storage: make storage servers declare oldest supported version == 1.0, and storage clients declare oldest supported version == 1.0
See comments in patch for intended semantics.
tests: don't use SignalMixin
It seems like we no longer need it, and it screws up something internal in
trial which causes trial's TestCase.mktemp() method to exhibit wrong behavior
(always using a certain test method name instead of using the current test
method name), and I wish to use TestCase.mktemp().
Of course, it is possible that the buildbot is about to tell me that we do
still require SignalMixin on some of our platforms...
setup: fix bug in Makefile -- ifeq, not ifneq -- so that now it sets poll reactor only if the user hasn't specified a REACTOR variable, instead of setting poll reactor only if the user has specified a REACTOR variable
setup: instead of setting --reactor=poll for trial in all cases (which fails on platforms that don't have poll reactor, such as Windows and some Mac OS X), just set --reactor=poll for linux2.
setup: pass --reactor=poll to trial unless REACTOR variable is set, in which case pass --reactor=$(REACTOR)
This hopefully works around the problem that Twisted v8.1.0 has a bug when used
with pyOpenSSL v0.7 which bug causes some unit tests to spuriously fail -- see
known_issues.txt r2788:
setup: require secure_connections from foolscap
This causes a problem on debian sid, since the pyOpenSSL v0.6 .deb doesn't come
with .egg-info, so setuptools will not know that it is already installed and
will try to install pyOpenSSL, and if it installs pyOpenSSL v0.7, then this
will trigger the bug in Twisted v8.1.0 when used with pyOpenSSL v0.7.
http://twistedmatrix.com/trac/ticket/3218
Now the comments in twisted #3218 suggest that it happens only with the select
reactor, so maybe using --reactor=poll will avoid it.
Brian Warner [Mon, 28 Jul 2008 19:44:21 +0000 (12:44 -0700)]
test/common.py: use pre-computed Tub certificates for the system-test mixin, to speed such tests up by maybe 15%. The goal is to encourage more full-grid tests.
immutable file download: make the ciphertext hash tree mandatory
This fixes #491 (URIs do not refer to unique files in Allmydata Tahoe).
Fortunately all of the versions of Tahoe currently in use are already producing
this ciphertext hash tree when uploading, so there is no
backwards-compatibility problem with having the downloader require it to be
present.
Brian Warner [Sat, 19 Jul 2008 01:06:50 +0000 (18:06 -0700)]
web: for GET save=true, don't interpret the filename= arg with any character set, just copy the bytes back into the Content-Disposition header. This seems to make it maximally compatible with Firefox and IE7
Brian Warner [Wed, 16 Jul 2008 00:23:25 +0000 (17:23 -0700)]
overhaul checker invocation
Removed the Checker service, removed checker results storage (both in-memory
and the tiny stub of sqlite-based storage). Added ICheckable, all
check/verify is now done by calling the check() method on filenodes and
dirnodes (immutable files, literal files, mutable files, and directory
instances).
Checker results are returned in a Results instance, with an html() method for
display. Checker results have been temporarily removed from the wui directory
listing until we make some other fixes.
Also fixed client.create_node_from_uri() to create LiteralFileNodes properly,
since they have different checking behavior. Previously we were creating full
FileNodes with LIT uris inside, which were downloadable but not checkable.
Brian Warner [Tue, 15 Jul 2008 22:23:41 +0000 (15:23 -0700)]
debian/changelog: remove all versions but 0.0.1, so debchange can always override it. The 'deb-X' Makefile targets should not be used; run 'make deb-X-head' instead
misc/simulate_load.py: use a more realistic distribution of file sizes, and automate the formatting of the ascii art so that it fits in 70 columns and the number of files uploaded is properly marked
setup: bundle darcsver-1.1.5 into misc/dependencies
Once we implement #249 (move bundled dependencies out of revision control history and make them optional) then we can stop doing this. Until then we try not to do it unless we really need to...
Brian Warner [Sat, 21 Jun 2008 04:55:25 +0000 (21:55 -0700)]
docs/accounts-pubkey.txt: when sending FURLs (signed or otherwise), also send
the object that it references, to bypass the Foolscap getReferenceByName
roundtrip that would otherwise be required. Recipients must still use the
FURL for security, but pre-populating the foolscap table is a good speed-up.