[PATCH OLK-5.10 v7 RESEND 00/12] arm64: Support xcall prefetch
Support xcall prefetch. Changes in v7: - Remove "echo @xxx > /proc/$PID/xcall" usage, replace with "echo 1 > /proc/$PID/prefetch". Now the xcall enable and prefetch can be enabled separately. Changes in v6: - Only use one reserved kabi in task_struct for xcall. - Free prefetch item after eventpoll_release_file() to avoid possible competition between ep_prefetch_item_enqueue() and pfi free. - Fix "echo > /proc/xcall/cpu_list" show problem. Changes in v5: - Fix dentry free before free_prefetch_item() and then UAF in prefetch work by cancel_work_sync() before free pfi and move free_prefetch_item() to the first of __fput(). - Fix __queue_work() oops bug. - /proc/xcall/mask_list -> /proc/xcall/cpu_list - Fix xcall_prefetch print data type. - Update prefetch tracepoint, move xcall points definition to separete header file. - Make xcall_prefetch depends on EPOLL. - Adjust xcall prefetch buf to 16KB by default. - Add xcall_prefetch_init() function. - Only use write lock in free_prefetch_item(). - Simplify xcall_read() and add some comment. - Update commit message. Changes in v4: - Enable FAST_SYSCALL/IRQ and XCALL_PREFETCH by default. - Fix xint sleeping function called from invalid context bug. - free_prefetch_item() in file_free() instead of filp_close(), which fix the alloc-queue and free competition. - Fix kernel_read() warning for FMODE_READ not set. - Add sock_from_file() limit for prefetch. - Check NULL for rc_work. - Simplfy the cpumask interface code, and rename to "/proc/xcall/mask_list". - Remove the xcall_cache_pages_order interface. - tracepoint update: fd -> file. - Simplify the xcall_read() function again. - Handle copy_to_user() return value. - Remove unused XCALL_CACHE_QUEUED. - Update the commit message. Changes in v3: - Add XCALL_PREFETCH config to isolate feature code. - Split the /proc/xxx interface code out to independent patches, which will make it clear. - Update the cpumask interface to "/proc/xcall/numa_mask", and it can set the numa mask of all numas one time. - Add xcall select count to make xcall_cache_pages_order adjust safe. - Introduce xcall_read_start/end() to make it clear. - Simplify the xcall_read() function. - Use cpumask_next() instead of get_nth_cpu_in_cpumask() function. - Use independent cpu select policy function. - Remove some unnecessary pr_err(). - Update the commit message. Changes in v2: - Upadte the xcall prefetch state machine, remove the queued state and add prefetch, cancel states. - Remove the pfi lock and use atomic variables. - Change the 'xcall select' semantics and simplify the code a lot. - Remove keep_running, remove unrelated code. - Remove the count in struct read_cache_entry, and use percpu hit/miss count. - Remove sync mode, so remove struct read_cache_entry in struct task_struct. - Use hash table to find prefetch item for a file, which will not change the file struct KABI. - Use rwlock instead of spinlock for hash table. - Use alloc_page() instead kmalloc() to align 4KB. - Update the commit message. Jinjie Ruan (9): arm64: Introduce Xint software solution arm64: Add debugfs dir for xint arm64: Reserve a kabi in task_struct exclusively for xcall arm64: Introduce xcall prefetch enable interface eventpoll: xcall: Support async prefetch data in epoll xcall: Add /proc/xcall/prefetch dir for performance tuning xcall: Add /proc/xcall/cpu_list for performance tuning xcall: eventpoll: add tracepoint config: Enable FAST_SYSCALL/IRQ and XCALL_PREFETCH by default Liao Chen (1): revert kpti bypass Yipeng Zou (2): arm64: Introduce xcall a faster svc exception handling arm64: Faster SVC exception handler with xcall arch/Kconfig | 81 +++++ arch/arm64/Kconfig | 2 + arch/arm64/configs/openeuler_defconfig | 7 + arch/arm64/include/asm/cpucaps.h | 2 + arch/arm64/include/asm/exception.h | 3 + arch/arm64/kernel/asm-offsets.c | 3 + arch/arm64/kernel/cpufeature.c | 56 ++++ arch/arm64/kernel/entry-common.c | 22 ++ arch/arm64/kernel/entry.S | 183 ++++++++++- arch/arm64/kernel/syscall.c | 57 ++++ drivers/irqchip/irq-gic-v3.c | 130 ++++++++ fs/eventpoll.c | 408 +++++++++++++++++++++++++ fs/file_table.c | 1 + fs/proc/base.c | 183 +++++++++++ fs/read_write.c | 10 +- include/linux/fs.h | 35 +++ include/linux/hardirq.h | 5 + include/linux/irqchip/arm-gic-v3.h | 13 + include/linux/sched.h | 4 + include/linux/xcall.h | 17 ++ include/trace/events/xcall.h | 106 +++++++ kernel/fork.c | 24 ++ kernel/irq/debugfs.c | 33 ++ kernel/irq/internals.h | 18 ++ kernel/irq/irqdesc.c | 19 ++ kernel/irq/proc.c | 10 + kernel/softirq.c | 73 +++++ 27 files changed, 1502 insertions(+), 3 deletions(-) create mode 100644 include/linux/xcall.h create mode 100644 include/trace/events/xcall.h -- 2.34.1
From: Yipeng Zou <zouyipeng@huawei.com> hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IB6JLE -------------------------------- The svc exception handling process in ARM64, which includes auxiliary functions for debug/trace and core functions like KPTI, has been identified as overly "lengthy". This inefficiency is particularly noticeable in short syscalls such as lseek() and getpid(), where the syscall function itself comprises a small percentage of the total instructions executed. To address this, we introduce the concept of xcall, a fast svc exception handling path that only considers necessary features such as security, context saving, and recovery. This approach can be seen as a high-speed syscall processing mechanism bridging the gap between vdso and traditional syscalls. We've implemented a per-task bitmap to enable or disable xcall for specific syscalls. Users can enable a syscall with the following command: echo $syscall_nr > /proc/$PID/xcall To disable a syscall, use: echo \!$syscall_nr > /proc/$PID/xcall The current status of enabled syscalls can be viewed by: cat /proc/$PID/xcall Finally, we've added a kernel boot parameter to control the xcall feature. To enable xcall, include "xcall" in the kernel boot command. By default, xcall is disabled. This patch introduces basic framework and have not modified to syscall path only copy to xcall. Signed-off-by: Yipeng Zou <zouyipeng@huawei.com> --- arch/Kconfig | 19 +++++ arch/arm64/Kconfig | 1 + arch/arm64/configs/openeuler_defconfig | 2 + arch/arm64/include/asm/cpucaps.h | 1 + arch/arm64/kernel/asm-offsets.c | 3 + arch/arm64/kernel/cpufeature.c | 28 +++++++ arch/arm64/kernel/entry.S | 59 +++++++++++++ fs/proc/base.c | 112 +++++++++++++++++++++++++ include/linux/sched.h | 4 + kernel/fork.c | 20 +++++ 10 files changed, 249 insertions(+) diff --git a/arch/Kconfig b/arch/Kconfig index 0fc9c6d591b8..48ef789de86b 100644 --- a/arch/Kconfig +++ b/arch/Kconfig @@ -1186,4 +1186,23 @@ source "kernel/gcov/Kconfig" source "scripts/gcc-plugins/Kconfig" +config ARCH_SUPPORTS_FAST_SYSCALL + bool + +config FAST_SYSCALL + bool "Fast Syscall support" + depends on ARCH_SUPPORTS_FAST_SYSCALL + default n + help + This enable support Fast syscall feature. + The svc exception handling process, which includes auxiliary + functions for debug/trace and core functions like + KPTI, has been identified as overly "lengthy". + This inefficiency is particularly noticeable in short syscalls + such as lseek() and getpid(), where the syscall function itself + comprises a small percentage of the total instructions executed. + To address this, we introduce the concept of fast syscall, a fast svc + exception handling path that only considers necessary features + such as security, context saving, and recovery. + endmenu diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig index 93ced97f8c6c..d19304745fee 100644 --- a/arch/arm64/Kconfig +++ b/arch/arm64/Kconfig @@ -219,6 +219,7 @@ config ARM64 select THREAD_INFO_IN_TASK select HAVE_LIVEPATCH_WO_FTRACE select THP_NUMA_CONTROL if ARM64_64K_PAGES && NUMA_BALANCING && TRANSPARENT_HUGEPAGE + select ARCH_SUPPORTS_FAST_SYSCALL if !ARM64_MTE && !KASAN_HW_TAGS help ARM 64-bit (AArch64) Linux support. diff --git a/arch/arm64/configs/openeuler_defconfig b/arch/arm64/configs/openeuler_defconfig index 5449de73fbbc..8f649fdedfea 100644 --- a/arch/arm64/configs/openeuler_defconfig +++ b/arch/arm64/configs/openeuler_defconfig @@ -921,6 +921,8 @@ CONFIG_ARCH_HAS_GCOV_PROFILE_ALL=y # end of GCOV-based kernel profiling CONFIG_HAVE_GCC_PLUGINS=y +CONFIG_ARCH_SUPPORTS_FAST_SYSCALL=y +# CONFIG_FAST_SYSCALL is not set # end of General architecture-dependent options CONFIG_RT_MUTEXES=y diff --git a/arch/arm64/include/asm/cpucaps.h b/arch/arm64/include/asm/cpucaps.h index ce9fbf260a3c..5c4e78ffa264 100644 --- a/arch/arm64/include/asm/cpucaps.h +++ b/arch/arm64/include/asm/cpucaps.h @@ -81,6 +81,7 @@ #define ARM64_HAS_PBHA_STAGE2 73 #define ARM64_SME 74 #define ARM64_SME_FA64 75 +#define ARM64_HAS_XCALL 76 #define ARM64_NCAPS 80 diff --git a/arch/arm64/kernel/asm-offsets.c b/arch/arm64/kernel/asm-offsets.c index c247e11130db..7c6ad4b1667b 100644 --- a/arch/arm64/kernel/asm-offsets.c +++ b/arch/arm64/kernel/asm-offsets.c @@ -26,6 +26,9 @@ int main(void) { +#ifdef CONFIG_FAST_SYSCALL + DEFINE(TSK_XCALL, offsetof(struct task_struct, xcall_enable)); +#endif DEFINE(TSK_ACTIVE_MM, offsetof(struct task_struct, active_mm)); BLANK(); DEFINE(TSK_TI_CPU, offsetof(struct task_struct, thread_info.cpu)); diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c index dee049d27c74..f5ef453593ff 100644 --- a/arch/arm64/kernel/cpufeature.c +++ b/arch/arm64/kernel/cpufeature.c @@ -2155,6 +2155,26 @@ static bool can_clearpage_use_stnp(const struct arm64_cpu_capabilities *entry, return use_clearpage_stnp && has_mor_nontemporal(entry); } +#ifdef CONFIG_FAST_SYSCALL +static bool is_xcall_support; +static int __init xcall_setup(char *str) +{ + is_xcall_support = true; + return 1; +} +__setup("xcall", xcall_setup); + +bool fast_syscall_enabled(void) +{ + return is_xcall_support; +} + +static bool has_xcall_support(const struct arm64_cpu_capabilities *entry, int __unused) +{ + return is_xcall_support; +} +#endif + static const struct arm64_cpu_capabilities arm64_features[] = { { .desc = "GIC system register CPU interface", @@ -2701,6 +2721,14 @@ static const struct arm64_cpu_capabilities arm64_features[] = { .cpu_enable = fa64_kernel_enable, }, #endif /* CONFIG_ARM64_SME */ +#ifdef CONFIG_FAST_SYSCALL + { + .desc = "Xcall Support", + .capability = ARM64_HAS_XCALL, + .type = ARM64_CPUCAP_SYSTEM_FEATURE, + .matches = has_xcall_support, + }, +#endif {}, }; diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S index 1290f36c8371..e49b5569bb97 100644 --- a/arch/arm64/kernel/entry.S +++ b/arch/arm64/kernel/entry.S @@ -675,11 +675,70 @@ SYM_CODE_START_LOCAL_NOALIGN(el1_irq) kernel_exit 1 SYM_CODE_END(el1_irq) +#ifdef CONFIG_FAST_SYSCALL + .macro check_esr_el1_ec_svc64 + /* Only support SVC64 for now */ + mrs x20, esr_el1 + lsr w20, w20, #ESR_ELx_EC_SHIFT + cmp x20, #ESR_ELx_EC_SVC64 + .endm + + .macro check_syscall_nr + cmp x8, __NR_syscalls + .endm + + .macro check_xcall_enable + /* x21 = task_struct->xcall_enable */ + ldr_this_cpu x20, __entry_task, x21 + ldr x21, [x20, #TSK_XCALL] + /* x20 = sc_no / 8 */ + lsr x20, x8, 3 + ldr x21, [x21, x20] + /* x8 = sc_no % 8 */ + and x8, x8, 7 + mov x20, 1 + lsl x20, x20, x8 + and x21, x21, x20 + cmp x21, 0 + .endm + + .macro check_xcall_pre_kernel_entry + stp x20, x21, [sp, #0] + /* is ESR_ELx_EC_SVC64 */ + check_esr_el1_ec_svc64 + bne .Lskip_xcall\@ + /* x8 >= __NR_syscalls */ + check_syscall_nr + bhs .Lskip_xcall\@ + str x8, [sp, #16] + /* is xcall enabled */ + check_xcall_enable + ldr x8, [sp, #16] + beq .Lskip_xcall\@ + ldp x20, x21, [sp, #0] + /* do xcall */ + kernel_entry 0, 64 + mov x0, sp + bl el0t_64_sync_handler + b ret_to_user +.Lskip_xcall\@: + ldp x20, x21, [sp, #0] + .endm +#endif + /* * EL0 mode handlers. */ .align 6 SYM_CODE_START_LOCAL_NOALIGN(el0_sync) +#ifdef CONFIG_FAST_SYSCALL + /* Only support el0 aarch64 sync exception */ + alternative_if_not ARM64_HAS_XCALL + b .Lret_to_kernel_entry + alternative_else_nop_endif + check_xcall_pre_kernel_entry + .Lret_to_kernel_entry: +#endif kernel_entry 0 mov x0, sp bl el0_sync_handler diff --git a/fs/proc/base.c b/fs/proc/base.c index 4e0054a37c4c..3206960c4bd7 100644 --- a/fs/proc/base.c +++ b/fs/proc/base.c @@ -3589,6 +3589,115 @@ static const struct file_operations proc_pid_sg_level_operations = { }; #endif +#ifdef CONFIG_FAST_SYSCALL +bool fast_syscall_enabled(void); + +static int xcall_show(struct seq_file *m, void *v) +{ + struct inode *inode = m->private; + struct task_struct *p; + unsigned int rs, re; + + if (!fast_syscall_enabled()) + return -EACCES; + + p = get_proc_task(inode); + if (!p) + return -ESRCH; + + if (!p->xcall_enable) + goto out; + + seq_printf(m, "Enabled Total[%d/%d]:", bitmap_weight(p->xcall_enable, __NR_syscalls), + __NR_syscalls); + + for (rs = 0, bitmap_next_set_region(p->xcall_enable, &rs, &re, __NR_syscalls); + rs < re; rs = re + 1, + bitmap_next_set_region(p->xcall_enable, &rs, &re, __NR_syscalls)) { + rs == (re - 1) ? seq_printf(m, "%d,", rs) : + seq_printf(m, "%d-%d,", rs, re - 1); + } + seq_puts(m, "\n"); +out: + put_task_struct(p); + + return 0; +} + +static int xcall_open(struct inode *inode, struct file *filp) +{ + return single_open(filp, xcall_show, inode); +} + +static int xcall_enable_one(struct task_struct *p, unsigned int sc_no) +{ + bitmap_set(p->xcall_enable, sc_no, 1); + return 0; +} + +static int xcall_disable_one(struct task_struct *p, unsigned int sc_no) +{ + bitmap_clear(p->xcall_enable, sc_no, 1); + return 0; +} + +static ssize_t xcall_write(struct file *file, const char __user *buf, + size_t count, loff_t *offset) +{ + struct inode *inode = file_inode(file); + struct task_struct *p; + char buffer[TASK_COMM_LEN]; + const size_t maxlen = sizeof(buffer) - 1; + unsigned int sc_no = __NR_syscalls; + int ret = 0; + int is_clear = 0; + + if (!fast_syscall_enabled()) + return -EACCES; + + memset(buffer, 0, sizeof(buffer)); + if (!count || copy_from_user(buffer, buf, count > maxlen ? maxlen : count)) + return -EFAULT; + + p = get_proc_task(inode); + if (!p || !p->xcall_enable) + return -ESRCH; + + if (buffer[0] == '!') + is_clear = 1; + + if (kstrtouint(buffer + is_clear, 10, &sc_no)) { + ret = -EINVAL; + goto out; + } + + if (sc_no >= __NR_syscalls) { + ret = -EINVAL; + goto out; + } + + if (!is_clear && !test_bit(sc_no, p->xcall_enable)) + ret = xcall_enable_one(p, sc_no); + else if (is_clear && test_bit(sc_no, p->xcall_enable)) + ret = xcall_disable_one(p, sc_no); + else + ret = -EINVAL; + +out: + put_task_struct(p); + + return ret ? ret : count; +} + +static const struct file_operations proc_pid_xcall_operations = { + .open = xcall_open, + .read = seq_read, + .write = xcall_write, + .llseek = seq_lseek, + .release = single_release, +}; +#endif + /* * Thread groups */ @@ -3615,6 +3724,9 @@ static const struct pid_entry tgid_base_stuff[] = { #ifdef CONFIG_QOS_SCHED_SMART_GRID REG("smart_grid_level", 0644, proc_pid_sg_level_operations), #endif +#ifdef CONFIG_FAST_SYSCALL + REG("xcall", 0644, proc_pid_xcall_operations), +#endif #ifdef CONFIG_SCHED_AUTOGROUP REG("autogroup", S_IRUGO|S_IWUSR, proc_pid_sched_autogroup_operations), #endif diff --git a/include/linux/sched.h b/include/linux/sched.h index e3170b7f81fa..18361e35a377 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -1477,7 +1477,11 @@ struct task_struct { #else KABI_RESERVE(14) #endif +#if defined(CONFIG_FAST_SYSCALL) + KABI_USE(15, unsigned long *xcall_enable) +#else KABI_RESERVE(15) +#endif KABI_RESERVE(16) KABI_AUX_PTR(task_struct) diff --git a/kernel/fork.c b/kernel/fork.c index 9b1ea79deaa5..bd7afeb364ab 100644 --- a/kernel/fork.c +++ b/kernel/fork.c @@ -479,6 +479,12 @@ void free_task(struct task_struct *tsk) #endif if (task_relationship_used()) sched_relationship_free(tsk); + +#ifdef CONFIG_FAST_SYSCALL + if (tsk->xcall_enable) + bitmap_free(tsk->xcall_enable); +#endif + free_task_struct(tsk); } EXPORT_SYMBOL(free_task); @@ -1007,6 +1013,11 @@ static struct task_struct *dup_task_struct(struct task_struct *orig, int node) #ifdef CONFIG_MEMCG tsk->active_memcg = NULL; #endif + +#ifdef CONFIG_FAST_SYSCALL + tsk->xcall_enable = NULL; +#endif + return tsk; free_stack: @@ -2085,6 +2096,15 @@ static __latent_entropy struct task_struct *copy_process( rt_mutex_init_task(p); +#ifdef CONFIG_FAST_SYSCALL + p->xcall_enable = bitmap_zalloc(__NR_syscalls, GFP_KERNEL); + if (!p->xcall_enable) + goto bad_fork_free; + + if (current->xcall_enable) + bitmap_copy(p->xcall_enable, current->xcall_enable, __NR_syscalls); +#endif + #ifdef CONFIG_QOS_SCHED_DYNAMIC_AFFINITY retval = sched_prefer_cpus_fork(p, current->prefer_cpus); if (retval) -- 2.34.1
From: Yipeng Zou <zouyipeng@huawei.com> hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IB6JLE -------------------------------- This patch is designed to optimize the performance of the SVC exception handler by simplifying its operation, which can lead to faster execution times. However, this optimization comes with the trade-off of reduced functionality, particularly in areas related to security and maintenance. When a task is executed with xcall, certain features that are crucial for robust system operation may not be available, which could impact the system's ability to perform essential tasks. Here's a breakdown of the potential impacts: 1. Memory Tagging Extension (MTE) 2. Process Trace (PTRACE) 3. System Call Trace (STRACE) 4. GNU Debugger (GDB) 5. Software single-stepping 6. Secure State Buffer Descriptor (SSBD) 7. Shadow Call Stack 8. Software Translation Table Buffer Zero Protection (SW_TTBR0_PAN) 9. Unmap Kernel at Exception Level 0 (UNMAP_KERNEL_AT_EL0) 10.ARM64_WORKAROUND_SPECULATIVE_UNPRIV_LOAD 11. GCC Plugin Stack Leak Detection (GCC_PLUGIN_STACKLEAK) 12. SYSCALL Trace Point In conclusion, while the patch is intended to enhance the performance of the SVC exception handler, it does so by sacrificing several important features that contribute to security, debugging, and overall system stability. It is imperative for developers and system administrators to be cognizant of these trade-offs and to plan for the potential effects on their applications and operational workflows. Signed-off-by: Yipeng Zou <zouyipeng@huawei.com> Signed-off-by: Liao chen <liaochen4@huawei.com> Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com> --- arch/Kconfig | 24 +++++++++++++ arch/arm64/include/asm/exception.h | 3 ++ arch/arm64/kernel/entry-common.c | 20 +++++++++++ arch/arm64/kernel/entry.S | 51 ++++++++++++++++++++++++-- arch/arm64/kernel/syscall.c | 57 ++++++++++++++++++++++++++++++ 5 files changed, 152 insertions(+), 3 deletions(-) diff --git a/arch/Kconfig b/arch/Kconfig index 48ef789de86b..b80a22ed5545 100644 --- a/arch/Kconfig +++ b/arch/Kconfig @@ -1205,4 +1205,28 @@ config FAST_SYSCALL exception handling path that only considers necessary features such as security, context saving, and recovery. +config DEBUG_FEATURE_BYPASS + bool "Bypass debug feature in fast syscall" + depends on FAST_SYSCALL + default y + help + This to bypass debug feature in fast syscall. + The svc exception handling process, which includes auxiliary + functions for debug/trace and core functions like + KPTI, has been identified as overly "lengthy". + In fast syscall we only considers necessary features. + Disable this config to keep debug feature in fast syscall. + +config SECURITY_FEATURE_BYPASS + bool "Bypass security feature in fast syscall" + depends on FAST_SYSCALL + default y + help + This to bypass security feature in fast syscall. + The svc exception handling process, which includes auxiliary + functions for debug/trace and core functions like + KPTI, has been identified as overly "lengthy". + In fast syscall we only considers necessary features. + Disable this config to keep security feature in fast syscall. + endmenu diff --git a/arch/arm64/include/asm/exception.h b/arch/arm64/include/asm/exception.h index d38d526d084e..4b7994bd2b94 100644 --- a/arch/arm64/include/asm/exception.h +++ b/arch/arm64/include/asm/exception.h @@ -48,6 +48,9 @@ void do_el0_sys(unsigned long esr, struct pt_regs *regs); void do_sp_pc_abort(unsigned long addr, unsigned int esr, struct pt_regs *regs); void bad_el0_sync(struct pt_regs *regs, int reason, unsigned int esr); void do_el0_cp15(unsigned long esr, struct pt_regs *regs); +#ifdef CONFIG_FAST_SYSCALL +void do_el0_xcall(struct pt_regs *regs); +#endif void do_el0_svc(struct pt_regs *regs); void do_el0_svc_compat(struct pt_regs *regs); void do_el0_fpac(struct pt_regs *regs, unsigned long esr); diff --git a/arch/arm64/kernel/entry-common.c b/arch/arm64/kernel/entry-common.c index 02cd5d57edb6..e567484fc662 100644 --- a/arch/arm64/kernel/entry-common.c +++ b/arch/arm64/kernel/entry-common.c @@ -388,6 +388,26 @@ static void noinstr el0_fpac(struct pt_regs *regs, unsigned long esr) do_el0_fpac(regs, esr); } +#ifdef CONFIG_FAST_SYSCALL +asmlinkage void noinstr fast_enter_from_user_mode(void) +{ +#ifndef CONFIG_DEBUG_FEATURE_BYPASS + lockdep_hardirqs_off(CALLER_ADDR0); + CT_WARN_ON(ct_state() != CONTEXT_USER); +#endif + user_exit_irqoff(); +#ifndef CONFIG_DEBUG_FEATURE_BYPASS + trace_hardirqs_off_finish(); +#endif +} + +asmlinkage void noinstr el0_xcall_handler(struct pt_regs *regs) +{ + fast_enter_from_user_mode(); + do_el0_xcall(regs); +} +#endif + asmlinkage void noinstr el0_sync_handler(struct pt_regs *regs) { unsigned long esr = read_sysreg(esr_el1); diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S index e49b5569bb97..e99a60fce105 100644 --- a/arch/arm64/kernel/entry.S +++ b/arch/arm64/kernel/entry.S @@ -182,7 +182,7 @@ alternative_else_nop_endif #endif .endm - .macro kernel_entry, el, regsize = 64 + .macro kernel_entry, el, regsize = 64, fast_mode = std .if \regsize == 32 mov w0, w0 // zero upper 32 bits of x0 .endif @@ -212,12 +212,19 @@ alternative_else_nop_endif * Ensure MDSCR_EL1.SS is clear, since we can unmask debug exceptions * when scheduling. */ + .if \fast_mode == std ldr x19, [tsk, #TSK_TI_FLAGS] disable_step_tsk x19, x20 + .endif /* Check for asynchronous tag check faults in user space */ + .if \fast_mode == std check_mte_async_tcf x22, x23 + .endif + + .if \fast_mode == std apply_ssbd 1, x22, x23 + .endif ptrauth_keys_install_kernel tsk, x20, x22, x23 @@ -243,9 +250,11 @@ alternative_else_nop_endif add x29, sp, #S_STACKFRAME #ifdef CONFIG_ARM64_SW_TTBR0_PAN +.if \fast_mode == std alternative_if_not ARM64_HAS_PAN bl __swpan_entry_el\el alternative_else_nop_endif +.endif #endif stp x22, x23, [sp, #S_PC] @@ -268,9 +277,11 @@ alternative_else_nop_endif /* Re-enable tag checking (TCO set on exception entry) */ #ifdef CONFIG_ARM64_MTE +.if \fast_mode == std alternative_if ARM64_MTE SET_PSTATE_TCO(0) alternative_else_nop_endif +.endif #endif /* @@ -283,7 +294,7 @@ alternative_else_nop_endif */ .endm - .macro kernel_exit, el + .macro kernel_exit, el, fast_mode = std .if \el != 0 disable_daif .endif @@ -303,14 +314,18 @@ alternative_else_nop_endif ldp x21, x22, [sp, #S_PC] // load ELR, SPSR #ifdef CONFIG_ARM64_SW_TTBR0_PAN +.if \fast_mode == std alternative_if_not ARM64_HAS_PAN bl __swpan_exit_el\el alternative_else_nop_endif +.endif #endif .if \el == 0 ldr x23, [sp, #S_SP] // load return stack pointer msr sp_el0, x23 + + .if \fast_mode == std tst x22, #PSR_MODE32_BIT // native task? b.eq 3f @@ -325,13 +340,17 @@ alternative_if ARM64_WORKAROUND_845719 alternative_else_nop_endif #endif 3: + .endif + scs_save tsk, x0 /* No kernel C function calls after this as user keys are set. */ ptrauth_keys_install_user tsk, x0, x1, x2 + .if \fast_mode == std apply_ssbd 0, x0, x1 .endif + .endif msr elr_el1, x21 // set up the return data msr spsr_el1, x22 @@ -352,6 +371,11 @@ alternative_else_nop_endif ldp x28, x29, [sp, #16 * 14] .if \el == 0 + .if \fast_mode != std + ldr lr, [sp, #S_LR] + add sp, sp, #S_FRAME_SIZE // restore sp + eret + .endif alternative_if_not ARM64_UNMAP_KERNEL_AT_EL0 ldr lr, [sp, #S_LR] add sp, sp, #S_FRAME_SIZE // restore sp @@ -717,10 +741,31 @@ SYM_CODE_END(el1_irq) beq .Lskip_xcall\@ ldp x20, x21, [sp, #0] /* do xcall */ +#ifdef CONFIG_SECURITY_FEATURE_BYPASS + kernel_entry 0, 64, xcall +#else kernel_entry 0, 64 +#endif mov x0, sp - bl el0t_64_sync_handler + bl el0_xcall_handler +#ifdef CONFIG_SECURITY_FEATURE_BYPASS + disable_daif + gic_prio_kentry_setup tmp=x3 + ldr x19, [tsk, #TSK_TI_FLAGS] + and x2, x19, #_TIF_WORK_MASK + cbnz x2, fast_work_pending\@ +fast_finish_ret_to_user\@: + user_enter_irqoff + kernel_exit 0 xcall +fast_work_pending\@: + mov x0, sp // 'regs' + mov x1, x19 + bl do_notify_resume + ldr x19, [tsk, #TSK_TI_FLAGS] // re-check for single-step + b fast_finish_ret_to_user\@ +#else b ret_to_user +#endif .Lskip_xcall\@: ldp x20, x21, [sp, #0] .endm diff --git a/arch/arm64/kernel/syscall.c b/arch/arm64/kernel/syscall.c index 9bd304568d90..2d73eaaf9bc2 100644 --- a/arch/arm64/kernel/syscall.c +++ b/arch/arm64/kernel/syscall.c @@ -106,6 +106,46 @@ static void cortex_a76_erratum_1463225_svc_handler(void) static void cortex_a76_erratum_1463225_svc_handler(void) { } #endif /* CONFIG_ARM64_ERRATUM_1463225 */ +#ifdef CONFIG_FAST_SYSCALL +static void el0_xcall_common(struct pt_regs *regs, int scno, int sc_nr, + const syscall_fn_t syscall_table[]) +{ + unsigned long flags = read_thread_flags(); + + regs->orig_x0 = regs->regs[0]; + regs->syscallno = scno; + +#ifndef CONFIG_SECURITY_FEATURE_BYPASS + cortex_a76_erratum_1463225_svc_handler(); +#endif + local_daif_restore(DAIF_PROCCTX); + + if (system_supports_mte() && (flags & _TIF_MTE_ASYNC_FAULT)) { + syscall_set_return_value(current, regs, -ERESTARTNOINTR, 0); + return; + } + + if (has_syscall_work(flags)) { + if (scno == NO_SYSCALL) + syscall_set_return_value(current, regs, -ENOSYS, 0); + scno = syscall_trace_enter(regs); + if (scno == NO_SYSCALL) + goto trace_exit; + } + + invoke_syscall(regs, scno, sc_nr, syscall_table); + + if (!has_syscall_work(flags) && !IS_ENABLED(CONFIG_DEBUG_RSEQ)) { + flags = read_thread_flags(); + if (!has_syscall_work(flags) && !(flags & _TIF_SINGLESTEP)) + return; + } + +trace_exit: + syscall_trace_exit(regs); +} +#endif + static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr, const syscall_fn_t syscall_table[]) { @@ -237,6 +277,23 @@ static inline void delouse_pt_regs(struct pt_regs *regs) } #endif +#ifdef CONFIG_FAST_SYSCALL +void do_el0_xcall(struct pt_regs *regs) +{ + const syscall_fn_t *t = sys_call_table; + +#ifdef CONFIG_ARM64_ILP32 + if (is_ilp32_compat_task()) { + t = ilp32_sys_call_table; + delouse_pt_regs(regs); + } +#endif + + fp_user_discard(); + el0_xcall_common(regs, regs->regs[8], __NR_syscalls, t); +} +#endif + void do_el0_svc(struct pt_regs *regs) { const syscall_fn_t *t = sys_call_table; -- 2.34.1
hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IB6JLE -------------------------------- Introduce xint software solution for kernel, it provides a lightweight interrupt processing framework for latency-sensitive interrupts, and enabled dynamically for each irq by /proc/irq/<irq>/xint interface. The main implementation schemes are as follows: 1. For a small number of latency-sensitive interrupts, it could be configured as xint state, and process irq by xint framework instead of the kernel general interrupt framework, so improve performance by remove unnecessary processes. It is not recommended to configure too many interrupts as xint in the system, as this will affect system stability to some extent. 2. For each SGI/PPI/SPI interrupts whoes irq numbers are consecutive and limited, use a bitmap to check whether a hwirq is xint. Signed-off-by: Zhang Jianhua <chris.zjh@huawei.com> Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com> Signed-off-by: Liao Chen <liaochen4@huawei.com> --- arch/Kconfig | 25 +++++- arch/arm64/Kconfig | 1 + arch/arm64/configs/openeuler_defconfig | 2 + arch/arm64/include/asm/cpucaps.h | 1 + arch/arm64/kernel/cpufeature.c | 26 ++++++ arch/arm64/kernel/entry-common.c | 4 +- arch/arm64/kernel/entry.S | 80 +++++++++++++++++ drivers/irqchip/irq-gic-v3.c | 117 +++++++++++++++++++++++++ include/linux/hardirq.h | 5 ++ include/linux/irqchip/arm-gic-v3.h | 13 +++ kernel/irq/irqdesc.c | 19 ++++ kernel/irq/proc.c | 10 +++ kernel/softirq.c | 73 +++++++++++++++ 13 files changed, 373 insertions(+), 3 deletions(-) diff --git a/arch/Kconfig b/arch/Kconfig index b80a22ed5545..6dc501a4afb1 100644 --- a/arch/Kconfig +++ b/arch/Kconfig @@ -1205,9 +1205,30 @@ config FAST_SYSCALL exception handling path that only considers necessary features such as security, context saving, and recovery. +config ARCH_SUPPORTS_FAST_IRQ + bool + +config FAST_IRQ + bool "Fast irq support" + depends on ARCH_SUPPORTS_FAST_IRQ + default n + help + The irq handling process, which includes auxiliary + functions for debug/trace and core functions like + KPTI, interrupt time record, interrupt processing as + a random number source, interrupt affinity + modification and interrupt processing race, as well as + spurious and unhandled interrupt debugging, has been + identified as overly "lengthy". + To address this, we introduce the concept of fast irq, + a fast interrupt handling path that only considers + necessary features such as security, context saving + and recovery, which adds an lightweight interrupt processing + framework for latency-sensitive interrupts. + config DEBUG_FEATURE_BYPASS bool "Bypass debug feature in fast syscall" - depends on FAST_SYSCALL + depends on FAST_SYSCALL || FAST_IRQ default y help This to bypass debug feature in fast syscall. @@ -1219,7 +1240,7 @@ config DEBUG_FEATURE_BYPASS config SECURITY_FEATURE_BYPASS bool "Bypass security feature in fast syscall" - depends on FAST_SYSCALL + depends on FAST_SYSCALL || FAST_IRQ default y help This to bypass security feature in fast syscall. diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig index d19304745fee..ad38ba5be590 100644 --- a/arch/arm64/Kconfig +++ b/arch/arm64/Kconfig @@ -220,6 +220,7 @@ config ARM64 select HAVE_LIVEPATCH_WO_FTRACE select THP_NUMA_CONTROL if ARM64_64K_PAGES && NUMA_BALANCING && TRANSPARENT_HUGEPAGE select ARCH_SUPPORTS_FAST_SYSCALL if !ARM64_MTE && !KASAN_HW_TAGS + select ARCH_SUPPORTS_FAST_IRQ if ARM_GIC_V3 && !ARM64_MTE && !KASAN_HW_TAGS help ARM 64-bit (AArch64) Linux support. diff --git a/arch/arm64/configs/openeuler_defconfig b/arch/arm64/configs/openeuler_defconfig index 8f649fdedfea..576da6c47caf 100644 --- a/arch/arm64/configs/openeuler_defconfig +++ b/arch/arm64/configs/openeuler_defconfig @@ -923,6 +923,8 @@ CONFIG_ARCH_HAS_GCOV_PROFILE_ALL=y CONFIG_HAVE_GCC_PLUGINS=y CONFIG_ARCH_SUPPORTS_FAST_SYSCALL=y # CONFIG_FAST_SYSCALL is not set +CONFIG_ARCH_SUPPORTS_FAST_IRQ=y +# CONFIG_FAST_IRQ is not set # end of General architecture-dependent options CONFIG_RT_MUTEXES=y diff --git a/arch/arm64/include/asm/cpucaps.h b/arch/arm64/include/asm/cpucaps.h index 5c4e78ffa264..e2a2b3e40c94 100644 --- a/arch/arm64/include/asm/cpucaps.h +++ b/arch/arm64/include/asm/cpucaps.h @@ -82,6 +82,7 @@ #define ARM64_SME 74 #define ARM64_SME_FA64 75 #define ARM64_HAS_XCALL 76 +#define ARM64_HAS_XINT 77 #define ARM64_NCAPS 80 diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c index f5ef453593ff..9b4a315e96bc 100644 --- a/arch/arm64/kernel/cpufeature.c +++ b/arch/arm64/kernel/cpufeature.c @@ -2175,6 +2175,24 @@ static bool has_xcall_support(const struct arm64_cpu_capabilities *entry, int __ } #endif +#ifdef CONFIG_FAST_IRQ +bool is_xint_support; +static int __init xint_setup(char *str) +{ + if (!cpus_have_cap(ARM64_HAS_SYSREG_GIC_CPUIF)) + return 1; + + is_xint_support = true; + return 1; +} +__setup("xint", xint_setup); + +static bool has_xint_support(const struct arm64_cpu_capabilities *entry, int __unused) +{ + return is_xint_support; +} +#endif + static const struct arm64_cpu_capabilities arm64_features[] = { { .desc = "GIC system register CPU interface", @@ -2728,6 +2746,14 @@ static const struct arm64_cpu_capabilities arm64_features[] = { .type = ARM64_CPUCAP_SYSTEM_FEATURE, .matches = has_xcall_support, }, +#endif +#ifdef CONFIG_FAST_IRQ + { + .desc = "Xint Support", + .capability = ARM64_HAS_XINT, + .type = ARM64_CPUCAP_SYSTEM_FEATURE, + .matches = has_xint_support, + }, #endif {}, }; diff --git a/arch/arm64/kernel/entry-common.c b/arch/arm64/kernel/entry-common.c index e567484fc662..3e59cbdedc4c 100644 --- a/arch/arm64/kernel/entry-common.c +++ b/arch/arm64/kernel/entry-common.c @@ -388,7 +388,7 @@ static void noinstr el0_fpac(struct pt_regs *regs, unsigned long esr) do_el0_fpac(regs, esr); } -#ifdef CONFIG_FAST_SYSCALL +#if defined(CONFIG_FAST_SYSCALL) || defined(CONFIG_FAST_IRQ) asmlinkage void noinstr fast_enter_from_user_mode(void) { #ifndef CONFIG_DEBUG_FEATURE_BYPASS @@ -400,7 +400,9 @@ asmlinkage void noinstr fast_enter_from_user_mode(void) trace_hardirqs_off_finish(); #endif } +#endif +#ifdef CONFIG_FAST_SYSCALL asmlinkage void noinstr el0_xcall_handler(struct pt_regs *regs) { fast_enter_from_user_mode(); diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S index e99a60fce105..664a0d3059ab 100644 --- a/arch/arm64/kernel/entry.S +++ b/arch/arm64/kernel/entry.S @@ -809,10 +809,90 @@ SYM_CODE_START_LOCAL_NOALIGN(el0_error_compat) kernel_entry 0, 32 b el0_error_naked SYM_CODE_END(el0_error_compat) +#endif + +#ifdef CONFIG_FAST_IRQ +.macro el0_xint_handler, handler:req +#if defined(CONFIG_CONTEXT_TRACKING) || defined(CONFIG_TRACE_IRQFLAGS) + bl fast_enter_from_user_mode +#endif + enable_da_f +#ifndef CONFIG_SECURITY_FEATURE_BYPASS + tbz x22, #55, 1f + bl do_el0_irq_bp_hardening +1: +#endif + irq_handler \handler +.endm + +.macro check_xint_pre_kernel_entry + stp x0, x1, [sp, #0] + stp x2, x3, [sp, #16] + + ldr x0, =irqnr_xint_map + /* get hpp irqnr */ + mrs_s x1, SYS_ICC_HPPIR1_EL1 + + /* xint hwirq can not exceed 1020 */ + cmp x1, 1020 + b.ge .Lskip_xint\@ + + /* x2 = irqnr % 8 */ + and x2, x1, #7 + /* x3 = irqnr / 8 */ + lsr x3, x1, #3 + /* x1 is the byte of irqnr in irqnr_xint_map */ + ldr x1, [x0, x3] + + /* Get the check mask */ + mov x3, #1 + /* x3 = 1 << (irqnr % 8) */ + lsl x3, x3, x2 + + /* x1 = x1 & x3 */ + ands x1, x1, x3 + b.eq .Lskip_xint\@ + + ldp x0, x1, [sp, #0] + ldp x2, x3, [sp, #16] +#ifdef CONFIG_SECURITY_FEATURE_BYPASS + kernel_entry 0, 64, xint + el0_xint_handler handle_arch_irq + disable_daif + gic_prio_kentry_setup tmp=x3 + ldr x19, [tsk, #TSK_TI_FLAGS] + and x2, x19, #_TIF_WORK_MASK + cbnz x2, xint_fast_work_pending\@ +xint_fast_finish_ret_to_user\@: + user_enter_irqoff + kernel_exit 0 xint +xint_fast_work_pending\@: + mov x0, sp // 'regs' + mov x1, x19 + bl do_notify_resume + b xint_fast_finish_ret_to_user\@ +#else + kernel_entry 0, 64 + el0_xint_handler handle_arch_irq + b ret_to_user +#endif + +.Lskip_xint\@: + ldp x0, x1, [sp, #0] + ldp x2, x3, [sp, #16] +.endm #endif .align 6 SYM_CODE_START_LOCAL_NOALIGN(el0_irq) +#ifdef CONFIG_FAST_IRQ + /* Only support el0 aarch64 irq */ + alternative_if_not ARM64_HAS_XINT + b .Lskip_check_xint + alternative_else_nop_endif + check_xint_pre_kernel_entry +.Lskip_check_xint: +#endif kernel_entry 0 el0_irq_naked: el0_interrupt_handler handle_arch_irq diff --git a/drivers/irqchip/irq-gic-v3.c b/drivers/irqchip/irq-gic-v3.c index 87af452d82dc..e2aef650a884 100644 --- a/drivers/irqchip/irq-gic-v3.c +++ b/drivers/irqchip/irq-gic-v3.c @@ -720,6 +720,123 @@ static asmlinkage void __exception_irq_entry gic_handle_irq(struct pt_regs *regs } } +#ifdef CONFIG_FAST_IRQ +DECLARE_BITMAP(irqnr_xint_map, 1024); + +static bool can_set_xint(unsigned int hwirq) +{ + if (__get_intid_range(hwirq) == SGI_RANGE || + __get_intid_range(hwirq) == SPI_RANGE) + return true; + + return false; +} + +static bool xint_transform(int irqno, enum xint_op op) +{ + struct irq_data *data = irq_get_irq_data(irqno); + int hwirq; + + while (data->parent_data) + data = data->parent_data; + + hwirq = data->hwirq; + + if (!can_set_xint(hwirq)) + return false; + + switch (op) { + case IRQ_TO_XINT: + set_bit(hwirq, irqnr_xint_map); + return true; + case XINT_TO_IRQ: + clear_bit(hwirq, irqnr_xint_map); + return false; + case XINT_SET_CHECK: + return test_bit(hwirq, irqnr_xint_map); + case XINT_RANGE_CHECK: + return true; + default: + return false; + } +} + +static ssize_t xint_proc_write(struct file *file, + const char __user *buffer, size_t count, loff_t *pos) +{ + int irq = (int)(long)PDE_DATA(file_inode(file)); + bool xint_state = false; + unsigned long val; + char *buf = NULL; + + if (!xint_transform(irq, XINT_RANGE_CHECK)) + return -EPERM; + + buf = memdup_user_nul(buffer, count); + if (IS_ERR(buf)) + return PTR_ERR(buf); + + if (kstrtoul(buf, 0, &val) || (val != 0 && val != 1)) { + kfree(buf); + return -EINVAL; + } + + xint_state = xint_transform(irq, XINT_SET_CHECK); + if (xint_state == val) { + kfree(buf); + return -EBUSY; + } + + local_irq_disable(); + disable_irq(irq); + + xint_transform(irq, xint_state ? XINT_TO_IRQ : IRQ_TO_XINT); + + enable_irq(irq); + local_irq_enable(); + + kfree(buf); + + return count; +} + +static int xint_proc_show(struct seq_file *m, void *v) +{ + seq_printf(m, "%d\n", xint_transform((long)m->private, XINT_SET_CHECK)); + return 0; +} + +static int xint_proc_open(struct inode *inode, struct file *file) +{ + return single_open(file, xint_proc_show, PDE_DATA(inode)); +} + +static const struct proc_ops xint_proc_ops = { + .proc_open = xint_proc_open, + .proc_read = seq_read, + .proc_lseek = seq_lseek, + .proc_release = single_release, + .proc_write = xint_proc_write, +}; + +void register_irqchip_proc(struct irq_desc *desc, void *irqp) +{ + if (!is_xint_support) + return; + + /* create /proc/irq/<irq>/xint */ + proc_create_data("xint", 0644, desc->dir, &xint_proc_ops, irqp); +} + +void unregister_irqchip_proc(struct irq_desc *desc) +{ + if (!is_xint_support) + return; + + remove_proc_entry("xint", desc->dir); +} +#endif /* CONFIG_FAST_IRQ */ + static u32 gic_get_pribits(void) { u32 pribits; diff --git a/include/linux/hardirq.h b/include/linux/hardirq.h index 754f67ac4326..ad08a37f3bc0 100644 --- a/include/linux/hardirq.h +++ b/include/linux/hardirq.h @@ -86,6 +86,11 @@ void irq_exit(void); */ void irq_exit_rcu(void); +#ifdef CONFIG_FAST_IRQ +void xint_enter(void); +void xint_exit(void); +#endif + #ifndef arch_nmi_enter #define arch_nmi_enter() do { } while (0) #define arch_nmi_exit() do { } while (0) diff --git a/include/linux/irqchip/arm-gic-v3.h b/include/linux/irqchip/arm-gic-v3.h index 88b02e3b81da..d94b013a091c 100644 --- a/include/linux/irqchip/arm-gic-v3.h +++ b/include/linux/irqchip/arm-gic-v3.h @@ -762,6 +762,19 @@ static inline enum gic_intid_range __get_intid_range(irq_hw_number_t hwirq) } } +#ifdef CONFIG_FAST_IRQ +extern bool is_xint_support; + +enum xint_op { + XINT_TO_IRQ, + IRQ_TO_XINT, + XINT_SET_CHECK, + XINT_RANGE_CHECK, +}; + +void register_irqchip_proc(struct irq_desc *desc, void *irqp); +void unregister_irqchip_proc(struct irq_desc *desc); +#endif #endif #endif diff --git a/kernel/irq/irqdesc.c b/kernel/irq/irqdesc.c index 6c009a033c73..5dc976d32c74 100644 --- a/kernel/irq/irqdesc.c +++ b/kernel/irq/irqdesc.c @@ -658,6 +658,10 @@ int generic_handle_irq(unsigned int irq) EXPORT_SYMBOL_GPL(generic_handle_irq); #ifdef CONFIG_HANDLE_DOMAIN_IRQ +#ifdef CONFIG_FAST_IRQ +extern DECLARE_BITMAP(irqnr_xint_map, 1024); +#endif + /** * __handle_domain_irq - Invoke the handler for a HW irq belonging to a domain * @domain: The domain where to perform the lookup @@ -673,8 +677,16 @@ int __handle_domain_irq(struct irq_domain *domain, unsigned int hwirq, struct pt_regs *old_regs = set_irq_regs(regs); unsigned int irq = hwirq; int ret = 0; +#ifdef CONFIG_FAST_IRQ + int is_xint = test_bit(hwirq, irqnr_xint_map); + if (is_xint) + xint_enter(); + else + irq_enter(); +#else irq_enter(); +#endif #ifdef CONFIG_IRQ_DOMAIN if (lookup) @@ -692,7 +704,14 @@ int __handle_domain_irq(struct irq_domain *domain, unsigned int hwirq, generic_handle_irq(irq); } +#ifdef CONFIG_FAST_IRQ + if (is_xint) + xint_exit(); + else + irq_exit(); +#else irq_exit(); +#endif set_irq_regs(old_regs); return ret; } diff --git a/kernel/irq/proc.c b/kernel/irq/proc.c index 0df62a3a1f37..64805a37c576 100644 --- a/kernel/irq/proc.c +++ b/kernel/irq/proc.c @@ -13,6 +13,10 @@ #include <linux/kernel_stat.h> #include <linux/mutex.h> +#ifdef CONFIG_FAST_IRQ +#include <linux/irqchip/arm-gic-v3.h> +#endif + #include "internals.h" /* @@ -331,6 +335,9 @@ void register_handler_proc(unsigned int irq, struct irqaction *action) action->dir = proc_mkdir(name, desc->dir); } +void __weak register_irqchip_proc(struct irq_desc *desc, void *irqp) { } +void __weak unregister_irqchip_proc(struct irq_desc *desc) { } + #undef MAX_NAMELEN #define MAX_NAMELEN 10 @@ -385,6 +392,7 @@ void register_irq_proc(unsigned int irq, struct irq_desc *desc) #endif proc_create_single_data("spurious", 0444, desc->dir, irq_spurious_proc_show, (void *)(long)irq); + register_irqchip_proc(desc, irqp); out_unlock: mutex_unlock(®ister_lock); @@ -408,6 +416,8 @@ void unregister_irq_proc(unsigned int irq, struct irq_desc *desc) #endif remove_proc_entry("spurious", desc->dir); + unregister_irqchip_proc(desc); + sprintf(name, "%u", irq); remove_proc_entry(name, root_irq_dir); } diff --git a/kernel/softirq.c b/kernel/softirq.c index 4196b9f84690..9fc69e6e2c11 100644 --- a/kernel/softirq.c +++ b/kernel/softirq.c @@ -345,6 +345,42 @@ asmlinkage __visible void do_softirq(void) local_irq_restore(flags); } +#ifdef CONFIG_FAST_IRQ +/** + * xint_enter_rcu - Copy from irq_enter_rcu + */ +void xint_enter_rcu(void) +{ + if (tick_nohz_full_cpu(smp_processor_id()) || + (is_idle_task(current) && !in_interrupt())) { + /* + * Prevent raise_softirq from needlessly waking up ksoftirqd + * here, as softirq will be serviced on return from interrupt. + */ + local_bh_disable(); + tick_irq_enter(); + _local_bh_enable(); + } + +#ifndef CONFIG_DEBUG_FEATURE_BYPASS + account_irq_enter_time(current); +#endif + preempt_count_add(HARDIRQ_OFFSET); +#ifndef CONFIG_DEBUG_FEATURE_BYPASS + lockdep_hardirq_enter(); +#endif +} + +/** + * irq_enter - Copy from irq_enter + */ +void xint_enter(void) +{ + rcu_irq_enter(); + xint_enter_rcu(); +} +#endif + /** * irq_enter_rcu - Enter an interrupt context with RCU watching */ @@ -411,6 +447,43 @@ static inline void tick_irq_exit(void) #endif } +#ifdef CONFIG_FAST_IRQ +static inline void __xint_exit_rcu(void) +{ +#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED + local_irq_disable(); +#else +#ifndef CONFIG_DEBUG_FEATURE_BYPASS + lockdep_assert_irqs_disabled(); +#endif +#endif + +#ifndef CONFIG_DEBUG_FEATURE_BYPASS + account_irq_exit_time(current); +#endif + preempt_count_sub(HARDIRQ_OFFSET); + if (!in_interrupt() && local_softirq_pending()) + invoke_softirq(); + + tick_irq_exit(); +} + +/** + * xint_exit - Copy from irq_exit + * + * Also processes softirqs if needed and possible. + */ +void xint_exit(void) +{ + __xint_exit_rcu(); + rcu_irq_exit(); + /* must be last! */ +#ifndef CONFIG_DEBUG_FEATURE_BYPASS + lockdep_hardirq_exit(); +#endif +} +#endif + static inline void __irq_exit_rcu(void) { #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED -- 2.34.1
hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IB6JLE -------------------------------- Add a debugfs dir for xint, so we can get the xint irq information such as 'which interrupts are currently in xint state' with following cmd: # ls /sys/kernel/debug/irq/xints Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com> --- drivers/irqchip/irq-gic-v3.c | 19 ++++++++++++++++--- kernel/irq/debugfs.c | 33 +++++++++++++++++++++++++++++++++ kernel/irq/internals.h | 18 ++++++++++++++++++ 3 files changed, 67 insertions(+), 3 deletions(-) diff --git a/drivers/irqchip/irq-gic-v3.c b/drivers/irqchip/irq-gic-v3.c index e2aef650a884..ef2f1ca54a42 100644 --- a/drivers/irqchip/irq-gic-v3.c +++ b/drivers/irqchip/irq-gic-v3.c @@ -31,6 +31,10 @@ #include "irq-gic-common.h" +#ifdef CONFIG_FAST_IRQ +#include "../../../kernel/irq/internals.h" +#endif + #define GICD_INT_NMI_PRI (GICD_INT_DEF_PRI & ~0x80) #define FLAGS_WORKAROUND_GICR_WAKER_MSM8996 (1ULL << 0) @@ -766,6 +770,7 @@ static ssize_t xint_proc_write(struct file *file, { int irq = (int)(long)PDE_DATA(file_inode(file)); bool xint_state = false; + enum xint_op switch_type; unsigned long val; char *buf = NULL; @@ -787,13 +792,21 @@ static ssize_t xint_proc_write(struct file *file, return -EBUSY; } - local_irq_disable(); + if (xint_state) { + switch_type = XINT_TO_IRQ; + xint_remove_debugfs_entry(irq); + } else { + switch_type = IRQ_TO_XINT; + xint_add_debugfs_entry(irq); + } + disable_irq(irq); + local_irq_disable(); - xint_transform(irq, xint_state ? XINT_TO_IRQ : IRQ_TO_XINT); + xint_transform(irq, switch_type); - enable_irq(irq); local_irq_enable(); + enable_irq(irq); kfree(buf); diff --git a/kernel/irq/debugfs.c b/kernel/irq/debugfs.c index e4cff358b437..a4a7f87eab39 100644 --- a/kernel/irq/debugfs.c +++ b/kernel/irq/debugfs.c @@ -236,6 +236,34 @@ void irq_add_debugfs_entry(unsigned int irq, struct irq_desc *desc) &dfs_irq_ops); } +#ifdef CONFIG_FAST_IRQ +static struct dentry *xint_dir; + +void xint_add_debugfs_entry(unsigned int irq) +{ + char name[10]; + char buf[100]; + + if (!xint_dir) + return; + + sprintf(name, "%d", irq); + sprintf(buf, "../irqs/%d", irq); + debugfs_create_symlink(name, xint_dir, buf); +} + +void xint_remove_debugfs_entry(unsigned int irq) +{ + char name[10]; + + if (!xint_dir) + return; + + sprintf(name, "%d", irq); + debugfs_lookup_and_remove(name, xint_dir); +} +#endif + static int __init irq_debugfs_init(void) { struct dentry *root_dir; @@ -247,6 +275,11 @@ static int __init irq_debugfs_init(void) irq_dir = debugfs_create_dir("irqs", root_dir); +#ifdef CONFIG_FAST_IRQ + if (is_xint_support) + xint_dir = debugfs_create_dir("xints", root_dir); +#endif + irq_lock_sparse(); for_each_active_irq(irq) irq_add_debugfs_entry(irq, irq_to_desc(irq)); diff --git a/kernel/irq/internals.h b/kernel/irq/internals.h index 48d6aa8cdbed..d725d8ef5ce7 100644 --- a/kernel/irq/internals.h +++ b/kernel/irq/internals.h @@ -492,6 +492,14 @@ static inline void irq_remove_debugfs_entry(struct irq_desc *desc) debugfs_remove(desc->debugfs_file); kfree(desc->dev_name); } + +#ifdef CONFIG_FAST_IRQ +extern bool is_xint_support; + +void xint_add_debugfs_entry(unsigned int irq); +void xint_remove_debugfs_entry(unsigned int irq); +#endif + void irq_debugfs_copy_devname(int irq, struct device *dev); # ifdef CONFIG_IRQ_DOMAIN void irq_domain_debugfs_init(struct dentry *root); @@ -507,6 +515,16 @@ static inline void irq_add_debugfs_entry(unsigned int irq, struct irq_desc *d) static inline void irq_remove_debugfs_entry(struct irq_desc *d) { } + +#ifdef CONFIG_FAST_IRQ +static inline void xint_add_debugfs_entry(unsigned int irq) +{ +} +static inline void xint_remove_debugfs_entry(unsigned int irq) +{ +} +#endif + static inline void irq_debugfs_copy_devname(int irq, struct device *dev) { } -- 2.34.1
From: Liao Chen <liaochen4@huawei.com> hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IB6JLE -------------------------------- revert kpti bypass in xint. Signed-off-by: Liao Chen <liaochen4@huawei.com> --- arch/arm64/kernel/entry.S | 5 ----- 1 file changed, 5 deletions(-) diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S index 664a0d3059ab..5ed8b8e8e58e 100644 --- a/arch/arm64/kernel/entry.S +++ b/arch/arm64/kernel/entry.S @@ -371,11 +371,6 @@ alternative_else_nop_endif ldp x28, x29, [sp, #16 * 14] .if \el == 0 - .if \fast_mode != std - ldr lr, [sp, #S_LR] - add sp, sp, #S_FRAME_SIZE // restore sp - eret - .endif alternative_if_not ARM64_UNMAP_KERNEL_AT_EL0 ldr lr, [sp, #S_LR] add sp, sp, #S_FRAME_SIZE // restore sp -- 2.34.1
hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IC9Q31 -------------------------------- Refactor the xcall_enable logic to ensure only one reserved kabi in task_struct is used by xcall, which facilitates future xcall expansion. Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com> --- arch/arm64/kernel/asm-offsets.c | 2 +- arch/arm64/kernel/entry.S | 2 +- fs/proc/base.c | 32 +++++++++++++++++++------------- include/linux/sched.h | 2 +- include/linux/xcall.h | 9 +++++++++ kernel/fork.c | 18 ++++++++++-------- 6 files changed, 41 insertions(+), 24 deletions(-) create mode 100644 include/linux/xcall.h diff --git a/arch/arm64/kernel/asm-offsets.c b/arch/arm64/kernel/asm-offsets.c index 7c6ad4b1667b..54a21afd4181 100644 --- a/arch/arm64/kernel/asm-offsets.c +++ b/arch/arm64/kernel/asm-offsets.c @@ -27,7 +27,7 @@ int main(void) { #ifdef CONFIG_FAST_SYSCALL - DEFINE(TSK_XCALL, offsetof(struct task_struct, xcall_enable)); + DEFINE(TSK_XCALL, offsetof(struct task_struct, xinfo)); #endif DEFINE(TSK_ACTIVE_MM, offsetof(struct task_struct, active_mm)); BLANK(); diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S index 5ed8b8e8e58e..aa3eac1eb96e 100644 --- a/arch/arm64/kernel/entry.S +++ b/arch/arm64/kernel/entry.S @@ -707,7 +707,7 @@ SYM_CODE_END(el1_irq) .endm .macro check_xcall_enable - /* x21 = task_struct->xcall_enable */ + /* x21 = task_struct->xinfo->xcall_enable */ ldr_this_cpu x20, __entry_task, x21 ldr x21, [x20, #TSK_XCALL] /* x20 = sc_no / 8 */ diff --git a/fs/proc/base.c b/fs/proc/base.c index 3206960c4bd7..924959c9a02b 100644 --- a/fs/proc/base.c +++ b/fs/proc/base.c @@ -3590,6 +3590,8 @@ static const struct file_operations proc_pid_sg_level_operations = { #endif #ifdef CONFIG_FAST_SYSCALL +#include <linux/xcall.h> + bool fast_syscall_enabled(void); static int xcall_show(struct seq_file *m, void *v) @@ -3597,6 +3599,7 @@ static int xcall_show(struct seq_file *m, void *v) struct inode *inode = m->private; struct task_struct *p; unsigned int rs, re; + struct xcall_info *xinfo; if (!fast_syscall_enabled()) return -EACCES; @@ -3605,15 +3608,16 @@ static int xcall_show(struct seq_file *m, void *v) if (!p) return -ESRCH; - if (!p->xcall_enable) + xinfo = p->xinfo; + if (!xinfo) goto out; - seq_printf(m, "Enabled Total[%d/%d]:", bitmap_weight(p->xcall_enable, __NR_syscalls), + seq_printf(m, "Enabled Total[%d/%d]:", bitmap_weight(xinfo->xcall_enable, __NR_syscalls), __NR_syscalls); - for (rs = 0, bitmap_next_set_region(p->xcall_enable, &rs, &re, __NR_syscalls); + for (rs = 0, bitmap_next_set_region(xinfo->xcall_enable, &rs, &re, __NR_syscalls); rs < re; rs = re + 1, - bitmap_next_set_region(p->xcall_enable, &rs, &re, __NR_syscalls)) { + bitmap_next_set_region(xinfo->xcall_enable, &rs, &re, __NR_syscalls)) { rs == (re - 1) ? seq_printf(m, "%d,", rs) : seq_printf(m, "%d-%d,", rs, re - 1); } @@ -3629,15 +3633,15 @@ static int xcall_open(struct inode *inode, struct file *filp) return single_open(filp, xcall_show, inode); } -static int xcall_enable_one(struct task_struct *p, unsigned int sc_no) +static int xcall_enable_one(struct xcall_info *xinfo, unsigned int sc_no) { - bitmap_set(p->xcall_enable, sc_no, 1); + test_and_set_bit(sc_no, xinfo->xcall_enable); return 0; } -static int xcall_disable_one(struct task_struct *p, unsigned int sc_no) +static int xcall_disable_one(struct xcall_info *xinfo, unsigned int sc_no) { - bitmap_clear(p->xcall_enable, sc_no, 1); + test_and_clear_bit(sc_no, xinfo->xcall_enable); return 0; } @@ -3651,6 +3655,7 @@ static ssize_t xcall_write(struct file *file, const char __user *buf, unsigned int sc_no = __NR_syscalls; int ret = 0; int is_clear = 0; + struct xcall_info *xinfo; if (!fast_syscall_enabled()) return -EACCES; @@ -3660,9 +3665,10 @@ static ssize_t xcall_write(struct file *file, const char __user *buf, return -EFAULT; p = get_proc_task(inode); - if (!p || !p->xcall_enable) + if (!p || !p->xinfo) return -ESRCH; + xinfo = p->xinfo; if (buffer[0] == '!') is_clear = 1; @@ -3676,10 +3682,10 @@ static ssize_t xcall_write(struct file *file, const char __user *buf, goto out; } - if (!is_clear && !test_bit(sc_no, p->xcall_enable)) - ret = xcall_enable_one(p, sc_no); - else if (is_clear && test_bit(sc_no, p->xcall_enable)) - ret = xcall_disable_one(p, sc_no); + if (!is_clear && !test_bit(sc_no, xinfo->xcall_enable)) + ret = xcall_enable_one(xinfo, sc_no); + else if (is_clear && test_bit(sc_no, xinfo->xcall_enable)) + ret = xcall_disable_one(xinfo, sc_no); else ret = -EINVAL; diff --git a/include/linux/sched.h b/include/linux/sched.h index 18361e35a377..fd975183ec26 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -1478,7 +1478,7 @@ struct task_struct { KABI_RESERVE(14) #endif #if defined(CONFIG_FAST_SYSCALL) - KABI_USE(15, unsigned long *xcall_enable) + KABI_USE(15, struct xcall_info *xinfo) #else KABI_RESERVE(15) #endif diff --git a/include/linux/xcall.h b/include/linux/xcall.h new file mode 100644 index 000000000000..f52f2d9685f0 --- /dev/null +++ b/include/linux/xcall.h @@ -0,0 +1,9 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _LINUX_XCALL_H +#define _LINUX_XCALL_H + +struct xcall_info { + /* Must be first! */ + DECLARE_BITMAP(xcall_enable, __NR_syscalls); +}; +#endif diff --git a/kernel/fork.c b/kernel/fork.c index bd7afeb364ab..429b909d093d 100644 --- a/kernel/fork.c +++ b/kernel/fork.c @@ -102,6 +102,9 @@ #include <linux/sched/grid_qos.h> #endif #include <linux/share_pool.h> +#ifdef CONFIG_FAST_SYSCALL +#include <linux/xcall.h> +#endif #include <asm/pgalloc.h> #include <linux/uaccess.h> #include <asm/mmu_context.h> @@ -481,8 +484,7 @@ void free_task(struct task_struct *tsk) sched_relationship_free(tsk); #ifdef CONFIG_FAST_SYSCALL - if (tsk->xcall_enable) - bitmap_free(tsk->xcall_enable); + kfree(tsk->xinfo); #endif free_task_struct(tsk); @@ -1015,7 +1017,7 @@ static struct task_struct *dup_task_struct(struct task_struct *orig, int node) #endif #ifdef CONFIG_FAST_SYSCALL - tsk->xcall_enable = NULL; + tsk->xinfo = NULL; #endif return tsk; @@ -2097,12 +2099,12 @@ static __latent_entropy struct task_struct *copy_process( rt_mutex_init_task(p); #ifdef CONFIG_FAST_SYSCALL - p->xcall_enable = bitmap_zalloc(__NR_syscalls, GFP_KERNEL); - if (!p->xcall_enable) + p->xinfo = kzalloc(sizeof(struct xcall_info), GFP_KERNEL); + if (!p->xinfo) goto bad_fork_free; - - if (current->xcall_enable) - bitmap_copy(p->xcall_enable, current->xcall_enable, __NR_syscalls); + if (current->xinfo) + bitmap_copy(p->xinfo->xcall_enable, current->xinfo->xcall_enable, + __NR_syscalls); #endif #ifdef CONFIG_QOS_SCHED_DYNAMIC_AFFINITY -- 2.34.1
hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IC9Q31 -------------------------------- Xcall prefetch implements customized epoll_wait() and read() system calls, which enable data prefetching. In scenarios where the number of network connections established by a service is large, the time consumed by the service in the read() system call can be reduced. Users can enable xcall prefetch with the following command: echo 1 > /proc/$PID/prefetch Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com> --- arch/Kconfig | 17 +++++++++++ fs/proc/base.c | 67 +++++++++++++++++++++++++++++++++++++++++++ include/linux/xcall.h | 1 + kernel/fork.c | 4 ++- 4 files changed, 88 insertions(+), 1 deletion(-) diff --git a/arch/Kconfig b/arch/Kconfig index 6dc501a4afb1..d27eb1800cd9 100644 --- a/arch/Kconfig +++ b/arch/Kconfig @@ -1205,6 +1205,23 @@ config FAST_SYSCALL exception handling path that only considers necessary features such as security, context saving, and recovery. +config XCALL_PREFETCH + bool "Xcall prefetch support" + depends on FAST_SYSCALL + depends on EPOLL + default n + help + This enable xcall prefetch feature. + Xcall prefetch feature implements customized epoll_wait() and + read() system calls, which enable data prefetching. + In high-concurrency connection scenarios, this improves + the parallel execution efficiency of the read() system call + and increases the system's business throughput. + The Xcall prefetch feature is suitable for business scenarios + where the epoll I/O multiplexing mechanism is used, the read() + system call takes up a large proportion of time, and the number + of concurrent connections is large. + config ARCH_SUPPORTS_FAST_IRQ bool diff --git a/fs/proc/base.c b/fs/proc/base.c index 924959c9a02b..3e19b7bcbf66 100644 --- a/fs/proc/base.c +++ b/fs/proc/base.c @@ -3704,6 +3704,70 @@ static const struct file_operations proc_pid_xcall_operations = { }; #endif +#ifdef CONFIG_XCALL_PREFETCH +static int xcall_prefetch_show(struct seq_file *m, void *v) +{ + struct inode *inode = m->private; + struct task_struct *p; + + if (!fast_syscall_enabled()) + return -EACCES; + + p = get_proc_task(inode); + if (!p) + return -ESRCH; + + if (p->xinfo) + seq_printf(m, "%d\n", p->xinfo->prefetch); + + put_task_struct(p); + + return 0; +} + +static int xcall_prefetch_open(struct inode *inode, struct file *filp) +{ + return single_open(filp, xcall_prefetch_show, inode); +} + +static ssize_t xcall_prefetch_write(struct file *file, const char __user *buf, + size_t count, loff_t *offset) +{ + struct inode *inode = file_inode(file); + struct task_struct *p; + char buffer[TASK_COMM_LEN]; + const size_t maxlen = sizeof(buffer) - 1; + bool prefetch_enable = true; + + memset(buffer, 0, sizeof(buffer)); + if (copy_from_user(buffer, buf, count > maxlen ? maxlen : count)) + return -EFAULT; + + p = get_proc_task(inode); + if (!p) + return -ESRCH; + + if (!p->xinfo || p->xinfo->prefetch || + kstrtobool(buffer, &prefetch_enable) || !prefetch_enable) { + put_task_struct(p); + return -EINVAL; + } + + p->xinfo->prefetch = true; + put_task_struct(p); + + return count; +} + +static const struct file_operations proc_pid_xcall_prefetch_operations = { + .open = xcall_prefetch_open, + .read = seq_read, + .write = xcall_prefetch_write, + .llseek = seq_lseek, + .release = single_release, +}; +#endif + /* * Thread groups */ @@ -3733,6 +3797,9 @@ static const struct pid_entry tgid_base_stuff[] = { #ifdef CONFIG_FAST_SYSCALL REG("xcall", 0644, proc_pid_xcall_operations), #endif +#ifdef CONFIG_XCALL_PREFETCH + REG("prefetch", 0644, proc_pid_xcall_prefetch_operations), +#endif #ifdef CONFIG_SCHED_AUTOGROUP REG("autogroup", S_IRUGO|S_IWUSR, proc_pid_sched_autogroup_operations), #endif diff --git a/include/linux/xcall.h b/include/linux/xcall.h index f52f2d9685f0..88ce88793940 100644 --- a/include/linux/xcall.h +++ b/include/linux/xcall.h @@ -5,5 +5,6 @@ struct xcall_info { /* Must be first! */ DECLARE_BITMAP(xcall_enable, __NR_syscalls); + bool prefetch; }; #endif diff --git a/kernel/fork.c b/kernel/fork.c index 429b909d093d..7295de34c015 100644 --- a/kernel/fork.c +++ b/kernel/fork.c @@ -2102,9 +2102,11 @@ static __latent_entropy struct task_struct *copy_process( p->xinfo = kzalloc(sizeof(struct xcall_info), GFP_KERNEL); if (!p->xinfo) goto bad_fork_free; - if (current->xinfo) + if (current->xinfo) { bitmap_copy(p->xinfo->xcall_enable, current->xinfo->xcall_enable, __NR_syscalls); + p->xinfo->prefetch = current->xinfo->prefetch; + } #endif #ifdef CONFIG_QOS_SCHED_DYNAMIC_AFFINITY -- 2.34.1
hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IC9Q31 -------------------------------- Add cache mode for fd wakeup in epoll_pwait. In epoll_pwait, read data from ready fd into pre-allocated kernel cache buffer. And then, in sys_read, read from cache buffer and copy to user. So, we can async prefetch read data in epoll_pwait. (lock state can not kernel_read()) +--------------CANCEL<------------------------------+ | ^ | | | | | (prefetch not | (kernel_read() done) | (cache buf | start) | (prefetching) | not empty) +--------------> NONE ---->PREFETCH ----> READY <---+ (vfs_read done or not need) Signed-off-by: Yipeng Zou <zouyipeng@huawei.com> Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com> --- arch/arm64/kernel/cpufeature.c | 2 + fs/eventpoll.c | 273 +++++++++++++++++++++++++++++++++ fs/file_table.c | 1 + fs/proc/base.c | 2 - fs/read_write.c | 10 +- include/linux/fs.h | 35 +++++ include/linux/xcall.h | 7 + 7 files changed, 327 insertions(+), 3 deletions(-) diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c index 9b4a315e96bc..4e12b39cc721 100644 --- a/arch/arm64/kernel/cpufeature.c +++ b/arch/arm64/kernel/cpufeature.c @@ -2156,6 +2156,8 @@ static bool can_clearpage_use_stnp(const struct arm64_cpu_capabilities *entry, } #ifdef CONFIG_FAST_SYSCALL +#include <linux/xcall.h> + static bool is_xcall_support; static int __init xcall_setup(char *str) { diff --git a/fs/eventpoll.c b/fs/eventpoll.c index 5ce1ea1f452b..6ead4b568b87 100644 --- a/fs/eventpoll.c +++ b/fs/eventpoll.c @@ -768,6 +768,268 @@ static void epi_rcu_free(struct rcu_head *head) kmem_cache_free(epi_cache, epi); } +#ifdef CONFIG_XCALL_PREFETCH +#include <linux/xcall.h> + +#define XCALL_CACHE_PAGE_ORDER 2 +#define XCALL_CACHE_BUF_SIZE ((1 << XCALL_CACHE_PAGE_ORDER) * PAGE_SIZE) +#define PREFETCH_ITEM_HASH_BITS 6 +static DEFINE_HASHTABLE(xcall_item_table, PREFETCH_ITEM_HASH_BITS); +static DEFINE_RWLOCK(xcall_table_lock); +static struct workqueue_struct *rc_work; + +static inline bool transition_state(struct prefetch_item *pfi, + enum cache_state old, enum cache_state new) +{ + return atomic_cmpxchg(&pfi->state, old, new) == old; +} + +static void xcall_prefetch_init(void) +{ + rc_work = alloc_workqueue("eventpoll_rc", 0, 0); + if (!rc_work) + pr_warn("alloc eventpoll_rc workqueue failed.\n"); + + hash_init(xcall_item_table); +} + +static struct prefetch_item *find_prefetch_item(struct file *file) +{ + unsigned int hash = hash_64((u64)file, PREFETCH_ITEM_HASH_BITS); + struct prefetch_item *found = NULL; + + read_lock(&xcall_table_lock); + hash_for_each_possible(xcall_item_table, found, node, hash) { + if (found->file == file) + break; + } + read_unlock(&xcall_table_lock); + + return found; +} + +static void prefetch_work_fn(struct work_struct *work) +{ + struct prefetch_item *pfi = container_of(work, struct prefetch_item, work); + + if (!transition_state(pfi, XCALL_CACHE_NONE, XCALL_CACHE_PREFETCH)) + return; + + pfi->pos = 0; + pfi->len = kernel_read(pfi->file, pfi->cache, + XCALL_CACHE_BUF_SIZE, &pfi->file->f_pos); + transition_state(pfi, XCALL_CACHE_PREFETCH, XCALL_CACHE_READY); +} + +static void set_prefetch_numa_cpu(struct prefetch_item *pfi, int fd) +{ + int cur_cpu = smp_processor_id(); + int cpu; + + cpumask_and(&pfi->related_cpus, cpu_cpu_mask(cur_cpu), cpu_online_mask); + cpu = cpumask_next(fd % cpumask_weight(&pfi->related_cpus), + &pfi->related_cpus); + if (cpu > cpumask_last(&pfi->related_cpus)) + cpu = cpumask_first(&pfi->related_cpus); + pfi->cpu = cpu; +} + +static struct prefetch_item *alloc_prefetch_item(struct epitem *epi) +{ + struct file *tfile = epi->ffd.file; + struct prefetch_item *pfi; + int fd = epi->ffd.fd; + unsigned int hash; + + pfi = kmalloc(sizeof(struct prefetch_item), GFP_KERNEL); + if (!pfi) + return NULL; + + pfi->cache_pages = alloc_pages(GFP_KERNEL_ACCOUNT | __GFP_ZERO, + XCALL_CACHE_PAGE_ORDER); + if (!pfi->cache_pages) { + kfree(pfi); + return NULL; + } + + pfi->cache = page_address(pfi->cache_pages); + atomic_set(&pfi->state, XCALL_CACHE_NONE); + INIT_WORK(&pfi->work, prefetch_work_fn); + INIT_HLIST_NODE(&pfi->node); + pfi->file = tfile; + pfi->len = 0; + pfi->pos = 0; + set_prefetch_numa_cpu(pfi, fd); + + write_lock(&xcall_table_lock); + hash = hash_64((u64)tfile, PREFETCH_ITEM_HASH_BITS); + hash_add(xcall_item_table, &pfi->node, hash); + write_unlock(&xcall_table_lock); + + return pfi; +} + +void free_prefetch_item(struct file *file) +{ + struct prefetch_item *pfi = NULL; + struct hlist_node *next; + bool has_pfi = false; + unsigned int hash; + + if (!fast_syscall_enabled()) + return; + + hash = hash_64((u64)file, PREFETCH_ITEM_HASH_BITS); + write_lock(&xcall_table_lock); + hash_for_each_possible_safe(xcall_item_table, pfi, next, node, hash) { + if (pfi->file == file) { + hlist_del_init(&pfi->node); + has_pfi = true; + break; + } + } + write_unlock(&xcall_table_lock); + if (!has_pfi) + return; + + cancel_work_sync(&pfi->work); + __free_pages(pfi->cache_pages, XCALL_CACHE_PAGE_ORDER); + pfi->cache = NULL; + kfree(pfi); +} + +static int xcall_read(struct prefetch_item *pfi, char __user *buf, size_t count) +{ + ssize_t copy_len = 0; + + /* + * Everytime it does the memcpy on prefetch buffer, it has to keep + * the state of pfi is "CANCEL" to avoid the race on the prefetch + * buffer from both the prefetch thread calling kernel_read() and + * other threads calling copy_to_user(), also avoid race on the + * prefetch file from both the prefetch thread calling kernel_read() + * and other threads calling vfs_read(). + */ + while (!transition_state(pfi, XCALL_CACHE_READY, XCALL_CACHE_CANCEL)) { + /* + * Once the prefetch thread read return error code or prefetch + * has not start, no need to waste CPU on waiting right here, + * it should do a slow vfs_read() to ensure no new arrival data. + */ + if (transition_state(pfi, XCALL_CACHE_NONE, XCALL_CACHE_CANCEL)) + goto slow_read; + } + + copy_len = pfi->len; + if (unlikely(copy_len < 0)) + goto slow_read; + + if (copy_len == 0) { + transition_state(pfi, XCALL_CACHE_CANCEL, XCALL_CACHE_NONE); + return 0; + } + + copy_len = (copy_len >= count) ? count : copy_len; + copy_len -= copy_to_user(buf, (void *)(pfi->cache + pfi->pos), copy_len); + pfi->len -= copy_len; + pfi->pos += copy_len; + if (pfi->len == 0) + transition_state(pfi, XCALL_CACHE_CANCEL, XCALL_CACHE_NONE); + else + transition_state(pfi, XCALL_CACHE_CANCEL, XCALL_CACHE_READY); + + return copy_len; + +slow_read: + pfi->len = 0; + pfi->pos = 0; + cancel_work(&pfi->work); + + return -EAGAIN; +} + +int xcall_read_begin(struct file *file, char __user *buf, size_t count) +{ + struct prefetch_item *pfi = NULL; + + if (!current_prefetch_enabled()) + return -EAGAIN; + + pfi = find_prefetch_item(file); + if (!pfi) + return -EAGAIN; + + return xcall_read(pfi, buf, count); +} + +void xcall_read_end(struct file *file) +{ + struct prefetch_item *pfi = NULL; + + if (!current_prefetch_enabled()) + return; + + pfi = find_prefetch_item(file); + if (!pfi) + return; + + transition_state(pfi, XCALL_CACHE_CANCEL, XCALL_CACHE_NONE); +} + +static int get_async_prefetch_cpu(struct prefetch_item *pfi) +{ + int cpu; + + if (pfi->cpu != smp_processor_id()) + return pfi->cpu; + + cpu = cpumask_next(pfi->cpu, &pfi->related_cpus); + if (cpu > cpumask_last(&pfi->related_cpus)) + cpu = cpumask_first(&pfi->related_cpus); + pfi->cpu = cpu; + return pfi->cpu; +} + +static void ep_prefetch_item_enqueue(struct eventpoll *ep, struct epitem *epi) +{ + struct prefetch_item *pfi; + int cpu, err; + + if (unlikely(!rc_work) || !current_prefetch_enabled()) + return; + + if (!(epi->event.events & EPOLLIN) || + !sock_from_file(epi->ffd.file, &err) || + !(epi->ffd.file->f_mode & FMODE_READ)) + return; + + pfi = find_prefetch_item(epi->ffd.file); + if (unlikely(!pfi)) { + pfi = alloc_prefetch_item(epi); + if (unlikely(!pfi)) + return; + } + + if (atomic_read(&pfi->state) != XCALL_CACHE_NONE) + return; + + cpu = get_async_prefetch_cpu(pfi); + queue_work_on(cpu, rc_work, &pfi->work); +} + +static void xcall_cancel_work(struct file *file) +{ + struct prefetch_item *pfi; + + if (!current_prefetch_enabled()) + return; + + pfi = find_prefetch_item(file); + if (pfi) + cancel_work_sync(&pfi->work); +} +#endif + /* * Removes a "struct epitem" from the eventpoll RB tree and deallocates * all the associated resources. Must be called with "mtx" held. @@ -782,6 +1044,9 @@ static int ep_remove(struct eventpoll *ep, struct epitem *epi) * Removes poll wait queue hooks. */ ep_unregister_pollwait(ep, epi); +#ifdef CONFIG_XCALL_PREFETCH + xcall_cancel_work(file); +#endif /* Remove the current item from the list of epoll hooks */ spin_lock(&file->f_lock); @@ -1751,6 +2016,10 @@ static __poll_t ep_send_events_proc(struct eventpoll *ep, struct list_head *head if (!revents) continue; +#ifdef CONFIG_XCALL_PREFETCH + ep_prefetch_item_enqueue(ep, epi); +#endif + if (__put_user(revents, &uevent->events) || __put_user(epi->event.data, &uevent->data)) { list_add(&epi->rdllink, head); @@ -2454,6 +2723,10 @@ static int __init eventpoll_init(void) pwq_cache = kmem_cache_create("eventpoll_pwq", sizeof(struct eppoll_entry), 0, SLAB_PANIC|SLAB_ACCOUNT, NULL); +#ifdef CONFIG_XCALL_PREFETCH + xcall_prefetch_init(); +#endif + return 0; } fs_initcall(eventpoll_init); diff --git a/fs/file_table.c b/fs/file_table.c index 542f4fddc0a0..935fdd668e05 100644 --- a/fs/file_table.c +++ b/fs/file_table.c @@ -301,6 +301,7 @@ static void __fput(struct file *file) * in the file cleanup chain. */ eventpoll_release(file); + free_prefetch_item(file); locks_remove_file(file); ima_file_free(file); diff --git a/fs/proc/base.c b/fs/proc/base.c index 3e19b7bcbf66..4f57e91f3e54 100644 --- a/fs/proc/base.c +++ b/fs/proc/base.c @@ -3592,8 +3592,6 @@ static const struct file_operations proc_pid_sg_level_operations = { #ifdef CONFIG_FAST_SYSCALL #include <linux/xcall.h> -bool fast_syscall_enabled(void); - static int xcall_show(struct seq_file *m, void *v) { struct inode *inode = m->private; diff --git a/fs/read_write.c b/fs/read_write.c index da03b3e65cf3..a850bb1b0a77 100644 --- a/fs/read_write.c +++ b/fs/read_write.c @@ -623,7 +623,13 @@ ssize_t ksys_read(unsigned int fd, char __user *buf, size_t count) ssize_t ret = -EBADF; if (f.file) { - loff_t pos, *ppos = file_ppos(f.file); + loff_t pos, *ppos; + + ret = xcall_read_begin(f.file, buf, count); + if (ret != -EAGAIN) + goto fdput; + + ppos = file_ppos(f.file); if (ppos) { pos = *ppos; ppos = &pos; @@ -631,6 +637,8 @@ ssize_t ksys_read(unsigned int fd, char __user *buf, size_t count) ret = vfs_read(f.file, buf, count, ppos); if (ret >= 0 && ppos) f.file->f_pos = pos; + xcall_read_end(f.file); +fdput: fdput_pos(f); } return ret; diff --git a/include/linux/fs.h b/include/linux/fs.h index a0ea6b64c45d..221b4d4e3889 100644 --- a/include/linux/fs.h +++ b/include/linux/fs.h @@ -3750,4 +3750,39 @@ static inline bool cachefiles_ondemand_is_enabled(void) } #endif +#ifdef CONFIG_XCALL_PREFETCH +enum cache_state { + XCALL_CACHE_NONE = 0, + XCALL_CACHE_PREFETCH, + XCALL_CACHE_READY, + XCALL_CACHE_CANCEL +}; + +struct prefetch_item { + struct file *file; + struct work_struct work; + int cpu; + cpumask_t related_cpus; + struct page *cache_pages; + char *cache; + ssize_t len; + /* cache state in epoll_wait */ + atomic_t state; + loff_t pos; + struct hlist_node node; +}; + +int xcall_read_begin(struct file *file, char __user *buf, size_t count); +void xcall_read_end(struct file *file); +void free_prefetch_item(struct file *file); +#else +static inline int xcall_read_begin(struct file *file, char __user *buf, + size_t count) +{ + return -EAGAIN; +} +static inline void xcall_read_end(struct file *file) {} +static inline void free_prefetch_item(struct file *file) {} +#endif + #endif /* _LINUX_FS_H */ diff --git a/include/linux/xcall.h b/include/linux/xcall.h index 88ce88793940..fc64def700e3 100644 --- a/include/linux/xcall.h +++ b/include/linux/xcall.h @@ -7,4 +7,11 @@ struct xcall_info { DECLARE_BITMAP(xcall_enable, __NR_syscalls); bool prefetch; }; + +bool fast_syscall_enabled(void); + +static inline bool current_prefetch_enabled(void) +{ + return current->xinfo && current->xinfo->prefetch; +} #endif -- 2.34.1
hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IC9Q31 -------------------------------- Add percpu cache hit/miss count to calculate the usage efficiency of xcall prefetch framework. Also Add "/proc/xcall/stats" dir, so we can get the xcall prefetch hit ratio on each CPU that initiates a read system call, which is important for performance tuning: cat /proc/xcall/prefetch Use "echo > /proc/xcall/prefetch" to clear the hit/miss count. Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com> --- fs/eventpoll.c | 75 ++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 75 insertions(+) diff --git a/fs/eventpoll.c b/fs/eventpoll.c index 6ead4b568b87..bc63a549ee4c 100644 --- a/fs/eventpoll.c +++ b/fs/eventpoll.c @@ -773,11 +773,83 @@ static void epi_rcu_free(struct rcu_head *head) #define XCALL_CACHE_PAGE_ORDER 2 #define XCALL_CACHE_BUF_SIZE ((1 << XCALL_CACHE_PAGE_ORDER) * PAGE_SIZE) +DEFINE_PER_CPU_ALIGNED(unsigned long, xcall_cache_hit); +DEFINE_PER_CPU_ALIGNED(unsigned long, xcall_cache_miss); + #define PREFETCH_ITEM_HASH_BITS 6 static DEFINE_HASHTABLE(xcall_item_table, PREFETCH_ITEM_HASH_BITS); static DEFINE_RWLOCK(xcall_table_lock); static struct workqueue_struct *rc_work; +static ssize_t xcall_prefetch_write(struct file *file, const char __user *buf, + size_t count, loff_t *pos) +{ + int cpu; + + for_each_cpu(cpu, cpu_online_mask) { + *per_cpu_ptr(&xcall_cache_hit, cpu) = 0; + *per_cpu_ptr(&xcall_cache_miss, cpu) = 0; + } + + return count; +} + +static int xcall_prefetch_show(struct seq_file *m, void *v) +{ + unsigned long hit = 0, miss = 0; + unsigned int cpu; + u64 percent; + + for_each_cpu(cpu, cpu_online_mask) { + hit = *per_cpu_ptr(&xcall_cache_hit, cpu); + miss = *per_cpu_ptr(&xcall_cache_miss, cpu); + + if (hit == 0 && miss == 0) + continue; + + percent = DIV_ROUND_CLOSEST(hit * 100ULL, hit + miss); + seq_printf(m, "cpu%u epoll cache_{hit,miss}: %lu,%lu, hit ratio: %llu%%\n", + cpu, hit, miss, percent); + } + return 0; +} + +static int xcall_prefetch_open(struct inode *inode, struct file *file) +{ + return single_open(file, xcall_prefetch_show, NULL); +} + +static const struct proc_ops xcall_prefetch_fops = { + .proc_open = xcall_prefetch_open, + .proc_read = seq_read, + .proc_write = xcall_prefetch_write, + .proc_lseek = seq_lseek, + .proc_release = single_release +}; + +static int __init init_xcall_prefetch_procfs(void) +{ + struct proc_dir_entry *xcall_proc_dir, *prefetch_dir; + + if (!fast_syscall_enabled()) + return 0; + + xcall_proc_dir = proc_mkdir("xcall", NULL); + if (!xcall_proc_dir) + return -ENOMEM; + prefetch_dir = proc_create("prefetch", 0644, xcall_proc_dir, &xcall_prefetch_fops); + if (!prefetch_dir) + goto rm_xcall_proc_dir; + + return 0; + +rm_xcall_proc_dir: + proc_remove(xcall_proc_dir); + return -ENOMEM; + +} +device_initcall(init_xcall_prefetch_procfs); + static inline bool transition_state(struct prefetch_item *pfi, enum cache_state old, enum cache_state new) { @@ -925,6 +997,7 @@ static int xcall_read(struct prefetch_item *pfi, char __user *buf, size_t count) goto slow_read; if (copy_len == 0) { + this_cpu_inc(xcall_cache_hit); transition_state(pfi, XCALL_CACHE_CANCEL, XCALL_CACHE_NONE); return 0; } @@ -938,9 +1011,11 @@ static int xcall_read(struct prefetch_item *pfi, char __user *buf, size_t count) else transition_state(pfi, XCALL_CACHE_CANCEL, XCALL_CACHE_READY); + this_cpu_inc(xcall_cache_hit); return copy_len; slow_read: + this_cpu_inc(xcall_cache_miss); pfi->len = 0; pfi->pos = 0; cancel_work(&pfi->work); -- 2.34.1
hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IC9Q31 -------------------------------- Add "/proc/xcall/cpu_list" to adjust the kernel async prefetch cpu range of the whole machine, default set to the online cpumask, use following command to set the cpu range: echo "$cpu_list" > /proc/xcall/cpu_list Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com> --- fs/eventpoll.c | 56 ++++++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 54 insertions(+), 2 deletions(-) diff --git a/fs/eventpoll.c b/fs/eventpoll.c index bc63a549ee4c..8e67dd8f5340 100644 --- a/fs/eventpoll.c +++ b/fs/eventpoll.c @@ -780,6 +780,48 @@ DEFINE_PER_CPU_ALIGNED(unsigned long, xcall_cache_miss); static DEFINE_HASHTABLE(xcall_item_table, PREFETCH_ITEM_HASH_BITS); static DEFINE_RWLOCK(xcall_table_lock); static struct workqueue_struct *rc_work; +static struct cpumask xcall_mask; + +static ssize_t xcall_mask_proc_write(struct file *file, const char __user *buf, + size_t count, loff_t *ppos) +{ + struct cpumask tmp; + int err; + + err = cpumask_parselist_user(buf, count, &tmp); + if (err) + return err; + + if (cpumask_empty(&tmp)) + return -EINVAL; + + if (!cpumask_intersects(&tmp, cpu_online_mask)) { + pr_warn("cpu %*pbl is not online.\n", cpumask_pr_args(&tmp)); + return -EINVAL; + } + + cpumask_copy(&xcall_mask, &tmp); + return count; +} + +static int xcall_mask_proc_show(struct seq_file *m, void *v) +{ + seq_printf(m, "%*pbl\n", cpumask_pr_args(&xcall_mask)); + return 0; +} + +static int xcall_mask_proc_open(struct inode *inode, struct file *file) +{ + return single_open(file, xcall_mask_proc_show, PDE_DATA(inode)); +} + +static const struct proc_ops xcall_mask_fops = { + .proc_open = xcall_mask_proc_open, + .proc_read = seq_read, + .proc_lseek = seq_lseek, + .proc_release = single_release, + .proc_write = xcall_mask_proc_write, +}; static ssize_t xcall_prefetch_write(struct file *file, const char __user *buf, size_t count, loff_t *pos) @@ -829,7 +871,7 @@ static const struct proc_ops xcall_prefetch_fops = { static int __init init_xcall_prefetch_procfs(void) { - struct proc_dir_entry *xcall_proc_dir, *prefetch_dir; + struct proc_dir_entry *xcall_proc_dir, *prefetch_dir, *xcall_mask_dir; if (!fast_syscall_enabled()) return 0; @@ -840,13 +882,19 @@ static int __init init_xcall_prefetch_procfs(void) prefetch_dir = proc_create("prefetch", 0644, xcall_proc_dir, &xcall_prefetch_fops); if (!prefetch_dir) goto rm_xcall_proc_dir; + xcall_mask_dir = proc_create("cpu_list", 0644, xcall_proc_dir, + &xcall_mask_fops); + if (!xcall_mask_dir) + goto rm_prefetch_dir; + cpumask_copy(&xcall_mask, cpu_online_mask); return 0; +rm_prefetch_dir: + proc_remove(prefetch_dir); rm_xcall_proc_dir: proc_remove(xcall_proc_dir); return -ENOMEM; - } device_initcall(init_xcall_prefetch_procfs); @@ -896,9 +944,13 @@ static void prefetch_work_fn(struct work_struct *work) static void set_prefetch_numa_cpu(struct prefetch_item *pfi, int fd) { int cur_cpu = smp_processor_id(); + struct cpumask tmp; int cpu; + cpumask_copy(&tmp, &xcall_mask); cpumask_and(&pfi->related_cpus, cpu_cpu_mask(cur_cpu), cpu_online_mask); + if (cpumask_intersects(&tmp, &pfi->related_cpus)) + cpumask_and(&pfi->related_cpus, &pfi->related_cpus, &tmp); cpu = cpumask_next(fd % cpumask_weight(&pfi->related_cpus), &pfi->related_cpus); if (cpu > cpumask_last(&pfi->related_cpus)) -- 2.34.1
hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IC9Q31 -------------------------------- For xcall prefetch different cache state, add tracepoints. Signed-off-by: Yipeng Zou <zouyipeng@huawei.com> Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com> --- fs/eventpoll.c | 8 +++ include/trace/events/xcall.h | 106 +++++++++++++++++++++++++++++++++++ 2 files changed, 114 insertions(+) create mode 100644 include/trace/events/xcall.h diff --git a/fs/eventpoll.c b/fs/eventpoll.c index 8e67dd8f5340..74f143684978 100644 --- a/fs/eventpoll.c +++ b/fs/eventpoll.c @@ -770,6 +770,8 @@ static void epi_rcu_free(struct rcu_head *head) #ifdef CONFIG_XCALL_PREFETCH #include <linux/xcall.h> +#define CREATE_TRACE_POINTS +#include <trace/events/xcall.h> #define XCALL_CACHE_PAGE_ORDER 2 #define XCALL_CACHE_BUF_SIZE ((1 << XCALL_CACHE_PAGE_ORDER) * PAGE_SIZE) @@ -935,10 +937,12 @@ static void prefetch_work_fn(struct work_struct *work) if (!transition_state(pfi, XCALL_CACHE_NONE, XCALL_CACHE_PREFETCH)) return; + trace_epoll_rc_prefetch(pfi->file); pfi->pos = 0; pfi->len = kernel_read(pfi->file, pfi->cache, XCALL_CACHE_BUF_SIZE, &pfi->file->f_pos); transition_state(pfi, XCALL_CACHE_PREFETCH, XCALL_CACHE_READY); + trace_epoll_rc_ready(pfi->file, pfi->len); } static void set_prefetch_numa_cpu(struct prefetch_item *pfi, int fd) @@ -1050,6 +1054,7 @@ static int xcall_read(struct prefetch_item *pfi, char __user *buf, size_t count) if (copy_len == 0) { this_cpu_inc(xcall_cache_hit); + trace_epoll_rc_hit(pfi->file, 0); transition_state(pfi, XCALL_CACHE_CANCEL, XCALL_CACHE_NONE); return 0; } @@ -1064,10 +1069,12 @@ static int xcall_read(struct prefetch_item *pfi, char __user *buf, size_t count) transition_state(pfi, XCALL_CACHE_CANCEL, XCALL_CACHE_READY); this_cpu_inc(xcall_cache_hit); + trace_epoll_rc_hit(pfi->file, copy_len); return copy_len; slow_read: this_cpu_inc(xcall_cache_miss); + trace_epoll_rc_miss(pfi->file); pfi->len = 0; pfi->pos = 0; cancel_work(&pfi->work); @@ -1142,6 +1149,7 @@ static void ep_prefetch_item_enqueue(struct eventpoll *ep, struct epitem *epi) cpu = get_async_prefetch_cpu(pfi); queue_work_on(cpu, rc_work, &pfi->work); + trace_epoll_rc_queue(pfi->file, cpu); } static void xcall_cancel_work(struct file *file) diff --git a/include/trace/events/xcall.h b/include/trace/events/xcall.h new file mode 100644 index 000000000000..524a21e433ba --- /dev/null +++ b/include/trace/events/xcall.h @@ -0,0 +1,106 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#undef TRACE_SYSTEM +#define TRACE_SYSTEM xcall + +#if !defined(_TRACE_XCALL_H) || defined(TRACE_HEADER_MULTI_READ) +#define _TRACE_XCALL_H + +#include <linux/types.h> +#include <linux/tracepoint.h> +#include <linux/fs.h> + +TRACE_EVENT(epoll_rc_queue, + + TP_PROTO(struct file *file, int cpu), + + TP_ARGS(file, cpu), + + TP_STRUCT__entry( + __field(struct file *, file) + __field(int, cpu) + ), + + TP_fast_assign( + __entry->file = file; + __entry->cpu = cpu; + ), + + TP_printk("0x%p on cpu %d", __entry->file, __entry->cpu) +); + +TRACE_EVENT(epoll_rc_prefetch, + + TP_PROTO(struct file *file), + + TP_ARGS(file), + + TP_STRUCT__entry( + __field(struct file *, file) + ), + + TP_fast_assign( + __entry->file = file; + ), + + TP_printk("0x%p", __entry->file) +); + +TRACE_EVENT(epoll_rc_ready, + + TP_PROTO(struct file *file, int len), + + TP_ARGS(file, len), + + TP_STRUCT__entry( + __field(struct file *, file) + __field(int, len) + ), + + TP_fast_assign( + __entry->file = file; + __entry->len = len; + ), + + TP_printk("0x%p, len %d", __entry->file, __entry->len) +); + +TRACE_EVENT(epoll_rc_hit, + + TP_PROTO(struct file *file, int len), + + TP_ARGS(file, len), + + TP_STRUCT__entry( + __field(struct file *, file) + __field(int, len) + ), + + TP_fast_assign( + __entry->file = file; + __entry->len = len; + ), + + TP_printk("0x%p, len: %d", __entry->file, __entry->len) +); + +TRACE_EVENT(epoll_rc_miss, + + TP_PROTO(struct file *file), + + TP_ARGS(file), + + TP_STRUCT__entry( + __field(struct file *, file) + ), + + TP_fast_assign( + __entry->file = file; + ), + + TP_printk("0x%p", __entry->file) +); + +#endif /* _TRACE_XCALL_H */ + +/* This part must be outside protection */ +#include <trace/define_trace.h> -- 2.34.1
hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/release-management/issues/IC9Q31 -------------------------------- Enable FAST_SYSCALL, FAST_IRQ and XCALL_PREFETCH in openeuler_defconfig. Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com> --- arch/arm64/configs/openeuler_defconfig | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/arch/arm64/configs/openeuler_defconfig b/arch/arm64/configs/openeuler_defconfig index 576da6c47caf..77f0a13a7543 100644 --- a/arch/arm64/configs/openeuler_defconfig +++ b/arch/arm64/configs/openeuler_defconfig @@ -922,9 +922,12 @@ CONFIG_ARCH_HAS_GCOV_PROFILE_ALL=y CONFIG_HAVE_GCC_PLUGINS=y CONFIG_ARCH_SUPPORTS_FAST_SYSCALL=y -# CONFIG_FAST_SYSCALL is not set +CONFIG_FAST_SYSCALL=y +CONFIG_XCALL_PREFETCH=y CONFIG_ARCH_SUPPORTS_FAST_IRQ=y -# CONFIG_FAST_IRQ is not set +CONFIG_FAST_IRQ=y +CONFIG_DEBUG_FEATURE_BYPASS=y +CONFIG_SECURITY_FEATURE_BYPASS=y # end of General architecture-dependent options CONFIG_RT_MUTEXES=y -- 2.34.1
participants (1)
-
Jinjie Ruan