April 2025 - Linux-kselftest-mirror

[PATCH v14 00/27] riscv control-flow integrity for usermode

by Deepak Gupta

Basics and overview =================== Software with larger attack surfaces (e.g. network facing apps like databases, browsers or apps relying on browser runtimes) suffer from memory corruption issues which can be utilized by attackers to bend control flow of the program to eventually gain control (by making their payload executable). Attackers are able to perform such attacks by leveraging call-sites which rely on indirect calls or return sites which rely on obtaining return address from stack memory. To mitigate such attacks, risc-v extension zicfilp enforces that all indirect calls must land on a landing pad instruction `lpad` else cpu will raise software check exception (a new cpu exception cause code on riscv). Similarly for return flow, risc-v extension zicfiss extends architecture with - `sspush` instruction to push return address on a shadow stack - `sspopchk` instruction to pop return address from shadow stack and compare with input operand (i.e. return address on stack) - `sspopchk` to raise software check exception if comparision above was a mismatch - Protection mechanism using which shadow stack is not writeable via regular store instructions More information an details can be found at extensions github repo [1]. Equivalent to landing pad (zicfilp) on x86 is `ENDBRANCH` instruction in Intel CET [3] and branch target identification (BTI) [4] on arm. Similarly x86's Intel CET has shadow stack [5] and arm64 has guarded control stack (GCS) [6] which are very similar to risc-v's zicfiss shadow stack. x86 and arm64 support for user mode shadow stack is already in mainline. Kernel awareness for user control flow integrity ================================================ This series picks up Samuel Holland's envcfg changes [2] as well. So if those are being applied independently, they should be removed from this series. Enabling: In order to maintain compatibility and not break anything in user mode, kernel doesn't enable control flow integrity cpu extensions on binary by default. Instead exposes a prctl interface to enable, disable and lock the shadow stack or landing pad feature for a task. This allows userspace (loader) to enumerate if all objects in its address space are compiled with shadow stack and landing pad support and accordingly enable the feature. Additionally if a subsequent `dlopen` happens on a library, user mode can take a decision again to disable the feature (if incoming library is not compiled with support) OR terminate the task (if user mode policy is strict to have all objects in address space to be compiled with control flow integirty cpu feature). prctl to enable shadow stack results in allocating shadow stack from virtual memory and activating for user address space. x86 and arm64 are also following same direction due to similar reason(s). clone/fork: On clone and fork, cfi state for task is inherited by child. Shadow stack is part of virtual memory and is a writeable memory from kernel perspective (writeable via a restricted set of instructions aka shadow stack instructions) Thus kernel changes ensure that this memory is converted into read-only when fork/clone happens and COWed when fault is taken due to sspush, sspopchk or ssamoswap. In case `CLONE_VM` is specified and shadow stack is to be enabled, kernel will automatically allocate a shadow stack for that clone call. map_shadow_stack: x86 introduced `map_shadow_stack` system call to allow user space to explicitly map shadow stack memory in its address space. It is useful to allocate shadow for different contexts managed by a single thread (green threads or contexts) risc-v implements this system call as well. signal management: If shadow stack is enabled for a task, kernel performs an asynchronous control flow diversion to deliver the signal and eventually expects userspace to issue sigreturn so that original execution can be resumed. Even though resume context is prepared by kernel, it is in user space memory and is subject to memory corruption and corruption bugs can be utilized by attacker in this race window to perform arbitrary sigreturn and eventually bypass cfi mechanism. Another issue is how to ensure that cfi related state on sigcontext area is not trampled by legacy apps or apps compiled with old kernel headers. In order to mitigate control-flow hijacting, kernel prepares a token and place it on shadow stack before signal delivery and places address of token in sigcontext structure. During sigreturn, kernel obtains address of token from sigcontext struture, reads token from shadow stack and validates it and only then allow sigreturn to succeed. Compatiblity issue is solved by adopting dynamic sigcontext management introduced for vector extension. This series re-factor the code little bit to allow future sigcontext management easy (as proposed by Andy Chiu from SiFive) config and compilation: Introduce a new risc-v config option `CONFIG_RISCV_USER_CFI`. Selecting this config option picks the kernel support for user control flow integrity. This optin is presented only if toolchain has shadow stack and landing pad support. And is on purpose guarded by toolchain support. Reason being that eventually vDSO also needs to be compiled in with shadow stack and landing pad support. vDSO compile patches are not included as of now because landing pad labeling scheme is yet to settle for usermode runtime. To get more information on kernel interactions with respect to zicfilp and zicfiss, patch series adds documentation for `zicfilp` and `zicfiss` in following: Documentation/arch/riscv/zicfiss.rst Documentation/arch/riscv/zicfilp.rst How to test this series ======================= Toolchain --------- $ git clone git@github.com:sifive/riscv-gnu-toolchain.git -b cfi-dev $ riscv-gnu-toolchain/configure --prefix=<path-to-where-to-build> --with-arch=rv64gc_zicfilp_zicfiss --enable-linux --disable-gdb --with-extra-multilib-test="rv64gc_zicfilp_zicfiss-lp64d:-static" $ make -j$(nproc) Qemu ---- Get the lastest qemu $ cd qemu $ mkdir build $ cd build $ ../configure --target-list=riscv64-softmmu $ make -j$(nproc) Opensbi ------- $ git clone git@github.com:deepak0414/opensbi.git -b v6_cfi_spec_split_opensbi $ make CROSS_COMPILE=<your riscv toolchain> -j$(nproc) PLATFORM=generic Linux ----- Running defconfig is fine. CFI is enabled by default if the toolchain supports it. $ make ARCH=riscv CROSS_COMPILE=<path-to-cfi-riscv-gnu-toolchain>/build/bin/riscv64-unknown-linux-gnu- -j$(nproc) defconfig $ make ARCH=riscv CROSS_COMPILE=<path-to-cfi-riscv-gnu-toolchain>/build/bin/riscv64-unknown-linux-gnu- -j$(nproc) In case you're building your own rootfs using toolchain, please make sure you pick following patch to ensure that vDSO compiled with lpad and shadow stack. "arch/riscv: compile vdso with landing pad" Branch where above patch can be picked https://github.com/deepak0414/linux-riscv-cfi/tree/vdso_user_cfi_v6.12-rc1 Running ------- Modify your qemu command to have: -bios <path-to-cfi-opensbi>/build/platform/generic/firmware/fw_dynamic.bin -cpu rv64,zicfilp=true,zicfiss=true,zimop=true,zcmop=true vDSO related Opens (in the flux) ================================= I am listing these opens for laying out plan and what to expect in future patch sets. And of course for the sake of discussion. Shadow stack and landing pad enabling in vDSO ---------------------------------------------- vDSO must have shadow stack and landing pad support compiled in for task to have shadow stack and landing pad support. This patch series doesn't enable that (yet). Enabling shadow stack support in vDSO should be straight forward (intend to do that in next versions of patch set). Enabling landing pad support in vDSO requires some collaboration with toolchain folks to follow a single label scheme for all object binaries. This is necessary to ensure that all indirect call-sites are setting correct label and target landing pads are decorated with same label scheme. How many vDSOs --------------- Shadow stack instructions are carved out of zimop (may be operations) and if CPU doesn't implement zimop, they're illegal instructions. Kernel could be running on a CPU which may or may not implement zimop. And thus kernel will have to carry 2 different vDSOs and expose the appropriate one depending on whether CPU implements zimop or not. References ========== [1] - https://github.com/riscv/riscv-cfi [2] - https://lore.kernel.org/all/20240814081126.956287-1-samuel.holland@sifive.c… [3] - https://lwn.net/Articles/889475/ [4] - https://developer.arm.com/documentation/109576/0100/Branch-Target-Identific… [5] - https://www.intel.com/content/dam/develop/external/us/en/documents/catc17-i… [6] - https://lwn.net/Articles/940403/ To: Thomas Gleixner <tglx(a)linutronix.de> To: Ingo Molnar <mingo(a)redhat.com> To: Borislav Petkov <bp(a)alien8.de> To: Dave Hansen <dave.hansen(a)linux.intel.com> To: x86(a)kernel.org To: H. Peter Anvin <hpa(a)zytor.com> To: Andrew Morton <akpm(a)linux-foundation.org> To: Liam R. Howlett <Liam.Howlett(a)oracle.com> To: Vlastimil Babka <vbabka(a)suse.cz> To: Lorenzo Stoakes <lorenzo.stoakes(a)oracle.com> To: Paul Walmsley <paul.walmsley(a)sifive.com> To: Palmer Dabbelt <palmer(a)dabbelt.com> To: Albert Ou <aou(a)eecs.berkeley.edu> To: Conor Dooley <conor(a)kernel.org> To: Rob Herring <robh(a)kernel.org> To: Krzysztof Kozlowski <krzk+dt(a)kernel.org> To: Arnd Bergmann <arnd(a)arndb.de> To: Christian Brauner <brauner(a)kernel.org> To: Peter Zijlstra <peterz(a)infradead.org> To: Oleg Nesterov <oleg(a)redhat.com> To: Eric Biederman <ebiederm(a)xmission.com> To: Kees Cook <kees(a)kernel.org> To: Jonathan Corbet <corbet(a)lwn.net> To: Shuah Khan <shuah(a)kernel.org> To: Jann Horn <jannh(a)google.com> To: Conor Dooley <conor+dt(a)kernel.org> To: Miguel Ojeda <ojeda(a)kernel.org> To: Alex Gaynor <alex.gaynor(a)gmail.com> To: Boqun Feng <boqun.feng(a)gmail.com> To: Gary Guo <gary(a)garyguo.net> To: Björn Roy Baron <bjorn3_gh(a)protonmail.com> To: Benno Lossin <benno.lossin(a)proton.me> To: Andreas Hindborg <a.hindborg(a)kernel.org> To: Alice Ryhl <aliceryhl(a)google.com> To: Trevor Gross <tmgross(a)umich.edu> Cc: linux-kernel(a)vger.kernel.org Cc: linux-fsdevel(a)vger.kernel.org Cc: linux-mm(a)kvack.org Cc: linux-riscv(a)lists.infradead.org Cc: devicetree(a)vger.kernel.org Cc: linux-arch(a)vger.kernel.org Cc: linux-doc(a)vger.kernel.org Cc: linux-kselftest(a)vger.kernel.org Cc: alistair.francis(a)wdc.com Cc: richard.henderson(a)linaro.org Cc: jim.shu(a)sifive.com Cc: andybnac(a)gmail.com Cc: kito.cheng(a)sifive.com Cc: charlie(a)rivosinc.com Cc: atishp(a)rivosinc.com Cc: evan(a)rivosinc.com Cc: cleger(a)rivosinc.com Cc: alexghiti(a)rivosinc.com Cc: samitolvanen(a)google.com Cc: broonie(a)kernel.org Cc: rick.p.edgecombe(a)intel.com Cc: rust-for-linux(a)vger.kernel.org changelog --------- v14: - rebased on top of palmer/sbi-v3. Thus dropped clement's FWFT patches Updated RISCV_ISA_EXT_XXXX in hwcap and hwprobe constants. - Took Radim's suggestions on bitfields. - Placed cfi_state at the end of thread_info block so that current situation is not disturbed with respect to member fields of thread_info in single cacheline. v13: - cpu_supports_shadow_stack/cpu_supports_indirect_br_lp_instr uses riscv_has_extension_unlikely() - uses nops(count) to create nop slide - RISCV_ACQUIRE_BARRIER is not needed in `amo_user_shstk`. Removed it - changed ternaries to simply use implicit casting to convert to bool. - kernel command line allows to disable zicfilp and zicfiss independently. updated kernel-parameters.txt. - ptrace user abi for cfi uses bitmasks instead of bitfields. Added ptrace kselftest. - cosmetic and grammatical changes to documentation. v12: - It seems like I had accidently squashed arch agnostic indirect branch tracking prctl and riscv implementation of those prctls. Split them again. - set_shstk_status/set_indir_lp_status perform CSR writes only when CPU support is available. As suggested by Zong Li. - Some minor clean up in kselftests as suggested by Zong Li. v11: - patch "arch/riscv: compile vdso with landing pad" was unconditionally selecting `_zicfilp` for vDSO compile. fixed that. Changed `lpad 1` to to `lpad 0`. v10: - dropped "mm: helper `is_shadow_stack_vma` to check shadow stack vma". This patch is not that interesting to this patch series for risc-v. There are instances in arch directories where VM_SHADOW_STACK flag is anyways used. Dropping this patch to expedite merging in riscv tree. - Took suggestions from `Clement` on "riscv: zicfiss / zicfilp enumeration" to validate presence of cfi based on config. - Added a patch for vDSO to have `lpad 0`. I had omitted this earlier to make sure we add single vdso object with cfi enabled. But a vdso object with scheme of zero labeled landing pad is least common denominator and should work with all objects of zero labeled as well as function-signature labeled objects. v9: - rebased on master (39a803b754d5 fix braino in "9p: fix ->rename_sem exclusion") - dropped "mm: Introduce ARCH_HAS_USER_SHADOW_STACK" (master has it from arm64/gcs) - dropped "prctl: arch-agnostic prctl for shadow stack" (master has it from arm64/gcs) v8: - rebased on palmer/for-next - dropped samuel holland's `envcfg` context switch patches. they are in parlmer/for-next v7: - Removed "riscv/Kconfig: enable HAVE_EXIT_THREAD for riscv" Instead using `deactivate_mm` flow to clean up. see here for more context https://lore.kernel.org/all/20230908203655.543765-1-rick.p.edgecombe@intel.… - Changed the header include in `kselftest`. Hopefully this fixes compile issue faced by Zong Li at SiFive. - Cleaned up an orphaned change to `mm/mmap.c` in below patch "riscv/mm : ensure PROT_WRITE leads to VM_READ | VM_WRITE" - Lock interfaces for shadow stack and indirect branch tracking expect arg == 0 Any future evolution of this interface should accordingly define how arg should be setup. - `mm/map.c` has an instance of using `VM_SHADOW_STACK`. Fixed it to use helper `is_shadow_stack_vma`. - Link to v6: https://lore.kernel.org/r/20241008-v5_user_cfi_series-v6-0-60d9fe073f37@riv… v6: - Picked up Samuel Holland's changes as is with `envcfg` placed in `thread` instead of `thread_info` - fixed unaligned newline escapes in kselftest - cleaned up messages in kselftest and included test output in commit message - fixed a bug in clone path reported by Zong Li - fixed a build issue if CONFIG_RISCV_ISA_V is not selected (this was introduced due to re-factoring signal context management code) v5: - rebased on v6.12-rc1 - Fixed schema related issues in device tree file - Fixed some of the documentation related issues in zicfilp/ss.rst (style issues and added index) - added `SHADOW_STACK_SET_MARKER` so that implementation can define base of shadow stack. - Fixed warnings on definitions added in usercfi.h when CONFIG_RISCV_USER_CFI is not selected. - Adopted context header based signal handling as proposed by Andy Chiu - Added support for enabling kernel mode access to shadow stack using FWFT (https://github.com/riscv-non-isa/riscv-sbi-doc/blob/master/src/ext-firmware…) - Link to v5: https://lore.kernel.org/r/20241001-v5_user_cfi_series-v1-0-3ba65b6e550f@riv… (Note: I had an issue in my workflow due to which version number wasn't picked up correctly while sending out patches) v4: - rebased on 6.11-rc6 - envcfg: Converged with Samuel Holland's patches for envcfg management on per- thread basis. - vma_is_shadow_stack is renamed to is_vma_shadow_stack - picked up Mark Brown's `ARCH_HAS_USER_SHADOW_STACK` patch - signal context: using extended context management to maintain compatibility. - fixed `-Wmissing-prototypes` compiler warnings for prctl functions - Documentation fixes and amending typos. - Link to v4: https://lore.kernel.org/all/20240912231650.3740732-1-debug@rivosinc.com/ v3: - envcfg logic to pick up base envcfg had a bug where `ENVCFG_CBZE` could have been picked on per task basis, even though CPU didn't implement it. Fixed in this series. - dt-bindings As suggested, split into separate commit. fixed the messaging that spec is in public review - arch_is_shadow_stack change arch_is_shadow_stack changed to vma_is_shadow_stack - hwprobe zicfiss / zicfilp if present will get enumerated in hwprobe - selftests As suggested, added object and binary filenames to .gitignore Selftest binary anyways need to be compiled with cfi enabled compiler which will make sure that landing pad and shadow stack are enabled. Thus removed separate enable/disable tests. Cleaned up tests a bit. - Link to v3: https://lore.kernel.org/lkml/20240403234054.2020347-1-debug@rivosinc.com/ v2: - Using config `CONFIG_RISCV_USER_CFI`, kernel support for riscv control flow integrity for user mode programs can be compiled in the kernel. - Enabling of control flow integrity for user programs is left to user runtime - This patch series introduces arch agnostic `prctls` to enable shadow stack and indirect branch tracking. And implements them on riscv. --- Changes in v14: - changelog posted just below cover letter - Link to v13: https://lore.kernel.org/r/20250424-v5_user_cfi_series-v13-0-971437de586a@ri… Changes in v13: - changelog posted just below cover letter - Link to v12: https://lore.kernel.org/r/20250314-v5_user_cfi_series-v12-0-e51202b53138@ri… Changes in v12: - changelog posted just below cover letter - Link to v11: https://lore.kernel.org/r/20250310-v5_user_cfi_series-v11-0-86b36cbfb910@ri… Changes in v11: - changelog posted just below cover letter - Link to v10: https://lore.kernel.org/r/20250210-v5_user_cfi_series-v10-0-163dcfa31c60@ri… --- Andy Chiu (1): riscv: signal: abstract header saving for setup_sigcontext Deepak Gupta (25): mm: VM_SHADOW_STACK definition for riscv dt-bindings: riscv: zicfilp and zicfiss in dt-bindings (extensions.yaml) riscv: zicfiss / zicfilp enumeration riscv: zicfiss / zicfilp extension csr and bit definitions riscv: usercfi state for task and save/restore of CSR_SSP on trap entry/exit riscv/mm : ensure PROT_WRITE leads to VM_READ | VM_WRITE riscv mm: manufacture shadow stack pte riscv mmu: teach pte_mkwrite to manufacture shadow stack PTEs riscv mmu: write protect and shadow stack riscv/mm: Implement map_shadow_stack() syscall riscv/shstk: If needed allocate a new shadow stack on clone riscv: Implements arch agnostic shadow stack prctls prctl: arch-agnostic prctl for indirect branch tracking riscv: Implements arch agnostic indirect branch tracking prctls riscv/traps: Introduce software check exception riscv/signal: save and restore of shadow stack for signal riscv/kernel: update __show_regs to print shadow stack register riscv/ptrace: riscv cfi status and state via ptrace and in core files riscv/hwprobe: zicfilp / zicfiss enumeration in hwprobe riscv: kernel command line option to opt out of user cfi riscv: enable kernel access to shadow stack memory via FWFT sbi call riscv: create a config for shadow stack and landing pad instr support riscv: Documentation for landing pad / indirect branch tracking riscv: Documentation for shadow stack on riscv kselftest/riscv: kselftest for user mode cfi Jim Shu (1): arch/riscv: compile vdso with landing pad Documentation/admin-guide/kernel-parameters.txt | 8 + Documentation/arch/riscv/index.rst | 2 + Documentation/arch/riscv/zicfilp.rst | 115 +++++ Documentation/arch/riscv/zicfiss.rst | 179 +++++++ .../devicetree/bindings/riscv/extensions.yaml | 14 + arch/riscv/Kconfig | 20 + arch/riscv/Makefile | 5 +- arch/riscv/include/asm/asm-prototypes.h | 1 + arch/riscv/include/asm/assembler.h | 44 ++ arch/riscv/include/asm/cpufeature.h | 12 + arch/riscv/include/asm/csr.h | 16 + arch/riscv/include/asm/entry-common.h | 2 + arch/riscv/include/asm/hwcap.h | 2 + arch/riscv/include/asm/mman.h | 25 + arch/riscv/include/asm/mmu_context.h | 7 + arch/riscv/include/asm/pgtable.h | 30 +- arch/riscv/include/asm/processor.h | 2 + arch/riscv/include/asm/thread_info.h | 3 + arch/riscv/include/asm/usercfi.h | 95 ++++ arch/riscv/include/asm/vector.h | 3 + arch/riscv/include/uapi/asm/hwprobe.h | 2 + arch/riscv/include/uapi/asm/ptrace.h | 34 ++ arch/riscv/include/uapi/asm/sigcontext.h | 1 + arch/riscv/kernel/Makefile | 1 + arch/riscv/kernel/asm-offsets.c | 8 + arch/riscv/kernel/cpufeature.c | 13 + arch/riscv/kernel/entry.S | 33 +- arch/riscv/kernel/head.S | 23 + arch/riscv/kernel/process.c | 26 +- arch/riscv/kernel/ptrace.c | 95 ++++ arch/riscv/kernel/signal.c | 148 +++++- arch/riscv/kernel/sys_hwprobe.c | 2 + arch/riscv/kernel/sys_riscv.c | 10 + arch/riscv/kernel/traps.c | 43 ++ arch/riscv/kernel/usercfi.c | 545 +++++++++++++++++++++ arch/riscv/kernel/vdso/Makefile | 6 + arch/riscv/kernel/vdso/flush_icache.S | 4 + arch/riscv/kernel/vdso/getcpu.S | 4 + arch/riscv/kernel/vdso/rt_sigreturn.S | 4 + arch/riscv/kernel/vdso/sys_hwprobe.S | 4 + arch/riscv/mm/init.c | 2 +- arch/riscv/mm/pgtable.c | 17 + include/linux/cpu.h | 4 + include/linux/mm.h | 7 + include/uapi/linux/elf.h | 2 + include/uapi/linux/prctl.h | 27 + kernel/sys.c | 30 ++ tools/testing/selftests/riscv/Makefile | 2 +- tools/testing/selftests/riscv/cfi/.gitignore | 3 + tools/testing/selftests/riscv/cfi/Makefile | 10 + tools/testing/selftests/riscv/cfi/cfi_rv_test.h | 82 ++++ tools/testing/selftests/riscv/cfi/riscv_cfi_test.c | 171 +++++++ tools/testing/selftests/riscv/cfi/shadowstack.c | 385 +++++++++++++++ tools/testing/selftests/riscv/cfi/shadowstack.h | 27 + 54 files changed, 2331 insertions(+), 29 deletions(-) --- base-commit: 4181f8ad7a1061efed0219951d608d4988302af7 change-id: 20240930-v5_user_cfi_series-3dc332f8f5b2 -- - debug

5 months, 3 weeks

1
27
0 0

[PATCH AUTOSEL 6.1 08/10] selftests/bpf: Mitigate sockmap_ktls disconnect_after_delete failure

by Sasha Levin

From: Ihor Solodrai <ihor.solodrai(a)linux.dev> [ Upstream commit f2858f308131a09e33afb766cd70119b5b900569 ] "sockmap_ktls disconnect_after_delete" test has been failing on BPF CI after recent merges from netdev: * https://github.com/kernel-patches/bpf/actions/runs/14458537639 * https://github.com/kernel-patches/bpf/actions/runs/14457178732 It happens because disconnect has been disabled for TLS [1], and it renders the test case invalid. Removing all the test code creates a conflict between bpf and bpf-next, so for now only remove the offending assert [2]. The test will be removed later on bpf-next. [1] https://lore.kernel.org/netdev/20250404180334.3224206-1-kuba@kernel.org/ [2] https://lore.kernel.org/bpf/cfc371285323e1a3f3b006bfcf74e6cf7ad65258@linux.… Signed-off-by: Ihor Solodrai <ihor.solodrai(a)linux.dev> Signed-off-by: Andrii Nakryiko <andrii(a)kernel.org> Reviewed-by: Jiayuan Chen <jiayuan.chen(a)linux.dev> Link: https://lore.kernel.org/bpf/20250416170246.2438524-1-ihor.solodrai@linux.dev Signed-off-by: Alexei Starovoitov <ast(a)kernel.org> Signed-off-by: Sasha Levin <sashal(a)kernel.org> --- tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c | 1 - 1 file changed, 1 deletion(-) diff --git a/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c b/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c index 2d0796314862a..0a99fd404f6dc 100644 --- a/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c +++ b/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c @@ -68,7 +68,6 @@ static void test_sockmap_ktls_disconnect_after_delete(int family, int map) goto close_cli; err = disconnect(cli); - ASSERT_OK(err, "disconnect"); close_cli: close(cli); -- 2.39.5

5 months, 3 weeks

1
0
0 0

[PATCH bpf-next v1 0/3] bpf, sockmap: Improve performance with CPU affinity

by Jiayuan Chen

Abstract === This patchset improves the performance of sockmap by providing CPU affinity, resulting in a 1-10x increase in throughput. Motivation === Traditional user-space reverse proxy: Reserve Proxy _________________ client -> | fd1 <-> fd2 | -> server |_________________| Using sockmap for reverse proxy: Reserve Proxy _________________ client -> | fd1 <-> fd2 | -> server | |_________________| | | | | | | _________ | | | sockmap | | --> |_________| --> By adding fds to sockmap and using a BPF program, we can quickly forward data and avoid data copying between user space and kernel space. Mainstream multi-process reverse proxy applications, such as Nginx and HAProxy, support CPU affinity settings, which allow each process to be pinned to a specific CPU, avoiding conflicts between data plane processes and other processes, especially in multi-tenant environments. Current Issues === The current design of sockmap uses a workqueue to forward ingress_skb and wakes up the workqueue without specifying a CPU (by calling schedule_delayed_work()). In the current implementation of schedule_delayed_work, it tends to run the workqueue on the current CPU. This approach has a high probability of running on the current CPU, which is the same CPU that handles the net rx soft interrupt, especially for programs that access each other using local interfaces. The loopback driver's transmit interface, loopback_xmit(), directly calls __netif_rx() on the current CPU, which means that the CPU handling sockmap's workqueue and the client's sending CPU are the same, resulting in contention. For a TCP flow, if the request or response is very large, the psock->ingress_skb queue can become very long. When the workqueue traverses this queue to forward the data, it can consume a significant amount of CPU time. Solution === Configuring RPS on a loopback interface can be useful, but it will trigger additional softirq, and furthermore, it fails to achieve our expected effect of CPU isolation from other processes. Instead, we provide a kfunc that allow users to specify the CPU on which the workqueue runs through a BPF program. We can use the existing benchmark to test the performance, which allows us to evaluate the effectiveness of this optimization. Because we use local interfaces for communication and the client consumes a significant amount of CPU when sending data, this prevents the workqueue from processing ingress_skb in a timely manner, ultimately causing the server to fail to read data quickly. Without cpu-affinity: ./bench sockmap -c 2 -p 1 -a --rx-verdict-ingress --no-verify Setting up benchmark 'sockmap'... create socket fd c1:14 p1:15 c2:16 p2:17 Benchmark 'sockmap' started. Iter 0 ( 36.031us): Send Speed 1143.693 MB/s ... Rcv Speed 109.572 MB/s Iter 1 ( 0.608us): Send Speed 1320.550 MB/s ... Rcv Speed 48.103 MB/s Iter 2 ( -5.448us): Send Speed 1314.790 MB/s ... Rcv Speed 47.842 MB/s Iter 3 ( -0.613us): Send Speed 1320.158 MB/s ... Rcv Speed 46.531 MB/s Iter 4 ( -3.441us): Send Speed 1319.375 MB/s ... Rcv Speed 46.662 MB/s Iter 5 ( 3.764us): Send Speed 1166.667 MB/s ... Rcv Speed 42.467 MB/s Iter 6 ( -4.404us): Send Speed 1319.508 MB/s ... Rcv Speed 47.973 MB/s Summary: total trans 7758 MB ± 1293.506 MB/s Without cpu-affinity(RPS enabled): ./bench sockmap -c 2 -p 1 -a --rx-verdict-ingress --no-verify Setting up benchmark 'sockmap'... create socket fd c1:14 p1:15 c2:16 p2:17 Benchmark 'sockmap' started. Iter 0 ( 28.925us): Send Speed 1630.357 MB/s ... Rcv Speed 850.960 MB/s Iter 1 ( -2.042us): Send Speed 1644.564 MB/s ... Rcv Speed 822.478 MB/s Iter 2 ( 0.754us): Send Speed 1644.297 MB/s ... Rcv Speed 850.787 MB/s Iter 3 ( 0.159us): Send Speed 1644.429 MB/s ... Rcv Speed 850.198 MB/s Iter 4 ( -2.898us): Send Speed 1646.924 MB/s ... Rcv Speed 830.867 MB/s Iter 5 ( -0.210us): Send Speed 1649.410 MB/s ... Rcv Speed 824.246 MB/s Iter 6 ( -1.448us): Send Speed 1650.723 MB/s ... Rcv Speed 808.256 MB/s With cpu-affinity(RPS disabled): ./bench sockmap -c 2 -p 1 -a --rx-verdict-ingress --no-verify --cpu-affinity Setting up benchmark 'sockmap'... create socket fd c1:14 p1:15 c2:16 p2:17 Benchmark 'sockmap' started. Iter 0 ( 36.051us): Send Speed 1883.437 MB/s ... Rcv Speed 1865.087 MB/s Iter 1 ( 1.246us): Send Speed 1900.542 MB/s ... Rcv Speed 1761.737 MB/s Iter 2 ( -8.595us): Send Speed 1883.128 MB/s ... Rcv Speed 1860.714 MB/s Iter 3 ( 7.033us): Send Speed 1890.831 MB/s ... Rcv Speed 1806.684 MB/s Iter 4 ( -8.397us): Send Speed 1884.700 MB/s ... Rcv Speed 1973.568 MB/s Iter 5 ( -1.822us): Send Speed 1894.125 MB/s ... Rcv Speed 1775.046 MB/s Iter 6 ( 4.936us): Send Speed 1877.597 MB/s ... Rcv Speed 1959.320 MB/s Summary: total trans 11328 MB ± 1888.507 MB/s Appendix === Through this optimization, we discovered that sk_mem_charge() and sk_mem_uncharge() have concurrency issues. The performance improvement brought by this optimization has made these concurrency issues more evident. This concurrency issue can cause the WARN_ON_ONCE(sk->sk_forward_alloc) check to be triggered when the socket is released. Since this patch is a feature-type patch and does not intend to fix this bug, I will provide additional patches to fix this issue later. Jiayuan Chen (3): bpf, sockmap: Introduce a new kfunc for sockmap bpf, sockmap: Affinitize workqueue to a specific CPU selftest/bpf/benchs: Add cpu-affinity for sockmap bench Documentation/bpf/map_sockmap.rst | 14 +++++++ include/linux/skmsg.h | 15 +++++++ kernel/bpf/btf.c | 3 ++ net/core/skmsg.c | 10 +++-- net/core/sock_map.c | 39 +++++++++++++++++++ .../selftests/bpf/benchs/bench_sockmap.c | 35 +++++++++++++++-- tools/testing/selftests/bpf/bpf_kfuncs.h | 6 +++ .../selftests/bpf/progs/bench_sockmap_prog.c | 7 ++++ 8 files changed, 122 insertions(+), 7 deletions(-) -- 2.47.1

5 months, 3 weeks

3
8
0 0

[PATCH AUTOSEL 6.6 19/21] selftests/bpf: Mitigate sockmap_ktls disconnect_after_delete failure

by Sasha Levin

From: Ihor Solodrai <ihor.solodrai(a)linux.dev> [ Upstream commit f2858f308131a09e33afb766cd70119b5b900569 ] "sockmap_ktls disconnect_after_delete" test has been failing on BPF CI after recent merges from netdev: * https://github.com/kernel-patches/bpf/actions/runs/14458537639 * https://github.com/kernel-patches/bpf/actions/runs/14457178732 It happens because disconnect has been disabled for TLS [1], and it renders the test case invalid. Removing all the test code creates a conflict between bpf and bpf-next, so for now only remove the offending assert [2]. The test will be removed later on bpf-next. [1] https://lore.kernel.org/netdev/20250404180334.3224206-1-kuba@kernel.org/ [2] https://lore.kernel.org/bpf/cfc371285323e1a3f3b006bfcf74e6cf7ad65258@linux.… Signed-off-by: Ihor Solodrai <ihor.solodrai(a)linux.dev> Signed-off-by: Andrii Nakryiko <andrii(a)kernel.org> Reviewed-by: Jiayuan Chen <jiayuan.chen(a)linux.dev> Link: https://lore.kernel.org/bpf/20250416170246.2438524-1-ihor.solodrai@linux.dev Signed-off-by: Alexei Starovoitov <ast(a)kernel.org> Signed-off-by: Sasha Levin <sashal(a)kernel.org> --- tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c | 1 - 1 file changed, 1 deletion(-) diff --git a/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c b/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c index 2d0796314862a..0a99fd404f6dc 100644 --- a/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c +++ b/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c @@ -68,7 +68,6 @@ static void test_sockmap_ktls_disconnect_after_delete(int family, int map) goto close_cli; err = disconnect(cli); - ASSERT_OK(err, "disconnect"); close_cli: close(cli); -- 2.39.5

5 months, 3 weeks

1
0
0 0

[PATCH AUTOSEL 6.12 33/37] selftests/bpf: Mitigate sockmap_ktls disconnect_after_delete failure

by Sasha Levin

From: Ihor Solodrai <ihor.solodrai(a)linux.dev> [ Upstream commit f2858f308131a09e33afb766cd70119b5b900569 ] "sockmap_ktls disconnect_after_delete" test has been failing on BPF CI after recent merges from netdev: * https://github.com/kernel-patches/bpf/actions/runs/14458537639 * https://github.com/kernel-patches/bpf/actions/runs/14457178732 It happens because disconnect has been disabled for TLS [1], and it renders the test case invalid. Removing all the test code creates a conflict between bpf and bpf-next, so for now only remove the offending assert [2]. The test will be removed later on bpf-next. [1] https://lore.kernel.org/netdev/20250404180334.3224206-1-kuba@kernel.org/ [2] https://lore.kernel.org/bpf/cfc371285323e1a3f3b006bfcf74e6cf7ad65258@linux.… Signed-off-by: Ihor Solodrai <ihor.solodrai(a)linux.dev> Signed-off-by: Andrii Nakryiko <andrii(a)kernel.org> Reviewed-by: Jiayuan Chen <jiayuan.chen(a)linux.dev> Link: https://lore.kernel.org/bpf/20250416170246.2438524-1-ihor.solodrai@linux.dev Signed-off-by: Alexei Starovoitov <ast(a)kernel.org> Signed-off-by: Sasha Levin <sashal(a)kernel.org> --- tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c | 1 - 1 file changed, 1 deletion(-) diff --git a/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c b/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c index 2d0796314862a..0a99fd404f6dc 100644 --- a/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c +++ b/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c @@ -68,7 +68,6 @@ static void test_sockmap_ktls_disconnect_after_delete(int family, int map) goto close_cli; err = disconnect(cli); - ASSERT_OK(err, "disconnect"); close_cli: close(cli); -- 2.39.5

5 months, 3 weeks

1
0
0 0

[PATCH AUTOSEL 6.14 34/39] selftests/bpf: Mitigate sockmap_ktls disconnect_after_delete failure

by Sasha Levin

From: Ihor Solodrai <ihor.solodrai(a)linux.dev> [ Upstream commit f2858f308131a09e33afb766cd70119b5b900569 ] "sockmap_ktls disconnect_after_delete" test has been failing on BPF CI after recent merges from netdev: * https://github.com/kernel-patches/bpf/actions/runs/14458537639 * https://github.com/kernel-patches/bpf/actions/runs/14457178732 It happens because disconnect has been disabled for TLS [1], and it renders the test case invalid. Removing all the test code creates a conflict between bpf and bpf-next, so for now only remove the offending assert [2]. The test will be removed later on bpf-next. [1] https://lore.kernel.org/netdev/20250404180334.3224206-1-kuba@kernel.org/ [2] https://lore.kernel.org/bpf/cfc371285323e1a3f3b006bfcf74e6cf7ad65258@linux.… Signed-off-by: Ihor Solodrai <ihor.solodrai(a)linux.dev> Signed-off-by: Andrii Nakryiko <andrii(a)kernel.org> Reviewed-by: Jiayuan Chen <jiayuan.chen(a)linux.dev> Link: https://lore.kernel.org/bpf/20250416170246.2438524-1-ihor.solodrai@linux.dev Signed-off-by: Alexei Starovoitov <ast(a)kernel.org> Signed-off-by: Sasha Levin <sashal(a)kernel.org> --- tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c | 1 - 1 file changed, 1 deletion(-) diff --git a/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c b/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c index 2d0796314862a..0a99fd404f6dc 100644 --- a/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c +++ b/tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c @@ -68,7 +68,6 @@ static void test_sockmap_ktls_disconnect_after_delete(int family, int map) goto close_cli; err = disconnect(cli); - ASSERT_OK(err, "disconnect"); close_cli: close(cli); -- 2.39.5

5 months, 3 weeks

1
0
0 0

[PATCH net 0/4] Fix Felix DSA taprio gates after clock jump

by Vladimir Oltean

Richie Pearn presented a reproducible situation where traffic would get blocked on the NXP LS1028A switch if a certain taprio schedule was applied, and stepping the PTP clock would take place. The latter event is an expected initial occurrence, but also at runtime, for example when transitioning from one grandmaster to another. The issue is completely described in patch 1/4, which also contains the fix, but it has left me with some doubts regarding the need for vsc9959_tas_clock_adjust() in general. In order to prove to myself that vsc9959_tas_clock_adjust() is needed in general, I have written a selftest for the tc-taprio data path in patch 4/4. On the LS1028A, we can clearly see the following failures without that function: INFO: Forcing a backward clock jump TEST: ping [FAIL] INFO: Setting up taprio after PTP TEST: In band with gate [FAIL] Reception of 100 packets failed TEST: Out of band with gate [FAIL] Reception of 100 packets failed As for testing my fix from patch 1/4, that was quite a bit more complex to do automatically. In fact, I couldn't find any other schedule that would fail to be updated by vsc9959_tas_clock_adjust() as cleanly as the schedule from Richie, so I've added that specific schedule as the test_clock_jump_backward() test. The test ordering is also (unfortunately) very strategic. Running the selftest to the end dirties the GCL RAM, and when running test_clock_jump_backward() once again, the GCL entries won't be all zeroes as they were the first time around. They will contain bits and pieces of old schedules, making it very challenging to make it fail. Thus, test_clock_jump_backward() is the first in the test suite, and without patch 1/4, it is only supposed to fail the _first_ time when running after a clean boot. Vladimir Oltean (4): net: dsa: felix: fix broken taprio gate states after clock jump selftests: net: tsn_lib: create common helper for counting received packets selftests: net: tsn_lib: add window_size argument to isochron_do() selftests: net: tc_taprio: new test drivers/net/dsa/ocelot/felix_vsc9959.c | 5 +- .../selftests/drivers/net/dsa/tc_taprio.sh | 1 + .../selftests/drivers/net/ocelot/psfp.sh | 8 +- .../selftests/net/forwarding/tc_taprio.sh | 421 ++++++++++++++++++ .../selftests/net/forwarding/tsn_lib.sh | 26 ++ 5 files changed, 454 insertions(+), 7 deletions(-) create mode 120000 tools/testing/selftests/drivers/net/dsa/tc_taprio.sh create mode 100755 tools/testing/selftests/net/forwarding/tc_taprio.sh -- 2.43.0

5 months, 3 weeks

2
5
0 0

[PATCH bpf-next] selftests/bpf: Fix compilation errors

by Feng Yang

From: Feng Yang <yangfeng(a)kylinos.cn> If the CONFIG_NET_SCH_BPF configuration is not enabled, the BPF test compilation will report the following error: In file included from progs/bpf_qdisc_fq.c:39: progs/bpf_qdisc_common.h:17:51: error: declaration of 'struct bpf_sk_buff_ptr' will not be visible outside of this function [-Werror,-Wvisibility] 17 | void bpf_qdisc_skb_drop(struct sk_buff *p, struct bpf_sk_buff_ptr *to_free) __ksym; | ^ progs/bpf_qdisc_fq.c:309:14: error: declaration of 'struct bpf_sk_buff_ptr' will not be visible outside of this function [-Werror,-Wvisibility] 309 | struct bpf_sk_buff_ptr *to_free) | ^ progs/bpf_qdisc_fq.c:309:14: error: declaration of 'struct bpf_sk_buff_ptr' will not be visible outside of this function [-Werror,-Wvisibility] progs/bpf_qdisc_fq.c:308:5: error: conflicting types for '____bpf_fq_enqueue' Fixes: 11c701639ba9 ("selftests/bpf: Add a basic fifo qdisc test") Signed-off-by: Feng Yang <yangfeng(a)kylinos.cn> --- tools/testing/selftests/bpf/progs/bpf_qdisc_common.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tools/testing/selftests/bpf/progs/bpf_qdisc_common.h b/tools/testing/selftests/bpf/progs/bpf_qdisc_common.h index 65a2c561c0bb..7e7f2fe04f22 100644 --- a/tools/testing/selftests/bpf/progs/bpf_qdisc_common.h +++ b/tools/testing/selftests/bpf/progs/bpf_qdisc_common.h @@ -12,6 +12,8 @@ #define private(name) SEC(".data." #name) __hidden __attribute__((aligned(8))) +struct bpf_sk_buff_ptr; + u32 bpf_skb_get_hash(struct sk_buff *p) __ksym; void bpf_kfree_skb(struct sk_buff *p) __ksym; void bpf_qdisc_skb_drop(struct sk_buff *p, struct bpf_sk_buff_ptr *to_free) __ksym; -- 2.43.0

5 months, 3 weeks

2
1
0 0

[PATCH] selftests/bpf: Fix kmem_cache iterator draining

by T.J. Mercier

The closing parentheses around the read syscall is misplaced, causing single byte reads from the iterator instead of buf sized reads. While the end result is the same, many more read calls than necessary are performed. $ tools/testing/selftests/bpf/vmtest.sh "./test_progs -t kmem_cache_iter" 145/1 kmem_cache_iter/check_task_struct:OK 145/2 kmem_cache_iter/check_slabinfo:OK 145/3 kmem_cache_iter/open_coded_iter:OK 145 kmem_cache_iter:OK Summary: 1/3 PASSED, 0 SKIPPED, 0 FAILED Fixes: a496d0cdc84d ("selftests/bpf: Add a test for kmem_cache_iter") Signed-off-by: T.J. Mercier <tjmercier(a)google.com> --- tools/testing/selftests/bpf/prog_tests/kmem_cache_iter.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tools/testing/selftests/bpf/prog_tests/kmem_cache_iter.c b/tools/testing/selftests/bpf/prog_tests/kmem_cache_iter.c index 8e13a3416a21..1de14b111931 100644 --- a/tools/testing/selftests/bpf/prog_tests/kmem_cache_iter.c +++ b/tools/testing/selftests/bpf/prog_tests/kmem_cache_iter.c @@ -104,7 +104,7 @@ void test_kmem_cache_iter(void) goto destroy; memset(buf, 0, sizeof(buf)); - while (read(iter_fd, buf, sizeof(buf) > 0)) { + while (read(iter_fd, buf, sizeof(buf)) > 0) { /* Read out all contents */ printf("%s", buf); } base-commit: b4432656b36e5cc1d50a1f2dc15357543add530e -- 2.49.0.906.g1f30a19c02-goog

5 months, 3 weeks

4
4
0 0

[PATCH net-next] selftests: net: exit cleanly on SIGTERM / timeout

by Jakub Kicinski

ksft runner sends 2 SIGTERMs in a row if a test runs out of time. Handle this in a similar way we handle SIGINT - cleanup and stop running further tests. Because we get 2 signals we need a bit of logic to ignore the subsequent one, they come immediately one after the other (due to commit 9616cb34b08e ("kselftest/runner.sh: Propagate SIGTERM to runner child")). This change makes sure we run cleanup (scheduled defer()s) and also print a stack trace on SIGTERM, which doesn't happen by default. Tests occasionally hang in NIPA and it's impossible to tell what they are waiting from or doing. Signed-off-by: Jakub Kicinski <kuba(a)kernel.org> --- CC: petrm(a)nvidia.com CC: willemb(a)google.com CC: sdf(a)fomichev.me CC: linux-kselftest(a)vger.kernel.org --- tools/testing/selftests/net/lib/py/ksft.py | 27 +++++++++++++++++++++- 1 file changed, 26 insertions(+), 1 deletion(-) diff --git a/tools/testing/selftests/net/lib/py/ksft.py b/tools/testing/selftests/net/lib/py/ksft.py index 3cfad0fd4570..73710634d457 100644 --- a/tools/testing/selftests/net/lib/py/ksft.py +++ b/tools/testing/selftests/net/lib/py/ksft.py @@ -3,6 +3,7 @@ import builtins import functools import inspect +import signal import sys import time import traceback @@ -26,6 +27,10 @@ KSFT_DISRUPTIVE = True pass +class KsftTerminate(KeyboardInterrupt): + pass + + def ksft_pr(*objs, **kwargs): print("#", *objs, **kwargs) @@ -193,6 +198,19 @@ KSFT_DISRUPTIVE = True return env +term_cnt = 0 + +def _ksft_intr(signum, frame): + # ksft runner.sh sends 2 SIGTERMs in a row on a timeout + # if we don't ignore the second one it will stop us from handling cleanup + global term_cnt + term_cnt += 1 + if term_cnt == 1: + raise KsftTerminate() + else: + ksft_pr(f"Ignoring SIGTERM (cnt: {term_cnt}), already exiting...") + + def ksft_run(cases=None, globs=None, case_pfx=None, args=()): cases = cases or [] @@ -205,6 +223,10 @@ KSFT_DISRUPTIVE = True cases.append(value) break + global term_cnt + term_cnt = 0 + prev_sigterm = signal.signal(signal.SIGTERM, _ksft_intr) + totals = {"pass": 0, "fail": 0, "skip": 0, "xfail": 0} print("TAP version 13") @@ -229,11 +251,12 @@ KSFT_DISRUPTIVE = True cnt_key = 'xfail' except BaseException as e: stop |= isinstance(e, KeyboardInterrupt) + stop |= isinstance(e, KsftTerminate) tb = traceback.format_exc() for line in tb.strip().split('\n'): ksft_pr("Exception|", line) if stop: - ksft_pr("Stopping tests due to KeyboardInterrupt.") + ksft_pr(f"Stopping tests due to {type(e).__name__}.") KSFT_RESULT = False cnt_key = 'fail' @@ -248,6 +271,8 @@ KSFT_DISRUPTIVE = True if stop: break + signal.signal(signal.SIGTERM, prev_sigterm) + print( f"# Totals: pass:{totals['pass']} fail:{totals['fail']} xfail:{totals['xfail']} xpass:0 skip:{totals['skip']} error:0" ) -- 2.49.0

5 months, 3 weeks

3
5
0 0

2025

2024

2023

2022

2021

2020

2019

2018

2017

Linux-kselftest-mirror April 2025