Enable VirtIO block to access hostOS /dev/ block devices

The user may not always have a disk image but might have a /dev/x block
device, such as a USB drive that they want to share with the guest OS.
So, allowing this type of virtio-blk source is intuitive. To support
this, ioctl is used to retrieve the actual size of the /dev/x block
device. This implementation supports both Apple and Linux platforms.

On Apple platforms, mmap() on block devices appears to be unsupported
with various flag combinations. To address this, a fallback mechanism is
added and used when mmap() fails, using malloc() along with pread(),
pwrite() and fsync() on the block device fd to emulate the behavior of
mmap().

Additionally, the initial fallback was incomplete, as it only allocated
heap memory without loading the block device's content into memory. This
commit resolves the issue by properly reading the device contents into
the allocated buffer.

Since there may be asynchronous exits, a new rv_fsync_device() function
is introduced to ensure the block device is properly synchronized during
such exits, and it can also be invoked by other asynchronous exit paths.

To fully support this fallback, disk_fd and disk_size are now stored in
the vblk state during its initialization.

Close #544

Author: ChinYikMing

src/devices/virtio-blk.c

@@ -5,14 +5,28 @@
 
 #include <assert.h>
 #include <fcntl.h>
+#include <libgen.h>
 #include <stdbool.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <sys/ioctl.h>
 #include <sys/mman.h>
 #include <sys/stat.h>
 #include <unistd.h>
 
+/*
+ * The /dev/ block devices cannot be embedded to the part of the wasm.
+ * Thus, no support access /dev/ block devices for wasm.
+ */
+#if !defined(__EMSCRIPTEN__)
+#if defined(__APPLE__)
+#include <sys/disk.h> /* DKIOCGETBLOCKCOUNT and DKIOCGETBLOCKSIZE */
+#else
+#include <linux/fs.h> /* BLKGETSIZE64 */
+#endif
+#endif
+
 #include "virtio.h"
 
 #define DISK_BLK_SIZE 512
@@ -97,12 +111,16 @@ static void virtio_blk_update_status(virtio_blk_state_t *vblk, uint32_t status)
     uint32_t device_features = vblk->device_features;
     uint32_t *ram = vblk->ram;
     uint32_t *disk = vblk->disk;
+    uint64_t disk_size = vblk->disk_size;
+    int disk_fd = vblk->disk_fd;
     void *priv = vblk->priv;
     uint32_t capacity = VBLK_PRIV(vblk)->capacity;
     memset(vblk, 0, sizeof(*vblk));
     vblk->device_features = device_features;
     vblk->ram = ram;
     vblk->disk = disk;
+    vblk->disk_size = disk_size;
+    vblk->disk_fd = disk_fd;
     vblk->priv = priv;
     VBLK_PRIV(vblk)->capacity = capacity;
 }
@@ -388,6 +406,9 @@ uint32_t *virtio_blk_init(virtio_blk_state_t *vblk,
         exit(EXIT_FAILURE);
     }
 
+    /* For mmap_fallback */
+    vblk->disk_fd = -1;
+
     /* Allocate memory for the private member */
     vblk->priv = &vblk_configs[vblk_dev_cnt++];
 
@@ -403,29 +424,81 @@ uint32_t *virtio_blk_init(virtio_blk_state_t *vblk,
     int disk_fd = open(disk_file, readonly ? O_RDONLY : O_RDWR);
     if (disk_fd < 0) {
         rv_log_error("Could not open %s", disk_file);
-        exit(EXIT_FAILURE);
+        goto fail;
     }
 
-    /* Get the disk image size */
-    struct stat st;
-    fstat(disk_fd, &st);
-    VBLK_PRIV(vblk)->disk_size = st.st_size;
+    const char *disk_file_dirname = dirname(disk_file);
+    if (!disk_file_dirname) {
+        rv_log_error("Fail dirname disk_file: %s", disk_file);
+        goto disk_size_fail;
+    }
+    /* Get the disk size */
+    uint64_t disk_size;
+    if (strcmp(disk_file_dirname, "/dev") ==
+        0) { /* from /dev/, leverage ioctl */
+#if !defined(__EMSCRIPTEN__)
+#if defined(__APPLE__)
+        uint32_t block_size;
+        uint64_t block_count;
+        if (ioctl(disk_fd, DKIOCGETBLOCKCOUNT, &block_count) == -1) {
+            rv_log_error("DKIOCGETBLOCKCOUNT failed");
+            goto disk_size_fail;
+        }
+        if (ioctl(disk_fd, DKIOCGETBLOCKSIZE, &block_size) == -1) {
+            rv_log_error("DKIOCGETBLOCKSIZE failed");
+            goto disk_size_fail;
+        }
+        disk_size = block_count * block_size;
+#else /* Linux */
+        if (ioctl(disk_fd, BLKGETSIZE64, &disk_size) == -1) {
+            rv_log_error("BLKGETSIZE64 failed");
+            goto disk_size_fail;
+        }
+#endif
+#endif       /* !defined(__EMSCRIPTEN__) */
+    } else { /* other path, stat it as normal file */
+        struct stat st;
+        if (fstat(disk_fd, &st) == -1) {
+            rv_log_error("fstat failed");
+            goto disk_size_fail;
+        }
+        disk_size = st.st_size;
+    }
+    VBLK_PRIV(vblk)->disk_size = disk_size;
 
     /* Set up the disk memory */
     uint32_t *disk_mem;
 #if HAVE_MMAP
     disk_mem = mmap(NULL, VBLK_PRIV(vblk)->disk_size,
                     readonly ? PROT_READ : (PROT_READ | PROT_WRITE), MAP_SHARED,
                     disk_fd, 0);
-    if (disk_mem == MAP_FAILED)
-        goto err;
-#else
+    if (disk_mem == MAP_FAILED) {
+        rv_log_trace(
+            "Fallback to malloc block device due to mmap failed or the "
+            "operation is unsupported");
+        goto mmap_fallback;
+    }
+    /*
+     * disk_fd should be closed on exit after flushing heap data back to the
+     * device when using mmap_fallback.
+     */
+    close(disk_fd);
+    goto disk_mem_ok;
+#endif
+
+mmap_fallback:
     disk_mem = malloc(VBLK_PRIV(vblk)->disk_size);
     if (!disk_mem)
-        goto err;
-#endif
+        goto disk_mem_err;
+    vblk->disk_fd = disk_fd;
+    vblk->disk_size = disk_size;
+    if (pread(disk_fd, disk_mem, disk_size, 0) == -1) {
+        rv_log_error("pread block device failed");
+        goto disk_mem_err;
+    }
+
+disk_mem_ok:
     assert(!(((uintptr_t) disk_mem) & 0b11));
-    close(disk_fd);
 
     vblk->disk = disk_mem;
     VBLK_PRIV(vblk)->capacity =
@@ -436,9 +509,14 @@ uint32_t *virtio_blk_init(virtio_blk_state_t *vblk,
 
     return disk_mem;
 
-err:
+disk_mem_err:
     rv_log_error("Could not map disk %s", disk_file);
-    return NULL;
+
+disk_size_fail:
+    close(disk_fd);
+
+fail:
+    exit(EXIT_FAILURE);
 }
 
 virtio_blk_state_t *vblk_new()
@@ -450,10 +528,11 @@ virtio_blk_state_t *vblk_new()
 
 void vblk_delete(virtio_blk_state_t *vblk)
 {
+    if (vblk->disk_fd != -1)
+        free(vblk->disk);
 #if HAVE_MMAP
-    munmap(vblk->disk, VBLK_PRIV(vblk)->disk_size);
-#else
-    free(vblk->disk);
+    else
+        munmap(vblk->disk, VBLK_PRIV(vblk)->disk_size);
 #endif
     free(vblk);
 }

src/devices/virtio.h

@@ -103,6 +103,8 @@ typedef struct {
     /* supplied by environment */
     uint32_t *ram;
     uint32_t *disk;
+    uint64_t disk_size;
+    int disk_fd;
     /* implementation-specific */
     void *priv;
 } virtio_blk_state_t;

src/main.c

@@ -298,6 +298,11 @@ int main(int argc, char **args)
 
     /* finalize the RISC-V runtime */
     rv_delete(rv);
+    /*
+     * Other translation units cannot update the pointer, update it here
+     * to prevent multiple atexit()'s callback be called.
+     */
+    rv = NULL;
     rv_log_info("RISC-V emulator is destroyed");
 
 end:

src/riscv.c

@@ -376,7 +376,35 @@ static void rv_async_block_clear()
     return;
 #endif /* !RV32_HAS(JIT) */
 }
-#endif
+
+static void rv_fsync_device()
+{
+    if (!rv)
+        return;
+
+    /* mmap_fallback, need to write and sync the device */
+    vm_attr_t *attr = PRIV(rv);
+    if (attr->vblk && attr->vblk->disk_fd >= 3) {
+        if (attr->vblk->device_features & VIRTIO_BLK_F_RO) /* readonly */
+            goto end;
+
+        if (pwrite(attr->vblk->disk_fd, attr->vblk->disk, attr->vblk->disk_size,
+                   0) == -1) {
+            rv_log_error("pwrite block device failed");
+            return;
+        }
+
+        if (fsync(attr->vblk->disk_fd) == -1) {
+            rv_log_error("fsync block device failed");
+            return;
+        }
+
+    end:
+        close(attr->vblk->disk_fd);
+        rv_log_info("Sync block devices OK");
+    }
+}
+#endif /* RV32_HAS(SYSTEM) && !RV32_HAS(ELF_LOADER) */
 
 riscv_t *rv_create(riscv_user_t rv_attr)
 {
@@ -388,6 +416,8 @@ riscv_t *rv_create(riscv_user_t rv_attr)
 #if RV32_HAS(SYSTEM) && !RV32_HAS(ELF_LOADER)
     /* register cleaning callback for CTRL+a+x exit */
     atexit(rv_async_block_clear);
+    /* register device sync callback for CTRL+a+x exit */
+    atexit(rv_fsync_device);
 #endif
 
     /* copy over the attr */
@@ -705,6 +735,8 @@ void rv_delete(riscv_t *rv)
 #if RV32_HAS(SYSTEM) && !RV32_HAS(ELF_LOADER)
     u8250_delete(attr->uart);
     plic_delete(attr->plic);
+    /* sync device before cleaning up */
+    rv_fsync_device();
     vblk_delete(attr->vblk);
 #endif
     free(rv);