A bus lock [1] is acquired through either split locked access to writeback (WB) memory or any locked access to non-WB memory. This is typically >1000 cycles slower than an atomic operation within a cache line. It also disrupts performance on other cores.
Although split lock can be detected by #AC trap, the trap is triggered before the instruction acquires bus lock. This makes it difficult to mitigate bus lock (e.g. throttle the user application). Some CPUs have ability to notify the kernel by an #DB trap after a user instruction acquires a bus lock and is executed. This allows the kernel to enforce user application throttling or mitigations. #DB for bus lock detect fixes issues in #AC for split lock detect: 1) It's architectural ... just need to look at one CPUID bit to know it exists 2) The IA32_DEBUGCTL MSR, which reports bus lock in #DB, is per-thread. So each process or guest can have different behavior. 3) It has support for VMM/guests (new VMEXIT codes, etc). 4) It detects not only split locks but also bus locks from non-WB. Hardware only generates #DB for bus lock detect when CPL>0 to avoid nested #DB from multiple bus locks while the first #DB is being handled. Use the existing kernel command line parameter "split_lock_detect=" to handle #DB for bus lock with an additional option "ratelimit=N" to set bus lock rate limit for a user. [1] Intel Instruction Set Extension Chapter 9: https://software.intel.com/content/dam/develop/public/us/en/documents/architecture-instruction-set-extensions-programming-reference.pdf Change Log: v5: Address all comments from Thomas: - In the cover letter, update the latest ISE link to include the #DB for bus lock spec. - In patch 1, add commit message for breakpoint and bus lock on the same instruction. - In patch 2, change warn to #AC if both #AC and #DB are supported, remove sld and bld variables, remove bus lock checking in handle_bus_lock() etc. - In patch 3 and 4, remove bld_ratelimit < HZ/2 check and define bld_ratelimit only for Intel CPUs. - Merge patch 2 and 3 into one patch for handling warn, fatal, and ratelimit. v4 is here: https://lore.kernel.org/lkml/20201124205245.4164633-2-fenghua...@intel.com/ v4: - Fix a ratelimit wording issue in the doc (Randy). - Patch 4 is acked by Randy (Randy). v3: - Enable Bus Lock Detection when fatal to handle bus lock from non-WB (PeterZ). - Add Acked-by: PeterZ in patch 2. v2: - Send SIGBUS in fatal case for bus lock #DB (PeterZ). v1: - Check bus lock bit by its positive polarity (Xiaoyao). - Fix a few wording issues in the documentation (Randy). [RFC v3 can be found at: https://lore.kernel.org/patchwork/cover/1329943/] RFC v3: - Remove DR6_RESERVED change (PeterZ). - Simplify the documentation (Randy). RFC v2: - Architecture changed based on feedback from Thomas and PeterZ. #DB is no longer generated for bus lock in ring0. - Split the one single patch into four patches. [RFC v1 can be found at: https://lore.kernel.org/lkml/1595021700-68460-1-git-send-email-fenghua...@intel.com/] Fenghua Yu (3): x86/cpufeatures: Enumerate #DB for bus lock detection x86/bus_lock: Handle #DB for bus lock Documentation/admin-guide: Change doc for split_lock_detect parameter .../admin-guide/kernel-parameters.txt | 30 +++- arch/x86/include/asm/cpu.h | 10 +- arch/x86/include/asm/cpufeatures.h | 1 + arch/x86/include/asm/msr-index.h | 1 + arch/x86/include/uapi/asm/debugreg.h | 1 + arch/x86/kernel/cpu/common.c | 2 +- arch/x86/kernel/cpu/intel.c | 148 +++++++++++++++--- arch/x86/kernel/traps.c | 7 + include/linux/sched/user.h | 5 +- kernel/user.c | 13 ++ 10 files changed, 187 insertions(+), 31 deletions(-) -- 2.30.2