]> git.ipfire.org Git - thirdparty/systemd.git/blobdiff - src/shared/loop-util.c
Merge pull request #16678 from poettering/loop-configure
[thirdparty/systemd.git] / src / shared / loop-util.c
index bbb85f9e6e0e78a97814d2a43a9bfdaf9abcbe0c..8db2fed66ff3ee4336e0a68678000e1f099998ad 100644 (file)
@@ -1,5 +1,9 @@
 /* SPDX-License-Identifier: LGPL-2.1+ */
 
+#if HAVE_VALGRIND_MEMCHECK_H
+#include <valgrind/memcheck.h>
+#endif
+
 #include <errno.h>
 #include <fcntl.h>
 #include <linux/blkpg.h>
 #include <linux/loop.h>
 #include <sys/file.h>
 #include <sys/ioctl.h>
+#include <unistd.h>
 
 #include "alloc-util.h"
+#include "blockdev-util.h"
+#include "errno-util.h"
 #include "fd-util.h"
 #include "fileio.h"
 #include "loop-util.h"
+#include "missing_loop.h"
 #include "parse-util.h"
 #include "stat-util.h"
 #include "stdio-util.h"
+#include "string-util.h"
+
+static void cleanup_clear_loop_close(int *fd) {
+        if (*fd < 0)
+                return;
 
-int loop_device_make_full(
+        (void) ioctl(*fd, LOOP_CLR_FD);
+        (void) safe_close(*fd);
+}
+
+static int loop_configure(int fd, const struct loop_config *c) {
+        int r;
+
+        assert(fd >= 0);
+        assert(c);
+
+        if (ioctl(fd, LOOP_CONFIGURE, c) < 0) {
+                /* Do fallback only if LOOP_CONFIGURE is not supported, propagate all other errors. Note that
+                 * the kernel is weird: non-existing ioctls currently return EINVAL rather than ENOTTY on
+                 * loopback block devices. They should fix that in the kernel, but in the meantime we accept
+                 * both here. */
+                if (!ERRNO_IS_NOT_SUPPORTED(errno) && errno != EINVAL)
+                        return -errno;
+        } else {
+                if (!FLAGS_SET(c->info.lo_flags, LO_FLAGS_PARTSCAN))
+                        return 0;
+
+                /* Kernel 5.8 vanilla doesn't properly propagate the partition scanning flag into the
+                 * block device. Let's hence verify if things work correctly here before returning. */
+
+                r = blockdev_partscan_enabled(fd);
+                if (r < 0)
+                        goto fail;
+                if (r > 0)
+                        return 0; /* All is good. */
+
+                /* Otherwise, undo the attachment and use the old APIs */
+                (void) ioctl(fd, LOOP_CLR_FD);
+        }
+
+        if (ioctl(fd, LOOP_SET_FD, c->fd) < 0)
+                return -errno;
+
+        if (ioctl(fd, LOOP_SET_STATUS64, &c->info) < 0) {
+                r = -errno;
+                goto fail;
+        }
+
+        return 0;
+
+fail:
+        (void) ioctl(fd, LOOP_CLR_FD);
+        return r;
+}
+
+int loop_device_make(
                 int fd,
                 int open_flags,
                 uint64_t offset,
@@ -24,10 +86,8 @@ int loop_device_make_full(
                 uint32_t loop_flags,
                 LoopDevice **ret) {
 
-        _cleanup_close_ int control = -1, loop = -1;
         _cleanup_free_ char *loopdev = NULL;
-        unsigned n_attempts = 0;
-        struct loop_info64 info;
+        struct loop_config config;
         LoopDevice *d = NULL;
         struct stat st;
         int nr = -1, r;
@@ -40,16 +100,21 @@ int loop_device_make_full(
                 return -errno;
 
         if (S_ISBLK(st.st_mode)) {
-                if (ioctl(loop, LOOP_GET_STATUS64, &info) >= 0) {
+                if (ioctl(fd, LOOP_GET_STATUS64, &config.info) >= 0) {
                         /* Oh! This is a loopback device? That's interesting! */
-                        nr = info.lo_number;
+
+#if HAVE_VALGRIND_MEMCHECK_H
+                        /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
+                        VALGRIND_MAKE_MEM_DEFINED(&config.info, sizeof(config.info));
+#endif
+                        nr = config.info.lo_number;
 
                         if (asprintf(&loopdev, "/dev/loop%i", nr) < 0)
                                 return -ENOMEM;
                 }
 
                 if (offset == 0 && IN_SET(size, 0, UINT64_MAX)) {
-                        int copy;
+                        _cleanup_close_ int copy = -1;
 
                         /* If this is already a block device, store a copy of the fd as it is */
 
@@ -60,9 +125,8 @@ int loop_device_make_full(
                         d = new(LoopDevice, 1);
                         if (!d)
                                 return -ENOMEM;
-
                         *d = (LoopDevice) {
-                                .fd = copy,
+                                .fd = TAKE_FD(copy),
                                 .nr = nr,
                                 .node = TAKE_PTR(loopdev),
                                 .relinquished = true, /* It's not allocated by us, don't destroy it when this object is freed */
@@ -77,13 +141,28 @@ int loop_device_make_full(
                         return r;
         }
 
+        _cleanup_close_ int control = -1;
+        _cleanup_(cleanup_clear_loop_close) int loop_with_fd = -1;
+
         control = open("/dev/loop-control", O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK);
         if (control < 0)
                 return -errno;
 
+        config = (struct loop_config) {
+                .fd = fd,
+                .info = {
+                        /* Use the specified flags, but configure the read-only flag from the open flags, and force autoclear */
+                        .lo_flags = (loop_flags & ~LO_FLAGS_READ_ONLY) | ((loop_flags & O_ACCMODE) == O_RDONLY ? LO_FLAGS_READ_ONLY : 0) | LO_FLAGS_AUTOCLEAR,
+                        .lo_offset = offset,
+                        .lo_sizelimit = size == UINT64_MAX ? 0 : size,
+                },
+        };
+
         /* Loop around LOOP_CTL_GET_FREE, since at the moment we attempt to open the returned device it might
          * be gone already, taken by somebody else racing against us. */
-        for (;;) {
+        for (unsigned n_attempts = 0;;) {
+                _cleanup_close_ int loop = -1;
+
                 nr = ioctl(control, LOOP_CTL_GET_FREE);
                 if (nr < 0)
                         return -errno;
@@ -92,69 +171,68 @@ int loop_device_make_full(
                         return -ENOMEM;
 
                 loop = open(loopdev, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|open_flags);
-                if (loop < 0)
-                        return -errno;
-                if (ioctl(loop, LOOP_SET_FD, fd) < 0) {
-                        if (errno != EBUSY)
+                if (loop < 0) {
+                        /* Somebody might've gotten the same number from the kernel, used the device,
+                         * and called LOOP_CTL_REMOVE on it. Let's retry with a new number. */
+                        if (errno != ENOENT)
                                 return -errno;
+                } else {
+                        r = loop_configure(loop, &config);
+                        if (r >= 0) {
+                                loop_with_fd = TAKE_FD(loop);
+                                break;
+                        }
+                        if (r != -EBUSY)
+                                return r;
+                }
 
-                        if (++n_attempts >= 64) /* Give up eventually */
-                                return -EBUSY;
-                } else
-                        break;
+                if (++n_attempts >= 64) /* Give up eventually */
+                        return -EBUSY;
 
                 loopdev = mfree(loopdev);
-                loop = safe_close(loop);
-        }
-
-        info = (struct loop_info64) {
-                /* Use the specified flags, but configure the read-only flag from the open flags, and force autoclear */
-                .lo_flags = (loop_flags & ~LO_FLAGS_READ_ONLY) | ((loop_flags & O_ACCMODE) == O_RDONLY ? LO_FLAGS_READ_ONLY : 0) | LO_FLAGS_AUTOCLEAR,
-                .lo_offset = offset,
-                .lo_sizelimit = size == UINT64_MAX ? 0 : size,
-        };
-
-        if (ioctl(loop, LOOP_SET_STATUS64, &info) < 0) {
-                r = -errno;
-                goto fail;
         }
 
         d = new(LoopDevice, 1);
-        if (!d) {
-                r = -ENOMEM;
-                goto fail;
-        }
-
+        if (!d)
+                return -ENOMEM;
         *d = (LoopDevice) {
-                .fd = TAKE_FD(loop),
+                .fd = TAKE_FD(loop_with_fd),
                 .node = TAKE_PTR(loopdev),
                 .nr = nr,
         };
 
         *ret = d;
-        return d->fd;
-
-fail:
-        if (fd >= 0)
-                (void) ioctl(fd, LOOP_CLR_FD);
-        if (d && d->fd >= 0)
-                (void) ioctl(d->fd, LOOP_CLR_FD);
-
-        return r;
+        return 0;
 }
 
 int loop_device_make_by_path(const char *path, int open_flags, uint32_t loop_flags, LoopDevice **ret) {
         _cleanup_close_ int fd = -1;
+        int r;
 
         assert(path);
         assert(ret);
-        assert(IN_SET(open_flags, O_RDWR, O_RDONLY));
+        assert(open_flags < 0 || IN_SET(open_flags, O_RDWR, O_RDONLY));
 
-        fd = open(path, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|open_flags);
-        if (fd < 0)
-                return -errno;
+        /* Passing < 0 as open_flags here means we'll try to open the device writable if we can, retrying
+         * read-only if we cannot. */
 
-        return loop_device_make(fd, open_flags, loop_flags, ret);
+        fd = open(path, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|(open_flags >= 0 ? open_flags : O_RDWR));
+        if (fd < 0) {
+                r = -errno;
+
+                /* Retry read-only? */
+                if (open_flags >= 0 || !(ERRNO_IS_PRIVILEGE(r) || r == -EROFS))
+                        return r;
+
+                fd = open(path, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|O_RDONLY);
+                if (fd < 0)
+                        return r; /* Propagate original error */
+
+                open_flags = O_RDONLY;
+        } else if (open_flags < 0)
+                open_flags = O_RDWR;
+
+        return loop_device_make(fd, open_flags, 0, 0, loop_flags, ret);
 }
 
 LoopDevice* loop_device_unref(LoopDevice *d) {
@@ -162,6 +240,9 @@ LoopDevice* loop_device_unref(LoopDevice *d) {
                 return NULL;
 
         if (d->fd >= 0) {
+                /* Implicitly sync the device, since otherwise in-flight blocks might not get written */
+                if (fsync(d->fd) < 0)
+                        log_debug_errno(errno, "Failed to sync loop block device, ignoring: %m");
 
                 if (d->nr >= 0 && !d->relinquished) {
                         if (ioctl(d->fd, LOOP_CLR_FD) < 0)
@@ -177,11 +258,19 @@ LoopDevice* loop_device_unref(LoopDevice *d) {
 
                 control = open("/dev/loop-control", O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK);
                 if (control < 0)
-                        log_debug_errno(errno, "Failed to open loop control device: %m");
-                else {
-                        if (ioctl(control, LOOP_CTL_REMOVE, d->nr) < 0)
-                                log_debug_errno(errno, "Failed to remove loop device: %m");
-                }
+                        log_warning_errno(errno,
+                                          "Failed to open loop control device, cannot remove loop device %s: %m",
+                                          strna(d->node));
+                else
+                        for (unsigned n_attempts = 0;;) {
+                                if (ioctl(control, LOOP_CTL_REMOVE, d->nr) >= 0)
+                                        break;
+                                if (errno != EBUSY || ++n_attempts >= 64) {
+                                        log_warning_errno(errno, "Failed to remove device %s: %m", strna(d->node));
+                                        break;
+                                }
+                                (void) usleep(50 * USEC_PER_MSEC);
+                        }
         }
 
         free(d->node);
@@ -217,9 +306,13 @@ int loop_device_open(const char *loop_path, int open_flags, LoopDevice **ret) {
         if (!S_ISBLK(st.st_mode))
                 return -ENOTBLK;
 
-        if (ioctl(loop_fd, LOOP_GET_STATUS64, &info) >= 0)
+        if (ioctl(loop_fd, LOOP_GET_STATUS64, &info) >= 0) {
+#if HAVE_VALGRIND_MEMCHECK_H
+                /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
+                VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info));
+#endif
                 nr = info.lo_number;
-        else
+        else
                 nr = -1;
 
         p = strdup(loop_path);
@@ -347,6 +440,11 @@ int loop_device_refresh_size(LoopDevice *d, uint64_t offset, uint64_t size) {
         if (ioctl(d->fd, LOOP_GET_STATUS64, &info) < 0)
                 return -errno;
 
+#if HAVE_VALGRIND_MEMCHECK_H
+        /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
+        VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info));
+#endif
+
         if (size == UINT64_MAX && offset == UINT64_MAX)
                 return 0;
         if (info.lo_sizelimit == size && info.lo_offset == offset)