]> arthur.barton.de Git - bup.git/blobdiff - lib/bup/_helpers.c
Rework write_sparsely() to fix in-buffer zero runs
[bup.git] / lib / bup / _helpers.c
index 6e93e35275bed693d372cb0f4f2c0a2d718bd38e..59f0fa381558e255a37f8b7059881f09feb73229 100644 (file)
 #define BUP_HAVE_FILE_ATTRS 1
 #endif
 
+/*
+ * Check for incomplete UTIMENSAT support (NetBSD 6), and if so,
+ * pretend we don't have it.
+ */
+#if !defined(AT_FDCWD) || !defined(AT_SYMLINK_NOFOLLOW)
+#undef HAVE_UTIMENSAT
+#endif
+
 #ifndef FS_NOCOW_FL
 // Of course, this assumes it's a bitfield value.
 #define FS_NOCOW_FL 0
 static int istty2 = 0;
 
 
+#ifndef htonll
+// This function should technically be macro'd out if it's going to be used
+// more than ocasionally.  As of this writing, it'll actually never be called
+// in real world bup scenarios (because our packs are < MAX_INT bytes).
+static uint64_t htonll(uint64_t value)
+{
+    static const int endian_test = 42;
+
+    if (*(char *)&endian_test == endian_test) // LSB-MSB
+       return ((uint64_t)htonl(value & 0xFFFFFFFF) << 32) | htonl(value >> 32);
+    return value; // already in network byte order MSB-LSB
+}
+#endif
+
+
+#define INTEGRAL_ASSIGNMENT_FITS(dest, src)                             \
+    ({                                                                  \
+        *(dest) = (src);                                                \
+        *(dest) == (src) && (*(dest) < 1) == ((src) < 1);               \
+    })
+
+
+// At the moment any code that calls INTGER_TO_PY() will have to
+// disable -Wtautological-compare for clang.  See below.
+
 #define INTEGER_TO_PY(x) \
     (((x) >= 0) ? PyLong_FromUnsignedLongLong(x) : PyLong_FromLongLong(x))
 
@@ -196,6 +229,135 @@ static void unpythonize_argv(void)
 #endif // not __WIN32__ or __CYGWIN__
 
 
+static unsigned long long count_leading_zeros(const unsigned char * const buf,
+                                              unsigned long long len)
+{
+    const unsigned char *cur = buf;
+    while(len-- && *cur == 0)
+        cur++;
+    return cur - buf;
+}
+
+
+static int write_all(int fd, const void *buf, const size_t count)
+{
+    size_t written = 0;
+    while (written < count)
+    {
+        const ssize_t rc = write(fd, buf + written, count - written);
+        if (rc == -1)
+            return -1;
+        written += rc;
+    }
+    return 0;
+}
+
+
+static int uadd(unsigned long long *dest,
+                const unsigned long long x,
+                const unsigned long long y)
+{
+    const unsigned long long result = x + y;
+    if (result < x || result < y)
+        return 0;
+    *dest = result;
+    return 1;
+}
+
+static PyObject *append_sparse_region(const int fd, unsigned long long n)
+{
+    while(n)
+    {
+        off_t new_off;
+        if (!INTEGRAL_ASSIGNMENT_FITS(&new_off, n))
+            new_off = INT_MAX;
+        const off_t off = lseek(fd, new_off, SEEK_CUR);
+        if (off == (off_t) -1)
+            return PyErr_SetFromErrno(PyExc_IOError);
+        n -= new_off;
+    }
+    return NULL;
+}
+
+
+static PyObject *bup_write_sparsely(PyObject *self, PyObject *args)
+{
+    int fd;
+    unsigned char *buf = NULL;
+    Py_ssize_t sbuf_len;
+    PyObject *py_min_sparse_len, *py_prev_sparse_len;
+    if (!PyArg_ParseTuple(args, "it#OO",
+                          &fd, &buf, &sbuf_len,
+                          &py_min_sparse_len, &py_prev_sparse_len))
+       return NULL;
+    unsigned long long min_sparse_len, prev_sparse_len, buf_len;
+    if (!bup_ullong_from_py(&min_sparse_len, py_min_sparse_len, "min_sparse_len"))
+        return NULL;
+    if (!bup_ullong_from_py(&prev_sparse_len, py_prev_sparse_len, "prev_sparse_len"))
+        return NULL;
+    if (sbuf_len < 0)
+        return PyErr_Format(PyExc_ValueError, "negative bufer length");
+    if (!INTEGRAL_ASSIGNMENT_FITS(&buf_len, sbuf_len))
+        return PyErr_Format(PyExc_OverflowError, "buffer length too large");
+
+    // The value of zeros_read indicates the number of zeros read from
+    // buf that haven't been accounted for yet (with respect to cur),
+    // while zeros indicates the total number of pending zeros, which
+    // could be larger in the first iteration if prev_sparse_len
+    // wasn't zero.
+    int rc;
+    unsigned long long unexamined = buf_len;
+    unsigned char *block_start = buf, *cur = buf;
+    unsigned long long zeros, zeros_read = count_leading_zeros(cur, unexamined);
+    assert(zeros_read <= unexamined);
+    unexamined -= zeros_read;
+    if (!uadd(&zeros, prev_sparse_len, zeros_read))
+    {
+        PyObject *err = append_sparse_region(fd, prev_sparse_len);
+        if (err != NULL)
+            return err;
+        zeros = zeros_read;
+    }
+
+    while(unexamined)
+    {
+        if (zeros < min_sparse_len)
+            cur += zeros_read;
+        else
+        {
+            rc = write_all(fd, block_start, cur - block_start);
+            if (rc)
+                return PyErr_SetFromErrno(PyExc_IOError);
+            PyObject *err = append_sparse_region(fd, zeros);
+            if (err != NULL)
+                return err;
+            cur += zeros_read;
+            block_start = cur;
+        }
+        // Pending zeros have ether been made sparse, or are going to
+        // be rolled into the next non-sparse block since we know we
+        // now have at least one unexamined non-zero byte.
+        assert(unexamined && *cur != 0);
+        zeros = zeros_read = 0;
+        while (unexamined && *cur != 0)
+        {
+            cur++; unexamined--;
+        }
+        if (unexamined)
+        {
+            zeros_read = count_leading_zeros(cur, unexamined);
+            assert(zeros_read <= unexamined);
+            unexamined -= zeros_read;
+            zeros = zeros_read;
+        }
+    }
+    rc = write_all(fd, block_start, cur - block_start);
+    if (rc)
+        return PyErr_SetFromErrno(PyExc_IOError);
+    return PyLong_FromUnsignedLongLong(zeros);
+}
+
+
 static PyObject *selftest(PyObject *self, PyObject *args)
 {
     if (!PyArg_ParseTuple(args, ""))
@@ -576,18 +738,6 @@ static PyObject *merge_into(PyObject *self, PyObject *args)
     return PyLong_FromUnsignedLong(count);
 }
 
-// This function should technically be macro'd out if it's going to be used
-// more than ocasionally.  As of this writing, it'll actually never be called
-// in real world bup scenarios (because our packs are < MAX_INT bytes).
-static uint64_t htonll(uint64_t value)
-{
-    static const int endian_test = 42;
-
-    if (*(char *)&endian_test == endian_test) // LSB-MSB
-       return ((uint64_t)htonl(value & 0xFFFFFFFF) << 32) | htonl(value >> 32);
-    return value; // already in network byte order MSB-LSB
-}
-
 #define FAN_ENTRIES 256
 
 static PyObject *write_idx(PyObject *self, PyObject *args)
@@ -801,11 +951,21 @@ static PyObject *fadvise_done(PyObject *self, PyObject *args)
 }
 
 
+// Currently the Linux kernel and FUSE disagree over the type for
+// FS_IOC_GETFLAGS and FS_IOC_SETFLAGS.  The kernel actually uses int,
+// but FUSE chose long (matching the declaration in linux/fs.h).  So
+// if you use int, and then traverse a FUSE filesystem, you may
+// corrupt the stack.  But if you use long, then you may get invalid
+// results on big-endian systems.
+//
+// For now, we just use long, and then disable Linux attrs entirely
+// (with a warning) in helpers.py on systems that are affected.
+
 #ifdef BUP_HAVE_FILE_ATTRS
 static PyObject *bup_get_linux_file_attr(PyObject *self, PyObject *args)
 {
     int rc;
-    unsigned int attr;
+    unsigned long attr;
     char *path;
     int fd;
 
@@ -816,16 +976,16 @@ static PyObject *bup_get_linux_file_attr(PyObject *self, PyObject *args)
     if (fd == -1)
         return PyErr_SetFromErrnoWithFilename(PyExc_OSError, path);
 
-    attr = 0;
+    attr = 0;  // Handle int/long mismatch (see above)
     rc = ioctl(fd, FS_IOC_GETFLAGS, &attr);
     if (rc == -1)
     {
         close(fd);
         return PyErr_SetFromErrnoWithFilename(PyExc_OSError, path);
     }
-
     close(fd);
-    return Py_BuildValue("I", attr);
+    assert(attr <= UINT_MAX);  // Kernel type is actually int
+    return PyLong_FromUnsignedLong(attr);
 }
 #endif /* def BUP_HAVE_FILE_ATTRS */
 
@@ -835,7 +995,8 @@ static PyObject *bup_get_linux_file_attr(PyObject *self, PyObject *args)
 static PyObject *bup_set_linux_file_attr(PyObject *self, PyObject *args)
 {
     int rc;
-    unsigned int orig_attr, attr;
+    unsigned long orig_attr;
+    unsigned int attr;
     char *path;
     PyObject *py_attr;
     int fd;
@@ -859,13 +1020,15 @@ static PyObject *bup_set_linux_file_attr(PyObject *self, PyObject *args)
     | FS_TOPDIR_FL | FS_NOCOW_FL;
 
     // The extents flag can't be removed, so don't (see chattr(1) and chattr.c).
+    orig_attr = 0; // Handle int/long mismatch (see above)
     rc = ioctl(fd, FS_IOC_GETFLAGS, &orig_attr);
+    assert(orig_attr <= UINT_MAX);  // Kernel type is actually int
     if (rc == -1)
     {
         close(fd);
         return PyErr_SetFromErrnoWithFilename(PyExc_OSError, path);
     }
-    attr |= (orig_attr & FS_EXTENT_FL);
+    attr |= ((unsigned int) orig_attr) & FS_EXTENT_FL;
 
     rc = ioctl(fd, FS_IOC_SETFLAGS, &attr);
     if (rc == -1)
@@ -889,25 +1052,17 @@ static PyObject *bup_set_linux_file_attr(PyObject *self, PyObject *args)
 #endif
 #endif
 
-
-#define INTEGRAL_ASSIGNMENT_FITS(dest, src)                             \
-    ({                                                                  \
-        *(dest) = (src);                                                \
-        *(dest) == (src) && (*(dest) < 1) == ((src) < 1);               \
-    })
-
-
 #define ASSIGN_PYLONG_TO_INTEGRAL(dest, pylong, overflow) \
     ({                                                     \
         int result = 0;                                                 \
         *(overflow) = 0;                                                \
-        const long long ltmp = PyLong_AsLongLong(pylong);               \
-        if (ltmp == -1 && PyErr_Occurred())                             \
+        const long long lltmp = PyLong_AsLongLong(pylong);              \
+        if (lltmp == -1 && PyErr_Occurred())                            \
         {                                                               \
             if (PyErr_ExceptionMatches(PyExc_OverflowError))            \
             {                                                           \
-                const unsigned long ultmp = PyLong_AsUnsignedLongLong(pylong); \
-                if (ultmp == (unsigned long long) -1 && PyErr_Occurred()) \
+                const unsigned long long ulltmp = PyLong_AsUnsignedLongLong(pylong); \
+                if (ulltmp == (unsigned long long) -1 && PyErr_Occurred()) \
                 {                                                       \
                     if (PyErr_ExceptionMatches(PyExc_OverflowError))    \
                     {                                                   \
@@ -915,7 +1070,7 @@ static PyObject *bup_set_linux_file_attr(PyObject *self, PyObject *args)
                         *(overflow) = 1;                                \
                     }                                                   \
                 }                                                       \
-                if (INTEGRAL_ASSIGNMENT_FITS((dest), ultmp))            \
+                if (INTEGRAL_ASSIGNMENT_FITS((dest), ulltmp))           \
                     result = 1;                                         \
                 else                                                    \
                     *(overflow) = 1;                                    \
@@ -923,7 +1078,7 @@ static PyObject *bup_set_linux_file_attr(PyObject *self, PyObject *args)
         }                                                               \
         else                                                            \
         {                                                               \
-            if (INTEGRAL_ASSIGNMENT_FITS((dest), ltmp))                 \
+            if (INTEGRAL_ASSIGNMENT_FITS((dest), lltmp))                \
                 result = 1;                                             \
             else                                                        \
                 *(overflow) = 1;                                        \
@@ -1065,6 +1220,9 @@ static PyObject *bup_lutimes(PyObject *self, PyObject *args)
 #endif
 
 
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wtautological-compare" // For INTEGER_TO_PY().
+
 static PyObject *stat_struct_to_py(const struct stat *st,
                                    const char *filename,
                                    int fd)
@@ -1090,6 +1248,7 @@ static PyObject *stat_struct_to_py(const struct stat *st,
                          (long) BUP_STAT_CTIME_NS(st));
 }
 
+#pragma clang diagnostic pop  // ignored "-Wtautological-compare"
 
 static PyObject *bup_stat(PyObject *self, PyObject *args)
 {
@@ -1139,6 +1298,8 @@ static PyObject *bup_fstat(PyObject *self, PyObject *args)
 
 
 static PyMethodDef helper_methods[] = {
+    { "write_sparsely", bup_write_sparsely, METH_VARARGS,
+      "Write buf excepting zeros at the end. Return trailing zero count." },
     { "selftest", selftest, METH_VARARGS,
        "Check that the rolling checksum rolls correctly (for unit tests)." },
     { "blobbits", blobbits, METH_VARARGS,
@@ -1211,11 +1372,20 @@ PyMODINIT_FUNC init_helpers(void)
     assert(sizeof(PY_LONG_LONG) <= sizeof(long long));
     assert(sizeof(unsigned PY_LONG_LONG) <= sizeof(unsigned long long));
 
+    if (sizeof(off_t) < sizeof(int))
+    {
+        // Originally required by append_sparse_region().
+        fprintf(stderr, "sizeof(off_t) < sizeof(int); please report.\n");
+        exit(1);
+    }
+
     char *e;
     PyObject *m = Py_InitModule("_helpers", helper_methods);
     if (m == NULL)
         return;
 
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wtautological-compare" // For INTEGER_TO_PY().
 #ifdef HAVE_UTIMENSAT
     {
         PyObject *value;
@@ -1230,6 +1400,19 @@ PyMODINIT_FUNC init_helpers(void)
         Py_DECREF(value);
     }
 #endif
+    {
+        PyObject *value;
+        const long arg_max = sysconf(_SC_ARG_MAX);
+        if (arg_max == -1)
+        {
+            fprintf(stderr, "Cannot find SC_ARG_MAX, please report a bug.\n");
+            exit(1);
+        }
+        value = INTEGER_TO_PY(arg_max);
+        PyObject_SetAttrString(m, "SC_ARG_MAX", value);
+        Py_DECREF(value);
+    }
+#pragma clang diagnostic pop  // ignored "-Wtautological-compare"
 
     e = getenv("BUP_FORCE_TTY");
     istty2 = isatty(2) || (atoi(e ? e : "0") & 2);