypto literature this function is usually called Keccak-f().
  */
@@ -937,12 +1008,179 @@ static void sha3_process_block72(uint64_t *state)
 {
         enum { NROUNDS = 24 };
 
-        /* Elements should be 64-bit, but top half is always zero or 0x80000000.
-         * We encode 63rd bits in a separate word below.
-         * Same is true for 31th bits, which lets us use 16-bit table instead of 64-bit.
-         * The speed penalty is lost in the noise.
-         */
+#if OPTIMIZE_SHA3_FOR_32
+        /*
+        static const uint32_t IOTA_CONST_0[NROUNDS] = {
+                0x00000001UL,
+                0x00000000UL,
+                0x00000000UL,
+                0x00000000UL,
+                0x00000001UL,
+                0x00000001UL,
+                0x00000001UL,
+                0x00000001UL,
+                0x00000000UL,
+                0x00000000UL,
+                0x00000001UL,
+                0x00000000UL,
+                0x00000001UL,
+                0x00000001UL,
+                0x00000001UL,
+                0x00000001UL,
+                0x00000000UL,
+                0x00000000UL,
+                0x00000000UL,
+                0x00000000UL,
+                0x00000001UL,
+                0x00000000UL,
+                0x00000001UL,
+                0x00000000UL,
+        };
+        ** bits are in lsb: 0101 0000 1111 0100 1111 0001
+        */
+        uint32_t IOTA_CONST_0bits = (uint32_t)(0x0050f4f1);
+        static const uint32_t IOTA_CONST_1[NROUNDS] = {
+                0x00000000UL,
+                0x00000089UL,
+                0x8000008bUL,
+                0x80008080UL,
+                0x0000008bUL,
+                0x00008000UL,
+                0x80008088UL,
+                0x80000082UL,
+                0x0000000bUL,
+                0x0000000aUL,
+                0x00008082UL,
+                0x00008003UL,
+                0x0000808bUL,
+                0x8000000bUL,
+                0x8000008aUL,
+                0x80000081UL,
+                0x80000081UL,
+                0x80000008UL,
+                0x00000083UL,
+                0x80008003UL,
+                0x80008088UL,
+                0x80000088UL,
+                0x00008000UL,
+                0x80008082UL,
+        };
+
+        uint32_t *const s32 = (uint32_t*)state;
+        unsigned round;
+
+        split_halves(state);
+
+        for (round = 0; round < NROUNDS; round++) {
+                unsigned x;
+
+                /* Theta */
+                {
+                        uint32_t BC[20];
+                        for (x = 0; x < 10; ++x) {
+                                BC[x+10] = BC[x] = s32[x]^s32[x+10]^s32[x+20]^s32[x+30]^s32[x+40];
+                        }
+                        for (x = 0; x < 10; x += 2) {
+                                uint32_t ta, tb;
+                                ta = BC[x+8] ^ rotl32(BC[x+3], 1);
+                                tb = BC[x+9] ^ BC[x+2];
+                                s32[x+0] ^= ta;
+                                s32[x+1] ^= tb;
+                                s32[x+10] ^= ta;
+                                s32[x+11] ^= tb;
+                                s32[x+20] ^= ta;
+                                s32[x+21] ^= tb;
+                                s32[x+30] ^= ta;
+                                s32[x+31] ^= tb;
+                                s32[x+40] ^= ta;
+                                s32[x+41] ^= tb;
+                        }
+                }
+                /* RhoPi */
+                {
+                        uint32_t t0a,t0b, t1a,t1b;
+                        t1a = s32[1*2+0];
+                        t1b = s32[1*2+1];
+
+#define RhoPi(PI_LANE, ROT_CONST) \
+        t0a = s32[PI_LANE*2+0];\
+        t0b = s32[PI_LANE*2+1];\
+        if (ROT_CONST & 1) {\
+                s32[PI_LANE*2+0] = rotl32(t1b, ROT_CONST/2+1);\
+                s32[PI_LANE*2+1] = ROT_CONST == 1 ? t1a : rotl32(t1a, ROT_CONST/2+0);\
+        } else {\
+                s32[PI_LANE*2+0] = rotl32(t1a, ROT_CONST/2);\
+                s32[PI_LANE*2+1] = rotl32(t1b, ROT_CONST/2);\
+        }\
+        t1a = t0a; t1b = t0b;
+
+                        RhoPi(10, 1)
+                        RhoPi( 7, 3)
+                        RhoPi(11, 6)
+                        RhoPi(17,10)
+                        RhoPi(18,15)
+                        RhoPi( 3,21)
+                        RhoPi( 5,28)
+                        RhoPi(16,36)
+                        RhoPi( 8,45)
+                        RhoPi(21,55)
+                        RhoPi(24, 2)
+                        RhoPi( 4,14)
+                        RhoPi(15,27)
+                        RhoPi(23,41)
+                        RhoPi(19,56)
+                        RhoPi(13, 8)
+                        RhoPi(12,25)
+                        RhoPi( 2,43)
+                        RhoPi(20,62)
+                        RhoPi(14,18)
+                        RhoPi(22,39)
+                        RhoPi( 9,61)
+                        RhoPi( 6,20)
+                        RhoPi( 1,44)
+#undef RhoPi
+                }
+                /* Chi */
+                for (x = 0; x <= 40;) {
+                        uint32_t BC0, BC1, BC2, BC3, BC4;
+                        BC0 = s32[x + 0*2];
+                        BC1 = s32[x + 1*2];
+                        BC2 = s32[x + 2*2];
+                        s32[x + 0*2] = BC0 ^ ((~BC1) & BC2);
+                        BC3 = s32[x + 3*2];
+                        s32[x + 1*2] = BC1 ^ ((~BC2) & BC3);
+                        BC4 = s32[x + 4*2];
+                        s32[x + 2*2] = BC2 ^ ((~BC3) & BC4);
+                        s32[x + 3*2] = BC3 ^ ((~BC4) & BC0);
+                        s32[x + 4*2] = BC4 ^ ((~BC0) & BC1);
+                        x++;
+                        BC0 = s32[x + 0*2];
+                        BC1 = s32[x + 1*2];
+                        BC2 = s32[x + 2*2];
+                        s32[x + 0*2] = BC0 ^ ((~BC1) & BC2);
+                        BC3 = s32[x + 3*2];
+                        s32[x + 1*2] = BC1 ^ ((~BC2) & BC3);
+                        BC4 = s32[x + 4*2];
+                        s32[x + 2*2] = BC2 ^ ((~BC3) & BC4);
+                        s32[x + 3*2] = BC3 ^ ((~BC4) & BC0);
+                        s32[x + 4*2] = BC4 ^ ((~BC0) & BC1);
+                        x += 9;
+                }
+                /* Iota */
+                s32[0] ^= IOTA_CONST_0bits & 1;
+                IOTA_CONST_0bits >>= 1;
+                s32[1] ^= IOTA_CONST_1[round];
+        }
+
+        combine_halves(state);
+#else
+        /* Native 64-bit algorithm */
         static const uint16_t IOTA_CONST[NROUNDS] = {
+                /* Elements should be 64-bit, but top half is always zero
+                 * or 0x80000000. We encode 63rd bits in a separate word below.
+                 * Same is true for 31th bits, which lets us use 16-bit table
+                 * instead of 64-bit. The speed penalty is lost in the noise.
+                 */
                 0x0001,
                 0x8082,
                 0x808a,
@@ -983,7 +1221,7 @@ static void sha3_process_block72(uint64_t *state)
         };
         /*static const uint8_t MOD5[10] = { 0, 1, 2, 3, 4, 0, 1, 2, 3, 4, };*/
 
-        unsigned x, y;
+        unsigned x;
         unsigned round;
 
         if (BB_BIG_ENDIAN) {
@@ -1045,22 +1283,62 @@ static void sha3_process_block72(uint64_t *state)
                         RhoPi_twice(20); RhoPi_twice(22);
 #undef RhoPi_twice
                 }
-
                 /* Chi */
-                for (y = 0; y <= 20; y += 5) {
+# if LONG_MAX > 0x7fffffff
+                for (x = 0; x <= 20; x += 5) {
                         uint64_t BC0, BC1, BC2, BC3, BC4;
-                        BC0 = state[y + 0];
-                        BC1 = state[y + 1];
-                        BC2 = state[y + 2];
-                        state[y + 0] = BC0 ^ ((~BC1) & BC2);
-                        BC3 = state[y + 3];
-                        state[y + 1] = BC1 ^ ((~BC2) & BC3);
-                        BC4 = state[y + 4];
-                        state[y + 2] = BC2 ^ ((~BC3) & BC4);
-                        state[y + 3] = BC3 ^ ((~BC4) & BC0);
-                        state[y + 4] = BC4 ^ ((~BC0) & BC1);
+                        BC0 = state[x + 0];
+                        BC1 = state[x + 1];
+                        BC2 = state[x + 2];
+                        state[x + 0] = BC0 ^ ((~BC1) & BC2);
+                        BC3 = state[x + 3];
+                        state[x + 1] = BC1 ^ ((~BC2) & BC3);
+                        BC4 = state[x + 4];
+                        state[x + 2] = BC2 ^ ((~BC3) & BC4);
+                        state[x + 3] = BC3 ^ ((~BC4) & BC0);
+                        state[x + 4] = BC4 ^ ((~BC0) & BC1);
                 }
-
+# else
+                /* Reduced register pressure version
+                 * for register-starved 32-bit arches
+                 * (i386: -95 bytes, and it is _faster_)
+                 */
+                for (x = 0; x <= 40;) {
+                        uint32_t BC0, BC1, BC2, BC3, BC4;
+                        uint32_t *const s32 = (uint32_t*)state;
+#  if SHA3_SMALL
+ do_half:
+#  endif
+                        BC0 = s32[x + 0*2];
+                        BC1 = s32[x + 1*2];
+                        BC2 = s32[x + 2*2];
+                        s32[x + 0*2] = BC0 ^ ((~BC1) & BC2);
+                        BC3 = s32[x + 3*2];
+                        s32[x + 1*2] = BC1 ^ ((~BC2) & BC3);
+                        BC4 = s32[x + 4*2];
+                        s32[x + 2*2] = BC2 ^ ((~BC3) & BC4);
+                        s32[x + 3*2] = BC3 ^ ((~BC4) & BC0);
+                        s32[x + 4*2] = BC4 ^ ((~BC0) & BC1);
+                        x++;
+#  if SHA3_SMALL
+                        if (x & 1)
+                                goto do_half;
+                        x += 8;
+#  else
+                        BC0 = s32[x + 0*2];
+                        BC1 = s32[x + 1*2];
+                        BC2 = s32[x + 2*2];
+                        s32[x + 0*2] = BC0 ^ ((~BC1) & BC2);
+                        BC3 = s32[x + 3*2];
+                        s32[x + 1*2] = BC1 ^ ((~BC2) & BC3);
+                        BC4 = s32[x + 4*2];
+                        s32[x + 2*2] = BC2 ^ ((~BC3) & BC4);
+                        s32[x + 3*2] = BC3 ^ ((~BC4) & BC0);
+                        s32[x + 4*2] = BC4 ^ ((~BC0) & BC1);
+                        x += 9;
+#  endif
+                }
+# endif /* long is 32-bit */
                 /* Iota */
                 state[0] ^= IOTA_CONST[round]
                         | (uint32_t)((IOTA_CONST_bit31 << round) & 0x80000000)
@@ -1072,6 +1350,7 @@ static void sha3_process_block72(uint64_t *state)
                         state[x] = SWAP_LE64(state[x]);
                 }
         }
+#endif
 }
 
 void FAST_FUNC sha3_begin(sha3_ctx_t *ctx)
diff --git a/libbb/xfuncs.c b/libbb/xfuncs.c
index 23f27516f..f25ce9446 100644
--- a/libbb/xfuncs.c
+++ b/libbb/xfuncs.c
@@ -25,20 +25,22 @@
 #include "libbb.h"
 
 /* Turn on nonblocking I/O on a fd */
-void FAST_FUNC ndelay_on(int fd)
+int FAST_FUNC ndelay_on(int fd)
 {
         int flags = fcntl(fd, F_GETFL);
         if (flags & O_NONBLOCK)
-                return;
+                return flags;
         fcntl(fd, F_SETFL, flags | O_NONBLOCK);
+        return flags;
 }
 
-void FAST_FUNC ndelay_off(int fd)
+int FAST_FUNC ndelay_off(int fd)
 {
         int flags = fcntl(fd, F_GETFL);
         if (!(flags & O_NONBLOCK))
-                return;
+                return flags;
         fcntl(fd, F_SETFL, flags & ~O_NONBLOCK);
+        return flags;
 }
 
 void FAST_FUNC close_on_exec_on(int fd)