We put some effort into testing the built copy rather than some
installed copy. On the other hand for people like packagers, testing
the installed copy is also of interest.
When NOTMUCH_TEST_INSTALLED is set to a nonempty value, tests do not
require a built notmuch tree or running configure.
Some of the tests marked as broken when running against installed
notmuch are probably fixable.
When running the test suite without building first, it is desirable to
have the tests consider these variables being undefined as equivalent
to the feature not being present, and in particular for the tests not
to generate errors.
This replaces two instances of Xapian::Query::MatchAll with the
equivalent but thread-safe alternative Xapian::Query(std::string()).
Xapian::Query::MatchAll maintains an internal pointer to a refcounted
Xapian::Internal::QueryTerm.
None of this is thread-safe but that wouldn't be an issue if
Xapian::Query::MatchAll wasn't static. Because it's static, the
refcounting goes awry when Notmuch is called from multiple threads.
This is actually documented by Xapian:
4715de3a9f/xapian-core/include/xapian/query.h (L65)
While static, Xapian::Query::MatchNothing is safe because it doesn't
maintain an internal object and as such, doesn't use references.
Two best-effort tests making use of TSan were added to showcase the
issue (I couldn't figure out a way to deterministically reproduce it
without making an unmaintainable mess).
First, when two databases are created in parallel, a query that uses
Xapian::Query::MatchAll is made (lib/query.cc), resulting in the
following backtrace on a segfault:
#0 0x00007ffff76822af in Xapian::Query::get_terms_begin (this=0x7fffe80137f0) at api/query.cc:141
#1 0x00007ffff7f933f5 in _notmuch_query_cache_terms (query=0x7fffe80137c0) at lib/query.cc:176
#2 0x00007ffff7f93784 in _notmuch_query_ensure_parsed_xapian (query=0x7fffe80137c0) at lib/query.cc:225
#3 0x00007ffff7f9381a in _notmuch_query_ensure_parsed (query=0x7fffe80137c0) at lib/query.cc:260
#4 0x00007ffff7f93bfe in _notmuch_query_search_documents (query=0x7fffe80137c0, type=0x7ffff7fa9b1e "mail", out=0x7ffff666da18) at lib/query.cc:361
#5 0x00007ffff7f93ba4 in notmuch_query_search_messages (query=0x7fffe80137c0, out=0x7ffff666da18) at lib/query.cc:349
#6 0x00007ffff7f83d98 in notmuch_database_upgrade (notmuch=0x7fffe8000bd0, progress_notify=0x0, closure=0x0) at lib/database.cc:934
#7 0x00007ffff7fa110f in notmuch_database_create_with_config (database_path=0x7ffff666dcb0 "/tmp/notmuch.MZ2AGr", config_path=0x7ffff7faab3c "", profile=0x0, database=0x0, status_string=0x7ffff666dc90) at lib/open.cc:754
#8 0x00007ffff7fa0d6f in notmuch_database_create_verbose (path=0x7ffff666dcb0 "/tmp/notmuch.MZ2AGr", database=0x0, status_string=0x7ffff666dc90) at lib/open.cc:653
#9 0x00007ffff7fa0ceb in notmuch_database_create (path=0x7ffff666dcb0 "/tmp/notmuch.MZ2AGr", database=0x0) at lib/open.cc:637
...
Second, some queries would make use of Xapian::Query::MatchAll
(lib/regexp-fields.cc), resulting in the following backtrace on a
segfault:
#0 0x00007f629828b690 in Xapian::Internal::QueryBranch::gather_terms (this=0x7f628800def0, void_terms=0x7f629726d5a0) at api/queryinternal.cc:1245
#1 0x00007f629828c260 in Xapian::Internal::QueryScaleWeight::gather_terms (this=0x7f628800df70, void_terms=0x7f629726d5a0) at api/queryinternal.cc:1434
#2 0x00007f629828b69f in Xapian::Internal::QueryBranch::gather_terms (this=0x7f628800dd90, void_terms=0x7f629726d5a0) at api/queryinternal.cc:1245
#3 0x00007f6298282571 in Xapian::Query::get_unique_terms_begin (this=0x7f628800dcd8) at api/query.cc:166
#4 0x00007f629841a59b in Xapian::Weight::Internal::accumulate_stats (this=0x7f628800dca0, subdb=..., rset=...) at weight/weightinternal.cc:86
#5 0x00007f62983c15ba in LocalSubMatch::prepare_match (this=0x7f628800df20, nowait=true, total_stats=...) at matcher/localsubmatch.cc:172
#6 0x00007f62983c8fcc in prepare_sub_matches (leaves=std::vector of length 1, capacity 1 = {...}, stats=...) at matcher/multimatch.cc:237
#7 0x00007f62983c98a3 in MultiMatch::MultiMatch (this=0x7f629726d9a0, db_=..., query_=..., qlen=3, omrset=0x0, collapse_max_=0, collapse_key_=4294967295, percent_cutoff_=0, weight_cutoff_=0, order_=Xapian::Enquire::ASCENDING, sort_key_=0, sort_by_=Xapian::Enquire::Internal::VAL, sort_value_forward_=true, time_limit_=0, stats=..., weight_=0x7f6288008d50, matchspies_=std::vector of length 0, capacity 0, have_sorter=false, have_mdecider=false) at matcher/multimatch.cc:353
#8 0x00007f629826fcba in Xapian::Enquire::Internal::get_mset (this=0x7f628800e0b0, first=0, maxitems=0, check_at_least=0, rset=0x0, mdecider=0x0) at api/omenquire.cc:569
#9 0x00007f629827181c in Xapian::Enquire::get_mset (this=0x7f629726db80, first=0, maxitems=0, check_at_least=0, rset=0x0, mdecider=0x0) at api/omenquire.cc:937
#10 0x00007f6298be529a in _notmuch_query_search_documents (query=0x7f6288009750, type=0x7f6298bfaafe "mail", out=0x7f629726dcc0) at lib/query.cc:447
#11 0x00007f6298be4ae8 in notmuch_query_search_messages (query=0x7f6288009750, out=0x7f629726dcc0) at lib/query.cc:349
...
Printing Xapian::Query::MatchAll->internal.px->_refs in these
circumstances can help quickly identifying this scenario.
This is motivated by some test frameworks (like Rust's Cargo) that
runs unit tests in parallel and would easily encounter this issue,
unless client code gates every call to Notmuch behind a lock.
This is what can be expected from the tests when they fail:
== stderr ==
+==================
+WARNING: ThreadSanitizer: data race (pid=207931)
+ Read of size 1 at 0x7b10000001a0 by thread T2:
+ #0 memcpy <null> (libtsan.so.2+0x62506)
+ #1 void std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >::_M_construct<char*>(char*, char*, std::forward_iterator_tag) [clone .isra.0] <null> (libxapian.so.30+0x872b3)
+
+ Previous write of size 8 at 0x7b10000001a0 by thread T1:
+ #0 operator new(unsigned long) <null> (libtsan.so.2+0x8ba83)
+ #1 Xapian::Query::Query(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, unsigned int, unsigned int) <null> (libxapian.so.30+0x855cd)
...
Commit [0] left the stemmer object accessible, but did not add
de-allocation code to notmuch_database_destroy. This commit corrects
that oversight.
Leak originally reported by Austin Ray [1].
[0]: 3202e0d1fe
[1]: id:20220105224538.m36lnjn7rf3ieonc@athena
