filelog: don't crash on invalid copy metadata (issue5748) "copy" and "copyrev" are both supposed to appear next to each other. However, a user report demonstrated a crash that indicates that something in the wild is producing "copy" without "copyrev" (probably `hg convert`). While we should definitely fix the source of the bad metadata, the bad code causing the crash is already in the wild and who knows how many repositories are impacted. So let's be more defensive when accessing the file revision metadata.

httppeer: detect redirect to URL without query string (issue5860) 197d10e157ce subtly changed the HTTP peer's handling of HTTP redirects. Before that changeset, we instantiated an HTTP peer instance and performed the capabilities lookup with that instance. The old code had the following relevant properties: 1) The HTTP request layer would automatically follow HTTP redirects. 2) An encountered HTTP redirect would update a peer instance variable pointing to the repo URL. 3) The peer would automagically perform a "capabilities" command request if a caller requested capabilities but capabilities were not yet defined. The first HTTP request issued by a peer is for ?cmd=capabilities. If the server responds with an HTTP redirect to a ?cmd=capabilities URL, the HTTP request layer automatically followed it, retrieved a valid capabilities response, and the peer's base URL was updated automatically so subsequent requests used the proper URL. In other words, things "just worked." In the case where the server redirected to a URL without the ?cmd=capabilities query string, the HTTP request layer would follow the redirect and likely encounter HTML. The peer's base URL would be updated and the unexpected Content-Type would raise a RepoError. We would catch RepoError and immediately call between() (testing the case for pre 0.9.1 servers not supporting the "capabilities" command). e.g. try: inst._fetchcaps() except error.RepoError: inst.between([(nullid, nullid)]) between() would eventually call into _callstream(). And _callstream() made a call to self.capable('httpheader'). capable() would call self.capabilities(), which would see that no capabilities were set (because HTML was returned for that request) and call the "capabilities" command to fetch capabilities. Because the base URL had been updated from the redirect, this 2nd "capabilities" command would succeed and the client would immediately call "between," which would also succeed. The legacy handshake succeeded. Only because "capabilities" was successfully executed as a side effect did the peer recognize that it was talking to a modern server. In other words, this all appeared to work accidentally. After 197d10e157ce, we stopped calling the "capabilities" command on the peer instance. Instead, we made the request via a low-level opener, detected the redirect as part of response handling code, and passed the redirected URL into the constructed peer instance. For cases where the redirected URL included the query string, this "just worked." But for cases where the redirected URL stripped the query string, we threw RepoError and because we removed the "between" handshake fallback, we fell through to the "is a static HTTP repo" check and performed an HTTP request for .hg/requires. While 197d10e157ce was marked as backwards incompatible, the only intended backwards incompatible behavior was not performing the "between" fallback. It was not realized that the "between" command had the side-effect of recovering from an errant redirect that dropped the query string. This commit restores the previous behavior and allows clients to handle a redirect that drops the query string. In the case where the request is redirected and the query string is dropped, we raise a special case of RepoError. We then catch this special exception in the handshake code and perform another "capabilities" request against the redirected URL. If that works, all is well. Otherwise, we fall back to the "is a static HTTP repo" check. The new code is arguably better than before 197d10e157ce, as it is explicit about the expected behavior and we avoid performing a "between" request, saving a server round trip. Differential Revision: https://phab.mercurial-scm.org/D3433

hgweb: prevent triggering dummy href="#" handler Follow up for the previous patch.

paper: add href="#" to links with click handlers This restores the styling that was accidentally removed by the previous change to these files. Differential Revision: https://phab.mercurial-scm.org/D3438

paper: don't register click handlers with inline javascript (issue5812) The use of inline href="javascript:" undermines CSP policies that don't allow inline javascript. This commit changes the registering of the diffstat and line wrapping toggle handlers to the the global DOMContentLoaded handler, thus eliminating all inline javascript from the paper template. Differential Revision: https://phab.mercurial-scm.org/D3437

hgweb: allow Content-Security-Policy header on 304 responses (issue5844) A side-effect of 98baf8dea553 was that the Content-Security-Policy header was set on all HTTP responses by default. This header wasn't in our list of allowed headers for HTTP 304 responses. This would trigger a ProgrammingError when a 304 response was issued via hgwebdir. This commit adds Content-Security-Policy to the allow list of headers for 304 responses so we no longer encounter the error. Differential Revision: https://phab.mercurial-scm.org/D3436

hgweb: discard Content-Type header for 304 responses (issue5844) A side-effect of 98baf8dea553 was that hgwebdir always sets a global default for the Content-Type header. HTTP 304 responses don't allow the Content-Type header. So a side-effect of this change was that HTTP 304 responses served via hgwebdir resulted in a ProgrammingError being raised. This commit teaches our 304 response issuing code to drop the Content-Type header. Differential Revision: https://phab.mercurial-scm.org/D3435

tests: add tests demonstrating ISE for HTTP 304 responses with hgwebdir There are two separate failures here. One for the Content-Type header. Another for the Content-Security-Policy header. Differential Revision: https://phab.mercurial-scm.org/D3434

hgweb: guard against empty Content-Length header Discussion in issue 5860 seems to indicate this can occur. Differential Revision: https://phab.mercurial-scm.org/D3432

test-push-http: do not clear pid file It's okay now, but we'll end up leaking daemon processes if we add some more.

debugcolor: fix crash by empty styles (issue5856)

tests: explicitly define compression engines for tests The zstd compression engine requires C extensions and isn't present in pure Python builds. The compression engine list leaks into the server capabilities string. Unless we're testing functionality specific to a compression format, the set of compression formats supported by a server doesn't matter much. So this commit explicitly defines the server's compression engines for some tests so behavior is consistent between pure and non-pure builds. Differential Revision: https://phab.mercurial-scm.org/D3431

tests: update no-zstd branch of test-treediscovery.t as in 330ada7e8ea5 This side of the test got overlooked. We should probably consider having a way to run some of our tests through a "no-zstd" case just like we run some things through a "no-obsmarkers" case, but that's not an appropriate thing for stable. Differential Revision: https://phab.mercurial-scm.org/D3430

tests: glob away content-length changes relating to missing zstd bindings This doesn't fix everything in these two tests around missing zstd: we still get some changes in the CBOR payload in ways that I think we probably shouldn't bother to glob around. Maybe we should just disable zstd support in some of these lower-level wireproto tests? Differential Revision: https://phab.mercurial-scm.org/D3429

revlog: make pure version of _partialmatch() support 40-byte hex nodeids Without this patch, test-histedit-arguments.t would fail when run with --pure. It turned out to be because the pure version of _partialmatch() does not support full 40-byte hex nodeids. When histedit's instructions include things like "pick tip", it resolves the "tip" revision early to a full nodeid (but plain hex nodeid prefixes are not resolved to full nodeids). Then the nodeid (full or not) is looked up using to a full nodeid later. This step is what fails in pure mode. It has been failing since my c4131138eadb (histedit: look up partial nodeid as partial nodeid, 2018-04-06). I haven't verified, but I suspect histedit instructions like "pick <full hex nodeid>" would have been failing before my commit too, though. The fix is trivial: change a "< 40" to "<= 40". Differential Revision: https://phab.mercurial-scm.org/D3428