Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Rip out all of our cancellation logic, use sd_event_exit
Originally, I wanted to address the following problem: When a user callback indicates failure, calling CancelAll() means that we end up (re)invoking a curl socket callback from a socket callback, leading to a double free (either in sd-event or curl). One possible backtrace looks like: (gdb) bt -> #0 aur::AurImpl::DispatchSocketCallback (this=0x618000000480, s=<optimized out>, action=4, io=<optimized out>) at ../src/aur/aur.cc:330 #1 0x00007ffff74fa4e1 in singlesocket () from /usr/lib/libcurl.so.4 #2 0x00007ffff74fe622 in curl_multi_remove_handle () from /usr/lib/libcurl.so.4 #3 0x000055555570e833 in aur::AurImpl::FinishRequest (this=<optimized out>, curl=0x623000005500, result=<optimized out>, dispatch_callback=<optimized out>) at ../src/aur/aur.cc:462 #4 0x0000555555708cc1 in std::__do_visit<std::__detail::__variant::__deduce_visit_result<void>, aur::AurImpl::Cancel(const value_type&)::Visitor, const std::variant<void*, sd_event_source*>&> (__visitor=...) at /usr/include/c++/10.2.0/variant:869 #5 std::visit<aur::AurImpl::Cancel(const value_type&)::Visitor, const std::variant<void*, sd_event_source*>&> (__visitor=...) at /usr/include/c++/10.2.0/variant:1710 #6 aur::AurImpl::Cancel (this=0x618000000480, request=...) at ../src/aur/aur.cc:291 #7 0x00005555557090f6 in aur::AurImpl::CancelAll (this=0x618000000480) at ../subprojects/abseil-cpp-20200225.2/absl/container/internal/raw_hash_set.h:311 #8 0x000055555570f08d in aur::AurImpl::CheckFinished (this=0x618000000480) at ../src/aur/aur.cc:485 -> #9 0x000055555570f341 in aur::AurImpl::DispatchSocketCallback (this=0x618000000480, s=<optimized out>, action=4, io=<optimized out>) at ../src/aur/aur.cc:332 #10 0x00007ffff74fb092 in Curl_multi_closed () from /usr/lib/libcurl.so.4 #11 0x00007ffff74cc631 in Curl_closesocket () from /usr/lib/libcurl.so.4 #12 0x00007ffff74df551 in Curl_disconnect () from /usr/lib/libcurl.so.4 #13 0x00007ffff74fc354 in multi_done () from /usr/lib/libcurl.so.4 #14 0x00007ffff74fcc91 in multi_runsingle () from /usr/lib/libcurl.so.4 #15 0x00007ffff74fe1d1 in multi_socket () from /usr/lib/libcurl.so.4 #16 0x00007ffff74fe354 in curl_multi_socket_action () from /usr/lib/libcurl.so.4 #17 0x000055555570f60d in aur::AurImpl::OnCurlTimer (userdata=0x618000000480) at ../src/aur/aur.cc:401 #18 0x00007ffff7466b3e in ?? () from /usr/lib/libsystemd.so.0 #19 0x00007ffff746821e in sd_event_dispatch () from /usr/lib/libsystemd.so.0 #20 0x00007ffff746a6a9 in sd_event_run () from /usr/lib/libsystemd.so.0 #21 0x0000555555704bd3 in aur::AurImpl::Wait (this=0x618000000480) at ../src/aur/aur.cc:495 #22 0x000055555563ee8a in auracle::Auracle::GetOutdatedPackages (this=<optimized out>, args=std::vector of length 0, capacity 0, packages=<optimized out>) at /usr/include/c++/10.2.0/bits/unique_ptr.h:421 #23 0x0000555555656816 in auracle::Auracle::Outdated (this=<optimized out>, args=..., options=...) at ../src/auracle/auracle.cc:564 #24 0x0000555555596c7f in main (argc=<optimized out>, argv=<optimized out>) at ../subprojects/abseil-cpp-20200225.2/absl/container/internal/raw_hash_set.h:311 We could do that by making CancelAll() merely schedule another event that performs the actual cancellation at a later point in order to avoid the recursion. However, our cancellation logic is all sorts of weird and makes assumptions about how events are dispatched (i.e. there might be multiple at a time). Let's just get rid of all of this and use the sd-event mechanism of sd_event_exit instead. This does, however (as did the original proposed solution), have the side effect of logging multiple times because we potentially open up to 5 connections to the AUR at once, e.g. $ build/auracle --baseurl http://129.168.255.1 outdated error: UNKNOWN: Connection timed out after 10000 milliseconds error: UNKNOWN: Connection timed out after 10000 milliseconds I suppose one way to fix this would be to do response merging on the backend to match the request splitting. That way, the frontend only gets one response. I think that comes with a lot of weird potential behaviors though (handling of partial failures, to mention one). Would be nicer if the AUR didn't have the crap behavior and could take POST requests in order to extend the arg limit. Whatever, this is a weird edge case. Fixes #82.
- Loading branch information