changes since v2: - drop debug error message for non-fatal case per Borislav's strong objections - rewrite cover letter by amending MCE and CMCI race background from Tony[1] - rewrite commit log - use is_copy_from_user() to determine copy-from-user context per Peter - keep comments of in kill_me_maybe() per Catalin - add ack-by tag for patch 3 from Miaohe Lin - Link: https://lore.kernel.org/lkml/20250217063335.22257-1-xueshuai@xxxxxxxxxxxxxxxxx/T/#m4d87f152a67e26f2aabb4cdf81e451a1e4c70094 changes singce v1: - Patch 1: Fix cur_sev and sev type to `int` per Tony - Patch 4: Fix return value to 0 for clean pages per Miaohe - Patch 5: pick return value comments of memory-failure() ## 1. What am I trying to do: This patch resolves two critical regressions related to memory failure handling that have appeared in the upstream kernel since version 5.17, as compared to 5.10 LTS. - copyin case: poison found in user page while kernel copying from user space - instr case: poison found while instruction fetching in user space ## 2. What is the expected outcome and why - For copyin case: Kernel can recover from poison found where kernel is doing get_user() or copy_from_user() if those places get an error return and the kernel return -EFAULT to the process instead of crashing. More specifily, MCE handler checks the fixup handler type to decide whether an in kernel #MC can be recovered. When EX_TYPE_UACCESS is found, the PC jumps to recovery code specified in _ASM_EXTABLE_FAULT() and return a -EFAULT to user space. - For instr case: If a poison found while instruction fetching in user space, full recovery is possible. User process takes #PF, Linux allocates a new page and fills by reading from storage. ## 3. What actually happens and why - For copyin case: kernel panic since v5.17 Commit 4c132d1d844a ("x86/futex: Remove .fixup usage") introduced a new extable fixup type, EX_TYPE_EFAULT_REG, and later patches updated the extable fixup type for copy-from-user operations, changing it from EX_TYPE_UACCESS to EX_TYPE_EFAULT_REG. It breaks previous EX_TYPE_UACCESS handling when posion found in get_user() or copy_from_user(). - For instr case: user process is killed by a SIGBUS signal due to #CMCI and #MCE race When an uncorrected memory error is consumed there is a race between the CMCI from the memory controller reporting an uncorrected error with a UCNA signature, and the core reporting and SRAR signature machine check when the data is about to be consumed. ### Background: why *UN*corrected errors tied to *C*MCI in Intel platform [1] Prior to Icelake memory controllers reported patrol scrub events that detected a previously unseen uncorrected error in memory by signaling a broadcast machine check with an SRAO (Software Recoverable Action Optional) signature in the machine check bank. This was overkill because it's not an urgent problem that no core is on the verge of consuming that bad data. It's also found that multi SRAO UCE may cause nested MCE interrupts and finally become an IERR. Hence, Intel downgrades the machine check bank signature of patrol scrub from SRAO to UCNA (Uncorrected, No Action required), and signal changed to #CMCI. Just to add to the confusion, Linux does take an action (in uc_decode_notifier()) to try to offline the page despite the UC*NA* signature name. ### Background: why #CMCI and #MCE race when poison is consuming in Intel platform [1] Having decided that CMCI/UCNA is the best action for patrol scrub errors, the memory controller uses it for reads too. But the memory controller is executing asynchronously from the core, and can't tell the difference between a "real" read and a speculative read. So it will do CMCI/UCNA if an error is found in any read. Thus: 1) Core is clever and thinks address A is needed soon, issues a speculative read. 2) Core finds it is going to use address A soon after sending the read request 3) The CMCI from the memory controller is in a race with MCE from the core that will soon try to retire the load from address A. Quite often (because speculation has got better) the CMCI from the memory controller is delivered before the core is committed to the instruction reading address A, so the interrupt is taken, and Linux offlines the page (marking it as poison). ## Why user process is killed for instr case Commit 046545a661af ("mm/hwpoison: fix error page recovered but reported "not recovered"") tries to fix noise message "Memory error not recovered" and skips duplicate SIGBUSs due to the race. But it also introduced a bug that kill_accessing_process() return -EHWPOISON for instr case, as result, kill_me_maybe() send a SIGBUS to user process. # 4. The fix, in my opinion, should be: - For copyin case: The key point is whether the error context is in a read from user memory. We do not care about the ex-type if we know its a MOV reading from userspace. is_copy_from_user() return true when both of the following two checks are true: - the current instruction is copy - source address is user memory If copy_user is true, we set m->kflags |= MCE_IN_KERNEL_COPYIN | MCE_IN_KERNEL_RECOV; Then do_machine_check() will try fixup_exception() first. - For instr case: let kill_accessing_process() return 0 to prevent a SIGBUS. - For patch 3: The return value of memory_failure() is quite important while discussed instr case regression with Tony and Miaohe for patch 4, so add comment about the return value. Shuai Xue (3): x86/mce: Use is_copy_from_user() to determine copy-from-user context mm/hwpoison: Do not send SIGBUS to processes with recovered clean pages mm: memory-failure: Enhance comments for return value of memory_failure() arch/x86/kernel/cpu/mce/severity.c | 11 +++++------ mm/memory-failure.c | 21 +++++++++++++++------ 2 files changed, 20 insertions(+), 12 deletions(-) -- 2.39.3
