During regular VM shutdown, the VM should sync() anyway. (And
admin.vm.volume.Import does fdatasync(), which is also fine.) But let's
be extra careful.
There were (at least) five ways for the volume's nominal size and the
volume image file's actual size to desynchronize:
- loading a stale qubes.xml if a crash happened right after resizing the
image but before saving the updated qubes.xml (-> previously fixed)
- restarting a snap_on_start volume after resizing the volume or its
source volume (-> previously fixed)
- reverting to a differently sized revision
- importing a volume
- user tinkering with image files
Rather than trying to fix these one by one and hoping that there aren't
any others, override the volume size getter itself to always update from
the image file size. (If the getter is called though the storage API, it
takes the volume lock to avoid clobbering the nominal size when resize()
is running concurrently.)
And change the volume lock from an asyncio.Lock to a threading.Lock -
locking is now handled before coroutinization.
This will allow the coroutinized resize() and a new *not* coroutinized
size() getter from one of the next commits ("storage/reflink: preferably
get volume size from image size") to both run under the volume lock.
Successfully resize volumes without any currently existing image file,
e.g. cleanly stopped volatile volumes: Just update the nominal size in
this case.
Disk usage may change dynamically not only at VM start/stop. Refresh the
size cache before checking usage property, but no more than once every
30sec (refresh interval of disk space widget)
FixesQubesOS/qubes-issues#4888
Return meaningful value for kernels_dir if VM has no 'kernel' volume.
Right now it's mostly useful for tests, but could be also used for new
VM classes which doesn't have modules.img, but still use dom0-provided
kernel.
Pool setup/destroy may be a time consuming operation, allow them to be
asynchronous. Fortunately add_pool and remove_pool are used only through
Admin API, so the change does not require modification of other
components.
some-vm-root is a valid VM name, and in that case it's volume can be
named some-vm-root-private. Do not let it confuse revision listing,
check for unexpected '-' in volume revision number.
The proper solution would be to use different separator, that is not
allowed in VM names. But that would require migration code that is
undesired in the middle of stable release life cycle.
FixesQubesOS/qubes-issues#4680
LVM operations can take significant amount of time. This is especially
visible when stopping a VM (`vm.storage.stop()`) - in that time the
whole qubesd freeze for about 2 seconds.
Fix this by making all the ThinVolume methods a coroutines (where
supported). Each public coroutine is also wrapped with locking on
volume._lock to avoid concurrency-related problems.
This all also require changing internal helper functions to
coroutines. There are two functions that still needs to be called from
non-coroutine call sites:
- init_cache/reset_cache (initial cache fill, ThinPool.setup())
- qubes_lvm (ThinVolume.export()
So, those two functions need to live in two variants. Extract its common
code to separate functions to reduce code duplications.
FixesQubesOS/qubes-issues#4283
On some storage pools this operation can also be time consuming - for
example require creating temporary volume, and volume.create() already
can be a coroutine.
This is also requirement for making common code used by start()/create()
etc be a coroutine, otherwise neither of them can be and will block
other operations.
Related to QubesOS/qubes-issues#4283
_wait_and_reraise() is similar to asyncio.gather(), but it preserves the
current behavior of waiting for all futures and only _then_ reraising
the first exception (if there is any) in line.
Also switch Storage.create() and Storage.clone() to _wait_and_reraise().
Previously, they called asyncio.wait() and implicitly swallowed all
exceptions.
With that syntax, the default timestamp would have been from the time of
the function's definition (not invocation). But all callers are passing
an explicit timestamp anyway.
Convert create(), verify(), remove(), start(), stop(), revert(),
resize(), and import_volume() into coroutine methods, via a decorator
that runs them in the event loop's thread-based default executor.
This reduces UI hangs by unblocking the event loop, and can e.g. speed
up VM starts by starting multiple volumes in parallel.
Instead of raising a NotImplementedError, just return self like 'file'
and lvm_thin. This is needed when Storage.clone() is modified in another
commit* to no longer swallow exceptions.
* "storage: factor out _wait_and_reraise(); fix clone/create"
Import volume data to a new _path_import (instead of _path_dirty) before
committing to _path_clean. In case the computer crashes while an import
operation is running, the partially written file should not be attached
to Xen on the next volume startup.
Use <name>-import.img as the filename like 'file' does, to be compatible
with qubes.tests.api_admin/TC_00_VMs/test_510_vm_volume_import.
When the AT_REPLACE flag for linkat() finally lands in the Linux kernel,
_replace_file() can be modified to use unnamed (O_TMPFILE) tempfiles.
Until then, make sure stale tempfiles from previous crashes can't hang
around for too long.
