linux/kernel/trace/rethook.c

331 lines
9.2 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0
#define pr_fmt(fmt) "rethook: " fmt
#include <linux/bug.h>
#include <linux/kallsyms.h>
#include <linux/kprobes.h>
#include <linux/preempt.h>
#include <linux/rethook.h>
#include <linux/slab.h>
/* Return hook list (shadow stack by list) */
/*
* This function is called from delayed_put_task_struct() when a task is
* dead and cleaned up to recycle any kretprobe instances associated with
* this task. These left over instances represent probed functions that
* have been called but will never return.
*/
void rethook_flush_task(struct task_struct *tk)
{
struct rethook_node *rhn;
struct llist_node *node;
node = __llist_del_all(&tk->rethooks);
while (node) {
rhn = container_of(node, struct rethook_node, llist);
node = node->next;
preempt_disable();
rethook_recycle(rhn);
preempt_enable();
}
}
static void rethook_free_rcu(struct rcu_head *head)
{
struct rethook *rh = container_of(head, struct rethook, rcu);
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
objpool_fini(&rh->pool);
}
fprobe: Ensure running fprobe_exit_handler() finished before calling rethook_free() Ensure running fprobe_exit_handler() has finished before calling rethook_free() in the unregister_fprobe() so that caller can free the fprobe right after unregister_fprobe(). unregister_fprobe() ensured that all running fprobe_entry/exit_handler() have finished by calling unregister_ftrace_function() which synchronizes RCU. But commit 5f81018753df ("fprobe: Release rethook after the ftrace_ops is unregistered") changed to call rethook_free() after unregister_ftrace_function(). So call rethook_stop() to make rethook disabled before unregister_ftrace_function() and ensure it again. Here is the possible code flow that can call the exit handler after unregister_fprobe(). ------ CPU1 CPU2 call unregister_fprobe(fp) ... __fprobe_handler() rethook_hook() on probed function unregister_ftrace_function() return from probed function rethook hooks find rh->handler == fprobe_exit_handler call fprobe_exit_handler() rethook_free(): set rh->handler = NULL; return from unreigster_fprobe; call fp->exit_handler() <- (*) ------ (*) At this point, the exit handler is called after returning from unregister_fprobe(). This fixes it as following; ------ CPU1 CPU2 call unregister_fprobe() ... rethook_stop(): set rh->handler = NULL; __fprobe_handler() rethook_hook() on probed function unregister_ftrace_function() return from probed function rethook hooks find rh->handler == NULL return from rethook rethook_free() return from unreigster_fprobe; ------ Link: https://lore.kernel.org/all/168873859949.156157.13039240432299335849.stgit@devnote2/ Fixes: 5f81018753df ("fprobe: Release rethook after the ftrace_ops is unregistered") Cc: stable@vger.kernel.org Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Reviewed-by: Steven Rostedt (Google) <rostedt@goodmis.org>
2023-07-07 14:03:19 +00:00
/**
* rethook_stop() - Stop using a rethook.
* @rh: the struct rethook to stop.
*
* Stop using a rethook to prepare for freeing it. If you want to wait for
* all running rethook handler before calling rethook_free(), you need to
* call this first and wait RCU, and call rethook_free().
*/
void rethook_stop(struct rethook *rh)
{
WRITE_ONCE(rh->handler, NULL);
}
/**
* rethook_free() - Free struct rethook.
* @rh: the struct rethook to be freed.
*
* Free the rethook. Before calling this function, user must ensure the
* @rh::data is cleaned if needed (or, the handler can access it after
* calling this function.) This function will set the @rh to be freed
* after all rethook_node are freed (not soon). And the caller must
* not touch @rh after calling this.
*/
void rethook_free(struct rethook *rh)
{
WRITE_ONCE(rh->handler, NULL);
call_rcu(&rh->rcu, rethook_free_rcu);
}
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
static int rethook_init_node(void *nod, void *context)
{
struct rethook_node *node = nod;
node->rethook = context;
return 0;
}
static int rethook_fini_pool(struct objpool_head *head, void *context)
{
kfree(context);
return 0;
}
/**
* rethook_alloc() - Allocate struct rethook.
* @data: a data to pass the @handler when hooking the return.
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
* @handler: the return hook callback function, must NOT be NULL
* @size: node size: rethook node and additional data
* @num: number of rethook nodes to be preallocated
*
* Allocate and initialize a new rethook with @data and @handler.
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
* Return pointer of new rethook, or error codes for failures.
*
* Note that @handler == NULL means this rethook is going to be freed.
*/
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
struct rethook *rethook_alloc(void *data, rethook_handler_t handler,
int size, int num)
{
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
struct rethook *rh;
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
if (!handler || num <= 0 || size < sizeof(struct rethook_node))
return ERR_PTR(-EINVAL);
rh = kzalloc(sizeof(struct rethook), GFP_KERNEL);
if (!rh)
return ERR_PTR(-ENOMEM);
rh->data = data;
rh->handler = handler;
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
/* initialize the objpool for rethook nodes */
if (objpool_init(&rh->pool, num, size, GFP_KERNEL, rh,
rethook_init_node, rethook_fini_pool)) {
kfree(rh);
return ERR_PTR(-ENOMEM);
}
return rh;
}
static void free_rethook_node_rcu(struct rcu_head *head)
{
struct rethook_node *node = container_of(head, struct rethook_node, rcu);
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
struct rethook *rh = node->rethook;
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
objpool_drop(node, &rh->pool);
}
/**
* rethook_recycle() - return the node to rethook.
* @node: The struct rethook_node to be returned.
*
* Return back the @node to @node::rethook. If the @node::rethook is already
* marked as freed, this will free the @node.
*/
void rethook_recycle(struct rethook_node *node)
{
lockdep_assert_preemption_disabled();
if (likely(READ_ONCE(node->rethook->handler)))
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
objpool_push(node, &node->rethook->pool);
else
call_rcu(&node->rcu, free_rethook_node_rcu);
}
NOKPROBE_SYMBOL(rethook_recycle);
/**
* rethook_try_get() - get an unused rethook node.
* @rh: The struct rethook which pools the nodes.
*
* Get an unused rethook node from @rh. If the node pool is empty, this
* will return NULL. Caller must disable preemption.
*/
struct rethook_node *rethook_try_get(struct rethook *rh)
{
rethook_handler_t handler = READ_ONCE(rh->handler);
lockdep_assert_preemption_disabled();
/* Check whether @rh is going to be freed. */
if (unlikely(!handler))
return NULL;
/*
* This expects the caller will set up a rethook on a function entry.
* When the function returns, the rethook will eventually be reclaimed
* or released in the rethook_recycle() with call_rcu().
* This means the caller must be run in the RCU-availabe context.
*/
if (unlikely(!rcu_is_watching()))
return NULL;
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 13:56:52 +00:00
return (struct rethook_node *)objpool_pop(&rh->pool);
}
NOKPROBE_SYMBOL(rethook_try_get);
/**
* rethook_hook() - Hook the current function return.
* @node: The struct rethook node to hook the function return.
* @regs: The struct pt_regs for the function entry.
* @mcount: True if this is called from mcount(ftrace) context.
*
* Hook the current running function return. This must be called when the
* function entry (or at least @regs must be the registers of the function
* entry.) @mcount is used for identifying the context. If this is called
* from ftrace (mcount) callback, @mcount must be set true. If this is called
* from the real function entry (e.g. kprobes) @mcount must be set false.
* This is because the way to hook the function return depends on the context.
*/
void rethook_hook(struct rethook_node *node, struct pt_regs *regs, bool mcount)
{
arch_rethook_prepare(node, regs, mcount);
__llist_add(&node->llist, &current->rethooks);
}
NOKPROBE_SYMBOL(rethook_hook);
/* This assumes the 'tsk' is the current task or is not running. */
static unsigned long __rethook_find_ret_addr(struct task_struct *tsk,
struct llist_node **cur)
{
struct rethook_node *rh = NULL;
struct llist_node *node = *cur;
if (!node)
node = tsk->rethooks.first;
else
node = node->next;
while (node) {
rh = container_of(node, struct rethook_node, llist);
if (rh->ret_addr != (unsigned long)arch_rethook_trampoline) {
*cur = node;
return rh->ret_addr;
}
node = node->next;
}
return 0;
}
NOKPROBE_SYMBOL(__rethook_find_ret_addr);
/**
* rethook_find_ret_addr -- Find correct return address modified by rethook
* @tsk: Target task
* @frame: A frame pointer
* @cur: a storage of the loop cursor llist_node pointer for next call
*
* Find the correct return address modified by a rethook on @tsk in unsigned
* long type.
* The @tsk must be 'current' or a task which is not running. @frame is a hint
* to get the currect return address - which is compared with the
* rethook::frame field. The @cur is a loop cursor for searching the
* kretprobe return addresses on the @tsk. The '*@cur' should be NULL at the
* first call, but '@cur' itself must NOT NULL.
*
* Returns found address value or zero if not found.
*/
unsigned long rethook_find_ret_addr(struct task_struct *tsk, unsigned long frame,
struct llist_node **cur)
{
struct rethook_node *rhn = NULL;
unsigned long ret;
if (WARN_ON_ONCE(!cur))
return 0;
if (WARN_ON_ONCE(tsk != current && task_is_running(tsk)))
return 0;
do {
ret = __rethook_find_ret_addr(tsk, cur);
if (!ret)
break;
rhn = container_of(*cur, struct rethook_node, llist);
} while (rhn->frame != frame);
return ret;
}
NOKPROBE_SYMBOL(rethook_find_ret_addr);
void __weak arch_rethook_fixup_return(struct pt_regs *regs,
unsigned long correct_ret_addr)
{
/*
* Do nothing by default. If the architecture which uses a
* frame pointer to record real return address on the stack,
* it should fill this function to fixup the return address
* so that stacktrace works from the rethook handler.
*/
}
/* This function will be called from each arch-defined trampoline. */
unsigned long rethook_trampoline_handler(struct pt_regs *regs,
unsigned long frame)
{
struct llist_node *first, *node = NULL;
unsigned long correct_ret_addr;
rethook_handler_t handler;
struct rethook_node *rhn;
correct_ret_addr = __rethook_find_ret_addr(current, &node);
if (!correct_ret_addr) {
pr_err("rethook: Return address not found! Maybe there is a bug in the kernel\n");
BUG_ON(1);
}
instruction_pointer_set(regs, correct_ret_addr);
/*
* These loops must be protected from rethook_free_rcu() because those
* are accessing 'rhn->rethook'.
*/
preempt_disable_notrace();
/*
* Run the handler on the shadow stack. Do not unlink the list here because
* stackdump inside the handlers needs to decode it.
*/
first = current->rethooks.first;
while (first) {
rhn = container_of(first, struct rethook_node, llist);
if (WARN_ON_ONCE(rhn->frame != frame))
break;
handler = READ_ONCE(rhn->rethook->handler);
if (handler)
handler(rhn, rhn->rethook->data,
correct_ret_addr, regs);
if (first == node)
break;
first = first->next;
}
/* Fixup registers for returning to correct address. */
arch_rethook_fixup_return(regs, correct_ret_addr);
/* Unlink used shadow stack */
first = current->rethooks.first;
current->rethooks.first = node->next;
node->next = NULL;
while (first) {
rhn = container_of(first, struct rethook_node, llist);
first = first->next;
rethook_recycle(rhn);
}
preempt_enable_notrace();
return correct_ret_addr;
}
NOKPROBE_SYMBOL(rethook_trampoline_handler);