** Description changed: + [ Impact ] + + * On 26.04, with use_pty (the default), suspending an editor that the + command under sudo starts as a child (sudo visudo, sudo sudo -e FILE) + and resuming it leaves the editor stopped. Run "sudo visudo" or "sudo + sudo -e FILE", suspend the editor, run fg, and the editor never comes + back. Ctrl-C still ends it. Other programs that sudo's command starts as + a child and that the user suspends with Ctrl-Z behave the same. A plain + "sudo nano FILE" is not affected, because there the editor is the + command itself. + + * Cause: on fg, the sudo-rs monitor resumes the command with + kill(command_pid, SIGCONT). That reaches the command (visudo, or the + inner sudo) but not the editor, its child in the same process group, + which stays in state T+. + + * Fix: upstream PR 1530 (commit 1c9a1eb2, in 0.2.14 and later) changes + the three kill() calls in MonitorClosure::send_signal + (src/exec/use_pty/monitor.rs) to killpg(command_pid, ...), as sudo.ws + does. The whole group gets SIGCONT, so Ctrl-Z then fg resumes the editor + again. The same change fixes LP: #2150073, which has the full + background. + + [ Test Plan ] + + Run this on a test machine or VM with a clean Ubuntu 26.04 install, + where /usr/bin/sudo is sudo-rs (the default) and /usr/bin/editor is nano + (the default). Use a real terminal (the local console, a desktop + terminal or "ssh -t") and a user that may run sudo. Open two terminals, + A and B, as that user. + + Steps: + + a. Do the setup below in terminals A and B. + b. Run tests 1 to 5 with 0.2.13-0ubuntu1.2 to see the bug. + c. Install the update from -proposed (below). + d. Run tests 1 to 5 again in the same terminals. + + Setup, in terminal A. First run this line on its own. It may ask for + your password: + + sudo -v + + sudo-rs keeps the password for 15 minutes, per terminal. If the tests + take longer, run "sudo -v" again before the next test, so that no + password prompt lands inside a test. Then run: + + sudo --version + printf 'test\n' | sudo tee /etc/lp2160212-test + + "sudo --version" prints sudo-rs 0.2.13-0ubuntu1.2 before the update. + + Setup, in terminal B. Run "sudo -v" there too. To look for a stuck + process, run this in terminal B. It lists every process in state T: + + ps -eo pid,ppid,stat,wchan:32,args | awk '$3 ~ /^T/' + + nano suspends with Ctrl-T and then Ctrl-Z. With vim, press Ctrl-Z. + + Bug tests, in terminal A: + + 1. The report's second case: + + sudo visudo + + Press Ctrl-T and then Ctrl-Z. The shell prints "Stopped" and the + prompt. Run fg. Expected: nano draws its screen again and takes keys. + Ctrl-X closes it, visudo prints "visudo: /tmp/sudoers-XXXXXX/sudoers + unchanged" (with a random name in place of XXXXXX), and the prompt comes + back. + + 2. The report's first case: + + sudo sudo -e /etc/lp2160212-test + + Press Ctrl-T and then Ctrl-Z, then run fg. Expected: nano comes + back. Ctrl-X closes it, sudo prints "sudo: /etc/lp2160212-test + unchanged", and the prompt comes back. + + 3. A child of the command that reads the terminal: + + sudo sh -c 'sh -c "read v && echo got-\$v"' + + Wait a second and press Ctrl-Z. The shell prints "Stopped". Run fg, + type hello and press Enter. Expected: got-hello, then the prompt. + + With 0.2.13-0ubuntu1.2 tests 1 to 3 hang after fg. The screen stays as + it is, keys do nothing, and terminal B lists the editor + (/usr/bin/editor) or "sh -c read v" in state T+. In test 2 terminal B + also lists the inner "sudo -e /etc/lp2160212-test" in state T+. To get + terminal A back, run "sudo kill -CONT PID" in terminal B for each PID + listed. The editor then takes keys and Ctrl-X closes it. + + With 0.2.13-0ubuntu1.3 tests 1 to 3 pass and terminal B lists nothing + after fg. + + Regression tests, in terminal A: + + 4. Ctrl-Z and fg while the command itself reads: + + sudo sh -c 'read v && echo got-$v' + + Press Ctrl-Z, run fg, type hello and press Enter. Expected on both + versions: got-hello. + + 5. Jobs of a root shell keep their state: + + sudo -s + + In the root shell: + + sleep 300 & + kill -STOP %1 + suspend + + At the user prompt, run fg. Back in the root shell: + + jobs -l + + Expected on both versions: sleep 300 is still listed as "Stopped + (signal)". Clean up with "kill -KILL %1" and "exit". + + Clean up at the end with "sudo rm /etc/lp2160212-test". + + Install the update: enable -proposed as + https://ubuntu.com/project/docs/SRU/reference/sru-verification/ + describes, then run: + + sudo apt install -t resolute-proposed sudo-rs + apt policy sudo-rs + sudo --version + + "sudo --version" must print sudo-rs 0.2.13-0ubuntu1.3. Run "sudo -v" + again in both terminals before step d. + + [ Where problems could occur ] + + * The upload changes how the monitor process of the use_pty path + delivers signals to the command. If the change is wrong, it would show + as commands under sudo on a terminal that do not come back after Ctrl-Z + and fg, that hang, or that get interrupted or killed when they did not + before. + + * Signals that sudo relays (SIGINT, SIGQUIT, SIGTSTP, SIGTERM, SIGHUP, + SIGUSR1, SIGUSR2) now reach every process in the command's process + group, not only the command. A "kill -HUP" or "kill -USR1" sent to sudo + now also hits helper children in that group, and the default action of + SIGHUP, SIGUSR1 and SIGUSR2 is to terminate. This is how sudo.ws (24.04) + has always relayed signals. + + * Ctrl-C: with no pipe, Ctrl-C already reached the whole group through + the pty. With stdout or stdin piped, the user's terminal stays cooked + until the first SIGTTOU or SIGTTIN stop, and until then sudo relays the + SIGINT from Ctrl-C. Before, only the command got it, so an apt-get or + dpkg run under "sudo ./script | tee log" did not get it. Now the whole + group gets it, as with sudo.ws, and Ctrl-C there now also reaches apt- + get or dpkg under the script. + + * SIGCONT now reaches group members that were not stopped, or that + something else stopped on purpose. Programs with a SIGCONT handler, such + as shells, editors, pagers and curses programs, run it, which for an + editor means a redraw. + + * Not changed: processes in other groups or sessions, such as the jobs + of a root shell under sudo -s (test 5). With Defaults !use_pty, or with + no terminal, sudo-rs takes src/exec/no_pty.rs, which this upload does + not touch. + + * The same change has shipped in upstream 0.2.14 and 0.2.15, and in + Ubuntu since 0.2.14-1ubuntu2 in 26.10. + + [ Other Info ] + + * Upstream: issue 1527 (this bug), PR 1530, commit 1c9a1eb2d4b746938338ff011c8b20c545188196, first released in 0.2.14. The upstream issue was closed by PR 1530. + https://github.com/trifectatechfoundation/sudo-rs/pull/1530 + + * The same upload fixes LP: #2150073 and LP: #2146860. The debdiff, the + patch details and the full test record are on LP: #2150073. Upstream's + ubuntu-raccoon-backport branch for 26.04 is at 0.2.13-0ubuntu1.2 and + does not have PR 1530. + + * Fixed in later releases: 26.10 (stonking) has the fix since + 0.2.14-1ubuntu2 and 0.2.14-1ubuntu4.1 in stonking-updates, and + tenacious, the current development series, has 0.2.14-1ubuntu4.1. So no + 26.10 task is needed. noble has rust-sudo-rs 0.2.2-1 in universe, which + is not the default sudo there. It has the same kill() code, but it was + not tested and no noble task is proposed. + + * Testing done: tests 1 and 2 ran in fresh ubuntu:26.04 containers on a + pty, typed into an interactive bash, as root and as a NOPASSWD user, + with nano as /usr/bin/editor, on three .deb builds: the archive + 0.2.13-0ubuntu1.2, a build with only the PR 1530 patch, and + 0.2.13-0ubuntu1.3 with both patches of this upload. Both tests hang on + 0.2.13-0ubuntu1.2 after fg, with the editor in T+ (in test 2 also the + inner "sudo -e" child), and kill -CONT on the listed PIDs brought the + editor back. Both pass on the two other builds, ending with "visudo: + /tmp/sudoers-XXXXXX/sudoers unchanged" and "sudo: /etc/FILE unchanged". + Tests 3 to 5 ran in a typed run of the plan as the NOPASSWD user: test 3 + hangs on 0.2.13-0ubuntu1.2 with "sh -c read v" in T+ and passes on + 0.2.13-0ubuntu1.3, and tests 4 and 5 pass on both. As a control, "sudo + nano FILE" with Ctrl-T Ctrl-Z, fg and Ctrl-X passes on 0.2.13-0ubuntu1.2 + and on 0.2.13-0ubuntu1.3. All of it ran once more on .deb packages built + from the attached debdiff: test 1 on the final build, and the rest on a + build whose debdiff differed only in changelog and DEP-3 text and whose + sudo, su and visudo binaries are byte identical. The results were the + same. Every run used .deb packages. The containers ran on a 6.18 kernel, + not on a 26.04 install with its 7.0 kernel. + + * Ctrl-C check, in the same containers: a script whose child traps INT, + Ctrl-C 2 seconds after the start. Under "sudo ./int-check.sh | cat" the + child never got SIGINT on 0.2.13-0ubuntu1.2 and got it on + 0.2.13-0ubuntu1.3. With no pipe, it got SIGINT on both. The full record + is on LP: #2150073. + + * The autopkgtest (debian/tests/run-cargo-tests.sh) runs cargo test, so + unit tests only. Upstream's compliance tests, which drive sudo on a + terminal, live in test-framework/, which is not in the crates.io tarball + the package builds from. No autopkgtest is added. + + [ Original Description ] + (Tracking bug from upstream: https://github.com/trifectatechfoundation/sudo-rs/issues/1527) In some combinations, resuming a suspended child process doesn't work. ## To Reproduce Make sure Defaults use_pty is turned on, then run $ sudo sudo -e /etc/file or $ sudo visudo and suspend the process by typing ^Z, then resume it by using fg. The process isn't "stuck" (Ctrl-C can break out of it). ## Expected behavior The editor is resumed.
-- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/2160212 Title: SIGCONT doesn't seem to propagate all the way through in some cases To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/rust-sudo-rs/+bug/2160212/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
