| Commit message (Collapse) | Author | Age |
|---|
| |
|
|
|
|
|
|
| |
Similar to TEST_COMMAND_RETURNS_ERROR_OR, but allows any command or sequence
to cause the error, without direct access to the command message.
Refactor so that TEST_COMMAND_RETURNS_ERROR_OR calls TEST_PROTOCOL_ERROR_OR
to make the assertion on the last error seen by the protocol.
|
| | |
|
| |
|
|
|
| |
Useful to help debug problems such as invalid file headers when the source
of the stream might not be too obvious.
|
| |
|
|
| |
This is hopefully the cause of the segfaults reported by Brendon Baumgartner.
|
| | |
|
| | |
|
| |
|
|
|
| |
Makes the daemon responsive to polling and commands, even during a file
upload operation.
|
| | |
|
| |
|
|
|
| |
It's incompatible with having a logger that logs everything, regardless of
the global log level.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When entries are added to a directory by a command (BackupStoreContext),
and when entries are removed from a directory (by housekeeping), update the
parent directory's entry for us if our size has changed.
Make BackupStoreCheck check for, report and fix errors when directory entries
are directories and the size is wrong (as well as files).
Conflicts:
test/backupstore/testbackupstore.cpp
Fix directories loaded without size being set, leading to warnings later.
We can't check that the old size in the parent entry matched the old real size
of the directory, unless we set the old real size in the directory.
And we don't need to pass the old directory size to
HousekeepStoreAccount::DeleteFile, because we can get it from the directory
itself.
|
| |
|
|
|
|
|
|
| |
This is useful to return from a test if an assertion/check fails, instead of
throwing an exception.
Also add logging of the actual error code received (with name) to
TEST_COMMAND_RETURNS_ERROR(_OR).
|
| | |
|
| |
|
|
| |
Returns the number of errors found and fixed.
|
| |
|
|
| |
It was hardcoded to 0x01234567 due to a copy and paste error.
|
| |
|
|
|
|
|
| |
Passing the wrong object ID would just result in an assertion being thrown
(in debug builds) or silently doing the wrong thing (in release builds). I
can't see any useful use for this extra parameter, since we can get the
container ID from the directory anyway.
|
| |
|
|
|
| |
An incorrect comment led me to believe that this parameter was the new
directory's AttributesModTime, but actually it was its hash.
|
| |
|
|
|
|
|
|
|
| |
If the old refcount database was shorter than the new one, then we could end
up trying to access a nonexistent refcount, triggering an exception, which
prevented all comparison of the old and new refcount databases.
Also, from now on, a mismatch between old and new refcounts is treated as a
housekeeping error, which is detectable in tests.
|
| | |
|
| |
|
|
|
| |
Using methods instead of macros where possible makes debugging easier.
Especially long macros like this one.
|
| |
|
|
|
|
|
|
|
|
| |
Otherwise the protocol might be broken and can't be used any more, even if
we made an effort to return an Error reply instead of throwing an exception.
This used to not be a problem because an Error reply would terminate the
connection anyway, but it no longer does. So if the client also didn't
terminate, but tried to handle the exception and keep using the connection,
then it might find that its next command fails because the protocol is broken.
|
| | |
|
| |
|
|
| |
Just like adding a file does.
|
| |
|
|
|
| |
Allows creating a directory with a specified modtime (in the parent directory entry).
Tests to come shortly.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
We pass a std::auto_ptr<SocketStream> to every Protocol subclass when we
construct it, and it takes control of this object. This reduces the risk of:
* accidentally reusing the same SocketStream for multiple Protocols
(it happened to me in testbackupstore);
* holding onto a reference to the SocketStream;
* allowing a locally-scoped SocketStream to go out of scope and be released
while still being referenced by a live Protocol.
|
| |
|
|
|
|
|
| |
Reading extended attributes may fail with an error if the file was removed
(race condition) or in tests when generating random nonexistent directory
entries. Handle this more gracefully by not backing up the extended attributes
instead of dying completely.
|
| |
|
|
|
|
| |
Rename mrDaemon to mpHousekeeping and make it a pointer, so that it can
officially be NULL, and don't crash if it is, and we fail to get a lock
on the account.
|
| | |
|
| |
|
|
| |
Was introduced by commit bc8261347880b813af58c4b873e1847cca03778d.
|
| | |
|
| |
|
|
| |
Fix previous changes by waiting for forked bbstored process during termination.
|
| | |
|
| | |
|
| |
|
|
| |
Useful in tests.
|
| |
|
|
| |
Saves ~2 lines each time it's used. And we use it a lot in testbackupstore.
|
| |
|
|
|
| |
There could be many reasons for this, including an invalid object being
deleted during the previous check run. And it's definitely not serious.
|
| |
|
|
|
| |
Also pass the actual reference count of an object to RaidFileWrite instead
of assuming that it's zero, to avoid accidentally deleting a referenced object.
|
| |
|
|
| |
Adding and deleting files was not always accounted properly before.
|
| |
|
|
| |
It's used in several places, and about to be one more.
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
| |
Will hopefully prevent future confusion about the meaning of this info
variable.
|
| |
|
|
|
|
| |
We don't have a test for it yet, and won't until snapshots are implemented,
but it's a bad idea to merge patches to remove an old version when another
directory is still holding a reference to the patch.
|
| |
|
|
|
|
| |
Fix accounting for current, old and deleted files and blocks. Previously there
was confusion over the meaning of NumFiles, now it's defined to mean the number
of current files (neither old nor deleted), similar to BlocksInCurrentFiles.
|
| |
|
|
| |
that we're doing so.
|
| | |
|
| |
|
|
|
|
|
|
|
|
| |
Caused by removal of BackupStoreRefCountDatabase::CreateForRegeneration.
It's not safe to do this in housekeeping anyway, because it might not
complete the scan, leaving an incomplete database, that might persuade
BackupStoreContext that it's OK to delete an object that still really has
references to it.
If your refcount db disappears or is damaged, you need to run a check fix.
|
| |
|
|
|
|
|
|
|
|
|
| |
Use a temporary refcount db for check instead of an in-memory vector. This
avoid the memory usage problems created by using the vector on large accounts,
but may require us to improve the efficiency of the refcount database itself
to avoid large numbers of small I/O operations. That is very doable now that
we're using a class for it.
Fix some inconsistencies and mistakes in handling reference counts and
info counters during account checks (more to come).
|
| |
|
|
|
| |
Caused by premature switch from GetNumFiles to GetNumCurrentFiles.
|
| |
|
|
|
| |
Allows us to get information about the account that's not available
using the old AccountUsage command. Currently only used in tests.
|
