bpf: implement bpf_send_signal() helper
Subsystems (8)
PATCH v2 2/3 tools/bpf: sync bpf uapi header bpf.h to tools directory
In reply to Yonghong SongThe bpf uapi header include/uapi/linux/bpf.h is sync'ed to tools/include/uapi/linux/bpf.h.
--- tools/include/uapi/linux/bpf.h | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-)
tools/include/uapi/linux/bpf.h +16 -1
@@ -2672,6 +2672,20 @@ union bpf_attr {* 0 on success. * * **-ENOENT** if the bpf-local-storage cannot be found. +* +* int bpf_send_signal(u32 sig) +* Description +* Send signal *sig* to the current task. +* Return +* 0 on success or successfully queued. +* +* **-EBUSY** if work queue under nmi is full. +* +* **-EINVAL** if *sig* is invalid. +* +* **-EPERM** if no permission to send the *sig*. +* +* **-EAGAIN** if bpf program can try again. */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \
@@ -2782,7 +2796,8 @@ union bpf_attr {FN (strtol), \FN (strtoul), \FN (sk_storage_get), \ -FN (sk_storage_delete), +FN (sk_storage_delete), \ +FN (send_signal),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call
Signature · 1 line
2.17.1
v2 cover bpf: implement bpf_send_signal() helper
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.
bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.
This patch implements bpf_send_signal() helper to send a signal to hhvm in real time, resulting in intended stack traces.
Patch #1 implemented the bpf_send_helper() in the kernel, Patch #2 synced uapi header bpf.h to tools directory. Patch #3 added a self test which covers tracepoint and perf_event bpf programs.
Changelogs:
RFC v1 => v2:
. previous version allows to send signal to an arbitrary
pid. This version just sends the signal to current task to avoid unstable pid and potential races between sending signals and task state changes for the pid.
Yonghong Song (3):
bpf: implement bpf_send_signal() helper tools/bpf: sync bpf uapi header bpf.h to tools directory tools/bpf: add a selftest for bpf_send_signal() helper
include/uapi/linux/bpf.h | 17 +- kernel/trace/bpf_trace.c | 67 ++++++ tools/include/uapi/linux/bpf.h | 17 +- tools/testing/selftests/bpf/Makefile | 3 +- tools/testing/selftests/bpf/bpf_helpers.h | 1 + .../bpf/progs/test_send_signal_kern.c | 51 +++++ .../selftests/bpf/test_send_signal_user.c | 212 ++++++++++++++++++ 7 files changed, 365 insertions(+), 3 deletions(-) create mode 100644 tools/testing/selftests/bpf/progs/test_send_signal_kern.c create mode 100644 tools/testing/selftests/bpf/test_send_signal_user.c
Signature · 1 line
2.17.1
PATCH v2 1/3 bpf: implement bpf_send_signal() helper
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.
bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.
Such a mechanism can be implemented in the following way:
. a perf ring buffer is created between bpf program
and tracing app.
. once a particular event happens, bpf program writes
to the ring buffer and the tracing app gets notified.
. the tracing app sends a signal SIGALARM to the hhvm.
But this method could have large delays and causing profiling results skewed.
This patch implements bpf_send_signal() helper to send a signal to hhvm in real time, resulting in intended stack traces.
--- include/uapi/linux/bpf.h | 17 +++++++++- kernel/trace/bpf_trace.c | 67 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 83 insertions(+), 1 deletion(-)
include/uapi/linux/bpf.h +16 -1
@@ -2672,6 +2672,20 @@ union bpf_attr {* 0 on success. * * **-ENOENT** if the bpf-local-storage cannot be found. +* +* int bpf_send_signal(u32 sig) +* Description +* Send signal *sig* to the current task. +* Return +* 0 on success or successfully queued. +* +* **-EBUSY** if work queue under nmi is full. +* +* **-EINVAL** if *sig* is invalid. +* +* **-EPERM** if no permission to send the *sig*. +* +* **-EAGAIN** if bpf program can try again. */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \
@@ -2782,7 +2796,8 @@ union bpf_attr {FN (strtol), \FN (strtoul), \FN (sk_storage_get), \ -FN (sk_storage_delete), +FN (sk_storage_delete), \ +FN (send_signal),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call
kernel/trace/bpf_trace.c +67 -0
84 lines not on this page · raw message
Signature · 1 line
2.17.1
PATCH v2 3/3 tools/bpf: add a selftest for bpf_send_signal() helper
In reply to Yonghong SongThe test covered both nmi and tracepoint perf events.
$ ./test_send_signal_user test_send_signal (tracepoint): OK test_send_signal (perf_event): OK
--- tools/testing/selftests/bpf/Makefile | 3 +- tools/testing/selftests/bpf/bpf_helpers.h | 1 + .../bpf/progs/test_send_signal_kern.c | 51 +++++ .../selftests/bpf/test_send_signal_user.c | 212 ++++++++++++++++++ 4 files changed, 266 insertions(+), 1 deletion(-) create mode 100644 tools/testing/selftests/bpf/progs/test_send_signal_kern.c create mode 100644 tools/testing/selftests/bpf/test_send_signal_user.c
tools/testing/selftests/bpf/Makefile +2 -1
@@ -23,7 +23,8 @@ TEST_GEN_PROGS = test_verifier test_tag test_maps test_lru_map test_lpm_map test test_align test_verifier_log test_dev_cgroup test_tcpbpf_user \ test_sock test_btf test_sockmap test_lirc_mode2_user get_cgroup_id_user \ test_socket_cookie test_cgroup_storage test_select_reuseport test_section_names \ - test_netcnt test_tcpnotify_user test_sock_fields test_sysctl + test_netcnt test_tcpnotify_user test_sock_fields test_sysctl \ + test_send_signal_user BPF_OBJ_FILES = $(patsubst %.c,%.o, $(notdir $(wildcard progs/*.c))) TEST_GEN_FILES = $(BPF_OBJ_FILES)
tools/testing/selftests/bpf/bpf_helpers.h +1 -0
@@ -216,6 +216,7 @@ static void *(*bpf_sk_storage_get)(void *map, struct bpf_sock *sk, (void *) BPF_FUNC_sk_storage_get;static int (*bpf_sk_storage_delete)(void *map,struct bpf_sock *sk) = (void *)BPF_FUNC_sk_storage_delete; +static int (*bpf_send_signal)(unsigned sig) = (void *)BPF_FUNC_send_signal;/* llvm builtin functions that eBPF C program may use to * emit BPF_LD_ABS and BPF_LD_IND instructions
tools/testing/selftests/bpf/progs/test_send_signal_kern.c new file +51 -0
51 lines not on this page · raw message
tools/testing/selftests/bpf/test_send_signal_user.c new file +212 -0
212 lines not on this page · raw message
Signature · 1 line
2.17.1
Quoted 19 lines · Yonghong Song
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function
get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.
[..]
Quoted 2 lines
This patch implements
bpf_send_signal()helper to send a signal to hhvm in real time, resulting in intended stack traces.
Series looks good. One minor nit/question: maybe rename bpf_send_signal to something like bpf_send_signal_to_current/bpf_current_send_signal/etc? bpf_send_signal is too generic now that you send the signal to the current process..
Quoted 31 lines
Patch #1 implemented the
bpf_send_helper()in the kernel, Patch #2 synced uapi headerbpf.hto tools directory. Patch #3 added a self test which covers tracepoint andperf_eventbpf programs.Changelogs:
RFC v1 => v2:
. previous version allows to send signal to an arbitrary
pid. This version just sends the signal to current task to avoid unstable pid and potential races between sending signals and task state changes for the pid.Yonghong Song (3):
bpf: implementbpf_send_signal()helpertools/bpf: sync bpf uapi headerbpf.hto tools directorytools/bpf: add a selftest forbpf_send_signal()helperinclude/uapi/linux/bpf.h | 17 +- kernel/trace/bpf_trace.c | 67 ++++++ tools/include/uapi/linux/bpf.h | 17 +- tools/testing/selftests/bpf/Makefile | 3 +- tools/testing/selftests/bpf/bpf_helpers.h | 1 + .../bpf/progs/test_send_signal_kern.c | 51 +++++ .../selftests/bpf/test_send_signal_user.c | 212 ++++++++++++++++++ 7 files changed, 365 insertions(+), 3 deletions(-) create mode 100644 tools/testing/selftests/bpf/progs/test_send_signal_kern.c create mode 100644 tools/testing/selftests/bpf/test_send_signal_user.cSignature · 1 line
2.17.1
Quoted 21 lines · Stanislav Fomichev
Quoted 10 lines · Yonghong Song
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function
get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.Quoted 9 lines
bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.
Quoted 7 lines
[..]
Quoted 2 lines
This patch implements
bpf_send_signal()helper to send a signal to hhvm in real time, resulting in intended stack traces.Series looks good. One minor nit/question: maybe rename
bpf_send_signalto something like bpf_send_signal_to_current/bpf_current_send_signal/etc?bpf_send_signalis too generic now that you send the signal to the current process..
bpf_send_signal_to_current was Yonghong's original name and I asked him to rename it to bpf_send_signal :) I don't see bpf sending signals to arbitrary processes in the near future.
Quoted 30 lines · Alexei Starovoitov
Quoted 21 lines · Stanislav Fomichev
Quoted 10 lines · Yonghong Song
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function
get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.Quoted 9 lines
bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.Quoted 7 lines
[..]
Quoted 2 lines
This patch implements
bpf_send_signal()helper to send a signal to hhvm in real time, resulting in intended stack traces.Series looks good. One minor nit/question: maybe rename
bpf_send_signalto something like bpf_send_signal_to_current/bpf_current_send_signal/etc?bpf_send_signalis too generic now that you send the signal to the current process..
Quoted 4 lines
bpf_send_signal_to_currentwas Yonghong's original name and I asked him to rename it tobpf_send_signal:) I don't see bpf sending signals to arbitrary processes in the near future.
:) I was thinking that it might be a bit more clear if we communicate target process in the name, but I don't feel strongly about it.
So it's not about the future. OTOH, who knows, I remember when people said there would not be any locking/loops in bpf. And here we are :)
Re: [PATCH bpf-next v2 3/3] tools/bpf: add a selftest for bpf_send_signal() helper
In reply to Yonghong SongQuoted 305 lines · Yonghong Song
The test covered both nmi and tracepoint perf events.
$ ./test_send_signal_user test_send_signal (tracepoint): OK test_send_signal (perf_event): OKSigned-off-by Yonghong Song <[email protected]>--- tools/testing/selftests/bpf/Makefile | 3 +- tools/testing/selftests/bpf/bpf_helpers.h | 1 + .../bpf/progs/test_send_signal_kern.c | 51 +++++ .../selftests/bpf/test_send_signal_user.c | 212 ++++++++++++++++++ 4 files changed, 266 insertions(+), 1 deletion(-) create mode 100644 tools/testing/selftests/bpf/progs/test_send_signal_kern.c create mode 100644 tools/testing/selftests/bpf/test_send_signal_user.ctools/testing/selftests/bpf/Makefile +2
-1@@ -23,7 +23,8 @@ TEST_GEN_PROGS = test_verifier test_tag test_maps test_lru_map test_lpm_map test test_align test_verifier_log test_dev_cgroup test_tcpbpf_user \ test_sock test_btf test_sockmap test_lirc_mode2_user get_cgroup_id_user \ test_socket_cookie test_cgroup_storage test_select_reuseport test_section_names \ - test_netcnt test_tcpnotify_user test_sock_fields test_sysctl + test_netcnt test_tcpnotify_user test_sock_fields test_sysctl \ + test_send_signal_user BPF_OBJ_FILES = $(patsubst %.c,%.o, $(notdir $(wildcard progs/*.c))) TEST_GEN_FILES = $(BPF_OBJ_FILES)tools/testing/selftests/bpf/bpf_helpers.h +1
-0@@ -216,6 +216,7 @@ static void *(*bpf_sk_storage_get)(void *map, struct bpf_sock *sk, (void *) BPF_FUNC_sk_storage_get;static int (*bpf_sk_storage_delete)(void *map,struct bpf_sock *sk) = (void *)BPF_FUNC_sk_storage_delete; +static int (*bpf_send_signal)(unsigned sig) = (void *)BPF_FUNC_send_signal;/* llvm builtin functions that eBPF C program may use to * emit BPF_LD_ABS and BPF_LD_IND instructions tools/testing/selftests/bpf/progs/test_send_signal_kern.c new file +51
-0@@ -0,0 +1,51 @@ +// SPDX-License-Identifier: GPL-2.0 +// Copyright (c) 2019 Facebook +#include <linux/bpf.h> +#include <linux/version.h> +#include "bpf_helpers.h" + +struct bpf_map_def SEC("maps") info_map = { + .type = BPF_MAP_TYPE_ARRAY, + .key_size = sizeof(__u32), + .value_size = sizeof(__u64), + .max_entries = 1, +}; + +BPF_ANNOTATE_KV_PAIR (info_map, __u32, __u64); + +struct bpf_map_def SEC("maps") status_map = { + .type = BPF_MAP_TYPE_ARRAY, + .key_size = sizeof(__u32), + .value_size = sizeof(__u64), + .max_entries = 1, +}; + +BPF_ANNOTATE_KV_PAIR (status_map, __u32, __u64); + +SEC("send_signal_demo") +int bpf_send_signal_test (void *ctx) +{ +__u64 *info_val, *status_val; +__u32 key =0 , pid, sig; +int ret; + + status_val =bpf_map_lookup_elem (&status_map, &key); +if (!status_val || *status_val !=0 ) +return 0 ; + + info_val =bpf_map_lookup_elem (&info_map, &key); +if (!info_val || *info_val ==0 ) +return 0 ; + + sig = *info_val >>32 ; + pid = *info_val &0xffffFFFF ; + +if ((bpf_get_current_pid_tgid () >>32 ) == pid) { + ret =bpf_send_signal (sig); +if (ret ==0 ) + *status_val =1 ; + } + +return 0 ; +} +char __license[] SEC("license") = "GPL";tools/testing/selftests/bpf/test_send_signal_user.c new file +200
-0@@ -0,0 +1,212 @@ +// SPDX-License-Identifier: GPL-2.0 +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <errno.h> +#include <signal.h> +#include <syscall.h> +#include <sys/ioctl.h> +#include <sys/types.h> +#include <sys/stat.h> +#include <sys/wait.h> +#include <fcntl.h> +#include <unistd.h> + +#include <linux/perf_event.h> +#include <bpf/bpf.h> +#include <bpf/libbpf.h> + +#include "bpf_rlimit.h" + +#define CHECK(condition, tag, format...) ({ \ + int __ret = !!(condition); \ + if (__ret) { \ + printf("%s(%d):FAIL:%s ", __func__, __LINE__, tag); \ + printf(format); \ + printf("\n"); \ + } \ + __ret; \ +}) + +static volatile int signal_received =0 ; + +static void sigusr1_handler (int signum) +{ + signal_received++; +} + +static void test_common (struct perf_event_attr *attr,int prog_type, +const char *test_name) +{ +int pmu_fd, prog_fd, info_map_fd, status_map_fd; +const char *file ="./test_send_signal_kern.o" ; +struct bpf_object *obj =NULL ; +int pipe_c2p[2 ], pipe_p2c[2 ]; +char buf[256 ]; +int err =0 ; +u32 key =0 ; +pid_t pid; +u64 val; + +if (CHECK (pipe (pipe_c2p), test_name, +"pipe pipe_c2p error: %s" ,strerror (errno))) +return ; + +if (CHECK (pipe (pipe_p2c), test_name, +"pipe pipe_p2c error: %s" ,strerror (errno))) { +close (pipe_c2p[0 ]); +close (pipe_c2p[1 ]); +return ; + } + + pid =fork (); +if (CHECK (pid <0 , test_name,"fork error: %s" ,strerror (errno))) { +close (pipe_c2p[0 ]); +close (pipe_c2p[1 ]); +close (pipe_p2c[0 ]); +close (pipe_p2c[1 ]); +return ; + } + +if (pid ==0 ) { +/* install signal handler and notify parent */ +signal (SIGUSR1 , sigusr1_handler); + +close (pipe_c2p[0 ]);/* close read */ +close (pipe_p2c[1 ]);/* close write */ + +/* notify parent signal handler is installed */ +write (pipe_c2p[1 ], buf,1 ); + +/* make sense parent enabled bpf program to send_signal */ +read (pipe_p2c[0 ], buf,1 ); + +/* wait a little for signal handler */ +sleep (1 ); + +if (signal_received) +write (pipe_c2p[1 ],"2" ,1 ); +else +write (pipe_c2p[1 ],"0" ,1 ); + +/* wait for parent notification and exit */ +read (pipe_p2c[0 ], buf,1 ); + +close (pipe_c2p[1 ]); +close (pipe_p2c[0 ]); +exit (0 ); + } + +close (pipe_c2p[1 ]);/* close write */ +close (pipe_p2c[0 ]);/* close read */ + + err =bpf_prog_load (file, prog_type, &obj, &prog_fd); +if (CHECK (err <0 , test_name,"bpf_prog_load error: %s" , +strerror (errno))) + goto prog_load_failure; + + pmu_fd =syscall (__NR_perf_event_open, attr, pid,-1 , +-1 /* group id */ ,0 /* flags */ ); +if (CHECK (pmu_fd <0 , test_name,"perf_event_open error: %s" , +strerror (errno))) + goto close_prog; + + err =ioctl (pmu_fd,PERF_EVENT_IOC_ENABLE ,0 ); +if (CHECK (err <0 , test_name,"ioctl perf_event_ioc_enable error: %s" , +strerror (errno))) + goto disable_pmu; + + err =ioctl (pmu_fd,PERF_EVENT_IOC_SET_BPF , prog_fd); +if (CHECK (err <0 , test_name,"ioctl perf_event_ioc_set_bpf error: %s" , +strerror (errno))) + goto disable_pmu; + + info_map_fd =bpf_object__find_map_fd_by_name (obj,"info_map" ); +if (CHECK (info_map_fd <0 , test_name,"find map %s error" ,"info_map" )) + goto disable_pmu; + + status_map_fd =bpf_object__find_map_fd_by_name (obj,"status_map" ); +if (CHECK (status_map_fd <0 , test_name,"find map %s error" ,"status_map" )) + goto disable_pmu; + +/* wait until child signal handler installed */ +read (pipe_c2p[0 ], buf,1 ); + +/* trigger the bpf send_signal */ + key =0 ; + val = (((u64)(SIGUSR1 )) <<32 ) | pid; +bpf_map_update_elem (info_map_fd, &key, &val,0 ); + +/* notify child that bpf program can send_signal now */ +write (pipe_p2c[1 ], buf,1 ); + +/* wait for result */ +read (pipe_c2p[0 ], buf,1 ); + +if (buf[0 ] =='2' ) +printf ("test_send_signal (%s): OK\n" , test_name); +else +printf ("test_send_signal (%s): FAIL\n" , test_name); + +/* notify child safe to exit */ +write (pipe_p2c[1 ], buf,1 ); + +disable_pmu: +close (pmu_fd); +close_prog: +bpf_object__close (obj); +prog_load_failure: +close (pipe_c2p[0 ]); +close (pipe_p2c[1 ]); +wait (NULL ); +} + +static void test_tracepoint (void ) +{ +struct perf_event_attr attr = { + .type =PERF_TYPE_TRACEPOINT , + .sample_type =PERF_SAMPLE_RAW |PERF_SAMPLE_CALLCHAIN , + .sample_period =1 , + .wakeup_events =1 , + }; +int bytes, efd; +char buf[256 ]; + +snprintf (buf,sizeof (buf), +"/sys/kernel/debug/tracing/events/syscalls/sys_enter_nanosleep/id" ); + efd =open (buf,O_RDONLY ,0 ); +if (CHECK (efd <0 ,"tracepoint" , +"open syscalls/sys_enter_nanosleep/id failure: %s" , +strerror (errno))) +return ; + + bytes =read (efd, buf,sizeof (buf)); +close (efd); +if (CHECK (bytes <=0 || bytes >=sizeof (buf),"tracepoint" , +"read syscalls/sys_enter_nanosleep/id failure: %s" , +strerror (errno))) +return ; + + attr.config =strtol (buf,NULL ,0 ); + +test_common (&attr,BPF_PROG_TYPE_TRACEPOINT ,"tracepoint" ); +} + +static void test_nmi_perf_event (void ) +{ +struct perf_event_attr attr = { + .sample_freq =50 , + .freq =1 , + .type =PERF_TYPE_HARDWARE ,
tools/testing/selftests/bpf/test_send_signal_user.c line 210
Quoted 10 lines
tools/testing/selftests/bpf/test_send_signal_user.c new file +10
-0+ .config = PERF_COUNT_HW_CPU_CYCLES, + }; + +test_common (&attr,BPF_PROG_TYPE_PERF_EVENT ,"perf_event" ); +} + +int main (void ) +{ +test_tracepoint (); +test_nmi_perf_event ();
Tests should probably propagate failure up to main() and return exit code != 0, if any of the tests failed.
Quoted 5 lines
tools/testing/selftests/bpf/test_send_signal_user.c new file +2
-1+return 0 ; +} --2.17.1
Re: [PATCH bpf-next v2 3/3] tools/bpf: add a selftest for bpf_send_signal() helper
In reply to Yonghong SongQuoted 102 lines · Yonghong Song
The test covered both nmi and tracepoint perf events.
$ ./test_send_signal_user test_send_signal (tracepoint): OK test_send_signal (perf_event): OKSigned-off-by Yonghong Song <[email protected]>--- tools/testing/selftests/bpf/Makefile | 3 +- tools/testing/selftests/bpf/bpf_helpers.h | 1 + .../bpf/progs/test_send_signal_kern.c | 51 +++++ .../selftests/bpf/test_send_signal_user.c | 212 ++++++++++++++++++ 4 files changed, 266 insertions(+), 1 deletion(-) create mode 100644 tools/testing/selftests/bpf/progs/test_send_signal_kern.c create mode 100644 tools/testing/selftests/bpf/test_send_signal_user.ctools/testing/selftests/bpf/Makefile +2
-1@@ -23,7 +23,8 @@ TEST_GEN_PROGS = test_verifier test_tag test_maps test_lru_map test_lpm_map test test_align test_verifier_log test_dev_cgroup test_tcpbpf_user \ test_sock test_btf test_sockmap test_lirc_mode2_user get_cgroup_id_user \ test_socket_cookie test_cgroup_storage test_select_reuseport test_section_names \ - test_netcnt test_tcpnotify_user test_sock_fields test_sysctl + test_netcnt test_tcpnotify_user test_sock_fields test_sysctl \ + test_send_signal_user BPF_OBJ_FILES = $(patsubst %.c,%.o, $(notdir $(wildcard progs/*.c))) TEST_GEN_FILES = $(BPF_OBJ_FILES)tools/testing/selftests/bpf/bpf_helpers.h +1
-0@@ -216,6 +216,7 @@ static void *(*bpf_sk_storage_get)(void *map, struct bpf_sock *sk, (void *) BPF_FUNC_sk_storage_get;static int (*bpf_sk_storage_delete)(void *map,struct bpf_sock *sk) = (void *)BPF_FUNC_sk_storage_delete; +static int (*bpf_send_signal)(unsigned sig) = (void *)BPF_FUNC_send_signal;/* llvm builtin functions that eBPF C program may use to * emit BPF_LD_ABS and BPF_LD_IND instructions tools/testing/selftests/bpf/progs/test_send_signal_kern.c new file +51
-0@@ -0,0 +1,51 @@ +// SPDX-License-Identifier: GPL-2.0 +// Copyright (c) 2019 Facebook +#include <linux/bpf.h> +#include <linux/version.h> +#include "bpf_helpers.h" + +struct bpf_map_def SEC("maps") info_map = { + .type = BPF_MAP_TYPE_ARRAY, + .key_size = sizeof(__u32), + .value_size = sizeof(__u64), + .max_entries = 1, +}; + +BPF_ANNOTATE_KV_PAIR (info_map, __u32, __u64); + +struct bpf_map_def SEC("maps") status_map = { + .type = BPF_MAP_TYPE_ARRAY, + .key_size = sizeof(__u32), + .value_size = sizeof(__u64), + .max_entries = 1, +}; + +BPF_ANNOTATE_KV_PAIR (status_map, __u32, __u64); + +SEC("send_signal_demo") +int bpf_send_signal_test (void *ctx) +{ +__u64 *info_val, *status_val; +__u32 key =0 , pid, sig; +int ret; + + status_val =bpf_map_lookup_elem (&status_map, &key); +if (!status_val || *status_val !=0 ) +return 0 ; + + info_val =bpf_map_lookup_elem (&info_map, &key); +if (!info_val || *info_val ==0 ) +return 0 ; + + sig = *info_val >>32 ; + pid = *info_val &0xffffFFFF ; + +if ((bpf_get_current_pid_tgid () >>32 ) == pid) { + ret =bpf_send_signal (sig); +if (ret ==0 ) + *status_val =1 ; + } + +return 0 ; +} +char __license[] SEC("license") = "GPL";tools/testing/selftests/bpf/test_send_signal_user.c new file +0
-1--- /dev/null
[..]
Quoted 1 line
+++
b/tools/testing/selftests/bpf/test_send_signal_user.c
Any reason you didn't put it under bpf/prog_tests? That way you don't need to define your own CHECK macro and care about the includes.
Quoted 216 lines
tools/testing/selftests/bpf/test_send_signal_user.c new file +212
-0@@ -0,0 +1,212 @@ +// SPDX-License-Identifier: GPL-2.0 +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <errno.h> +#include <signal.h> +#include <syscall.h> +#include <sys/ioctl.h> +#include <sys/types.h> +#include <sys/stat.h> +#include <sys/wait.h> +#include <fcntl.h> +#include <unistd.h> + +#include <linux/perf_event.h> +#include <bpf/bpf.h> +#include <bpf/libbpf.h> + +#include "bpf_rlimit.h" + +#define CHECK(condition, tag, format...) ({ \ + int __ret = !!(condition); \ + if (__ret) { \ + printf("%s(%d):FAIL:%s ", __func__, __LINE__, tag); \ + printf(format); \ + printf("\n"); \ + } \ + __ret; \ +}) + +static volatile int signal_received =0 ; + +static void sigusr1_handler (int signum) +{ + signal_received++; +} + +static void test_common (struct perf_event_attr *attr,int prog_type, +const char *test_name) +{ +int pmu_fd, prog_fd, info_map_fd, status_map_fd; +const char *file ="./test_send_signal_kern.o" ; +struct bpf_object *obj =NULL ; +int pipe_c2p[2 ], pipe_p2c[2 ]; +char buf[256 ]; +int err =0 ; +u32 key =0 ; +pid_t pid; +u64 val; + +if (CHECK (pipe (pipe_c2p), test_name, +"pipe pipe_c2p error: %s" ,strerror (errno))) +return ; + +if (CHECK (pipe (pipe_p2c), test_name, +"pipe pipe_p2c error: %s" ,strerror (errno))) { +close (pipe_c2p[0 ]); +close (pipe_c2p[1 ]); +return ; + } + + pid =fork (); +if (CHECK (pid <0 , test_name,"fork error: %s" ,strerror (errno))) { +close (pipe_c2p[0 ]); +close (pipe_c2p[1 ]); +close (pipe_p2c[0 ]); +close (pipe_p2c[1 ]); +return ; + } + +if (pid ==0 ) { +/* install signal handler and notify parent */ +signal (SIGUSR1 , sigusr1_handler); + +close (pipe_c2p[0 ]);/* close read */ +close (pipe_p2c[1 ]);/* close write */ + +/* notify parent signal handler is installed */ +write (pipe_c2p[1 ], buf,1 ); + +/* make sense parent enabled bpf program to send_signal */ +read (pipe_p2c[0 ], buf,1 ); + +/* wait a little for signal handler */ +sleep (1 ); + +if (signal_received) +write (pipe_c2p[1 ],"2" ,1 ); +else +write (pipe_c2p[1 ],"0" ,1 ); + +/* wait for parent notification and exit */ +read (pipe_p2c[0 ], buf,1 ); + +close (pipe_c2p[1 ]); +close (pipe_p2c[0 ]); +exit (0 ); + } + +close (pipe_c2p[1 ]);/* close write */ +close (pipe_p2c[0 ]);/* close read */ + + err =bpf_prog_load (file, prog_type, &obj, &prog_fd); +if (CHECK (err <0 , test_name,"bpf_prog_load error: %s" , +strerror (errno))) + goto prog_load_failure; + + pmu_fd =syscall (__NR_perf_event_open, attr, pid,-1 , +-1 /* group id */ ,0 /* flags */ ); +if (CHECK (pmu_fd <0 , test_name,"perf_event_open error: %s" , +strerror (errno))) + goto close_prog; + + err =ioctl (pmu_fd,PERF_EVENT_IOC_ENABLE ,0 ); +if (CHECK (err <0 , test_name,"ioctl perf_event_ioc_enable error: %s" , +strerror (errno))) + goto disable_pmu; + + err =ioctl (pmu_fd,PERF_EVENT_IOC_SET_BPF , prog_fd); +if (CHECK (err <0 , test_name,"ioctl perf_event_ioc_set_bpf error: %s" , +strerror (errno))) + goto disable_pmu; + + info_map_fd =bpf_object__find_map_fd_by_name (obj,"info_map" ); +if (CHECK (info_map_fd <0 , test_name,"find map %s error" ,"info_map" )) + goto disable_pmu; + + status_map_fd =bpf_object__find_map_fd_by_name (obj,"status_map" ); +if (CHECK (status_map_fd <0 , test_name,"find map %s error" ,"status_map" )) + goto disable_pmu; + +/* wait until child signal handler installed */ +read (pipe_c2p[0 ], buf,1 ); + +/* trigger the bpf send_signal */ + key =0 ; + val = (((u64)(SIGUSR1 )) <<32 ) | pid; +bpf_map_update_elem (info_map_fd, &key, &val,0 ); + +/* notify child that bpf program can send_signal now */ +write (pipe_p2c[1 ], buf,1 ); + +/* wait for result */ +read (pipe_c2p[0 ], buf,1 ); + +if (buf[0 ] =='2' ) +printf ("test_send_signal (%s): OK\n" , test_name); +else +printf ("test_send_signal (%s): FAIL\n" , test_name); + +/* notify child safe to exit */ +write (pipe_p2c[1 ], buf,1 ); + +disable_pmu: +close (pmu_fd); +close_prog: +bpf_object__close (obj); +prog_load_failure: +close (pipe_c2p[0 ]); +close (pipe_p2c[1 ]); +wait (NULL ); +} + +static void test_tracepoint (void ) +{ +struct perf_event_attr attr = { + .type =PERF_TYPE_TRACEPOINT , + .sample_type =PERF_SAMPLE_RAW |PERF_SAMPLE_CALLCHAIN , + .sample_period =1 , + .wakeup_events =1 , + }; +int bytes, efd; +char buf[256 ]; + +snprintf (buf,sizeof (buf), +"/sys/kernel/debug/tracing/events/syscalls/sys_enter_nanosleep/id" ); + efd =open (buf,O_RDONLY ,0 ); +if (CHECK (efd <0 ,"tracepoint" , +"open syscalls/sys_enter_nanosleep/id failure: %s" , +strerror (errno))) +return ; + + bytes =read (efd, buf,sizeof (buf)); +close (efd); +if (CHECK (bytes <=0 || bytes >=sizeof (buf),"tracepoint" , +"read syscalls/sys_enter_nanosleep/id failure: %s" , +strerror (errno))) +return ; + + attr.config =strtol (buf,NULL ,0 ); + +test_common (&attr,BPF_PROG_TYPE_TRACEPOINT ,"tracepoint" ); +} + +static void test_nmi_perf_event (void ) +{ +struct perf_event_attr attr = { + .sample_freq =50 , + .freq =1 , + .type =PERF_TYPE_HARDWARE , + .config =PERF_COUNT_HW_CPU_CYCLES , + }; + +test_common (&attr,BPF_PROG_TYPE_PERF_EVENT ,"perf_event" ); +} + +int main (void ) +{ +test_tracepoint (); +test_nmi_perf_event (); +return 0 ; +}Signature · 1 line
2.17.1
Re: [PATCH bpf-next v2 3/3] tools/bpf: add a selftest for bpf_send_signal() helper
In reply to Andrii NakryikoQuoted 35 lines · Andrii Nakryiko
Quoted 34 lines · Yonghong Song
The test covered both nmi and tracepoint perf events.
$ ./test_send_signal_user test_send_signal (tracepoint): OK test_send_signal (perf_event): OKSigned-off-by Yonghong Song <[email protected]>--- tools/testing/selftests/bpf/Makefile | 3 +- tools/testing/selftests/bpf/bpf_helpers.h | 1 + .../bpf/progs/test_send_signal_kern.c | 51 +++++ .../selftests/bpf/test_send_signal_user.c | 212 ++++++++++++++++++ 4 files changed, 266 insertions(+), 1 deletion(-)create mode 100644
tools/testing/selftests/bpf/progs/test_send_signal_kern.ccreate mode 100644tools/testing/selftests/bpf/test_send_signal_user.ctools/testing/selftests/bpf/Makefile +2
-1@@ -23,7 +23,8 @@ TEST_GEN_PROGS = test_verifier test_tag test_maps test_lru_map test_lpm_map test test_align test_verifier_log test_dev_cgroup test_tcpbpf_user \ test_sock test_btf test_sockmap test_lirc_mode2_user get_cgroup_id_user \ test_socket_cookie test_cgroup_storage test_select_reuseport test_section_names \ - test_netcnt test_tcpnotify_user test_sock_fields test_sysctl + test_netcnt test_tcpnotify_user test_sock_fields test_sysctl \ + test_send_signal_user BPF_OBJ_FILES = $(patsubst %.c,%.o, $(notdir $(wildcard progs/*.c))) TEST_GEN_FILES = $(BPF_OBJ_FILES)tools/testing/selftests/bpf/bpf_helpers.h +0
-0
[...]
Quoted 8 lines
tools/testing/selftests/bpf/bpf_helpers.h
Quoted 5 lines
tools/testing/selftests/bpf/bpf_helpers.h +5
-0+ +int main (void ) +{ +test_tracepoint (); +test_nmi_perf_event ();Tests should probably propagate failure up to
main()and return exit code != 0, if any of the tests failed.
Good catch! Will fix it in the next revision.
Quoted 6 lines
Quoted 6 lines
tools/testing/selftests/bpf/bpf_helpers.h +2
-1+return 0 ; +} --2.17.1
Re: [PATCH bpf-next v2 3/3] tools/bpf: add a selftest for bpf_send_signal() helper
In reply to Stanislav FomichevQuoted 104 lines · Stanislav Fomichev
Quoted 102 lines · Yonghong Song
The test covered both nmi and tracepoint perf events.
$ ./test_send_signal_user test_send_signal (tracepoint): OK test_send_signal (perf_event): OKSigned-off-by Yonghong Song <[email protected]>--- tools/testing/selftests/bpf/Makefile | 3 +- tools/testing/selftests/bpf/bpf_helpers.h | 1 + .../bpf/progs/test_send_signal_kern.c | 51 +++++ .../selftests/bpf/test_send_signal_user.c | 212 ++++++++++++++++++ 4 files changed, 266 insertions(+), 1 deletion(-)create mode 100644
tools/testing/selftests/bpf/progs/test_send_signal_kern.ccreate mode 100644tools/testing/selftests/bpf/test_send_signal_user.ctools/testing/selftests/bpf/Makefile +2
-1@@ -23,7 +23,8 @@ TEST_GEN_PROGS = test_verifier test_tag test_maps test_lru_map test_lpm_map test test_align test_verifier_log test_dev_cgroup test_tcpbpf_user \ test_sock test_btf test_sockmap test_lirc_mode2_user get_cgroup_id_user \ test_socket_cookie test_cgroup_storage test_select_reuseport test_section_names \ - test_netcnt test_tcpnotify_user test_sock_fields test_sysctl + test_netcnt test_tcpnotify_user test_sock_fields test_sysctl \ + test_send_signal_user BPF_OBJ_FILES = $(patsubst %.c,%.o, $(notdir $(wildcard progs/*.c))) TEST_GEN_FILES = $(BPF_OBJ_FILES)tools/testing/selftests/bpf/bpf_helpers.h +1
-0@@ -216,6 +216,7 @@ static void *(*bpf_sk_storage_get)(void *map, struct bpf_sock *sk, (void *) BPF_FUNC_sk_storage_get;static int (*bpf_sk_storage_delete)(void *map,struct bpf_sock *sk) = (void *)BPF_FUNC_sk_storage_delete; +static int (*bpf_send_signal)(unsigned sig) = (void *)BPF_FUNC_send_signal;/* llvm builtin functions that eBPF C program may use to * emit BPF_LD_ABS and BPF_LD_IND instructions tools/testing/selftests/bpf/progs/test_send_signal_kern.c new file +51
-0@@ -0,0 +1,51 @@ +// SPDX-License-Identifier: GPL-2.0 +// Copyright (c) 2019 Facebook +#include <linux/bpf.h> +#include <linux/version.h> +#include "bpf_helpers.h" + +struct bpf_map_def SEC("maps") info_map = { + .type = BPF_MAP_TYPE_ARRAY, + .key_size = sizeof(__u32), + .value_size = sizeof(__u64), + .max_entries = 1, +}; + +BPF_ANNOTATE_KV_PAIR (info_map, __u32, __u64); + +struct bpf_map_def SEC("maps") status_map = { + .type = BPF_MAP_TYPE_ARRAY, + .key_size = sizeof(__u32), + .value_size = sizeof(__u64), + .max_entries = 1, +}; + +BPF_ANNOTATE_KV_PAIR (status_map, __u32, __u64); + +SEC("send_signal_demo") +int bpf_send_signal_test (void *ctx) +{ +__u64 *info_val, *status_val; +__u32 key =0 , pid, sig; +int ret; + + status_val =bpf_map_lookup_elem (&status_map, &key); +if (!status_val || *status_val !=0 ) +return 0 ; + + info_val =bpf_map_lookup_elem (&info_map, &key); +if (!info_val || *info_val ==0 ) +return 0 ; + + sig = *info_val >>32 ; + pid = *info_val &0xffffFFFF ; + +if ((bpf_get_current_pid_tgid () >>32 ) == pid) { + ret =bpf_send_signal (sig); +if (ret ==0 ) + *status_val =1 ; + } + +return 0 ; +} +char __license[] SEC("license") = "GPL";tools/testing/selftests/bpf/test_send_signal_user.c new file +0
-1--- /dev/null[..]
Quoted 2 lines
Quoted 1 line
+++
b/tools/testing/selftests/bpf/test_send_signal_user.cAny reason you didn't put it under bpf/prog_tests?
The only reason I put it as a standalone test is that the program receives signals and tries to minimize potential impact on other tests.
But it might be okay as the signal is sent to child process.
Quoted 2 lines
That way you don't need to define your own CHECK macro and care about the includes.
Right. I will try to use bpf/prog_tests infrastructure in the next revision.
Daniel Borkmann
maintainerQuoted 103 lines · Yonghong Song
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function
get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.Such a mechanism can be implemented in the following way:
. a perf ring buffer is created between bpf program
and tracing app.
. once a particular event happens, bpf program writes
to the ring buffer and the tracing app gets notified.
. the tracing app sends a signal SIGALARM to the hhvm.But this method could have large delays and causing profiling results skewed.
This patch implements
bpf_send_signal()helper to send a signal to hhvm in real time, resulting in intended stack traces.Signed-off-by Yonghong Song <[email protected]>--- include/uapi/linux/bpf.h | 17 +++++++++- kernel/trace/bpf_trace.c | 67 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 83 insertions(+), 1 deletion(-)include/uapi/linux/bpf.h +16
-1@@ -2672,6 +2672,20 @@ union bpf_attr {* 0 on success. * * **-ENOENT** if the bpf-local-storage cannot be found. +* +* int bpf_send_signal(u32 sig) +* Description +* Send signal *sig* to the current task. +* Return +* 0 on success or successfully queued. +* +* **-EBUSY** if work queue under nmi is full. +* +* **-EINVAL** if *sig* is invalid. +* +* **-EPERM** if no permission to send the *sig*. +* +* **-EAGAIN** if bpf program can try again. */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \@@ -2782,7 +2796,8 @@ union bpf_attr {FN (strtol), \FN (strtoul), \FN (sk_storage_get), \ -FN (sk_storage_delete), +FN (sk_storage_delete), \ +FN (send_signal),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call kernel/trace/bpf_trace.c +22
-0@@ -567,6 +567,58 @@ static const struct bpf_func_proto bpf_probe_read_str_proto = { .arg3_type = ARG_ANYTHING, }; +struct send_signal_irq_work { +struct irq_work irq_work; +u32 sig; +}; + +static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work); + +static void do_bpf_send_signal (struct irq_work *entry) +{ +struct send_signal_irq_work *work; + + work = container_of(entry, struct send_signal_irq_work, irq_work); +group_send_sig_info (work->sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +} + +BPF_CALL_1(bpf_send_signal, u32, sig) +{ +struct send_signal_irq_work *work =NULL ; + +/* Similar to bpf_probe_write_user, task needs to be +* in a sound condition and kernel memory access be +* permitted in order to send signal to the current
kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 601
Quoted 10 lines
kernel/trace/bpf_trace.c +10
-0+* task. +*/ +if (unlikely (current->flags & (PF_KTHREAD |PF_EXITING ))) +return -EPERM ; +if (unlikely (uaccess_kernel ())) +return -EPERM ; +if (unlikely (!nmi_uaccess_okay ())) +return -EPERM ; + +if (in_nmi ()) {
Hm, bit confused, can't this only be done out of process context in general since only there current points to e.g. hhvm? I'm probably missing something. Could you elaborate?
kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 612
Quoted 11 lines
kernel/trace/bpf_trace.c +11
-0+ work =this_cpu_ptr (&send_signal_work); +if (work->irq_work.flags &IRQ_WORK_BUSY ) +return -EBUSY ; + + work->sig = sig; +irq_work_queue (&work->irq_work); +return 0 ; + } + +return group_send_sig_info (sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +
Nit: extra newline slipped in
Quoted 41 lines
kernel/trace/bpf_trace.c +24
-0+} + +static const struct bpf_func_proto bpf_send_signal_proto = { + .func = bpf_send_signal, + .gpl_only = false, + .ret_type =RET_INTEGER , + .arg1_type =ARG_ANYTHING , +}; +static const struct bpf_func_proto *tracing_func_proto (enum bpf_func_id func_id,const struct bpf_prog *prog) {@@ -617,6 +669,8 @@ tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)case BPF_FUNC_get_current_cgroup_id:return &bpf_get_current_cgroup_id_proto;#endif +case BPF_FUNC_send_signal: +return &bpf_send_signal_proto;default :return NULL ; }@@ -1343,5 +1397,18 @@ static int __init bpf_event_init(void) return 0; } +static int __init send_signal_irq_work_init (void ) +{ +int cpu; +struct send_signal_irq_work *work; + +for_each_possible_cpu (cpu) { + work =per_cpu_ptr (&send_signal_work, cpu); +init_irq_work (&work->irq_work, do_bpf_send_signal); + } +return 0 ; +} +fs_initcall (bpf_event_init); +subsys_initcall (send_signal_irq_work_init);#endif /* CONFIG_MODULES */
Quoted 110 lines · Daniel Borkmann
Quoted 109 lines · Yonghong Song
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function
get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.Such a mechanism can be implemented in the following way:
. a perf ring buffer is created between bpf program
and tracing app.
. once a particular event happens, bpf program writes
to the ring buffer and the tracing app gets notified.
. the tracing app sends a signal SIGALARM to the hhvm.But this method could have large delays and causing profiling results skewed.
This patch implements
bpf_send_signal()helper to send a signal to hhvm in real time, resulting in intended stack traces.Signed-off-by Yonghong Song <[email protected]>--- include/uapi/linux/bpf.h | 17 +++++++++- kernel/trace/bpf_trace.c | 67 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 83 insertions(+), 1 deletion(-)include/uapi/linux/bpf.h +16
-1@@ -2672,6 +2672,20 @@ union bpf_attr {* 0 on success. * * **-ENOENT** if the bpf-local-storage cannot be found. +* +* int bpf_send_signal(u32 sig) +* Description +* Send signal *sig* to the current task. +* Return +* 0 on success or successfully queued. +* +* **-EBUSY** if work queue under nmi is full. +* +* **-EINVAL** if *sig* is invalid. +* +* **-EPERM** if no permission to send the *sig*. +* +* **-EAGAIN** if bpf program can try again. */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \@@ -2782,7 +2796,8 @@ union bpf_attr {FN (strtol), \FN (strtoul), \FN (sk_storage_get), \ -FN (sk_storage_delete), +FN (sk_storage_delete), \ +FN (send_signal),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call kernel/trace/bpf_trace.c +28
-0@@ -567,6 +567,58 @@ static const struct bpf_func_proto bpf_probe_read_str_proto = { .arg3_type = ARG_ANYTHING, }; +struct send_signal_irq_work { +struct irq_work irq_work; +u32 sig; +}; + +static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work); + +static void do_bpf_send_signal (struct irq_work *entry) +{ +struct send_signal_irq_work *work; + + work = container_of(entry, struct send_signal_irq_work, irq_work); +group_send_sig_info (work->sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +} + +BPF_CALL_1(bpf_send_signal, u32, sig) +{ +struct send_signal_irq_work *work =NULL ; + +/* Similar to bpf_probe_write_user, task needs to be +* in a sound condition and kernel memory access be +* permitted in order to send signal to the current +* task. +*/ +if (unlikely (current->flags & (PF_KTHREAD |PF_EXITING ))) +return -EPERM ; +if (unlikely (uaccess_kernel ())) +return -EPERM ;
Quoted 8 lines
kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 601
Quoted 4 lines
kernel/trace/bpf_trace.c +4
-0+if (unlikely (!nmi_uaccess_okay ())) +return -EPERM ; + +if (in_nmi ()) {Hm, bit confused, can't this only be done out of process context in general since only there current points to e.g. hhvm? I'm probably missing something. Could you elaborate?
That is true. If in nmi, it is out of process context and in nmi context, we use an irq_work here since group_send_sig_info() has spinlock inside. The bpf program (e.g., a perf_event program) needs to check it is with right current (e.g., by pid) before calling this helper.
Does this address your question?
Quoted 14 lines
kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 612
Quoted 12 lines
kernel/trace/bpf_trace.c +11
-0+ work =this_cpu_ptr (&send_signal_work); +if (work->irq_work.flags &IRQ_WORK_BUSY ) +return -EBUSY ; + + work->sig = sig; +irq_work_queue (&work->irq_work); +return 0 ; + } + +return group_send_sig_info (sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +Nit: extra newline slipped in
Thanks. Will remove this in the next revision.
Quoted 43 lines
Quoted 42 lines
kernel/trace/bpf_trace.c +24
-0+} + +static const struct bpf_func_proto bpf_send_signal_proto = { + .func = bpf_send_signal, + .gpl_only = false, + .ret_type =RET_INTEGER , + .arg1_type =ARG_ANYTHING , +}; +static const struct bpf_func_proto *tracing_func_proto (enum bpf_func_id func_id,const struct bpf_prog *prog) {@@ -617,6 +669,8 @@ tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)case BPF_FUNC_get_current_cgroup_id:return &bpf_get_current_cgroup_id_proto;#endif +case BPF_FUNC_send_signal: +return &bpf_send_signal_proto;default :return NULL ; }@@ -1343,5 +1397,18 @@ static int __init bpf_event_init(void) return 0; } +static int __init send_signal_irq_work_init (void ) +{ +int cpu; +struct send_signal_irq_work *work; + +for_each_possible_cpu (cpu) { + work =per_cpu_ptr (&send_signal_work, cpu); +init_irq_work (&work->irq_work, do_bpf_send_signal); + } +return 0 ; +} +fs_initcall (bpf_event_init); +subsys_initcall (send_signal_irq_work_init);#endif /* CONFIG_MODULES */
Daniel Borkmann
maintainerQuoted 120 lines · Yonghong Song
Quoted 110 lines · Daniel Borkmann
Quoted 109 lines · Yonghong Song
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function
get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.Such a mechanism can be implemented in the following way:
. a perf ring buffer is created between bpf program
and tracing app.
. once a particular event happens, bpf program writes
to the ring buffer and the tracing app gets notified.
. the tracing app sends a signal SIGALARM to the hhvm.But this method could have large delays and causing profiling results skewed.
This patch implements
bpf_send_signal()helper to send a signal to hhvm in real time, resulting in intended stack traces.Signed-off-by Yonghong Song <[email protected]>--- include/uapi/linux/bpf.h | 17 +++++++++- kernel/trace/bpf_trace.c | 67 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 83 insertions(+), 1 deletion(-)include/uapi/linux/bpf.h +16
-1@@ -2672,6 +2672,20 @@ union bpf_attr {* 0 on success. * * **-ENOENT** if the bpf-local-storage cannot be found. +* +* int bpf_send_signal(u32 sig) +* Description +* Send signal *sig* to the current task. +* Return +* 0 on success or successfully queued. +* +* **-EBUSY** if work queue under nmi is full. +* +* **-EINVAL** if *sig* is invalid. +* +* **-EPERM** if no permission to send the *sig*. +* +* **-EAGAIN** if bpf program can try again. */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \@@ -2782,7 +2796,8 @@ union bpf_attr {FN (strtol), \FN (strtoul), \FN (sk_storage_get), \ -FN (sk_storage_delete), +FN (sk_storage_delete), \ +FN (send_signal),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call kernel/trace/bpf_trace.c +28
-0@@ -567,6 +567,58 @@ static const struct bpf_func_proto bpf_probe_read_str_proto = { .arg3_type = ARG_ANYTHING, }; +struct send_signal_irq_work { +struct irq_work irq_work; +u32 sig; +}; + +static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work); + +static void do_bpf_send_signal (struct irq_work *entry) +{ +struct send_signal_irq_work *work; + + work = container_of(entry, struct send_signal_irq_work, irq_work); +group_send_sig_info (work->sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +} + +BPF_CALL_1(bpf_send_signal, u32, sig) +{ +struct send_signal_irq_work *work =NULL ; + +/* Similar to bpf_probe_write_user, task needs to be +* in a sound condition and kernel memory access be +* permitted in order to send signal to the current +* task. +*/ +if (unlikely (current->flags & (PF_KTHREAD |PF_EXITING ))) +return -EPERM ; +if (unlikely (uaccess_kernel ())) +return -EPERM ;Quoted 8 lines
kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 601
Quoted 4 lines
kernel/trace/bpf_trace.c +4
-0+if (unlikely (!nmi_uaccess_okay ())) +return -EPERM ; + +if (in_nmi ()) {Hm, bit confused, can't this only be done out of process context in general since only there current points to e.g. hhvm? I'm probably missing something. Could you elaborate?
Quoted 7 lines
That is true. If in nmi, it is out of process context and in nmi context, we use an
irq_workhere sincegroup_send_sig_info()has spinlock inside. The bpf program (e.g., aperf_eventprogram) needs to check it is with right current (e.g., by pid) before calling this helper.Does this address your question?
Hm, but how is it guaranteed that 'current' inside the callback is still the very same you intend to send the signal to?
What happens if you're in softirq and send SIGKILL to yourself? Is this ignored/handled gracefully in such case?
I think some more elaborate comment in the code would definitely be help.
Btw, you probably need to wrap it under #ifdef CONFIG_IRQ_WORK.
Quoted 57 lines
Quoted 5 lines
Quoted 5 lines
kernel/trace/bpf_trace.c +5
-0+ work =this_cpu_ptr (&send_signal_work); +if (work->irq_work.flags &IRQ_WORK_BUSY ) +return -EBUSY ; + + work->sig = sig;Quoted 8 lines
kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 612
Quoted 6 lines
kernel/trace/bpf_trace.c +6
-0+irq_work_queue (&work->irq_work); +return 0 ; + } + +return group_send_sig_info (sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +Nit: extra newline slipped in
Thanks. Will remove this in the next revision.
Quoted 43 lines
Quoted 42 lines
kernel/trace/bpf_trace.c +24
-0+} + +static const struct bpf_func_proto bpf_send_signal_proto = { + .func = bpf_send_signal, + .gpl_only = false, + .ret_type =RET_INTEGER , + .arg1_type =ARG_ANYTHING , +}; +static const struct bpf_func_proto *tracing_func_proto (enum bpf_func_id func_id,const struct bpf_prog *prog) {@@ -617,6 +669,8 @@ tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)case BPF_FUNC_get_current_cgroup_id:return &bpf_get_current_cgroup_id_proto;#endif +case BPF_FUNC_send_signal: +return &bpf_send_signal_proto;default :return NULL ; }@@ -1343,5 +1397,18 @@ static int __init bpf_event_init(void) return 0; } +static int __init send_signal_irq_work_init (void ) +{ +int cpu; +struct send_signal_irq_work *work; + +for_each_possible_cpu (cpu) { + work =per_cpu_ptr (&send_signal_work, cpu); +init_irq_work (&work->irq_work, do_bpf_send_signal); + } +return 0 ; +} +fs_initcall (bpf_event_init); +subsys_initcall (send_signal_irq_work_init);#endif /* CONFIG_MODULES */
Quoted 121 lines · Daniel Borkmann
Quoted 120 lines · Yonghong Song
Quoted 118 lines · Daniel Borkmann
Quoted 103 lines · Yonghong Song
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function
get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.Such a mechanism can be implemented in the following way:
. a perf ring buffer is created between bpf program
and tracing app.
. once a particular event happens, bpf program writes
to the ring buffer and the tracing app gets notified.
. the tracing app sends a signal SIGALARM to the hhvm.But this method could have large delays and causing profiling results skewed.
This patch implements
bpf_send_signal()helper to send a signal to hhvm in real time, resulting in intended stack traces.Signed-off-by Yonghong Song <[email protected]>--- include/uapi/linux/bpf.h | 17 +++++++++- kernel/trace/bpf_trace.c | 67 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 83 insertions(+), 1 deletion(-)include/uapi/linux/bpf.h +16
-1@@ -2672,6 +2672,20 @@ union bpf_attr {* 0 on success. * * **-ENOENT** if the bpf-local-storage cannot be found. +* +* int bpf_send_signal(u32 sig) +* Description +* Send signal *sig* to the current task. +* Return +* 0 on success or successfully queued. +* +* **-EBUSY** if work queue under nmi is full. +* +* **-EINVAL** if *sig* is invalid. +* +* **-EPERM** if no permission to send the *sig*. +* +* **-EAGAIN** if bpf program can try again. */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \@@ -2782,7 +2796,8 @@ union bpf_attr {FN (strtol), \FN (strtoul), \FN (sk_storage_get), \ -FN (sk_storage_delete), +FN (sk_storage_delete), \ +FN (send_signal),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call kernel/trace/bpf_trace.c +22
-0@@ -567,6 +567,58 @@ static const struct bpf_func_proto bpf_probe_read_str_proto = { .arg3_type = ARG_ANYTHING, }; +struct send_signal_irq_work { +struct irq_work irq_work; +u32 sig; +}; + +static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work); + +static void do_bpf_send_signal (struct irq_work *entry) +{ +struct send_signal_irq_work *work; + + work = container_of(entry, struct send_signal_irq_work, irq_work); +group_send_sig_info (work->sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +} + +BPF_CALL_1(bpf_send_signal, u32, sig) +{ +struct send_signal_irq_work *work =NULL ; + +/* Similar to bpf_probe_write_user, task needs to be +* in a sound condition and kernel memory access be +* permitted in order to send signal to the current kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 601
Quoted 10 lines
kernel/trace/bpf_trace.c +10
-0+* task. +*/ +if (unlikely (current->flags & (PF_KTHREAD |PF_EXITING ))) +return -EPERM ; +if (unlikely (uaccess_kernel ())) +return -EPERM ; +if (unlikely (!nmi_uaccess_okay ())) +return -EPERM ; + +if (in_nmi ()) {Hm, bit confused, can't this only be done out of process context in general since only there current points to e.g. hhvm? I'm probably missing something. Could you elaborate?
Quoted 7 lines
Quoted 7 lines
That is true. If in nmi, it is out of process context and in nmi context, we use an
irq_workhere sincegroup_send_sig_info()has spinlock inside. The bpf program (e.g., aperf_eventprogram) needs to check it is with right current (e.g., by pid) before calling this helper.Does this address your question?
Thanks, Daniel. The below are really good questions which I did not really think through with irq_work.
Quoted 3 lines
Hm, but how is it guaranteed that 'current' inside the callback is still the very same you intend to send the signal to?
I went through irq_work infrastructure. It looks that "current" may change. irq_work is registered as an interrupt on x86.
After nmi, some lower priority
interrupts get chances to execute including irq_work. But there are some other interrupts like timer_interrupt and reschedule_interrupt may change "current". But since we are still in interrupt context, task should not be destroyed, so the task structure pointer is still valid.
I will pass "current" task struct pointer to irq_work as well. This is similar to what we did in stackmap.c:
work->sem = ¤t->mm->mmap_sem; irq_work_queue(&work->irq_work);
At irq_work_run() time, the previous "current" in nmi should still be valid.
Quoted 3 lines
What happens if you're in softirq and send SIGKILL to yourself? Is this ignored/handled gracefully in such case?
It is not ignored. It handled gracefully in this case. I tried my example to send SIGKILL. The call stack looks below.
[ 24.211943] bpf_send_signal+0x9/0xb0 [ 24.212427] bpf_prog_fec6e7cc664d5b91_bpf_send_signal_test+0x228/0x1000 [ 24.213249] ? bpf_overflow_handler+0xb7/0x180 [ 24.213853] ? __perf_event_overflow+0x51/0xe0 [ 24.214385] ? perf_swevent_hrtimer+0x10a/0x160 [ 24.214965] ? __update_load_avg_cfs_rq+0x1a9/0x1c0 [ 24.215609] ? task_tick_fair+0x50/0x690 [ 24.216104] ? run_timer_softirq+0x208/0x490 [ 24.216637] ? timerqueue_del+0x1e/0x40 [ 24.217111] ? task_clock_event_del+0x10/0x10 [ 24.217658] ? __hrtimer_run_queues+0x10d/0x2c0 [ 24.218217] ? hrtimer_interrupt+0x122/0x270 [ 24.218765] ? rcu_irq_enter+0x31/0x110 [ 24.219223] ? smp_apic_timer_interrupt+0x67/0x160 [ 24.219842] ? apic_timer_interrupt+0xf/0x20 [ 24.220383] </IRQ> [ 24.220655] ? event_sched_out.isra.108+0x150/0x150 [ 24.221271] ? smp_call_function_single+0xdc/0x100 [ 24.221898] ? perf_event_sysfs_show+0x20/0x20 [ 24.222469] ? cpu_function_call+0x42/0x60 [ 24.222982] ? cpu_clock_event_read+0x10/0x10 [ 24.223525] ? event_function_call+0xe6/0xf0 [ 24.224053] ? event_sched_out.isra.108+0x150/0x150 [ 24.224657] ? perf_remove_from_context+0x20/0x70 [ 24.225262] ? perf_event_release_kernel+0x106/0x2e0 [ 24.225896] ? perf_release+0xc/0x10 [ 24.226347] ? __fput+0xc9/0x230 [ 24.226767] ? task_work_run+0x83/0xb0 [ 24.227243] ? do_exit+0x300/0xc50 [ 24.227674] ? syscall_trace_enter+0x1c9/0x2d0 [ 24.228223] ? do_group_exit+0x39/0xb0 [ 24.228695] ? __x64_sys_exit_group+0x14/0x20 [ 24.229270] ? do_syscall_64+0x49/0x130 [ 24.229762] ? entry_SYSCALL_64_after_hwframe+0x44/0xa9
I see the task is killed and other process is not impacted and no kernel crash/warning.
Quoted 2 lines
I think some more elaborate comment in the code would definitely be help.
Definitely will add some comments.
Quoted 2 lines
Btw, you probably need to wrap it under #ifdef
CONFIG_IRQ_WORK.
I will check this. stackmaps.c use irq_work as well and did not have CONFIG_IRQ_WORK. Maybe we are missing there as well.
Quoted 59 lines
Quoted 58 lines
Quoted 14 lines
kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 612
Quoted 12 lines
kernel/trace/bpf_trace.c +11
-0+ work =this_cpu_ptr (&send_signal_work); +if (work->irq_work.flags &IRQ_WORK_BUSY ) +return -EBUSY ; + + work->sig = sig; +irq_work_queue (&work->irq_work); +return 0 ; + } + +return group_send_sig_info (sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +Nit: extra newline slipped in
Thanks. Will remove this in the next revision.
Quoted 43 lines
Quoted 42 lines
kernel/trace/bpf_trace.c +24
-0+} + +static const struct bpf_func_proto bpf_send_signal_proto = { + .func = bpf_send_signal, + .gpl_only = false, + .ret_type =RET_INTEGER , + .arg1_type =ARG_ANYTHING , +}; +static const struct bpf_func_proto *tracing_func_proto (enum bpf_func_id func_id,const struct bpf_prog *prog) {@@ -617,6 +669,8 @@ tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)case BPF_FUNC_get_current_cgroup_id:return &bpf_get_current_cgroup_id_proto;#endif +case BPF_FUNC_send_signal: +return &bpf_send_signal_proto;default :return NULL ; }@@ -1343,5 +1397,18 @@ static int __init bpf_event_init(void) return 0; } +static int __init send_signal_irq_work_init (void ) +{ +int cpu; +struct send_signal_irq_work *work; + +for_each_possible_cpu (cpu) { + work =per_cpu_ptr (&send_signal_work, cpu); +init_irq_work (&work->irq_work, do_bpf_send_signal); + } +return 0 ; +} +fs_initcall (bpf_event_init); +subsys_initcall (send_signal_irq_work_init);#endif /* CONFIG_MODULES */
Quoted 202 lines · Yonghong Song
Quoted 121 lines · Daniel Borkmann
Quoted 120 lines · Yonghong Song
Quoted 118 lines · Daniel Borkmann
Quoted 103 lines · Yonghong Song
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function
get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.Such a mechanism can be implemented in the following way:
. a perf ring buffer is created between bpf program
and tracing app.
. once a particular event happens, bpf program writes
to the ring buffer and the tracing app gets notified.
. the tracing app sends a signal SIGALARM to the hhvm.But this method could have large delays and causing profiling results skewed.
This patch implements
bpf_send_signal()helper to send a signal to hhvm in real time, resulting in intended stack traces.Signed-off-by Yonghong Song <[email protected]>--- include/uapi/linux/bpf.h | 17 +++++++++- kernel/trace/bpf_trace.c | 67 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 83 insertions(+), 1 deletion(-)include/uapi/linux/bpf.h +16
-1@@ -2672,6 +2672,20 @@ union bpf_attr {* 0 on success. * * **-ENOENT** if the bpf-local-storage cannot be found. +* +* int bpf_send_signal(u32 sig) +* Description +* Send signal *sig* to the current task. +* Return +* 0 on success or successfully queued. +* +* **-EBUSY** if work queue under nmi is full. +* +* **-EINVAL** if *sig* is invalid. +* +* **-EPERM** if no permission to send the *sig*. +* +* **-EAGAIN** if bpf program can try again. */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \@@ -2782,7 +2796,8 @@ union bpf_attr {FN (strtol), \FN (strtoul), \FN (sk_storage_get), \ -FN (sk_storage_delete), +FN (sk_storage_delete), \ +FN (send_signal),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call kernel/trace/bpf_trace.c +22
-0@@ -567,6 +567,58 @@ static const struct bpf_func_proto bpf_probe_read_str_proto = { .arg3_type = ARG_ANYTHING, }; +struct send_signal_irq_work { +struct irq_work irq_work; +u32 sig; +}; + +static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work); + +static void do_bpf_send_signal (struct irq_work *entry) +{ +struct send_signal_irq_work *work; + + work = container_of(entry, struct send_signal_irq_work, irq_work); +group_send_sig_info (work->sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +} + +BPF_CALL_1(bpf_send_signal, u32, sig) +{ +struct send_signal_irq_work *work =NULL ; + +/* Similar to bpf_probe_write_user, task needs to be +* in a sound condition and kernel memory access be +* permitted in order to send signal to the current kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 601
Quoted 10 lines
kernel/trace/bpf_trace.c +10
-0+* task. +*/ +if (unlikely (current->flags & (PF_KTHREAD |PF_EXITING ))) +return -EPERM ; +if (unlikely (uaccess_kernel ())) +return -EPERM ; +if (unlikely (!nmi_uaccess_okay ())) +return -EPERM ; + +if (in_nmi ()) {Hm, bit confused, can't this only be done out of process context in general since only there current points to e.g. hhvm? I'm probably missing something. Could you elaborate?
Quoted 7 lines
Quoted 7 lines
That is true. If in nmi, it is out of process context and in nmi context, we use an
irq_workhere sincegroup_send_sig_info()has spinlock inside. The bpf program (e.g., aperf_eventprogram) needs to check it is with right current (e.g., by pid) before calling this helper.Does this address your question?
Thanks, Daniel. The below are really good questions which I did not really think through with
irq_work.Quoted 3 lines
Hm, but how is it guaranteed that 'current' inside the callback is still the very same you intend to send the signal to?
I went through
irq_workinfrastructure. It looks that "current" may change.irq_workis registered as an interrupt on x86.
After nmi, some lower priority
interrupts get chances to execute includingirq_work. But there are some other interrupts liketimer_interruptandreschedule_interruptmay change "current". But since we are still in interrupt context, task should not be destroyed, so the task structure pointer is still valid.I will pass "current" task struct pointer to
irq_workas well. This is similar to what we did instackmap.c:work->sem = ¤t->mm->mmap_sem; irq_work_queue(&work->irq_work);At
irq_work_run()time, the previous "current" in nmi should still be valid.Quoted 3 lines
What happens if you're in softirq and send SIGKILL to yourself? Is this ignored/handled gracefully in such case?
It is not ignored. It handled gracefully in this case. I tried my example to send SIGKILL. The call stack looks below.
[ 24.211943] bpf_send_signal+0x9/0xb0 [ 24.212427] bpf_prog_fec6e7cc664d5b91_bpf_send_signal_test+0x228/0x1000 [ 24.213249] ? bpf_overflow_handler+0xb7/0x180 [ 24.213853] ? __perf_event_overflow+0x51/0xe0 [ 24.214385] ? perf_swevent_hrtimer+0x10a/0x160 [ 24.214965] ? __update_load_avg_cfs_rq+0x1a9/0x1c0 [ 24.215609] ? task_tick_fair+0x50/0x690 [ 24.216104] ? run_timer_softirq+0x208/0x490 [ 24.216637] ? timerqueue_del+0x1e/0x40 [ 24.217111] ? task_clock_event_del+0x10/0x10 [ 24.217658] ? __hrtimer_run_queues+0x10d/0x2c0 [ 24.218217] ? hrtimer_interrupt+0x122/0x270 [ 24.218765] ? rcu_irq_enter+0x31/0x110 [ 24.219223] ? smp_apic_timer_interrupt+0x67/0x160 [ 24.219842] ? apic_timer_interrupt+0xf/0x20 [ 24.220383] </IRQ> [ 24.220655] ? event_sched_out.isra.108+0x150/0x150 [ 24.221271] ? smp_call_function_single+0xdc/0x100 [ 24.221898] ? perf_event_sysfs_show+0x20/0x20 [ 24.222469] ? cpu_function_call+0x42/0x60 [ 24.222982] ? cpu_clock_event_read+0x10/0x10 [ 24.223525] ? event_function_call+0xe6/0xf0 [ 24.224053] ? event_sched_out.isra.108+0x150/0x150 [ 24.224657] ? perf_remove_from_context+0x20/0x70 [ 24.225262] ? perf_event_release_kernel+0x106/0x2e0 [ 24.225896] ? perf_release+0xc/0x10 [ 24.226347] ? __fput+0xc9/0x230 [ 24.226767] ? task_work_run+0x83/0xb0 [ 24.227243] ? do_exit+0x300/0xc50 [ 24.227674] ? syscall_trace_enter+0x1c9/0x2d0 [ 24.228223] ? do_group_exit+0x39/0xb0 [ 24.228695] ? __x64_sys_exit_group+0x14/0x20 [ 24.229270] ? do_syscall_64+0x49/0x130 [ 24.229762] ? entry_SYSCALL_64_after_hwframe+0x44/0xa9I see the task is killed and other process is not impacted and no kernel crash/warning.
Quoted 2 lines
I think some more elaborate comment in the code would definitely be help.
Quoted 7 lines
Definitely will add some comments.
Quoted 2 lines
Btw, you probably need to wrap it under #ifdef
CONFIG_IRQ_WORK.I will check this.
stackmaps.cuseirq_workas well and did not haveCONFIG_IRQ_WORK. Maybe we are missing there as well.
Looks like we do not need CONFIG_IRQ_WORK.
We have:
obj-$(CONFIG_BPF_EVENTS) += bpf_trace.o
config BPF_EVENTS
depends on BPF_SYSCALL
depends on (KPROBE_EVENTS || UPROBE_EVENTS) && PERF_EVENTS
config PERF_EVENTS
bool "Kernel performance events and counters" default y if PROFILING
depends on HAVE_PERF_EVENTS
select IRQ_WORK
Quoted 60 lines
Quoted 60 lines
Quoted 59 lines
Quoted 15 lines
kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 612
Quoted 13 lines
kernel/trace/bpf_trace.c +11
-0+ work =this_cpu_ptr (&send_signal_work); +if (work->irq_work.flags &IRQ_WORK_BUSY ) +return -EBUSY ; + + work->sig = sig; +irq_work_queue (&work->irq_work); +return 0 ; + } + +return group_send_sig_info (sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +Nit: extra newline slipped in
Thanks. Will remove this in the next revision.
Quoted 43 lines
Quoted 42 lines
kernel/trace/bpf_trace.c +24
-0+} + +static const struct bpf_func_proto bpf_send_signal_proto = { + .func = bpf_send_signal, + .gpl_only = false, + .ret_type =RET_INTEGER , + .arg1_type =ARG_ANYTHING , +}; +static const struct bpf_func_proto *tracing_func_proto (enum bpf_func_id func_id,const struct bpf_prog *prog) {@@ -617,6 +669,8 @@ tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)case BPF_FUNC_get_current_cgroup_id:return &bpf_get_current_cgroup_id_proto;#endif +case BPF_FUNC_send_signal: +return &bpf_send_signal_proto;default :return NULL ; }@@ -1343,5 +1397,18 @@ static int __init bpf_event_init(void) return 0; } +static int __init send_signal_irq_work_init (void ) +{ +int cpu; +struct send_signal_irq_work *work; + +for_each_possible_cpu (cpu) { + work =per_cpu_ptr (&send_signal_work, cpu); +init_irq_work (&work->irq_work, do_bpf_send_signal); + } +return 0 ; +} +fs_initcall (bpf_event_init); +subsys_initcall (send_signal_irq_work_init);#endif /* CONFIG_MODULES */
Daniel Borkmann
maintainerQuoted 193 lines · Yonghong Song
Quoted 192 lines · Yonghong Song
Quoted 121 lines · Daniel Borkmann
Quoted 120 lines · Yonghong Song
Quoted 118 lines · Daniel Borkmann
Quoted 103 lines · Yonghong Song
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function
get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.Such a mechanism can be implemented in the following way:
. a perf ring buffer is created between bpf program
and tracing app.
. once a particular event happens, bpf program writes
to the ring buffer and the tracing app gets notified.
. the tracing app sends a signal SIGALARM to the hhvm.But this method could have large delays and causing profiling results skewed.
This patch implements
bpf_send_signal()helper to send a signal to hhvm in real time, resulting in intended stack traces.Signed-off-by Yonghong Song <[email protected]>--- include/uapi/linux/bpf.h | 17 +++++++++- kernel/trace/bpf_trace.c | 67 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 83 insertions(+), 1 deletion(-)include/uapi/linux/bpf.h +16
-1@@ -2672,6 +2672,20 @@ union bpf_attr {* 0 on success. * * **-ENOENT** if the bpf-local-storage cannot be found. +* +* int bpf_send_signal(u32 sig) +* Description +* Send signal *sig* to the current task. +* Return +* 0 on success or successfully queued. +* +* **-EBUSY** if work queue under nmi is full. +* +* **-EINVAL** if *sig* is invalid. +* +* **-EPERM** if no permission to send the *sig*. +* +* **-EAGAIN** if bpf program can try again. */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \@@ -2782,7 +2796,8 @@ union bpf_attr {FN (strtol), \FN (strtoul), \FN (sk_storage_get), \ -FN (sk_storage_delete), +FN (sk_storage_delete), \ +FN (send_signal),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call kernel/trace/bpf_trace.c +22
-0@@ -567,6 +567,58 @@ static const struct bpf_func_proto bpf_probe_read_str_proto = { .arg3_type = ARG_ANYTHING, }; +struct send_signal_irq_work { +struct irq_work irq_work; +u32 sig; +}; + +static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work); + +static void do_bpf_send_signal (struct irq_work *entry) +{ +struct send_signal_irq_work *work; + + work = container_of(entry, struct send_signal_irq_work, irq_work); +group_send_sig_info (work->sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +} + +BPF_CALL_1(bpf_send_signal, u32, sig) +{ +struct send_signal_irq_work *work =NULL ; + +/* Similar to bpf_probe_write_user, task needs to be +* in a sound condition and kernel memory access be +* permitted in order to send signal to the current kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 601
Quoted 10 lines
kernel/trace/bpf_trace.c +10
-0+* task. +*/ +if (unlikely (current->flags & (PF_KTHREAD |PF_EXITING ))) +return -EPERM ; +if (unlikely (uaccess_kernel ())) +return -EPERM ; +if (unlikely (!nmi_uaccess_okay ())) +return -EPERM ; + +if (in_nmi ()) {Hm, bit confused, can't this only be done out of process context in general since only there current points to e.g. hhvm? I'm probably missing something. Could you elaborate?
Quoted 7 lines
Quoted 7 lines
That is true. If in nmi, it is out of process context and in nmi context, we use an
irq_workhere sincegroup_send_sig_info()has spinlock inside. The bpf program (e.g., aperf_eventprogram) needs to check it is with right current (e.g., by pid) before calling this helper.Does this address your question?
Thanks, Daniel. The below are really good questions which I did not really think through with
irq_work.Quoted 2 lines
Hm, but how is it guaranteed that 'current' inside the callback is still the very same you intend to send the signal to?
I went through
irq_workinfrastructure. It looks that "current" may change.irq_workis registered as an interrupt on x86.
After nmi, some lower priority
interrupts get chances to execute includingirq_work. But there are some other interrupts liketimer_interruptandreschedule_interruptmay change "current". But since we are still in interrupt context, task should not be destroyed, so the task structure pointer is still valid.I will pass "current" task struct pointer to
irq_workas well. This is similar to what we did instackmap.c:work->sem = ¤t->mm->mmap_sem; irq_work_queue(&work->irq_work);At
irq_work_run()time, the previous "current" in nmi should still be valid.Quoted 2 lines
What happens if you're in softirq and send SIGKILL to yourself? Is this ignored/handled gracefully in such case?
It is not ignored. It handled gracefully in this case. I tried my example to send SIGKILL. The call stack looks below.
[ 24.211943] bpf_send_signal+0x9/0xb0 [ 24.212427] bpf_prog_fec6e7cc664d5b91_bpf_send_signal_test+0x228/0x1000 [ 24.213249] ? bpf_overflow_handler+0xb7/0x180 [ 24.213853] ? __perf_event_overflow+0x51/0xe0 [ 24.214385] ? perf_swevent_hrtimer+0x10a/0x160 [ 24.214965] ? __update_load_avg_cfs_rq+0x1a9/0x1c0 [ 24.215609] ? task_tick_fair+0x50/0x690 [ 24.216104] ? run_timer_softirq+0x208/0x490 [ 24.216637] ? timerqueue_del+0x1e/0x40 [ 24.217111] ? task_clock_event_del+0x10/0x10 [ 24.217658] ? __hrtimer_run_queues+0x10d/0x2c0 [ 24.218217] ? hrtimer_interrupt+0x122/0x270 [ 24.218765] ? rcu_irq_enter+0x31/0x110 [ 24.219223] ? smp_apic_timer_interrupt+0x67/0x160 [ 24.219842] ? apic_timer_interrupt+0xf/0x20 [ 24.220383] </IRQ> [ 24.220655] ? event_sched_out.isra.108+0x150/0x150 [ 24.221271] ? smp_call_function_single+0xdc/0x100 [ 24.221898] ? perf_event_sysfs_show+0x20/0x20 [ 24.222469] ? cpu_function_call+0x42/0x60 [ 24.222982] ? cpu_clock_event_read+0x10/0x10 [ 24.223525] ? event_function_call+0xe6/0xf0 [ 24.224053] ? event_sched_out.isra.108+0x150/0x150 [ 24.224657] ? perf_remove_from_context+0x20/0x70 [ 24.225262] ? perf_event_release_kernel+0x106/0x2e0 [ 24.225896] ? perf_release+0xc/0x10 [ 24.226347] ? __fput+0xc9/0x230 [ 24.226767] ? task_work_run+0x83/0xb0 [ 24.227243] ? do_exit+0x300/0xc50 [ 24.227674] ? syscall_trace_enter+0x1c9/0x2d0 [ 24.228223] ? do_group_exit+0x39/0xb0 [ 24.228695] ? __x64_sys_exit_group+0x14/0x20 [ 24.229270] ? do_syscall_64+0x49/0x130 [ 24.229762] ? entry_SYSCALL_64_after_hwframe+0x44/0xa9
Quoted 2 lines
Quoted 2 lines
I see the task is killed and other process is not impacted and no kernel crash/warning.
Hm, but I rather meant when you have the case that we're in_serving_softirq() e.g. when processing packets on rx and you send a signal to yourself. Shouldn't we bail out from the helper in such situation as well?
Thanks,
Daniel
Quoted 194 lines · Daniel Borkmann
Quoted 193 lines · Yonghong Song
Quoted 192 lines · Yonghong Song
Quoted 121 lines · Daniel Borkmann
Quoted 120 lines · Yonghong Song
Quoted 118 lines · Daniel Borkmann
Quoted 103 lines · Yonghong Song
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function
get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.Such a mechanism can be implemented in the following way:
. a perf ring buffer is created between bpf program
and tracing app.
. once a particular event happens, bpf program writes
to the ring buffer and the tracing app gets notified.
. the tracing app sends a signal SIGALARM to the hhvm.But this method could have large delays and causing profiling results skewed.
This patch implements
bpf_send_signal()helper to send a signal to hhvm in real time, resulting in intended stack traces.Signed-off-by Yonghong Song <[email protected]>--- include/uapi/linux/bpf.h | 17 +++++++++- kernel/trace/bpf_trace.c | 67 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 83 insertions(+), 1 deletion(-)include/uapi/linux/bpf.h +16
-1@@ -2672,6 +2672,20 @@ union bpf_attr {* 0 on success. * * **-ENOENT** if the bpf-local-storage cannot be found. +* +* int bpf_send_signal(u32 sig) +* Description +* Send signal *sig* to the current task. +* Return +* 0 on success or successfully queued. +* +* **-EBUSY** if work queue under nmi is full. +* +* **-EINVAL** if *sig* is invalid. +* +* **-EPERM** if no permission to send the *sig*. +* +* **-EAGAIN** if bpf program can try again. */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \@@ -2782,7 +2796,8 @@ union bpf_attr {FN (strtol), \FN (strtoul), \FN (sk_storage_get), \ -FN (sk_storage_delete), +FN (sk_storage_delete), \ +FN (send_signal),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call kernel/trace/bpf_trace.c +22
-0@@ -567,6 +567,58 @@ static const struct bpf_func_proto bpf_probe_read_str_proto = { .arg3_type = ARG_ANYTHING, }; +struct send_signal_irq_work { +struct irq_work irq_work; +u32 sig; +}; + +static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work); + +static void do_bpf_send_signal (struct irq_work *entry) +{ +struct send_signal_irq_work *work; + + work = container_of(entry, struct send_signal_irq_work, irq_work); +group_send_sig_info (work->sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +} + +BPF_CALL_1(bpf_send_signal, u32, sig) +{ +struct send_signal_irq_work *work =NULL ; + +/* Similar to bpf_probe_write_user, task needs to be +* in a sound condition and kernel memory access be +* permitted in order to send signal to the current kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 601
Quoted 10 lines
kernel/trace/bpf_trace.c +10
-0+* task. +*/ +if (unlikely (current->flags & (PF_KTHREAD |PF_EXITING ))) +return -EPERM ; +if (unlikely (uaccess_kernel ())) +return -EPERM ; +if (unlikely (!nmi_uaccess_okay ())) +return -EPERM ; + +if (in_nmi ()) {Hm, bit confused, can't this only be done out of process context in general since only there current points to e.g. hhvm? I'm probably missing something. Could you elaborate?
Quoted 7 lines
Quoted 7 lines
That is true. If in nmi, it is out of process context and in nmi context, we use an
irq_workhere sincegroup_send_sig_info()has spinlock inside. The bpf program (e.g., aperf_eventprogram) needs to check it is with right current (e.g., by pid) before calling this helper.Does this address your question?
Thanks, Daniel. The below are really good questions which I did not really think through with
irq_work.Quoted 2 lines
Hm, but how is it guaranteed that 'current' inside the callback is still the very same you intend to send the signal to?
I went through
irq_workinfrastructure. It looks that "current" may change.irq_workis registered as an interrupt on x86.
After nmi, some lower priority
interrupts get chances to execute includingirq_work. But there are some other interrupts liketimer_interruptandreschedule_interruptmay change "current". But since we are still in interrupt context, task should not be destroyed, so the task structure pointer is still valid.I will pass "current" task struct pointer to
irq_workas well. This is similar to what we did instackmap.c:work->sem = ¤t->mm->mmap_sem; irq_work_queue(&work->irq_work);At
irq_work_run()time, the previous "current" in nmi should still be valid.Quoted 2 lines
What happens if you're in softirq and send SIGKILL to yourself? Is this ignored/handled gracefully in such case?
It is not ignored. It handled gracefully in this case. I tried my example to send SIGKILL. The call stack looks below.
[ 24.211943] bpf_send_signal+0x9/0xb0 [ 24.212427] bpf_prog_fec6e7cc664d5b91_bpf_send_signal_test+0x228/0x1000 [ 24.213249] ? bpf_overflow_handler+0xb7/0x180 [ 24.213853] ? __perf_event_overflow+0x51/0xe0 [ 24.214385] ? perf_swevent_hrtimer+0x10a/0x160 [ 24.214965] ? __update_load_avg_cfs_rq+0x1a9/0x1c0 [ 24.215609] ? task_tick_fair+0x50/0x690 [ 24.216104] ? run_timer_softirq+0x208/0x490 [ 24.216637] ? timerqueue_del+0x1e/0x40 [ 24.217111] ? task_clock_event_del+0x10/0x10 [ 24.217658] ? __hrtimer_run_queues+0x10d/0x2c0 [ 24.218217] ? hrtimer_interrupt+0x122/0x270 [ 24.218765] ? rcu_irq_enter+0x31/0x110 [ 24.219223] ? smp_apic_timer_interrupt+0x67/0x160 [ 24.219842] ? apic_timer_interrupt+0xf/0x20 [ 24.220383] </IRQ> [ 24.220655] ? event_sched_out.isra.108+0x150/0x150 [ 24.221271] ? smp_call_function_single+0xdc/0x100 [ 24.221898] ? perf_event_sysfs_show+0x20/0x20 [ 24.222469] ? cpu_function_call+0x42/0x60 [ 24.222982] ? cpu_clock_event_read+0x10/0x10 [ 24.223525] ? event_function_call+0xe6/0xf0 [ 24.224053] ? event_sched_out.isra.108+0x150/0x150 [ 24.224657] ? perf_remove_from_context+0x20/0x70 [ 24.225262] ? perf_event_release_kernel+0x106/0x2e0 [ 24.225896] ? perf_release+0xc/0x10 [ 24.226347] ? __fput+0xc9/0x230 [ 24.226767] ? task_work_run+0x83/0xb0 [ 24.227243] ? do_exit+0x300/0xc50 [ 24.227674] ? syscall_trace_enter+0x1c9/0x2d0 [ 24.228223] ? do_group_exit+0x39/0xb0 [ 24.228695] ? __x64_sys_exit_group+0x14/0x20 [ 24.229270] ? do_syscall_64+0x49/0x130 [ 24.229762] ? entry_SYSCALL_64_after_hwframe+0x44/0xa9
Quoted 6 lines
Quoted 2 lines
Quoted 2 lines
I see the task is killed and other process is not impacted and no kernel crash/warning.
Hm, but I rather meant when you have the case that we're
in_serving_softirq()e.g. when processing packets on rx and you send a signal to yourself. Shouldn't we bail out from the helper in such situation as well?
Just want to clarify. Are you concerned with safety or correctness?
For safety, if we do send signal here, we may wreck the system?
For correctness, you mean the information we got to send a signal to process is not quite right if in_serving_softirq()? F.e, the performance counter overflow may be caused by softirq rather the process itself? So in this case, we should only send signal to process when in process context, and in nmi (not serving softirq)?
If for correctness, do you think we should add a "flags" parameter to the bpf_send_signal() helper such that:
. default not checking is_serving_softirq()
. bit0: if set, bail out if is_serving_softirq()
. other bits: reserved
Quoted 4 lines
Thanks,
Daniel
Daniel Borkmann
maintainerQuoted 214 lines · Yonghong Song
Quoted 194 lines · Daniel Borkmann
Quoted 193 lines · Yonghong Song
Quoted 192 lines · Yonghong Song
Quoted 121 lines · Daniel Borkmann
Quoted 120 lines · Yonghong Song
Quoted 118 lines · Daniel Borkmann
Quoted 103 lines · Yonghong Song
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces through kernel function
get_perf_callchain()
when certain events happens (e.g., cache miss counter or cpu clock counter overflows). But such stack traces are not enough for jitted programs, e.g., hhvm (jited php). To get real stack trace, jit engine internal data structures need to be traversed in order to get the real user functions.bpf program itself may not be the best place to traverse the jit engine as the traversing logic could be complex and it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which it can dump stack traces. When it receives a signal, it will dump the stack in next such program location.Such a mechanism can be implemented in the following way:
. a perf ring buffer is created between bpf program
and tracing app.
. once a particular event happens, bpf program writes
to the ring buffer and the tracing app gets notified.
. the tracing app sends a signal SIGALARM to the hhvm.But this method could have large delays and causing profiling results skewed.
This patch implements
bpf_send_signal()helper to send a signal to hhvm in real time, resulting in intended stack traces.Signed-off-by Yonghong Song <[email protected]>--- include/uapi/linux/bpf.h | 17 +++++++++- kernel/trace/bpf_trace.c | 67 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 83 insertions(+), 1 deletion(-)include/uapi/linux/bpf.h +16
-1@@ -2672,6 +2672,20 @@ union bpf_attr {* 0 on success. * * **-ENOENT** if the bpf-local-storage cannot be found. +* +* int bpf_send_signal(u32 sig) +* Description +* Send signal *sig* to the current task. +* Return +* 0 on success or successfully queued. +* +* **-EBUSY** if work queue under nmi is full. +* +* **-EINVAL** if *sig* is invalid. +* +* **-EPERM** if no permission to send the *sig*. +* +* **-EAGAIN** if bpf program can try again. */ #define __BPF_FUNC_MAPPER (FN ) \ FN(unspec), \@@ -2782,7 +2796,8 @@ union bpf_attr {FN (strtol), \FN (strtoul), \FN (sk_storage_get), \ -FN (sk_storage_delete), +FN (sk_storage_delete), \ +FN (send_signal),/* integer value in 'imm' field of BPF_CALL instruction selects which helper * function eBPF program intends to call kernel/trace/bpf_trace.c +22
-0@@ -567,6 +567,58 @@ static const struct bpf_func_proto bpf_probe_read_str_proto = { .arg3_type = ARG_ANYTHING, }; +struct send_signal_irq_work { +struct irq_work irq_work; +u32 sig; +}; + +static DEFINE_PER_CPU(struct send_signal_irq_work, send_signal_work); + +static void do_bpf_send_signal (struct irq_work *entry) +{ +struct send_signal_irq_work *work; + + work = container_of(entry, struct send_signal_irq_work, irq_work); +group_send_sig_info (work->sig,SEND_SIG_PRIV , current,PIDTYPE_TGID ); +} + +BPF_CALL_1(bpf_send_signal, u32, sig) +{ +struct send_signal_irq_work *work =NULL ; + +/* Similar to bpf_probe_write_user, task needs to be +* in a sound condition and kernel memory access be +* permitted in order to send signal to the current kernel/trace/bpf_trace.c static const struct bpf_func_proto bpf_probe_read_str_proto = { · line 601
Quoted 10 lines
kernel/trace/bpf_trace.c +10
-0+* task. +*/ +if (unlikely (current->flags & (PF_KTHREAD |PF_EXITING ))) +return -EPERM ; +if (unlikely (uaccess_kernel ())) +return -EPERM ; +if (unlikely (!nmi_uaccess_okay ())) +return -EPERM ; + +if (in_nmi ()) {Hm, bit confused, can't this only be done out of process context in general since only there current points to e.g. hhvm? I'm probably missing something. Could you elaborate?
Quoted 7 lines
Quoted 7 lines
That is true. If in nmi, it is out of process context and in nmi context, we use an
irq_workhere sincegroup_send_sig_info()has spinlock inside. The bpf program (e.g., aperf_eventprogram) needs to check it is with right current (e.g., by pid) before calling this helper.Does this address your question?
Thanks, Daniel. The below are really good questions which I did not really think through with
irq_work.Quoted 2 lines
Hm, but how is it guaranteed that 'current' inside the callback is still the very same you intend to send the signal to?
I went through
irq_workinfrastructure. It looks that "current" may change.irq_workis registered as an interrupt on x86.
After nmi, some lower priority
interrupts get chances to execute includingirq_work. But there are some other interrupts liketimer_interruptandreschedule_interruptmay change "current". But since we are still in interrupt context, task should not be destroyed, so the task structure pointer is still valid.I will pass "current" task struct pointer to
irq_workas well. This is similar to what we did instackmap.c:work->sem = ¤t->mm->mmap_sem; irq_work_queue(&work->irq_work);At
irq_work_run()time, the previous "current" in nmi should still be valid.Quoted 2 lines
What happens if you're in softirq and send SIGKILL to yourself? Is this ignored/handled gracefully in such case?
It is not ignored. It handled gracefully in this case. I tried my example to send SIGKILL. The call stack looks below.
[ 24.211943] bpf_send_signal+0x9/0xb0 [ 24.212427] bpf_prog_fec6e7cc664d5b91_bpf_send_signal_test+0x228/0x1000 [ 24.213249] ? bpf_overflow_handler+0xb7/0x180 [ 24.213853] ? __perf_event_overflow+0x51/0xe0 [ 24.214385] ? perf_swevent_hrtimer+0x10a/0x160 [ 24.214965] ? __update_load_avg_cfs_rq+0x1a9/0x1c0 [ 24.215609] ? task_tick_fair+0x50/0x690 [ 24.216104] ? run_timer_softirq+0x208/0x490 [ 24.216637] ? timerqueue_del+0x1e/0x40 [ 24.217111] ? task_clock_event_del+0x10/0x10 [ 24.217658] ? __hrtimer_run_queues+0x10d/0x2c0 [ 24.218217] ? hrtimer_interrupt+0x122/0x270 [ 24.218765] ? rcu_irq_enter+0x31/0x110 [ 24.219223] ? smp_apic_timer_interrupt+0x67/0x160 [ 24.219842] ? apic_timer_interrupt+0xf/0x20 [ 24.220383] </IRQ> [ 24.220655] ? event_sched_out.isra.108+0x150/0x150 [ 24.221271] ? smp_call_function_single+0xdc/0x100 [ 24.221898] ? perf_event_sysfs_show+0x20/0x20 [ 24.222469] ? cpu_function_call+0x42/0x60 [ 24.222982] ? cpu_clock_event_read+0x10/0x10 [ 24.223525] ? event_function_call+0xe6/0xf0 [ 24.224053] ? event_sched_out.isra.108+0x150/0x150 [ 24.224657] ? perf_remove_from_context+0x20/0x70 [ 24.225262] ? perf_event_release_kernel+0x106/0x2e0 [ 24.225896] ? perf_release+0xc/0x10 [ 24.226347] ? __fput+0xc9/0x230 [ 24.226767] ? task_work_run+0x83/0xb0 [ 24.227243] ? do_exit+0x300/0xc50 [ 24.227674] ? syscall_trace_enter+0x1c9/0x2d0 [ 24.228223] ? do_group_exit+0x39/0xb0 [ 24.228695] ? __x64_sys_exit_group+0x14/0x20 [ 24.229270] ? do_syscall_64+0x49/0x130 [ 24.229762] ? entry_SYSCALL_64_after_hwframe+0x44/0xa9Quoted 6 lines
Quoted 2 lines
Quoted 2 lines
I see the task is killed and other process is not impacted and no kernel crash/warning.
Hm, but I rather meant when you have the case that we're
in_serving_softirq()e.g. when processing packets on rx and you send a signal to yourself. Shouldn't we bail out from the helper in such situation as well?Just want to clarify. Are you concerned with safety or correctness?
For safety, if we do send signal here, we may wreck the system?
For correctness, you mean the information we got to send a signal to process is not quite right if
in_serving_softirq()? F.e, the performance counter overflow may be caused by softirq rather the process itself? So in this case, we should only send signal to process when in process context, and in nmi (not serving softirq)?
Quoted 5 lines
If for correctness, do you think we should add a "flags" parameter to the
bpf_send_signal()helper such that:
. default not checkingis_serving_softirq()
. bit0: if set, bail out ifis_serving_softirq()
. other bits: reserved
Scratch my thought, we do bail out in case of PF_KTHREAD, so should be okay. Was thinking in terms of both, not wrecking the system / messing with kthreads and with regards to correctness.