Attached are debdiffs for the SRU:

- boost1.90, resolute: boost1.90_1.90.0-6ubuntu1.1.debdiff
- boost1.90, stonking: boost1.90_1.90.0-6.1ubuntu1.1.debdiff
- boost1.88, resolute: boost1.88_1.88.0-1.4ubuntu4.1.debdiff

All three add the upstream fix as d/p/context-manage-exception-state-
noinline.patch, at the end of the series, with a changelog entry, and
run update-maintainer. All patches apply cleanly with quilt, and the
unpacked sources contain the patched header.

boost1.88 is affected as well: manage_exception_state was introduced in
Boost 1.88 without BOOST_NOINLINE. 1.83 and 1.87 do not have it, and
1.92 has the fix. In resolute the only user of libboost-context1.88.0
outside Boost itself is icinga2 2.15.1-1, which is built against
boost1.88. The attached reproducer, built purely against the boost1.88
sources, fails the same way: 2.7M of 11M catch blocks without a current
exception, and "terminate called without an active exception" on
rethrow. With the patch it reports 0. A small objdump-based check also
finds 27 affected fiber switches in icinga2-bin 2.15.1-1 (amd64); I can
attach it if that helps. I have added a boost1.88 task for this.

The development release is already fixed: tenacious uses Boost 1.92
(boost-defaults 1.92.0.1ubuntu2), and nothing in tenacious depends on
libboost-context1.90.0 any more apart from boost1.90's own packages. So
I don't think an upload of boost1.90 to tenacious is needed. If you want
one anyway, I can attach that debdiff too (1.90.0-6.1ubuntu2).

Since the bug is in a header, the affected packages need no-change
rebuilds after the fix:

- resolute: ceph (radosgw, librgw2, ceph-common), pdns-recursor, nix, simgrid, 
graph-tool, wesnoth-1.18 and yuzu, all on boost1.90, and icinga2, on boost1.88
- stonking: the same set on boost1.90, including icinga2

Setting the status back to New.


** Patch added: "boost1.90_1.90.0-6.1ubuntu1.1.debdiff"
   
https://bugs.launchpad.net/ubuntu/+source/boost1.90/+bug/2173538/+attachment/6017072/+files/boost1.90_1.90.0-6.1ubuntu1.1.debdiff

** Changed in: boost1.90 (Ubuntu)
       Status: Incomplete => New

** Description changed:

  [ Impact ]
  
- Since Boost 1.90, boost::context fibers can overwrite the libstdc++
- exception state (__cxa_get_globals) of another thread when a fiber
- resumes on a different thread. manage_exception_state in
- boost/context/fiber_fcontext.hpp lost its BOOST_NOINLINE, so gcc caches
- the TLS address across the fiber switch.
+ Boost.Context 1.88 to 1.91 save and restore the libstdc++ exception
+ state (__cxa_eh_globals) around each fiber switch, in
+ manage_exception_state in boost/context/fiber_fcontext.hpp.
+ __cxa_get_globals() is declared __attribute__((const)). Once the
+ constructor and destructor are inlined, gcc computes the TLS address
+ once before the switch and reuses it afterwards.
  
- User-visible result: `throw;` inside a catch block in a
- boost::asio::spawn coroutine aborts with "terminate called without an
- active exception". With Icinga 2 on 26.04, about 1 in 240 cancelled
- connections crashed an I/O thread; with the patched header, 0 in 12,675.
- radosgw (ceph) is affected in the archive: it runs asio::spawn
- coroutines on strands over a pool of rgw_thread_pool_size (default 512)
- threads, and its amd64 binaries contain the broken code (see Other Info).
+ If the fiber is resumed on a different thread, it writes the saved state
+ into the thread it came from, while that thread is running other code. A
+ thread inside a catch block then loses its current exception: throw;
+ calls std::terminate() ("terminate called without an active exception")
+ and std::current_exception() returns null. This is easy to hit with
+ boost::asio::spawn on an io_context run by several threads.
  
- This is a regression from noble: Boost 1.83 has no
- manage_exception_state, so 24.04 is not affected.
+ Upstream issue: https://github.com/boostorg/context/issues/323. It is
+ fixed in Boost 1.92 by commit 0921b9fd5c
+ (https://github.com/boostorg/context/pull/331), which re-adds
+ BOOST_NOINLINE to both functions.
  
- The fix is the upstream commit 0921b9fd5c
- (https://github.com/boostorg/context/pull/331), in Boost 1.92. It
- restores BOOST_NOINLINE. The attached
- context-manage-exception-state-noinline.patch is that commit as a quilt
- patch for the package layout and applies cleanly on 1.90.0-6ubuntu1.
+ Affected source packages:
+ 
+ - resolute: boost1.90 (1.90.0-6ubuntu1) and boost1.88 (1.88.0-1.4ubuntu4)
+ - stonking: boost1.90 (1.90.0-6.1ubuntu1)
+ 
+ The development release is not affected: tenacious defaults to Boost
+ 1.92, and nothing there uses libboost-context1.90.0 apart from
+ boost1.90's own packages.
+ 
+ Real-world impact: Icinga 2 on 26.04 (the packages.icinga.com build,
+ against boost1.90). On a production monitoring master, 17 of 4,026
+ cancelled outgoing connections crashed an I/O thread within 21.6 hours.
+ After rebuilding the same sources against the patched header there were
+ 0 of 12,675. Ubuntu's own icinga2 is affected too: 2.15.1-1 in resolute
+ against boost1.88, and 2.16.5-1 in stonking against boost1.90.
+ 
+ The bug is in a header, so it is compiled into the binaries that use it.
+ Those need no-change rebuilds after the update to pick up the fix:
+ 
+ - resolute: ceph (radosgw, librgw2, ceph-common), pdns-recursor, nix, 
simgrid, graph-tool, wesnoth-1.18 and yuzu, all on boost1.90, and icinga2, on 
boost1.88
+ - stonking: the same set on boost1.90, including icinga2
  
  [ Test Plan ]
  
- 1. On resolute (amd64 and arm64), install libboost-context1.90-dev and
-    build the attached reproducer:
-      g++ -O2 -std=c++17 -pthread -o repro repro.cpp -lboost_context
- 2. Run: ./repro rethrow 10
-    - Without the fix: aborts within milliseconds with
-      "terminate called without an active exception".
-    - With the fix: runs for 10 seconds and exits 0.
- 3. Optionally run: ./repro check 10
-    - Without the fix: reports a large number of catch blocks without
-      a current exception and exits 1.
-    - With the fix: reports 0 and exits 0.
- 4. Run the boost1.90 autopkgtests and check for regressions.
+ Build the attached repro.cpp against the installed libboost-context-dev
+ and run it:
+ 
+     g++ -O2 -std=c++17 -pthread -o repro repro.cpp -lboost_context
+     ./repro check 10
+     ./repro rethrow 10
+ 
+ With the current packages, check reports a large number of catch blocks
+ without a current exception (millions within 10 s) and exits 1. rethrow
+ aborts within milliseconds with "terminate called without an active
+ exception" (exit 134).
+ 
+ With the fixed packages, rebuild repro after upgrading. check then
+ reports 0 and exits 0, and rethrow runs the full 10 s and exits 0.
+ 
+ Do this for boost1.90 in resolute and stonking. For boost1.88 in
+ resolute, build against libboost1.88-dev / libboost-context1.88-dev
+ instead.
+ 
+ Results so far, using the patched header via an include overlay:
+ 
+     boost1.90, amd64 2 vCPU, check 10 2 8    unpatched 1845846 of 15032671 
bad   patched 0
+     boost1.90, arm64, check 10               unpatched 3739253 of 12523609 
bad   patched 0
+     boost1.88 sources, arm64, check 10       unpatched 2733288 of 11038347 
bad   patched 0
+     rethrow (all of the above)               unpatched terminate              
 patched 10 s, exit 0
+ 
+ For rebuilt reverse dependencies, we can provide a small objdump-based
+ check on request. It counts fiber switches that write through a cached
+ eh_globals address: currently 27 for icinga2-bin 2.15.1-1 in resolute,
+ and 0 is expected after the rebuild.
  
  [ Where problems could occur ]
  
- The patch only adds a noinline attribute to one header-only function,
- so it does not change ABI. The worst case is a small slowdown on each
- fiber switch, because the function is no longer inlined.
+ The patch adds BOOST_NOINLINE to the constructor and destructor of
+ boost::context::detail::manage_exception_state. Each of them copies 16
+ bytes. Nothing else changes, and there is no change to ABI, API or the
+ shared libraries.
  
- The fix sits in a header, so already-built packages stay affected until
- they are rebuilt. In resolute, only ceph needs a no-change rebuild
- (radosgw, librgw2, ceph-common). That rebuild could show unrelated
- FTBFS or pick up other changes in ceph's build dependencies.
+ A fiber switch costs two extra calls, which is negligible next to the
+ switch itself. A regression would show up in programs using
+ Boost.Context fibers (asio::spawn, Coroutine2, Fiber), as unexpected
+ exception behaviour around a switch. The identical change is part of
+ Boost 1.92.0.
  
- The other resolute packages that depend on libboost-context1.90.0 do
- not need a rebuild:
-   - pdns-recursor, simgrid, yuzu: call jump_fcontext/make_fcontext
-     directly and never use fiber_fcontext.hpp.
-   - wesnoth-1.18 (server): asio::spawn, but io_context only runs on the
-     main thread.
-   - nix: coroutine2 sources/sinks are deliberately kept on the calling
-     thread (see the comment in src/libstore/filetransfer.cc).
-   - graph-tool: coroutine2 generators exposed to Python; only affected
-     if one is resumed from another thread while the first thread is
-     inside a C++ catch block, which the GIL makes impractical.
-   - libboost-fiber1.90.0 is built from boost1.90 itself and gets fixed
-     by this upload.
- Third-party binaries built against the 1.90 headers (e.g. Icinga 2 from
- packages.icinga.com) need to be rebuilt by their vendors.
+ Binaries built against the old header stay affected until they are
+ rebuilt. The update does not make them worse.
  
  [ Other Info ]
  
- Upstream: https://github.com/boostorg/context/issues/323
- 
- Affected versions (resolute, amd64 and arm64):
-   libboost-context1.90-dev / libboost1.90-dev 1.90.0-6ubuntu1
-   g++ 15.2.0-16ubuntu1
- Icinga 2 was installed from packages.icinga.com.
- 
- Affected binaries in ceph 20.2.0-0ubuntu2.1 (amd64): radosgw,
- radosgw-admin, radosgw-object-expirer, librgw.so.2, denc-mod-rgw.so and
- libceph_crypto_qat.so. In radosgw,
- boost::asio::detail::spawned_fiber_thread::resume() keeps the address
- returned by __cxa_get_globals() in a register across jump_fcontext and
- writes the saved state back through it:
- 
-   7e91b1:  call   __cxa_get_globals@plt
-   7e91c3:  mov    (%rax),%r14
-   7e91c6:  mov    0x8(%rax),%r13d
-   7e91ca:  mov    %rax,%r12
-   7e91cd:  call   boost::context::detail::jump_fcontext@plt
-   7e91d2:  mov    %r14,(%r12)        <- stale TLS address after the switch
-   7e91da:  mov    %r13d,0x8(%r12)
- 
- spawned_fiber_thread::suspend_with() at 0x7f114d has the same sequence.
+ - Upstream: https://github.com/boostorg/context/issues/323, fix 
https://github.com/boostorg/context/pull/331 (0921b9fd5c). The earlier attempt, 
#324, was reverted.
+ - Boost 1.83 and 1.87 do not have manage_exception_state and are not 
affected. 25.10 (Boost 1.88) is affected but end of life.
+ - Downstream: https://github.com/Icinga/icinga2/issues/11160

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2173538

Title:
  [SRU] boost1.90: fibers clobber the libstdc++ exception state of
  another thread (upstream boostorg/context#323)

To manage notifications about this bug go to:
https://bugs.launchpad.net/boost/+bug/2173538/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to