From a5f7a5bb3b7f28ba7e4fa246775b29a0e5537255 Mon Sep 17 00:00:00 2001 From: Jinke Han Date: Tue, 8 Sep 2026 15:37:42 +0800 Subject: x86/kprobes: Fix crash when probing CS CALL instructions When using eBPF to probe CS CALL instructions within a function, a crash can be triggered. The eBPF tool probes offset 257 of the __hrtimer_run_queues() function: <__hrtimer_run_queues+249>: nopl 0x0(%rax,%rax,1) <__hrtimer_run_queues+254>: mov %r14,%rdi <__hrtimer_run_queues+257>: cs call <__x86_indirect_thunk_r12> <__hrtimer_run_queues+263>: mov %eax,%r12d <__hrtimer_run_queues+266>: xchg %ax,%ax <__hrtimer_run_queues+268>: mov %r13,%rdi Which triggers this crash: BUG: unable to handle page fault for address: 00000000000f41c9 #PF: supervisor write access in kernel mode #PF: error_code(0x0002) - not-present page PGD 0 P4D 0 Oops: 0002 [#1] SMP NOPTI CPU: 1 PID: 0 Comm: swapper/1 Kdump: loaded Tainted: P RIP: 0010:__hrtimer_run_queues+0x106/0x230 Note that __hrtimer_run_queues+0x106 is __hrtimer_run_queues+262, which is at the 6th byte of the above CS CALL instruction. Since the CS CALL instruction occupies 6 bytes, the exception occurred in the middle of that call instruction. The root cause is that when using eBPF tools to probe in the middle of a function, a kprobe with INT3 is used as the underlying implementation. During single-step emulation of the original CALL instruction, int3_emulate_call() assumes that the probed CALL instruction is 5 bytes long. However, the actual CS-prefixed CALL instruction occupies 6 bytes, so it constructs an incorrect exception return address. When the CPU returns from the kprobe handler, the next instruction to be executed is at the address of the last byte of that CS CALL instruction. Coincidentally, starting from that address, the CPU fetches and decodes a completely different instruction, which ultimately triggers a kernel crash. Fix the issue by using the actual instruction length obtained from the instruction decoder when constructing the exception return address, rather than relying on the hardcoded CALL_INSN_SIZE macro. [ mingo: Refined the changelog ] Fixes: 6256e668b7af ("x86/kprobes: Use int3 instead of debug trap for single-step") Suggested-by: Masami Hiramatsu (Google) Signed-off-by: Jinke Han Signed-off-by: Ingo Molnar Reviewed-by: Masami Hiramatsu (Google) Acked-by: Yafang Shao Acked-by: Borislav Petkov Cc: Peter Zijlstra Link: https://patch.msgid.link/20260908073742.GA10517@didi-ThinkCentre-M920t-N000 --- Documentation/admin-guide/bug-bisect.rst | 165 +++++++++++++++++++++++++++++++ 1 file changed, 165 insertions(+) create mode 100644 Documentation/admin-guide/bug-bisect.rst (limited to 'Documentation/admin-guide/bug-bisect.rst') diff --git a/Documentation/admin-guide/bug-bisect.rst b/Documentation/admin-guide/bug-bisect.rst new file mode 100644 index 000000000..f4f867cab --- /dev/null +++ b/Documentation/admin-guide/bug-bisect.rst @@ -0,0 +1,165 @@ +.. SPDX-License-Identifier: (GPL-2.0+ OR CC-BY-4.0) +.. [see the bottom of this file for redistribution information] + +====================== +Bisecting a regression +====================== + +This document describes how to use a ``git bisect`` to find the source code +change that broke something -- for example when some functionality stopped +working after upgrading from Linux 6.0 to 6.1. + +The text focuses on the gist of the process. If you are new to bisecting the +kernel, better follow Documentation/admin-guide/verify-bugs-and-bisect-regressions.rst +instead: it depicts everything from start to finish while covering multiple +aspects even kernel developers occasionally forget. This includes detecting +situations early where a bisection would be a waste of time, as nobody would +care about the result -- for example, because the problem happens after the +kernel marked itself as 'tainted', occurs in an abandoned version, was already +fixed, or is caused by a .config change you or your Linux distributor performed. + +Finding the change causing a kernel issue using a bisection +=========================================================== + +*Note: the following process assumes you prepared everything for a bisection. +This includes having a Git clone with the appropriate sources, installing the +software required to build and install kernels, as well as a .config file stored +in a safe place (the following example assumes '~/prepared_kernel_.config') to +use as pristine base at each bisection step; ideally, you have also worked out +a fully reliable and straight-forward way to reproduce the regression, too.* + +* Preparation: start the bisection and tell Git about the points in the history + you consider to be working and broken, which Git calls 'good' and 'bad':: + + git bisect start + git bisect good v6.0 + git bisect bad v6.1 + + Instead of Git tags like 'v6.0' and 'v6.1' you can specify commit-ids, too. + +1. Copy your prepared .config into the build directory and adjust it to the + needs of the codebase Git checked out for testing:: + + cp ~/prepared_kernel_.config .config + make olddefconfig + +2. Now build, install, and boot a kernel. This might fail for unrelated reasons, + for example, when a compile error happens at the current stage of the + bisection a later change resolves. In such cases run ``git bisect skip`` and + go back to step 1. + +3. Check if the functionality that regressed works in the kernel you just built. + + If it works, execute:: + + git bisect good + + If it is broken, run:: + + git bisect bad + + Note, getting this wrong just once will send the rest of the bisection + totally off course. To prevent having to start anew later you thus want to + ensure what you tell Git is correct; it is thus often wise to spend a few + minutes more on testing in case your reproducer is unreliable. + + After issuing one of these two commands, Git will usually check out another + bisection point and print something like 'Bisecting: 675 revisions left to + test after this (roughly 10 steps)'. In that case go back to step 1. + + If Git instead prints something like 'cafecaca0c0dacafecaca0c0dacafecaca0c0da + is the first bad commit', then you have finished the bisection. In that case + move to the next point below. Note, right after displaying that line Git will + show some details about the culprit including its patch description; this can + easily fill your terminal, so you might need to scroll up to see the message + mentioning the culprit's commit-id. + + In case you missed Git's output, you can always run ``git bisect log`` to + print the status: it will show how many steps remain or mention the result of + the bisection. + +* Recommended complementary task: put the bisection log and the current .config + file aside for the bug report; furthermore tell Git to reset the sources to + the state before the bisection:: + + git bisect log > ~/bisection-log + cp .config ~/bisection-config-culprit + git bisect reset + +* Recommended optional task: try reverting the culprit on top of the latest + codebase and check if that fixes your bug; if that is the case, it validates + the bisection and enables developers to resolve the regression through a + revert. + + To try this, update your clone and check out latest mainline. Then tell Git + to revert the change by specifying its commit-id:: + + git revert --no-edit cafec0cacaca0 + + Git might reject this, for example when the bisection landed on a merge + commit. In that case, abandon the attempt. Do the same, if Git fails to revert + the culprit on its own because later changes depend on it -- at least unless + you bisected a stable or longterm kernel series, in which case you want to + check out its latest codebase and try a revert there. + + If a revert succeeds, build and test another kernel to check if reverting + resolved your regression. + +With that the process is complete. Now report the regression as described by +Documentation/admin-guide/reporting-issues.rst. + +Bisecting linux-next +-------------------- + +If you face a problem only happening in linux-next, bisect between the +linux-next branches 'stable' and 'master'. The following commands will start +the process for a linux-next tree you added as a remote called 'next':: + + git bisect start + git bisect good next/stable + git bisect bad next/master + +The 'stable' branch refers to the state of linux-mainline that the current +linux-next release (found in the 'master' branch) is based on -- the former +thus should be free of any problems that show up in -next, but not in Linus' +tree. + +This will bisect across a wide range of changes, some of which you might have +used in earlier linux-next releases without problems. Sadly there is no simple +way to avoid checking them: bisecting from one linux-next release to a later +one (say between 'next-20241020' and 'next-20241021') is impossible, as they +share no common history. + +Additional reading material +--------------------------- + +* The `man page for 'git bisect' `_ and + `fighting regressions with 'git bisect' `_ + in the Git documentation. +* `Working with git bisect `_ + from kernel developer Nathan Chancellor. +* `Using Git bisect to figure out when brokenness was introduced `_. +* `Fully automated bisecting with 'git bisect run' `_. + +.. + end-of-content +.. + This document is maintained by Thorsten Leemhuis . If + you spot a typo or small mistake, feel free to let him know directly and + he'll fix it. You are free to do the same in a mostly informal way if you + want to contribute changes to the text -- but for copyright reasons please CC + linux-doc@vger.kernel.org and 'sign-off' your contribution as + Documentation/process/submitting-patches.rst explains in the section 'Sign + your work - the Developer's Certificate of Origin'. +.. + This text is available under GPL-2.0+ or CC-BY-4.0, as stated at the top + of the file. If you want to distribute this text under CC-BY-4.0 only, + please use 'The Linux kernel development community' for author attribution + and link this as source: + https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/plain/Documentation/admin-guide/bug-bisect.rst + +.. + Note: Only the content of this RST file as found in the Linux kernel sources + is available under CC-BY-4.0, as versions of this text that were processed + (for example by the kernel's build system) might contain content taken from + files which use a more restrictive license. -- cgit v1.3.1