On linux, write tests are failing because data written to fuse isn't showing
up in tahoe by the time it's checked. it's not clear where this is originating,
since the fuse implementation [should be] waiting for completion of tahoe
operations before returning from its calls. This adds an option to control the
duration of a pause between the fuse write and the check of tahoe, which is by
default set to 2s on linux, which - somewhat inexplicably - seems to 'fix' the
problem, in as far as it allows tests to complete.
previously, upon opening a file for reading, the open() call would block
while the entire file was retrieved from tahoe into the cache directory.
This change adds a DownloaderWithReadQueue class, and associated plumbing,
such that an open() will return promptly with the download initiated 'in
the background'. Subsequent read() operations will block until enough
data has been downloaded to satisfy that request. This provides a behaviour
similar to streaming, i.e. the client application will be able to read
data from the fuse interface while the remainder of the file is still being
downloaded.
this test uploads a test file to tahoe, and then reads the file from fuse,
but reads the blocks of the file in a random order; this is designed to
exercise the asynchronous download feature of blackmatch - where the file
is downloaded from tahoe asynchronously, and rather than blocking open()
for the entirety of the download, instead individual read() calls are
blocked until enough of the file has been downloaded to satisfy them
the code had a 'fullcleanup' flag internally which controlled whether
working directories were cleaned up. this promotes that to a command
line option (negated) '--no-cleanup' defaulting to False, i.e. do cleanup
this avoids dumping the repr of 1Mb of random data to stdout in the event
of a test failure, but rather just dumps the start/end of the errant strings
if the amount of data is > 200 chars repr'd
the repr_flags debug/logging function had a list of fields from the os
module that might be passed into an open() call, but it included at
least one which was available on the mac but not on linux. symmetrically
linux has numerous flags which are not present on the mac. the repr_flags
function is now tolerant of flags not being present, and has an expanded
list of flags
since the current tests assume that the implementation responds to changes made
to tahoe after mount, and impl_b prefetches and cached directory data, impl_b
fails the current 'read' test suite.
rather than reflect that problem in the overall failure of the runtests exit
code, this adds a 'todo' flag to the implementations table, and sets the todo
flag for impl_b. Thus errors will therein be reported in output, but not cause
a failing exit code.
previously the runtests suite removed the webport file created by
tahoe create-client in all but the first node. now that the node config
is in tahoe.cfg by default this file might not exist.
simplejson v2 returns strings as either unicode or str, depending upon its
mood. thus the interpretation of the node's json repr of a directory, and
the serialisation of strings in the json based rpc both exploded when built
against simplejson v2. this makes both of these places liberal in their
acceptance of either str or unicode.
humphf. my build runs the fuse stuff fine, but the build from the buildslave
doesn't seem to start up properly. hopefully this will elicit some useful info
adds a couple of functions to unpack 'mode' and 'flags' for open() calls, to
facilitate debugging.
adds a fix to ensure that all tmp files created for writing are opened with
permissions 0600 - one problem I had with testing with the Finder was that
files were being opened write only (0200) and were then failing to upload
to tahoe due to internal permission denied errors.
there remain a variety of problems with finder access which I'm unable to
comprehend at this time. sometimes copies to tahoe will work fine, sometimes
they yield "the finder cannot complete the operation because some data ...
could not be read or written. (Error code -36)" sometimes "You may need to
enter the name and password for an administrator on this computer to change
the item" sometimes "The operation cannot be completed because an item with
the name ... already exists." and sometimes "The operation cannot be completed
because the item ... is locked." What seems to be absent is rhyme or reason.
unix operations (cp, mv) work fine, rsync works fine.
the daemonize() function imported from twisted was causing problems when
run from a frozen (py2app) build. I simply copied the daemonize function
into this file, and that fixes the problem.
also removed a couple of lines of debugging spam that slipped through.
though it seemed to work before the 'fstype' passed to fuse of 'allmydata' was
today throwing errors that len(fstype) must be at most 7.
fixed a typo in changes to 'mount_filesystem()' args
bumped the delay between mounting a filesystem and 'open'ing it in Finder to
4s, as it seems to take a little longer to mount now the client and server
fuse processes need to coordinate.
This implements a client/server split for blackmatch, where the client
implements the fuse_main bindings and a simple blocking rpc client mechanism.
The server implements the other half of that rpc mechanism, and contains all
the actual logic for interpreting fuse requests in the context of the on disk
cache and requests to the tahoe node. The server is based on a twisted reactor.
The rpc mechanism implements a simple method dispatch including marshalling,
using json, of basic inert data types, in a flat namespace (no objects).
The client side is written in a blocking idiom, to interface with the threading
model used by the fuse_main bindings, whereas the server side is written for a
twisted reactor-based environment, intended to facilitate implementing more
sophisticated logic in that paradigm. The two communicate over a unix domain
socket, allocated within the nodedir.
Command line usage is unchanged; the server is launched automatically by the
client. The server daemonizes itself, to avoid preventing the original parent
process (e.g. 'runtests') from waiting upon the server exiting.
The client keeps open a 'keepalive' connection to the server; upon loss thereof
the server will exit. This addresses the fact that the python-fuse bindings
provide no notification of exit of the client process upon unmount.
The client thus provides a relatively thin 'shim' proxying requests from the
fuse_main bindings across the rpc to the server process, which handles the
logic behind each request.
For the time being, a '--no-split' option is provided to surpress the splitting
into client/server, yielding the prior behaviour. Once the server logic gets
more complex and more entrenched in a twisted idiom, this might be removed.
The 'runtests' test harness currently tests both modes, as 'impl_c' and
'impl_c_no_split'
the previous revision of blackmatch used a file_class to delegate all fuse
api operations on files to a specific per-file class, which is an option
given by the python-fuse bindings.
this is a pre-cursor to the 'split' client/server version, which uses a
simple, moreover flat, rpc mechanism to broker access to methods.
this tests opening a file for update, overwriting a small part of it, and
ensuring that the end result constitutes an overwrite of the original file.
This tests, e.g. the implementation doesn' open a 'fresh' file but does in
fact initialise the file to be uploaded with the contents of any extant
file before applying updates
changed the --tests option to be --suites, as it takes a prefix, e.g. 'read'
'write' (or 'all', the default) and runs those suites which are applicable to
each implementation being tested.
added a --tests option, which takes a list of tests, e.g. 'read_file_contents'
'write_overlapping_large_writes' and runs all tests specified without regard
to whether the implementation(s) under test are declared to support them.
this is basically to allow a specific test or two to be run, saving time
during development and debugging by not running the entire suite
this writes the test file in a randomised order, with randomly sized writes.
also for each 'slice' of the file written, a randomly chosen overlapping
write is also made to the file. this ensures that the file will be written
in its entirety in a thoroughly random order, with many overlapping writes.
using both small and large blocksizes for writes, write a 1Mb file to fuse
where every write overlaps another.
This serves a useful purpose - in manual testing of blackmatch some time ago
most operations e.g. bulk copies, worked fine, but using rsync caused data
corruption on most files. it turned out to be that rsync writes in 64K blocks,
but rather than making the last block short, the last block instead overlaps
the preceding (already written) block. This revealed a problem where cache
files were being opened 'append' rather than 'write' and hence the overlapping
write to the fuse layer caused the overlapping portion of the file to be
duplicated in cache, leading to oversized and corrupt files being uploaded.
unit tests to test writing contiguous blocks linearly through the file,
for a variety of block sizes; 'tiny_file' is an entire file fitting within
a single io block / write operation. 'linear_{small,large}_writes' test
a 1Mb file written with each write operation containing significantly less
or more, respecitvely, data than fuse will pass into the implementation as
a single operation (which on the mac at least is 64Kib)
this performs a very simple write through the fuse layer and confirms that
the file is stored correctly into the tahoe mesh. ('simple' in the sense
that the entire file body fits trivially in a single write() operation,
disk block etc)
similar to the --debug-wait option which causes the test harness to
pause at various stages of the process to facilitate debugging, this
option simplifies that debugging by automatically opening a web browser
to the root dir of that implementation's tests when tests are commenced.
in addition, if --web-open is specfied but --debug-wait is not, the
harness will still pause after running tests but before tearing down
the tahoe grid - this allows all tests to run to completion, but
provide a debugging hook to investigate the end state of the grid's
contents thereafter.
from my examination of the tahoe_fuse ('impl_a') code, it looks like
the intention is to cache the file contents in memory while it's open,
since it does in fact do that. however it looks like it also ignored
that cache entirely, and made an individual tahoe webapi GET request
for each and every read() operation regardless of the relative size of
the read block and the file in question.
this changes that to make read() use the data in memory rather than
fetch the data over again. if there's something more subtle going
on, please let me know.
a handful of code cleanup, renaming and refactoring. basically consolidating
'application logic' (mount/unmount fs) into the 'MacGuiApp' class (the wx.App)
and cleaning up various scoping things around that. renamed all references to
'app' to refer more clearly to the 'AppContainer' or to the guiapp.
globally renamed basedir -> nodedir
also made the guiapp keep a note of each filesystem it mounts, and unmount
them upon 'quit' so as to cleanup the user's environment before the tahoe node
vanishes from out underneath the orphaned tahoe fuse processes
this changes the 'open webroot' menu item to be a submenu listing all aliases
defined in ~/.tahoe. Note that the dock menu does not support submenus, so it
only offers a single 'open webroot' option for the default tahoe: alias.
I had trouble with this at first and concluded that the submenus didn't work,
and made it a distinct 'WebUI' menu in it's own right. on further inspection,
there are still problems but they seem to be something like once the dock menu
has been used, sometimes the app's main menubar menus will cease to function,
and this happens regardless of whether submenus or plain simple menus are used.
I have no idea what the peoblem is, but it's not submenu specific.
These constraints were originally intended to protect against attacks on the
storage server protocol layer which exhaust memory in the peer. However,
defending against that sort of DoS is hard -- probably it isn't completely
achieved -- and it costs development time to think about it, and it sometimes
imposes limits on legitimate users which we don't necessarily want to impose.
So, for now we forget about limiting the amount of RAM that a foolscap peer can
cause you to start using.
