Skip to contentSkip to conversation

sys_enter tracepoint ctx structure

DISCUSSION Yauheni Kaliuta 2 participants 10 messages Apr 20, 2023 – Apr 23, 2023 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Hi!

Should perf_call_bpf_enter/exit (kernel/trace/trace_syscalls.c) use struct trace_event_raw_sys_enter/exit instead of locally crafted struct syscall_tp_t nowadays? test_progs's vmlinux test expects it as the context.

Or at least use struct trace_entry instead of struct pt_regs?

I have a problem with one RT patch with extends trace_entry.

Signature · 2 lines
WBR,
Yauheni Kaliuta

A message that is not in the archive (sent off-list or not yet received): CAADnVQ+JdPGV95Y30PskgdOomU2K0UXsoCydgqaJfJ5j4S8BtQ@mail.gmail.com

#2 lore raw
Quoted 6 lines · Yauheni Kaliuta

Hi!

Should perf_call_bpf_enter/exit (kernel/trace/trace_syscalls.c) use struct trace_event_raw_sys_enter/exit instead of locally crafted struct syscall_tp_t nowadays?

No. It needs syscall_tp_t.

Quoted 2 lines

test_progs's vmlinux test
expects it as the context.

what do you mean? Pls share a code pointer?

Quoted 3 lines

Or at least use struct trace_entry instead of struct pt_regs?

no. It needs a pointer to pt_regs.
See all of the pe_* flavor of helpers.

Quoted 3 lines

I have a problem with one RT patch with extends trace_entry.

Just extend it. It shouldn't matter.
I'm likely missing something.

#3 lore raw
In reply to a message not in the archive

Hi, Alexei!

Quoted 10 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 5 lines · Yauheni Kaliuta

Hi!

Should perf_call_bpf_enter/exit (kernel/trace/trace_syscalls.c) use struct trace_event_raw_sys_enter/exit instead of locally crafted struct syscall_tp_t nowadays?

Quoted 7 lines

No. It needs syscall_tp_t.

test_progs's vmlinux test

Quoted 2 lines

expects it as the context.

what do you mean? Pls share a code pointer?

https://github.com/torvalds/linux/blob/master/tools/testing/selftests/bpf/progs/test_vmlinux.c#L19

SEC("tp/syscalls/sys_enter_nanosleep")
int handle__tp(struct trace_event_raw_sys_enter *args)

So, should it use different structure then? syscall_tp_t to be declared publicly?

Quoted 9 lines
Quoted 3 lines

Or at least use struct trace_entry instead of struct pt_regs?

no. It needs a pointer to pt_regs.
See all of the pe_* flavor of helpers.

Quoted 1 line
Quoted 5 lines
Quoted 2 lines

I have a problem with one RT patch with extends trace_entry.

Just extend it. It shouldn't matter.
I'm likely missing something.

Or me. Let's figure out :)

Signature · 2 lines
WBR,
Yauheni Kaliuta
#4 lore raw
Quoted 19 lines · Yauheni Kaliuta

Hi, Alexei!

Quoted 15 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 5 lines · Yauheni Kaliuta

Hi!

Should perf_call_bpf_enter/exit (kernel/trace/trace_syscalls.c) use struct trace_event_raw_sys_enter/exit instead of locally crafted struct syscall_tp_t nowadays?

No. It needs syscall_tp_t.

test_progs's vmlinux test

Quoted 2 lines

expects it as the context.

Quoted 6 lines
Quoted 1 line

what do you mean? Pls share a code pointer?

https://github.com/torvalds/linux/blob/master/tools/testing/selftests/bpf/progs/test_vmlinux.c#L19

SEC("tp/syscalls/sys_enter_nanosleep")
int handle__tp(struct trace_event_raw_sys_enter *args)

I see. That bit is correct and that's what bpftrace is doing when attaching to syscalls.
What do you see in your patched RT kernel when you do:
cat /sys/kernel/debug/tracing/events/syscalls/sys_enter_nanosleep/format ?
Depending on the answer we might need to fix
the kernel side that has to use struct trace_entry
in syscall_tp_t instead of plain long long.

#5 lore raw

Hi, Alexei!

Quoted 22 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 13 lines · Yauheni Kaliuta
Quoted 12 lines
Quoted 5 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov

Should perf_call_bpf_enter/exit (kernel/trace/trace_syscalls.c) use struct trace_event_raw_sys_enter/exit instead of locally crafted struct syscall_tp_t nowadays?

No. It needs syscall_tp_t.

test_progs's vmlinux test

Quoted 2 lines

expects it as the context.

Quoted 6 lines
Quoted 1 line

what do you mean? Pls share a code pointer?

https://github.com/torvalds/linux/blob/master/tools/testing/selftests/bpf/progs/test_vmlinux.c#L19

SEC("tp/syscalls/sys_enter_nanosleep")
int handle__tp(struct trace_event_raw_sys_enter *args)

Quoted 8 lines

I see. That bit is correct and that's what bpftrace is doing when attaching to syscalls.
What do you see in your patched RT kernel when you do:
cat /sys/kernel/debug/tracing/events/syscalls/sys_enter_nanosleep/format ?
Depending on the answer we might need to fix
the kernel side that has to use struct trace_entry
in syscall_tp_t instead of plain long long.

 # cat /sys/kernel/debug/tracing/events/syscalls/sys_enter_nanosleep/format
name: sys_enter_nanosleep
ID: 374
format:
        field:unsigned short common_type;       offset:0;       size:2; signed:0;
        field:unsigned char common_flags;       offset:2;       size:1; signed:0;
        field:unsigned char common_preempt_count;       offset:3;       size:1; signed:0;
        field:int common_pid;   offset:4;       size:4; signed:1;
        field:unsigned char common_preempt_lazy_count;  offset:8;       size:1; signed:0;
        field:int __syscall_nr; offset:12;      size:4; signed:1;
        field:struct __kernel_timespec * rqtp;  offset:16;      size:8; signed:0;
        field:struct __kernel_timespec * rmtp;  offset:24;      size:8; signed:0;

print fmt: "rqtp: 0x%08lx, rmtp: 0x%08lx", ((unsigned long)(REC->rqtp)), ((unsigned long)(REC->rmtp))

Signature · 2 lines
WBR,
Yauheni Kaliuta
#6 lore raw
Quoted 44 lines · Yauheni Kaliuta

Hi, Alexei!

Quoted 22 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 13 lines · Yauheni Kaliuta
Quoted 12 lines
Quoted 5 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov

Should perf_call_bpf_enter/exit (kernel/trace/trace_syscalls.c) use struct trace_event_raw_sys_enter/exit instead of locally crafted struct syscall_tp_t nowadays?

No. It needs syscall_tp_t.

test_progs's vmlinux test

Quoted 2 lines

expects it as the context.

Quoted 6 lines
Quoted 1 line

what do you mean? Pls share a code pointer?

https://github.com/torvalds/linux/blob/master/tools/testing/selftests/bpf/progs/test_vmlinux.c#L19

SEC("tp/syscalls/sys_enter_nanosleep")
int handle__tp(struct trace_event_raw_sys_enter *args)

Quoted 8 lines

I see. That bit is correct and that's what bpftrace is doing when attaching to syscalls.
What do you see in your patched RT kernel when you do:
cat /sys/kernel/debug/tracing/events/syscalls/sys_enter_nanosleep/format ?
Depending on the answer we might need to fix
the kernel side that has to use struct trace_entry
in syscall_tp_t instead of plain long long.

 # cat /sys/kernel/debug/tracing/events/syscalls/sys_enter_nanosleep/format
name: sys_enter_nanosleep
ID: 374
format:
        field:unsigned short common_type;       offset:0;       size:2; signed:0;
        field:unsigned char common_flags;       offset:2;       size:1; signed:0;
        field:unsigned char common_preempt_count;       offset:3;       size:1; signed:0;
        field:int common_pid;   offset:4;       size:4; signed:1;
        field:unsigned char common_preempt_lazy_count;  offset:8;       size:1; signed:0;
Quoted 5 lines
        field:int __syscall_nr; offset:12;      size:4; signed:1;
        field:struct __kernel_timespec * rqtp;  offset:16;      size:8; signed:0;
        field:struct __kernel_timespec * rmtp;  offset:24;      size:8; signed:0;

print fmt: "rqtp: 0x%08lx, rmtp: 0x%08lx", ((unsigned long)(REC->rqtp)), ((unsigned long)(REC->rmtp))

Lol.
Jiri even fixed the issue with this format in bpftrace 3 years ago:
https://github.com/iovisor/bpftrace/commit/a2e3d5dbc03ceb49b776cf5602d31896158844a7

Let's fix the kernel side too. Something like this should do it:

kernel/trace/trace_syscalls.c +0 -0
index 942ddbdace4a..7aa1f4299486 100644
@@ -555,7 +555,7 @@ static int perf_call_bpf_enter(struct

trace_event_call *call, struct pt_regs *re
struct syscall_trace_enter *rec)

 {
        struct syscall_tp_t {
-               unsigned long long regs;
+               struct trace_entry ent;
                unsigned long syscall_nr;
                unsigned long args[SYSCALL_DEFINE_MAXARGS];
        } param;
@@ -657,7 +657,7 @@ static int perf_call_bpf_exit(struct
trace_event_call *call, struct pt_regs *reg
                              struct syscall_trace_exit *rec)
 {
        struct syscall_tp_t {

- unsigned long long regs;
+ struct trace_entry ent;

pls add build_bug_on that sizeof(ent) >= sizeof(void*).

#7 lore raw

Hi, Alexei!

Quoted 47 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 45 lines · Yauheni Kaliuta
Quoted 22 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 13 lines · Yauheni Kaliuta
Quoted 12 lines
Quoted 5 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov

Should perf_call_bpf_enter/exit (kernel/trace/trace_syscalls.c) use struct trace_event_raw_sys_enter/exit instead of locally crafted struct syscall_tp_t nowadays?

No. It needs syscall_tp_t.

test_progs's vmlinux test

Quoted 2 lines

expects it as the context.

Quoted 6 lines
Quoted 1 line

what do you mean? Pls share a code pointer?

https://github.com/torvalds/linux/blob/master/tools/testing/selftests/bpf/progs/test_vmlinux.c#L19

SEC("tp/syscalls/sys_enter_nanosleep")
int handle__tp(struct trace_event_raw_sys_enter *args)

Quoted 8 lines

I see. That bit is correct and that's what bpftrace is doing when attaching to syscalls.
What do you see in your patched RT kernel when you do:
cat /sys/kernel/debug/tracing/events/syscalls/sys_enter_nanosleep/format ?
Depending on the answer we might need to fix
the kernel side that has to use struct trace_entry
in syscall_tp_t instead of plain long long.

# cat /sys/kernel/debug/tracing/events/syscalls/sys_enter_nanosleep/format
name: sys_enter_nanosleep
ID: 374
format:
field:unsigned short common_type;       offset:0;       size:2; signed:0;
field:unsigned char common_flags;       offset:2;       size:1; signed:0;
field:unsigned char common_preempt_count;       offset:3;       size:1; signed:0;
field:int common_pid;   offset:4;       size:4; signed:1;
field:unsigned char common_preempt_lazy_count;  offset:8;       size:1; signed:0;
field:int __syscall_nr; offset:12;      size:4; signed:1;
field:struct __kernel_timespec * rqtp;  offset:16;      size:8; signed:0;
field:struct __kernel_timespec * rmtp;  offset:24;      size:8; signed:0;
Quoted 6 lines
Quoted 1 line

print fmt: "rqtp: 0x%08lx, rmtp: 0x%08lx", ((unsigned long)(REC->rqtp)), ((unsigned long)(REC->rmtp))

Lol.
Jiri even fixed the issue with this format in bpftrace 3 years ago:
https://github.com/iovisor/bpftrace/commit/a2e3d5dbc03ceb49b776cf5602d31896158844a7

Hehe :)

Quoted 25 lines

Let's fix the kernel side too. Something like this should do it:

kernel/trace/trace_syscalls.c +0 -0
index 942ddbdace4a..7aa1f4299486 100644
@@ -555,7 +555,7 @@ static int perf_call_bpf_enter(struct

trace_event_call *call, struct pt_regs *re
struct syscall_trace_enter *rec)

 {
        struct syscall_tp_t {
kernel/trace/trace_syscalls.c +1 -1
-               unsigned long long regs;
+               struct trace_entry ent;
                unsigned long syscall_nr;
                unsigned long args[SYSCALL_DEFINE_MAXARGS];
        } param;
@@ -657,7 +657,7 @@ static int perf_call_bpf_exit(struct

trace_event_call *call, struct pt_regs *reg
struct syscall_trace_exit *rec)

 {
        struct syscall_tp_t {
kernel/trace/trace_syscalls.c +1 -1
-               unsigned long long regs;
+               struct trace_entry ent;
Quoted 1 line

pls add build_bug_on that sizeof(ent) >= sizeof(void*).

Ok. Should the line *(struct pt_regs **)&param = regs; be commented somehow?

Signature · 2 lines
WBR,
Yauheni Kaliuta
#8 lore raw
Quoted 84 lines · Yauheni Kaliuta

Hi, Alexei!

Quoted 47 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 45 lines · Yauheni Kaliuta
Quoted 22 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 13 lines · Yauheni Kaliuta
Quoted 12 lines
Quoted 5 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov

Should perf_call_bpf_enter/exit (kernel/trace/trace_syscalls.c) use struct trace_event_raw_sys_enter/exit instead of locally crafted struct syscall_tp_t nowadays?

No. It needs syscall_tp_t.

test_progs's vmlinux test

Quoted 2 lines

expects it as the context.

Quoted 6 lines
Quoted 1 line

what do you mean? Pls share a code pointer?

https://github.com/torvalds/linux/blob/master/tools/testing/selftests/bpf/progs/test_vmlinux.c#L19

SEC("tp/syscalls/sys_enter_nanosleep")
int handle__tp(struct trace_event_raw_sys_enter *args)

Quoted 8 lines

I see. That bit is correct and that's what bpftrace is doing when attaching to syscalls.
What do you see in your patched RT kernel when you do:
cat /sys/kernel/debug/tracing/events/syscalls/sys_enter_nanosleep/format ?
Depending on the answer we might need to fix
the kernel side that has to use struct trace_entry
in syscall_tp_t instead of plain long long.

# cat /sys/kernel/debug/tracing/events/syscalls/sys_enter_nanosleep/format
name: sys_enter_nanosleep
ID: 374
format:
field:unsigned short common_type;       offset:0;       size:2; signed:0;
field:unsigned char common_flags;       offset:2;       size:1; signed:0;
field:unsigned char common_preempt_count;       offset:3;       size:1; signed:0;
field:int common_pid;   offset:4;       size:4; signed:1;
field:unsigned char common_preempt_lazy_count;  offset:8;       size:1; signed:0;
field:int __syscall_nr; offset:12;      size:4; signed:1;
field:struct __kernel_timespec * rqtp;  offset:16;      size:8; signed:0;
field:struct __kernel_timespec * rmtp;  offset:24;      size:8; signed:0;
Quoted 6 lines
Quoted 1 line

print fmt: "rqtp: 0x%08lx, rmtp: 0x%08lx", ((unsigned long)(REC->rqtp)), ((unsigned long)(REC->rmtp))

Lol.
Jiri even fixed the issue with this format in bpftrace 3 years ago:
https://github.com/iovisor/bpftrace/commit/a2e3d5dbc03ceb49b776cf5602d31896158844a7

Hehe :)

Quoted 23 lines

Let's fix the kernel side too. Something like this should do it:

kernel/trace/trace_syscalls.c +0 -0
index 942ddbdace4a..7aa1f4299486 100644
@@ -555,7 +555,7 @@ static int perf_call_bpf_enter(struct

trace_event_call *call, struct pt_regs *re
struct syscall_trace_enter *rec)

 {
        struct syscall_tp_t {
kernel/trace/trace_syscalls.c +1 -1
-               unsigned long long regs;
+               struct trace_entry ent;
                unsigned long syscall_nr;
                unsigned long args[SYSCALL_DEFINE_MAXARGS];
        } param;
@@ -657,7 +657,7 @@ static int perf_call_bpf_exit(struct

trace_event_call *call, struct pt_regs *reg
struct syscall_trace_exit *rec)

 {
        struct syscall_tp_t {
kernel/trace/trace_syscalls.c +1 -1
-               unsigned long long regs;
+               struct trace_entry ent;
Quoted 3 lines
Quoted 1 line

pls add build_bug_on that sizeof(ent) >= sizeof(void*).

Ok. Should the line *(struct pt_regs **)&param = regs; be commented somehow?

commented out? No. It's mandatory.
And the reason for build_bug_on existence... to make sure that there is enough space there.

#9 lore raw

Hi, Alexei!

Quoted 7 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 5 lines · Yauheni Kaliuta
Quoted 5 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 3 lines · Yauheni Kaliuta
Quoted 3 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 1 line · Yauheni Kaliuta
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov

[...]

Quoted 25 lines
Quoted 25 lines
Quoted 23 lines

Let's fix the kernel side too. Something like this should do it:

kernel/trace/trace_syscalls.c +0 -0
index 942ddbdace4a..7aa1f4299486 100644
@@ -555,7 +555,7 @@ static int perf_call_bpf_enter(struct

trace_event_call *call, struct pt_regs *re
struct syscall_trace_enter *rec)

 {
        struct syscall_tp_t {
kernel/trace/trace_syscalls.c +1 -1
-               unsigned long long regs;
+               struct trace_entry ent;
                unsigned long syscall_nr;
                unsigned long args[SYSCALL_DEFINE_MAXARGS];
        } param;
@@ -657,7 +657,7 @@ static int perf_call_bpf_exit(struct

trace_event_call *call, struct pt_regs *reg
struct syscall_trace_exit *rec)

 {
        struct syscall_tp_t {
kernel/trace/trace_syscalls.c +1 -1
-               unsigned long long regs;
+               struct trace_entry ent;
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

pls add build_bug_on that sizeof(ent) >= sizeof(void*).

Ok. Should the line *(struct pt_regs **)&param = regs; be commented somehow?

commented out?

No, no :)

Quoted 3 lines

No. It's mandatory.
And the reason for build_bug_on existence... to make sure that there is enough space there.

Yes, it's clear for sure.

It can be not obvious why basically 'ent' is inited with 'regs'. Before it was called 'regs' at least.

Signature · 2 lines
WBR,
Yauheni Kaliuta
#10 lore raw
Quoted 46 lines · Yauheni Kaliuta

Hi, Alexei!

Quoted 7 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 5 lines · Yauheni Kaliuta
Quoted 5 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 3 lines · Yauheni Kaliuta
Quoted 3 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov
Quoted 1 line · Yauheni Kaliuta
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Alexei Starovoitov

[...]

Quoted 25 lines
Quoted 25 lines
Quoted 23 lines

Let's fix the kernel side too. Something like this should do it:

kernel/trace/trace_syscalls.c +0 -0
index 942ddbdace4a..7aa1f4299486 100644
@@ -555,7 +555,7 @@ static int perf_call_bpf_enter(struct

trace_event_call *call, struct pt_regs *re
struct syscall_trace_enter *rec)

 {
        struct syscall_tp_t {
kernel/trace/trace_syscalls.c +1 -1
-               unsigned long long regs;
+               struct trace_entry ent;
                unsigned long syscall_nr;
                unsigned long args[SYSCALL_DEFINE_MAXARGS];
        } param;
@@ -657,7 +657,7 @@ static int perf_call_bpf_exit(struct

trace_event_call *call, struct pt_regs *reg
struct syscall_trace_exit *rec)

 {
        struct syscall_tp_t {
kernel/trace/trace_syscalls.c +1 -1
-               unsigned long long regs;
+               struct trace_entry ent;
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

pls add build_bug_on that sizeof(ent) >= sizeof(void*).

Ok. Should the line *(struct pt_regs **)&param = regs; be commented somehow?

commented out?

No, no :)

Quoted 8 lines
Quoted 3 lines

No. It's mandatory.
And the reason for build_bug_on existence... to make sure that there is enough space there.

Yes, it's clear for sure.

It can be not obvious why basically 'ent' is inited with 'regs'. Before it was called 'regs' at least.

Got it :) Yeah. A comment describing the intent would be nice.