Non-resident COW files #273
Conversation
Cow files can now exist off of the disk it is cow-ing. This comes with a few limitations: 1. You cannot unmount a device with an active cow file on it. You would have to destroy or unmount the device being snapshotted by that cow file. DattoBD will hold the file handle to the cow file and will prevent the unmount. 3. Don't create loops with the snapshotted devices. If device A's cow file exists on device B, and device B's cow file is on device A, then you wouldn't be able to unmount either device or shutdown the machine. You'd have to destroy either snapshot to be able to unmount. This commit also adds some fields to the /proc/datto-info device. The `resident` field is a 0/1 false/true value that answers the question of whether the cow file is on the device it is tracking. The `full_cow_path` field represents the full path to the cow device, rather than `cow_file`'s relative-to-mountpoint path. You are encouraged to use the new `full_cow_path` field; `cow_file` will be kept, but deprecated, for backwards compatibility.
FUSE mounts cannot have resident cow files. FUSE mounts the device first, then sets up its ioctl callbacks (open, close, read, write). This becomes a problem on a remount, since we'd try to reopen the cow file during the mount syscall, which will hang since open hasn't been set by FUSE yet.
src/snap_device.h
Outdated
| @@ -31,7 +35,8 @@ struct snap_device { | |||
| struct block_device *sd_base_dev; // device being snapshot | |||
| char *sd_bdev_path; // base device file path | |||
| struct cow_manager *sd_cow; // cow manager | |||
| char *sd_cow_path; // cow file path | |||
| char *sd_cow_path; // cow file path (for resident cow files) | |||
There was a problem hiding this comment.
Should this be marked as deprecated in a comment?
src/tracer.c
Outdated
| @@ -1608,13 +1624,13 @@ void __tracer_dormant_to_active(struct snap_device *dev, | |||
| set_bit(ACTIVE, &dev->sd_state); | |||
| clear_bit(UNVERIFIED, &dev->sd_state); | |||
|
|
|||
| kfree(cow_path); | |||
| kfree(resident_cow_path); | |||
There was a problem hiding this comment.
Can this be freed immediately after __tracer_setup_cow_reopen() and then the extra test/free in the error: case could be removed?
| if (0 == strncmp(bdev->bd_super->s_type->name, fuseblk_name, | ||
| FUSEBLK_MNT_LEN)) { | ||
| ret = -EDEADLOCK; | ||
| LOG_ERROR(ret, "Cow file cannot be on same volume for" |
There was a problem hiding this comment.
Feel free to ignore but as a user of this driver would it be helpful to know slightly more about why the cow file cannot be on the same block device without having to consult the code? The comment above is great. I'm just thinking about having a bit more information so that I could at least google an answer to why dattobd is telling me I cannot do what I want to do. We should probably consider adding enough information in our log messages so that steps to correct are at least Googleable. As I said, feel free to ignore, I'm just thinking about the experience of trying to figure out why something, that seems on the surface to be sane, does not work.
|
Hi, [Discuss] |
|
@nickchen-cpu You're correct in that we wouldn't need to worry about the deleted data on a separate disk, but we'd run into a similar problem as my second limitation in the first comment. The issue would show up if the device you put the cow file on was snapshotted afterwards. It depends if we were to disallow this behavior entirely and tbh I don't think we would, so we're back the same unused block problem as before. I've found some problems in this MRs implementation at the moment and am still working to resolve them. Going forward, I will keep your feature in mind. |
|
Does seem work for me. Would love to see this merged. |
Cow files can now exist off of the disk it is cow-ing. This comes with a
few limitations:
have to destroy or unmount the device being snapshotted by that cow
file. DattoBD will hold the file handle to the cow file and will
prevent the unmount.
file exists on device B, and device B's cow file is on device A, then
you wouldn't be able to unmount either device or shutdown the
machine. You'd have to destroy either snapshot to be able to unmount.
This commit also adds some fields to the /proc/datto-info device. The
residentfield is a 0/1 false/true value that answers the question ofwhether the cow file is on the device it is tracking. The
full_cow_pathfield represents the full path to the cow device, ratherthan
cow_file's relative-to-mountpoint path. You are encouraged to usethe new
full_cow_pathfield;cow_filewill be kept, but deprecated,for backwards compatibility.
Both of these limitations intend to be addressed in the future.