ftrace: Add support to keep some functions out of ftrace
Subsystems (10)
Jiri Olsa
reviewerRFC 1/1 ftrace: Add support to keep some functions out of ftrace
hi,
we recently hit bug where ftrace update raced with bpf_dispatcher_update that calls directly bpf_arch_text_poke [1].
The bpf_dispatcher_update creates special trampoline and attaches it to designated bpf_dispatcher_xdp_func function, which is run for xdp bpf programs from several places.
After discussion with Alexei we'd rather keep this code update out of ftrace, because it's already slow and had troubles with CI because of that.
This patch is presenting the idea to allow some functions not to be managed by ftrace by marking them with NOFTRACE_SYMBOL macro and such symbols will not be added to ftrace_pages on the kernel start.
Please note it's RFC so I did not bother with some fast search for is_noftrace_function function.
Perhaps we could use existing NOKPROBE_SYMBOL for this? but I'm not sure you can (or want) to run function trace on such symbols.
thoughts? thanks
jirka
[1] https://lore.kernel.org/bpf/[email protected]/
--- include/asm-generic/vmlinux.lds.h | 10 ++++++ include/linux/bpf.h | 2 ++ include/linux/ftrace.h | 10 ++++++ kernel/trace/ftrace.c | 53 +++++++++++++++++++++++++++++++ 4 files changed, 75 insertions(+)
include/asm-generic/vmlinux.lds.h +10 -0
@@ -210,6 +210,15 @@#define KPROBE_BLACKLIST ()#endif +#ifdef CONFIG_FTRACE +#define NOFTRACE () . = ALIGN(8); \ + __start_noftrace = .; \ + KEEP(*(_no_ftrace)) \ + __stop_noftrace = .; +#else +#define NOFTRACE () +#endif +#ifdef CONFIG_FUNCTION_ERROR_INJECTION #define ERROR_INJECT_WHITELIST () STRUCT_ALIGN(); \ __start_error_injection_whitelist = .; \
@@ -705,6 +714,7 @@ FTRACE_EVENTS() \ TRACE_SYSCALLS() \ KPROBE_BLACKLIST() \ + NOFTRACE() \ ERROR_INJECT_WHITELIST() \ MEM_DISCARD(init.rodata) \CLK_OF_TABLES () \
include/linux/bpf.h +2 -0
@@ -27,6 +27,7 @@#include <linux/bpfptr.h> #include <linux/btf.h> #include <linux/rcupdate_trace.h> +#include <linux/ftrace.h> struct bpf_verifier_env ;struct bpf_verifier_log ;
@@ -919,6 +920,7 @@ int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs);return bpf_func (ctx, insnsi); \ } \ EXPORT_SYMBOL(bpf_dispatcher_##name##_func); \ + NOFTRACE_SYMBOL(bpf_dispatcher_##name##_func); \ struct bpf_dispatcher bpf_dispatcher_##name = \ BPF_DISPATCHER_INIT(bpf_dispatcher_##name);#define DECLARE_BPF_DISPATCHER (name) \
include/linux/ftrace.h +10 -0
@@ -1141,4 +1141,14 @@ unsigned long arch_syscall_addr(int nr);#endif /* CONFIG_FTRACE_SYSCALLS */ +#ifdef CONFIG_FUNCTION_TRACER +#define __NOFTRACE_SYMBOL (fname) \ +static unsigned long __used \ + __section("_no_ftrace") \ + _noftrace_addr_##fname = (unsigned long)fname; +#define NOFTRACE_SYMBOL (fname) __NOFTRACE_SYMBOL(fname) +#else +#define NOFTRACE_SYMBOL (fname) +#endif /* CONFIG_FUNCTION_TRACER */ +#endif /* _LINUX_FTRACE_H */
kernel/trace/ftrace.c +53 -0
71 lines not on this page · raw message
Signature · 1 line
2.35.3
Steven Rostedt
maintainerQuoted 8 lines · On Fri, 22 Jul 2022 13:08:11 +0200 Jiri Olsa <[email protected]> wrote:
hi,
we recently hit bug where ftrace update raced withbpf_dispatcher_updatethat calls directlybpf_arch_text_poke[1].The
bpf_dispatcher_updatecreates special trampoline and attaches it to designatedbpf_dispatcher_xdp_funcfunction, which is run for xdp bpf programs from several places.
Quoted 7 lines
After discussion with Alexei we'd rather keep this code update out of ftrace, because it's already slow and had troubles with CI because of that.
This patch is presenting the idea to allow some functions not to be managed by ftrace by marking them with
NOFTRACE_SYMBOLmacro and such symbols will not be added toftrace_pageson the kernel start.
NACK on any generic way to hide mcount/fentry functions from ftrace.
There's a lot of infrastructure to see what functions are being modified, as the user should know. (See tracefs/enabled_functions).
There's no need for a generic way to hide functions. Once that happens, it will grow and then it will be more confusing to why some functions are traced while others are not.
Quoted 6 lines
Please note it's RFC so I did not bother with some fast search for
is_noftrace_functionfunction.Perhaps we could use existing
NOKPROBE_SYMBOLfor this? but I'm not sure you can (or want) to run function trace on such symbols.
I trace those functions all the time. Yes, I want to continue doing so.
-- Steve
Quoted 20 lines · Steven Rostedt
Quoted 8 lines · On Fri, 22 Jul 2022 13:08:11 +0200 Jiri Olsa <[email protected]> wrote:
hi,
we recently hit bug where ftrace update raced withbpf_dispatcher_updatethat calls directlybpf_arch_text_poke[1].The
bpf_dispatcher_updatecreates special trampoline and attaches it to designatedbpf_dispatcher_xdp_funcfunction, which is run for xdp bpf programs from several places.Quoted 7 lines
After discussion with Alexei we'd rather keep this code update out of ftrace, because it's already slow and had troubles with CI because of that.
This patch is presenting the idea to allow some functions not to be managed by ftrace by marking them with
NOFTRACE_SYMBOLmacro and such symbols will not be added toftrace_pageson the kernel start.
Quoted 8 lines
NACK on any generic way to hide mcount/fentry functions from ftrace.
There's a lot of infrastructure to see what functions are being modified, as the user should know. (See tracefs/enabled_functions).
There's no need for a generic way to hide functions. Once that happens, it will grow and then it will be more confusing to why some functions are traced while others are not.
Steven,
ftrace must not peek into bpf specific functions. Currently ftrace is causing the kernel to crash. What Jiri is proposing is to fix ftrace bug. And you're saying nack? let ftrace be broken ?
If you don't like Jiri's approach please propose something else.
Steven Rostedt
maintainerQuoted 6 lines · On Fri, 22 Jul 2022 09:04:29 -0700 Alexei Starovoitov…
ftrace must not peek into bpf specific functions. Currently ftrace is causing the kernel to crash. What Jiri is proposing is to fix ftrace bug. And you're saying nack? let ftrace be broken ?
If you don't like Jiri's approach please propose something else.
So, why not mark it as notrace? That will prevent ftrace from looking at it.
-- Steve
Steven Rostedt
maintainerQuoted 7 lines · On Fri, 22 Jul 2022 12:08:54 -0400 Steven Rostedt <[email protected]> wrote:
Quoted 4 lines · On Fri, 22 Jul 2022 09:04:29 -0700 Alexei Starovoitov…
ftrace must not peek into bpf specific functions. Currently ftrace is causing the kernel to crash. What Jiri is proposing is to fix ftrace bug. And you're saying nack? let ftrace be broken ?
Sounds like a BPF bug to me. Ftrace did nothing to cause this breakage. It was something BPF must have done. What exactly is BPF doing to ftrace locations anyway?
Quoted 5 lines
Quoted 2 lines
If you don't like Jiri's approach please propose something else.
So, why not mark it as notrace? That will prevent ftrace from looking at it.
And if for some strange reason you need the mcount/fentry on some internal BPF infrastructure, the work around is to register two ftrace_ops() that have filters to that function. In which case ftrace will force the call to the ftrace iterator loop, and any more ops attached will simply be added to that loop, and ftrace will no longer touch that location.
Then you can do whatever you want to it without fear of racing with ftrace.
But other than that, we don't need infrastructure to hide any mcount/fentry locations from ftrace. Those were add *for* ftrace.
-- Steve
Quoted 7 lines · Steven Rostedt
Quoted 3 lines · On Fri, 22 Jul 2022 12:08:54 -0400 Steven Rostedt <[email protected]> wrote:
On Fri, 22 Jul 2022 09:04:29 -0700
Alexei Starovoitov <[email protected]> wrote:
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
ftrace must not peek into bpf specific functions. Currently ftrace is causing the kernel to crash. What Jiri is proposing is to fix ftrace bug. And you're saying nack? let ftrace be broken ?
Sounds like a BPF bug to me. Ftrace did nothing to cause this breakage. It was something BPF must have done. What exactly is BPF doing to ftrace locations anyway?
ftrace location?
fentry != ftrace.
nop5 in the beginning of the function or in the middle of it doesn't mean that it's safe for ftrace to attach there. In some cases bpf has custom calling convention like it preserves %rax.
In other cases there will be multiple nop5 locations through the function where special care needs to be taken to attach.
Quoted 6 lines
Quoted 5 lines
Quoted 2 lines
If you don't like Jiri's approach please propose something else.
So, why not mark it as notrace? That will prevent ftrace from looking at it.
Quoted 7 lines
And if for some strange reason you need the mcount/fentry on some internal BPF infrastructure, the work around is to register two
ftrace_ops()that have filters to that function. In which case ftrace will force the call to the ftrace iterator loop, and any more ops attached will simply be added to that loop, and ftrace will no longer touch that location.Then you can do whatever you want to it without fear of racing with ftrace.
Jiri,
that sounds like a workable solution.
wdyt?
Quoted 2 lines
But other than that, we don't need infrastructure to hide any mcount/fentry locations from ftrace. Those were add *for* ftrace.
fentry != ftrace.
Steven Rostedt
maintainerQuoted 7 lines · On Fri, 22 Jul 2022 09:53:32 -0700 Alexei Starovoitov…
Quoted 4 lines
Sounds like a BPF bug to me. Ftrace did nothing to cause this breakage. It was something BPF must have done. What exactly is BPF doing to ftrace locations anyway?
ftrace location?
fentry != ftrace.
It was the entire reason to add that. Basically, Yes, fentry == ftrace! BPF can not steal it. Look at the history of it. Everything to do with fentry being added to Linux was for ftrace, and ftrace was designed specifically for fentry.
The ftrace infrastructure was designed around fentry/mcount. The rest of the tracing infrastructure was called "ftrace" for the same reason people call distros Linux (and why FSF hates that). Just because it was built around the ftrace infrastructure, not the other way around. This is why I renamed all the ftrace.h files in include/trace/ to trace_event.h. Because calling it ftrace.h was a misnomer.
Quoted 2 lines
nop5 in the beginning of the function or in the middle of it doesn't mean that it's safe for ftrace to attach there.
How did that nop5 get put there?
Before compilers added support for doing that at compile time (which they added *FOR* ftrace), it was ftrace that converted the calls to mcount/fentry to nops. In fact, the fentry trampoline exists in ftrace_64.S.
Quoted 20 lines
In some cases bpf has custom calling convention like it preserves %rax.
In other cases there will be multiple nop5 locations through the function where special care needs to be taken to attach.Quoted 6 lines
Quoted 5 lines
Quoted 2 lines
If you don't like Jiri's approach please propose something else.
So, why not mark it as notrace? That will prevent ftrace from looking at it.
Quoted 7 lines
And if for some strange reason you need the mcount/fentry on some internal BPF infrastructure, the work around is to register two
ftrace_ops()that have filters to that function. In which case ftrace will force the call to the ftrace iterator loop, and any more ops attached will simply be added to that loop, and ftrace will no longer touch that location.Then you can do whatever you want to it without fear of racing with ftrace.
Quoted 8 lines
Jiri,
that sounds like a workable solution.
wdyt?Quoted 2 lines
But other than that, we don't need infrastructure to hide any mcount/fentry locations from ftrace. Those were add *for* ftrace.
fentry != ftrace.
I 100% disagree with the above statement.
-- Steve
Quoted 21 lines · Steven Rostedt
Quoted 7 lines · On Fri, 22 Jul 2022 12:08:54 -0400 Steven Rostedt <[email protected]> wrote:
Quoted 4 lines · On Fri, 22 Jul 2022 09:04:29 -0700 Alexei Starovoitov…
ftrace must not peek into bpf specific functions. Currently ftrace is causing the kernel to crash. What Jiri is proposing is to fix ftrace bug. And you're saying nack? let ftrace be broken ?
Sounds like a BPF bug to me. Ftrace did nothing to cause this breakage. It was something BPF must have done. What exactly is BPF doing to ftrace locations anyway?
Quoted 5 lines
Quoted 2 lines
If you don't like Jiri's approach please propose something else.
So, why not mark it as notrace? That will prevent ftrace from looking at it.
Quoted 7 lines
And if for some strange reason you need the mcount/fentry on some internal BPF infrastructure, the work around is to register two
ftrace_ops()that have filters to that function. In which case ftrace will force the call to the ftrace iterator loop, and any more ops attached will simply be added to that loop, and ftrace will no longer touch that location.Then you can do whatever you want to it without fear of racing with ftrace.
ok, I think we could use that, I'll check
Quoted 3 lines
But other than that, we don't need infrastructure to hide any mcount/fentry locations from ftrace. Those were add *for* ftrace.
I think I understand the fentry/ftrace equivalence you see, I remember the perl mcount script ;-)
still I think we should be able to define function that has fentry profile call and be able to manage it without ftrace
one other thought.. how about adding function that would allow to disable function in ftrace, with existing FTRACE_FL_DISABLED or some new flag
that way ftrace still keeps track of it, but won't allow to use it in ftrace infra
jirka
Steven Rostedt
maintainerQuoted 8 lines · On Fri, 22 Jul 2022 23:05:19 +0200 Jiri Olsa <[email protected]> wrote:
ok, I think we could use that, I'll check
Quoted 3 lines
But other than that, we don't need infrastructure to hide any mcount/fentry locations from ftrace. Those were add *for* ftrace.
I think I understand the fentry/ftrace equivalence you see, I remember the perl mcount script ;-)
It's even more than that. We worked with the compiler folks to get fentry for ftrace purposes (namely to speed it up, and not rely on frame pointers, which mcount did). fentry never existed until then. Like I said. fentry was created *for* ftrace. And currently it's x86 specific, as it relies on the calling convention that a call does both, push the return address onto the stack, and jump to a function. The blr (branch-link-register) method is more complex, which is where the "patchable" work comes from.
Quoted 9 lines
still I think we should be able to define function that has fentry profile call and be able to manage it without ftrace
one other thought.. how about adding function that would allow to disable function in ftrace, with existing
FTRACE_FL_DISABLEDor some new flagthat way ftrace still keeps track of it, but won't allow to use it in ftrace infra
Another way is to remove it at compile time from the mcount_loc table, and add it to your own table. I take it, this is for bpf infrastructure code and not for any code that's in the day to day processing of the kernel, right?
-- Steve
Steven Rostedt
maintainerQuoted 11 lines · On Fri, 22 Jul 2022 17:41:20 -0400 Steven Rostedt <[email protected]> wrote:
Quoted 2 lines
I think I understand the fentry/ftrace equivalence you see, I remember the perl mcount script ;-)
It's even more than that. We worked with the compiler folks to get fentry for ftrace purposes (namely to speed it up, and not rely on frame pointers, which mcount did). fentry never existed until then. Like I said. fentry was created *for* ftrace. And currently it's x86 specific, as it relies on the calling convention that a call does both, push the return address onto the stack, and jump to a function. The blr (branch-link-register) method is more complex, which is where the "patchable" work comes from.
If you are interested in more details about the birth of fentry, here's the email that started it all:
https://lore.kernel.org/lkml/[email protected]/
-- Steve
Steven Rostedt
maintainerQuoted 4 lines · On Fri, 22 Jul 2022 23:53:58 -0400 Steven Rostedt <[email protected]> wrote:
If you are interested in more details about the birth of fentry, here's the email that started it all:
And I was the one to suggest the "fentry" name as well.
https://lore.kernel.org/lkml/[email protected]/
-- Steve
Quoted 11 lines · Steven Rostedt
Quoted 4 lines · On Fri, 22 Jul 2022 23:53:58 -0400 Steven Rostedt <[email protected]> wrote:
If you are interested in more details about the birth of fentry, here's the email that started it all:
And I was the one to suggest the "fentry" name as well.
ah, did not know that, thanks for sharing
jirka
Quoted 25 lines · Steven Rostedt
Quoted 8 lines · On Fri, 22 Jul 2022 23:05:19 +0200 Jiri Olsa <[email protected]> wrote:
ok, I think we could use that, I'll check
Quoted 3 lines
But other than that, we don't need infrastructure to hide any mcount/fentry locations from ftrace. Those were add *for* ftrace.
I think I understand the fentry/ftrace equivalence you see, I remember the perl mcount script ;-)
It's even more than that. We worked with the compiler folks to get fentry for ftrace purposes (namely to speed it up, and not rely on frame pointers, which mcount did). fentry never existed until then. Like I said. fentry was created *for* ftrace. And currently it's x86 specific, as it relies on the calling convention that a call does both, push the return address onto the stack, and jump to a function. The blr (branch-link-register) method is more complex, which is where the "patchable" work comes from.
Quoted 4 lines
still I think we should be able to define function that has fentry profile call and be able to manage it without ftrace
Quoted 8 lines
Quoted 5 lines
one other thought.. how about adding function that would allow to disable function in ftrace, with existing
FTRACE_FL_DISABLEDor some new flagthat way ftrace still keeps track of it, but won't allow to use it in ftrace infra
Another way is to remove it at compile time from the
mcount_loctable, and add it to your own table. I take it, this is for bpf infrastructure code
hm, perhaps we could move it to separate object and switch off -mrecord-mcount for it, I'll check
Quoted 2 lines
and not for any code that's in the day to day processing of the kernel, right?
yes, it's bpf specific code
thanks,
jirka
Quoted 29 lines · Jiri Olsa
Quoted 28 lines · Steven Rostedt
Quoted 8 lines · On Fri, 22 Jul 2022 23:05:19 +0200 Jiri Olsa <[email protected]> wrote:
ok, I think we could use that, I'll check
Quoted 3 lines
But other than that, we don't need infrastructure to hide any mcount/fentry locations from ftrace. Those were add *for* ftrace.
I think I understand the fentry/ftrace equivalence you see, I remember the perl mcount script ;-)
It's even more than that. We worked with the compiler folks to get fentry for ftrace purposes (namely to speed it up, and not rely on frame pointers, which mcount did). fentry never existed until then. Like I said. fentry was created *for* ftrace. And currently it's x86 specific, as it relies on the calling convention that a call does both, push the return address onto the stack, and jump to a function. The blr (branch-link-register) method is more complex, which is where the "patchable" work comes from.
Quoted 7 lines
still I think we should be able to define function that has fentry profile call and be able to manage it without ftrace
one other thought.. how about adding function that would allow to disable function in ftrace, with existing
FTRACE_FL_DISABLEDor some new flag
Quoted 8 lines
Quoted 5 lines
Quoted 2 lines
that way ftrace still keeps track of it, but won't allow to use it in ftrace infra
Another way is to remove it at compile time from the
mcount_loctable, and add it to your own table. I take it, this is for bpf infrastructure codehm, perhaps we could move it to separate object and switch off -mrecord-mcount for it, I'll check
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
this seems to work, but it depends on -mrecord-mcount support in gcc and it's x86 specific... other archs seems to use -fpatchable-function-entry, which does not seem to have option to omit symbol from being collected to the section
disabling specific ftrace symbol with FTRACE_FL_DISABLED flag seems to be easir and generic solution.. I'll send RFC for that
jirka
---
net/core/Makefile +5 -1
@@ -12,10 +12,14 @@ obj-$(CONFIG_SYSCTL) += sysctl_net_core.o obj-y += dev.o dev_addr_lists.o dst.o netevent.o \ neighbour.o rtnetlink.o utils.o link_watch.o filter.o \ sock_diag.o dev_ioctl.o tso.o sock_reuseport.o \ - fib_notifier.o xdp.o flow_offload.o gro.o + fib_notifier.o xdp.o flow_offload.o gro.o \ + dispatcher.o obj-$(CONFIG_NETDEV_ADDR_LIST_TEST) += dev_addr_lists_test.o +# remove dispatcher function from ftrace sight +CFLAGS_REMOVE_dispatcher.o = -mrecord-mcount + obj-y += net-sysfs.o obj-$(CONFIG_PAGE_POOL) += page_pool.o obj-$(CONFIG_PROC_FS) += net-procfs.o
net/core/dispatcher.c new file +3 -0
@@ -0,0 +1,3 @@ +#include <linux/bpf.h> + +DEFINE_BPF_DISPATCHER (xdp)
net/core/filter.c +0 -2
@@ -11333,8 +11333,6 @@ const struct bpf_verifier_ops sk_lookup_verifier_ops = {#endif /* CONFIG_INET */ -DEFINE_BPF_DISPATCHER (xdp) -void bpf_prog_change_xdp (struct bpf_prog *prev_prog,struct bpf_prog *prog) {bpf_dispatcher_change_prog (BPF_DISPATCHER_PTR (xdp), prev_prog, prog);
Steven Rostedt
maintainerQuoted 11 lines · On Fri, 12 Aug 2022 23:18:15 +0200 Jiri Olsa <[email protected]> wrote:
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
this seems to work, but it depends on -mrecord-mcount support in gcc and it's x86 specific... other archs seems to use -fpatchable-function-entry, which does not seem to have option to omit symbol from being collected to the section
disabling specific ftrace symbol with
FTRACE_FL_DISABLEDflag seems to be easir and generic solution.. I'll send RFC for that
No, please don't.
I could help you with recordmcount (which creates the .mcount_loc locations when -mrecordmcount is not enabled) and remove it for you.
I still want this solution over the easy-way-out that can lead to half of the kernel being hidden from ftrace.
This is the point that I made about fentry == ftrace. Because we did the hard part to make fentry work. That included creating sections that point to them.
All your patch needs is to tell the build not to run recordmcount on the file. Remember that perl script I wrote? It (and the C version) is what creates the mcount_loc location that you need to hide these files from.
-- Steve
Steven Rostedt
maintainerQuoted 8 lines · On Fri, 12 Aug 2022 23:18:15 +0200 Jiri Olsa <[email protected]> wrote:
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
this seems to work, but it depends on -mrecord-mcount support in gcc and it's x86 specific... other archs seems to use -fpatchable-function-entry, which does not seem to have option to omit symbol from being collected to the section
As I stated. the __mcount_loc section was created by ftrace. It has nothing to do with -fpatchable-function-entry. It's just that the archs that use that, do not have a gcc that creates the __mcount_loc section.
Quoted 3 lines
disabling specific ftrace symbol with
FTRACE_FL_DISABLEDflag seems to be easir and generic solution.. I'll send RFC for that
It's not easier.
Here, I have a POC for you and some more history.
The recordmcount.pl Perl script was the first to create the __mcount_loc section in all objects that ftrace needed to hook to. It did an objdump, found the locations of the calls to mcount, created another file that had a section __mcount_loc that referenced all those locations. Compiled and relinked that back into the original object.
This was performed on all object files during the build, and had an impact on build times. But this is when I also created and introduced "make localmodconfig", which shrunk the build times for many people, so nobody noticed the build time increase! :-)
Then John Reiser sent me a patch that created recordmcount.c that did the same work but directly modified the ELF object files without having to run scripts. This got rid of that horrible overhead from the scripts.
Then, the gcc folks decided to be helpful here as well and created the --mrecord-mcount option that would create the __mcount_loc section for us, where we no longer needed the recordmcount scripts/C program. But is not available across the board.
Today, objtool has also got involved, and added an "--mcount" option that will create the section too.
Below is a patch that extends yours by adding a NO_MCOUNT_FILES list, that you add the object file to and it will prevent the other methods from adding an mcount_loc location.
I'm adding the objtool folks to make sure this is fine with them. Again, this is a proof of concept, but works. It may need to be cleaned a bit before it is final.
-- Steve
scripts/Makefile.build +3 -1
@@ -206,8 +206,10 @@ sub_cmd_record_mcount = perl $(srctree)/ "$(if $(part-of-module),1,0)" "$(@)"; recordmcount_source := $(srctree)/scripts/recordmcount.pl endif # BUILD_C_RECORDMCOUNT -cmd_record_mcount = $(if $(findstring $(strip $(CC_FLAGS_FTRACE)),$(_c_flags)), \ +chk_sub_cmd_record_mcount = $(if $(filter $(shell basename $@),$(NO_MCOUNT_FILES)),, \ $(sub_cmd_record_mcount)) +cmd_record_mcount = $(if $(findstring $(strip $(CC_FLAGS_FTRACE)),$(_c_flags)), \ + $(chk_sub_cmd_record_mcount)) endif # CONFIG_FTRACE_MCOUNT_USE_RECORDMCOUNT # 'OBJECT_FILES_NON_STANDARD := y': skip objtool checking for a directory
scripts/Makefile.lib +2 -1
@@ -233,7 +233,8 @@ objtool_args = \ $(if $(CONFIG_HAVE_JUMP_LABEL_HACK), --hacks=jump_label) \ $(if $(CONFIG_HAVE_NOINSTR_HACK), --hacks=noinstr) \ $(if $(CONFIG_X86_KERNEL_IBT), --ibt) \ - $(if $(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL), --mcount) \ + $(if $(filter $(shell basename $@),$(NO_MCOUNT_FILES)),, \ + $(if $(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL), --mcount)) \ $(if $(CONFIG_UNWINDER_ORC), --orc) \ $(if $(CONFIG_RETPOLINE), --retpoline) \ $(if $(CONFIG_SLS), --sls) \
net/core/Makefile +6 -1
@@ -11,10 +11,15 @@ obj-$(CONFIG_SYSCTL) += sysctl_net_core. obj-y += dev.o dev_addr_lists.o dst.o netevent.o \ neighbour.o rtnetlink.o utils.o link_watch.o filter.o \ sock_diag.o dev_ioctl.o tso.o sock_reuseport.o \ - fib_notifier.o xdp.o flow_offload.o gro.o + fib_notifier.o xdp.o flow_offload.o gro.o \ + dispatcher.o obj-$(CONFIG_NETDEV_ADDR_LIST_TEST) += dev_addr_lists_test.o +# remove dispatcher function from ftrace sight +CFLAGS_REMOVE_dispatcher.o = -mrecord-mcount +NO_MCOUNT_FILES += dispatcher.o + obj-y += net-sysfs.o obj-$(CONFIG_PAGE_POOL) += page_pool.o obj-$(CONFIG_PROC_FS) += net-procfs.o
net/core/dispatcher.c new file +3 -0
@@ -0,0 +1,3 @@ +#include <linux/bpf.h> + +DEFINE_BPF_DISPATCHER (xdp)
net/core/filter.c +0 -2
@@ -11162,8 +11162,6 @@ const struct bpf_verifier_ops sk_lookup_#endif /* CONFIG_INET */ -DEFINE_BPF_DISPATCHER (xdp) -void bpf_prog_change_xdp (struct bpf_prog *prev_prog,struct bpf_prog *prog) {bpf_dispatcher_change_prog (BPF_DISPATCHER_PTR (xdp), prev_prog, prog);
Steven Rostedt
maintainerscripts/Makefile.lib objtool_args = \ · line 237
Quoted 11 lines · On Sat, 13 Aug 2022 15:02:52 -0400 Steven Rostedt <[email protected]> wrote:
Index: linux-trace.git/scripts/Makefile.lib
===================================================================scripts/Makefile.lib +2
-1@@ -233,7 +233,8 @@ objtool_args = \ $(if $(CONFIG_HAVE_JUMP_LABEL_HACK), --hacks=jump_label) \ $(if $(CONFIG_HAVE_NOINSTR_HACK), --hacks=noinstr) \ $(if $(CONFIG_X86_KERNEL_IBT), --ibt) \ - $(if $(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL), --mcount) \ + $(if $(filter $(shell basename $@),$(NO_MCOUNT_FILES)),, \ + $(if $(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL), --mcount)) \
I believe there's some security and other validations that objtool does that requires it to know about the mcount locations.
If BPF is doing something unique, and modifying code as well (outside the jump label and ftrace work), does objtool need to know about that too?
-- Steve
Quoted 3 lines
scripts/Makefile.lib +0
-0$(if $(CONFIG_UNWINDER_ORC), --orc) \ $(if $(CONFIG_RETPOLINE), --retpoline) \ $(if $(CONFIG_SLS), --sls) \
Quoted 39 lines · Steven Rostedt
Quoted 8 lines · On Fri, 12 Aug 2022 23:18:15 +0200 Jiri Olsa <[email protected]> wrote:
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
this seems to work, but it depends on -mrecord-mcount support in gcc and it's x86 specific... other archs seems to use -fpatchable-function-entry, which does not seem to have option to omit symbol from being collected to the section
As I stated. the
__mcount_locsection was created by ftrace. It has nothing to do with -fpatchable-function-entry. It's just that the archs that use that, do not have a gcc that creates the__mcount_locsection.Quoted 3 lines
disabling specific ftrace symbol with
FTRACE_FL_DISABLEDflag seems to be easir and generic solution.. I'll send RFC for thatIt's not easier.
Here, I have a POC for you and some more history.
The recordmcount.pl Perl script was the first to create the
__mcount_locsection in all objects that ftrace needed to hook to. It did an objdump, found the locations of the calls to mcount, created another file that had a section__mcount_locthat referenced all those locations. Compiled and relinked that back into the original object.This was performed on all object files during the build, and had an impact on build times. But this is when I also created and introduced "make localmodconfig", which shrunk the build times for many people, so nobody noticed the build time increase! :-)
Then John Reiser sent me a patch that created
recordmcount.cthat did the same work but directly modified the ELF object files without having to run scripts. This got rid of that horrible overhead from the scripts.
Quoted 7 lines
Then, the gcc folks decided to be helpful here as well and created the --mrecord-mcount option that would create the
__mcount_locsection for us, where we no longer needed the recordmcountscripts/Cprogram. But is not available across the board.Today, objtool has also got involved, and added an "--mcount" option that will create the section too.
I overlooked that objtool is involved as well,
will check on that
Quoted 4 lines
Below is a patch that extends yours by adding a
NO_MCOUNT_FILESlist, that you add the object file to and it will prevent the other methods from adding anmcount_loclocation.
thanks,
jirka
Quoted 80 lines
I'm adding the objtool folks to make sure this is fine with them. Again, this is a proof of concept, but works. It may need to be cleaned a bit before it is final.
-- Steve
Index: linux-trace.git/scripts/Makefile.build
===================================================================scripts/Makefile.build +3
-1@@ -206,8 +206,10 @@ sub_cmd_record_mcount = perl $(srctree)/ "$(if $(part-of-module),1,0)" "$(@)"; recordmcount_source := $(srctree)/scripts/recordmcount.pl endif # BUILD_C_RECORDMCOUNT -cmd_record_mcount = $(if $(findstring $(strip $(CC_FLAGS_FTRACE)),$(_c_flags)), \ +chk_sub_cmd_record_mcount = $(if $(filter $(shell basename $@),$(NO_MCOUNT_FILES)),, \ $(sub_cmd_record_mcount)) +cmd_record_mcount = $(if $(findstring $(strip $(CC_FLAGS_FTRACE)),$(_c_flags)), \ + $(chk_sub_cmd_record_mcount)) endif # CONFIG_FTRACE_MCOUNT_USE_RECORDMCOUNT # 'OBJECT_FILES_NON_STANDARD := y': skip objtool checking for a directoryIndex: linux-trace.git/scripts/Makefile.lib
===================================================================scripts/Makefile.lib +2
-1@@ -233,7 +233,8 @@ objtool_args = \ $(if $(CONFIG_HAVE_JUMP_LABEL_HACK), --hacks=jump_label) \ $(if $(CONFIG_HAVE_NOINSTR_HACK), --hacks=noinstr) \ $(if $(CONFIG_X86_KERNEL_IBT), --ibt) \ - $(if $(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL), --mcount) \ + $(if $(filter $(shell basename $@),$(NO_MCOUNT_FILES)),, \ + $(if $(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL), --mcount)) \ $(if $(CONFIG_UNWINDER_ORC), --orc) \ $(if $(CONFIG_RETPOLINE), --retpoline) \ $(if $(CONFIG_SLS), --sls) \Index: linux-trace.git/net/core/Makefile
===================================================================net/core/Makefile +6
-1@@ -11,10 +11,15 @@ obj-$(CONFIG_SYSCTL) += sysctl_net_core. obj-y += dev.o dev_addr_lists.o dst.o netevent.o \ neighbour.o rtnetlink.o utils.o link_watch.o filter.o \ sock_diag.o dev_ioctl.o tso.o sock_reuseport.o \ - fib_notifier.o xdp.o flow_offload.o gro.o + fib_notifier.o xdp.o flow_offload.o gro.o \ + dispatcher.o obj-$(CONFIG_NETDEV_ADDR_LIST_TEST) += dev_addr_lists_test.o +# remove dispatcher function from ftrace sight +CFLAGS_REMOVE_dispatcher.o = -mrecord-mcount +NO_MCOUNT_FILES += dispatcher.o + obj-y += net-sysfs.o obj-$(CONFIG_PAGE_POOL) += page_pool.o obj-$(CONFIG_PROC_FS) += net-procfs.oIndex:
linux-trace.git/net/core/dispatcher.c
===================================================================net/core/dispatcher.c new file +3
-0@@ -0,0 +1,3 @@ +#include <linux/bpf.h> + +DEFINE_BPF_DISPATCHER (xdp)Index:
linux-trace.git/net/core/filter.c
===================================================================net/core/filter.c +0
-2@@ -11162,8 +11162,6 @@ const struct bpf_verifier_ops sk_lookup_#endif /* CONFIG_INET */ -DEFINE_BPF_DISPATCHER (xdp) -void bpf_prog_change_xdp (struct bpf_prog *prev_prog,struct bpf_prog *prog) {bpf_dispatcher_change_prog (BPF_DISPATCHER_PTR (xdp), prev_prog, prog);
Quoted 42 lines · Jiri Olsa
Quoted 41 lines · Steven Rostedt
Quoted 8 lines · On Fri, 12 Aug 2022 23:18:15 +0200 Jiri Olsa <[email protected]> wrote:
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
this seems to work, but it depends on -mrecord-mcount support in gcc and it's x86 specific... other archs seems to use -fpatchable-function-entry, which does not seem to have option to omit symbol from being collected to the section
As I stated. the
__mcount_locsection was created by ftrace. It has nothing to do with -fpatchable-function-entry. It's just that the archs that use that, do not have a gcc that creates the__mcount_locsection.Quoted 2 lines
disabling specific ftrace symbol with
FTRACE_FL_DISABLEDflag seems to be easir and generic solution.. I'll send RFC for thatIt's not easier.
Here, I have a POC for you and some more history.
The recordmcount.pl Perl script was the first to create the
__mcount_locsection in all objects that ftrace needed to hook to. It did an objdump, found the locations of the calls to mcount, created another file that had a section__mcount_locthat referenced all those locations. Compiled and relinked that back into the original object.This was performed on all object files during the build, and had an impact on build times. But this is when I also created and introduced "make localmodconfig", which shrunk the build times for many people, so nobody noticed the build time increase! :-)
Then John Reiser sent me a patch that created
recordmcount.cthat did the same work but directly modified the ELF object files without having to run scripts. This got rid of that horrible overhead from the scripts.Then, the gcc folks decided to be helpful here as well and created the --mrecord-mcount option that would create the
__mcount_locsection for us, where we no longer needed the recordmcountscripts/Cprogram. But is not available across the board.
Quoted 4 lines
Quoted 2 lines
Today, objtool has also got involved, and added an "--mcount" option that will create the section too.
I overlooked that objtool is involved as well,
will check on that
objtool --mcount option only involves mcount_loc generation (see annotate_call_site) and other validation check call destination directly (see is_fentry_call).
Some simply removing --mcount option dose work for this.
Another question, it seems we can export and use DEFINE_BPF_DISPATCHER out of kernel, does that means we should add NO_MCOUNT_FILES for these single uages as well?
I dont think it can be made automatically. If ignored, this can be a trouble.
Best,
Chen
Quoted 85 lines
Quoted 3 lines
Below is a patch that extends yours by adding a
NO_MCOUNT_FILESlist, that you add the object file to and it will prevent the other methods from adding anmcount_loclocation.thanks,
jirkaQuoted 79 lines
I'm adding the objtool folks to make sure this is fine with them. Again, this is a proof of concept, but works. It may need to be cleaned a bit before it is final.
-- Steve
Index: linux-trace.git/scripts/Makefile.build
===================================================================scripts/Makefile.build +3
-1@@ -206,8 +206,10 @@ sub_cmd_record_mcount = perl $(srctree)/ "$(if $(part-of-module),1,0)" "$(@)"; recordmcount_source := $(srctree)/scripts/recordmcount.pl endif # BUILD_C_RECORDMCOUNT -cmd_record_mcount = $(if $(findstring $(strip $(CC_FLAGS_FTRACE)),$(_c_flags)), \ +chk_sub_cmd_record_mcount = $(if $(filter $(shell basename $@),$(NO_MCOUNT_FILES)),, \ $(sub_cmd_record_mcount)) +cmd_record_mcount = $(if $(findstring $(strip $(CC_FLAGS_FTRACE)),$(_c_flags)), \ + $(chk_sub_cmd_record_mcount)) endif # CONFIG_FTRACE_MCOUNT_USE_RECORDMCOUNT # 'OBJECT_FILES_NON_STANDARD := y': skip objtool checking for a directoryIndex: linux-trace.git/scripts/Makefile.lib
===================================================================scripts/Makefile.lib +2
-1@@ -233,7 +233,8 @@ objtool_args = \ $(if $(CONFIG_HAVE_JUMP_LABEL_HACK), --hacks=jump_label) \ $(if $(CONFIG_HAVE_NOINSTR_HACK), --hacks=noinstr) \ $(if $(CONFIG_X86_KERNEL_IBT), --ibt) \ - $(if $(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL), --mcount) \ + $(if $(filter $(shell basename $@),$(NO_MCOUNT_FILES)),, \ + $(if $(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL), --mcount)) \ $(if $(CONFIG_UNWINDER_ORC), --orc) \ $(if $(CONFIG_RETPOLINE), --retpoline) \ $(if $(CONFIG_SLS), --sls) \Index: linux-trace.git/net/core/Makefile
===================================================================net/core/Makefile +6
-1@@ -11,10 +11,15 @@ obj-$(CONFIG_SYSCTL) += sysctl_net_core. obj-y += dev.o dev_addr_lists.o dst.o netevent.o \ neighbour.o rtnetlink.o utils.o link_watch.o filter.o \ sock_diag.o dev_ioctl.o tso.o sock_reuseport.o \ - fib_notifier.o xdp.o flow_offload.o gro.o + fib_notifier.o xdp.o flow_offload.o gro.o \ + dispatcher.o obj-$(CONFIG_NETDEV_ADDR_LIST_TEST) += dev_addr_lists_test.o +# remove dispatcher function from ftrace sight +CFLAGS_REMOVE_dispatcher.o = -mrecord-mcount +NO_MCOUNT_FILES += dispatcher.o + obj-y += net-sysfs.o obj-$(CONFIG_PAGE_POOL) += page_pool.o obj-$(CONFIG_PROC_FS) += net-procfs.oIndex:
linux-trace.git/net/core/dispatcher.c
===================================================================net/core/dispatcher.c new file +3
-0@@ -0,0 +1,3 @@ +#include <linux/bpf.h> + +DEFINE_BPF_DISPATCHER (xdp)Index:
linux-trace.git/net/core/filter.c
===================================================================net/core/filter.c +0
-2@@ -11162,8 +11162,6 @@ const struct bpf_verifier_ops sk_lookup_#endif /* CONFIG_INET */ -DEFINE_BPF_DISPATCHER (xdp) -void bpf_prog_change_xdp (struct bpf_prog *prev_prog,struct bpf_prog *prog) {bpf_dispatcher_change_prog (BPF_DISPATCHER_PTR (xdp), prev_prog, prog);
Quoted 53 lines · Chen Zhongjin
Quoted 42 lines · Jiri Olsa
Quoted 41 lines · Steven Rostedt
Quoted 8 lines · On Fri, 12 Aug 2022 23:18:15 +0200 Jiri Olsa <[email protected]> wrote:
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
this seems to work, but it depends on -mrecord-mcount support in gcc and it's x86 specific... other archs seems to use -fpatchable-function-entry, which does not seem to have option to omit symbol from being collected to the section
As I stated. the
__mcount_locsection was created by ftrace. It has nothing to do with -fpatchable-function-entry. It's just that the archs that use that, do not have a gcc that creates the__mcount_locsection.Quoted 2 lines
disabling specific ftrace symbol with
FTRACE_FL_DISABLEDflag seems to be easir and generic solution.. I'll send RFC for thatIt's not easier.
Here, I have a POC for you and some more history.
The recordmcount.pl Perl script was the first to create the
__mcount_locsection in all objects that ftrace needed to hook to. It did an objdump, found the locations of the calls to mcount, created another file that had a section__mcount_locthat referenced all those locations. Compiled and relinked that back into the original object.This was performed on all object files during the build, and had an impact on build times. But this is when I also created and introduced "make localmodconfig", which shrunk the build times for many people, so nobody noticed the build time increase! :-)
Then John Reiser sent me a patch that created
recordmcount.cthat did the same work but directly modified the ELF object files without having to run scripts. This got rid of that horrible overhead from the scripts.Then, the gcc folks decided to be helpful here as well and created the --mrecord-mcount option that would create the
__mcount_locsection for us, where we no longer needed the recordmcountscripts/Cprogram. But is not available across the board.Quoted 4 lines
Quoted 2 lines
Today, objtool has also got involved, and added an "--mcount" option that will create the section too.
I overlooked that objtool is involved as well,
will check on thatobjtool --mcount option only involves
mcount_locgeneration (seeannotate_call_site) and other validation check call destination directly (seeis_fentry_call).
Quoted 6 lines
Some simply removing --mcount option dose work for this.
Another question, it seems we can export and use
DEFINE_BPF_DISPATCHERout of kernel, does that means we should addNO_MCOUNT_FILESfor these single uages as well?
yes, cc-ing Björn to make sure it's valid use case for dispatcher
jirka
Quoted 90 lines
I dont think it can be made automatically. If ignored, this can be a trouble.
Best,
Chen
Quoted 80 lines
Quoted 3 lines
Below is a patch that extends yours by adding a
NO_MCOUNT_FILESlist, that you add the object file to and it will prevent the other methods from adding anmcount_loclocation.thanks,
jirkaQuoted 74 lines
I'm adding the objtool folks to make sure this is fine with them. Again, this is a proof of concept, but works. It may need to be cleaned a bit before it is final.
-- Steve
Index: linux-trace.git/scripts/Makefile.build
===================================================================scripts/Makefile.build +3
-1@@ -206,8 +206,10 @@ sub_cmd_record_mcount = perl $(srctree)/ "$(if $(part-of-module),1,0)" "$(@)"; recordmcount_source := $(srctree)/scripts/recordmcount.pl endif # BUILD_C_RECORDMCOUNT -cmd_record_mcount = $(if $(findstring $(strip $(CC_FLAGS_FTRACE)),$(_c_flags)), \ +chk_sub_cmd_record_mcount = $(if $(filter $(shell basename $@),$(NO_MCOUNT_FILES)),, \ $(sub_cmd_record_mcount)) +cmd_record_mcount = $(if $(findstring $(strip $(CC_FLAGS_FTRACE)),$(_c_flags)), \ + $(chk_sub_cmd_record_mcount)) endif # CONFIG_FTRACE_MCOUNT_USE_RECORDMCOUNT # 'OBJECT_FILES_NON_STANDARD := y': skip objtool checking for a directoryIndex: linux-trace.git/scripts/Makefile.lib
===================================================================scripts/Makefile.lib +2
-1@@ -233,7 +233,8 @@ objtool_args = \ $(if $(CONFIG_HAVE_JUMP_LABEL_HACK), --hacks=jump_label) \ $(if $(CONFIG_HAVE_NOINSTR_HACK), --hacks=noinstr) \ $(if $(CONFIG_X86_KERNEL_IBT), --ibt) \ - $(if $(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL), --mcount) \ + $(if $(filter $(shell basename $@),$(NO_MCOUNT_FILES)),, \ + $(if $(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL), --mcount)) \ $(if $(CONFIG_UNWINDER_ORC), --orc) \ $(if $(CONFIG_RETPOLINE), --retpoline) \ $(if $(CONFIG_SLS), --sls) \Index: linux-trace.git/net/core/Makefile
===================================================================net/core/Makefile +6
-1@@ -11,10 +11,15 @@ obj-$(CONFIG_SYSCTL) += sysctl_net_core. obj-y += dev.o dev_addr_lists.o dst.o netevent.o \ neighbour.o rtnetlink.o utils.o link_watch.o filter.o \ sock_diag.o dev_ioctl.o tso.o sock_reuseport.o \ - fib_notifier.o xdp.o flow_offload.o gro.o + fib_notifier.o xdp.o flow_offload.o gro.o \ + dispatcher.o obj-$(CONFIG_NETDEV_ADDR_LIST_TEST) += dev_addr_lists_test.o +# remove dispatcher function from ftrace sight +CFLAGS_REMOVE_dispatcher.o = -mrecord-mcount +NO_MCOUNT_FILES += dispatcher.o + obj-y += net-sysfs.o obj-$(CONFIG_PAGE_POOL) += page_pool.o obj-$(CONFIG_PROC_FS) += net-procfs.oIndex:
linux-trace.git/net/core/dispatcher.c
===================================================================net/core/dispatcher.c new file +3
-0@@ -0,0 +1,3 @@ +#include <linux/bpf.h> + +DEFINE_BPF_DISPATCHER (xdp)Index:
linux-trace.git/net/core/filter.c
===================================================================net/core/filter.c +0
-2@@ -11162,8 +11162,6 @@ const struct bpf_verifier_ops sk_lookup_#endif /* CONFIG_INET */ -DEFINE_BPF_DISPATCHER (xdp) -void bpf_prog_change_xdp (struct bpf_prog *prev_prog,struct bpf_prog *prog) {bpf_dispatcher_change_prog (BPF_DISPATCHER_PTR (xdp), prev_prog, prog);
On Mon, 15 Aug 2022 at 10:04, Jiri Olsa <[email protected]> wrote:
[...]
Quoted 10 lines
Quoted 10 lines
Quoted 5 lines
Quoted 3 lines
Today, objtool has also got involved, and added an "--mcount" option that will create the section too.
I overlooked that objtool is involved as well,
will check on thatobjtool --mcount option only involves
mcount_locgeneration (seeannotate_call_site) and other validation check call destination directly (seeis_fentry_call).
Quoted 9 lines
Quoted 6 lines
Some simply removing --mcount option dose work for this.
Another question, it seems we can export and use
DEFINE_BPF_DISPATCHERout of kernel, does that means we should addNO_MCOUNT_FILESfor these single uages as well?yes, cc-ing Björn to make sure it's valid use case for dispatcher
Hmm, could you expand a bit on how this would work?
Quoted 15 lines · Björn Töpel
On Mon, 15 Aug 2022 at 10:04, Jiri Olsa <[email protected]> wrote:
[...]Quoted 13 lines
Quoted 13 lines
Quoted 5 lines
Quoted 3 lines
Today, objtool has also got involved, and added an "--mcount" option that will create the section too.
I overlooked that objtool is involved as well,
will check on thatobjtool --mcount option only involves
mcount_locgeneration (seeannotate_call_site) and other validation check call destination directly (seeis_fentry_call).Some simply removing --mcount option dose work for this.
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines
Another question, it seems we can export and use
DEFINE_BPF_DISPATCHERout of kernel, does that means we should addNO_MCOUNT_FILESfor these single uages as well?yes, cc-ing Björn to make sure it's valid use case for dispatcher
Hmm, could you expand a bit on how this would work?
the goal here is to remove bpf_dispatcher_<FUNC>_func functions from ftrace, because it's updated by dispatcher code with bpf_arch_text_poke, but it's also visible and attachable to ftrace.. and will cause problems when these 2 updates will race
question was if DEFINE_BPF_DISPATCHER can be used in kernel module, which would bring another realm of problems ;-)
jirka
Quoted 26 lines · Jiri Olsa
Quoted 15 lines · Björn Töpel
On Mon, 15 Aug 2022 at 10:04, Jiri Olsa <[email protected]> wrote:
[...]Quoted 13 lines
Quoted 13 lines
Quoted 5 lines
Quoted 3 lines
Today, objtool has also got involved, and added an "--mcount" option that will create the section too.
I overlooked that objtool is involved as well,
will check on thatobjtool --mcount option only involves
mcount_locgeneration (seeannotate_call_site) and other validation check call destination directly (seeis_fentry_call).Some simply removing --mcount option dose work for this.
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines
Another question, it seems we can export and use
DEFINE_BPF_DISPATCHERout of kernel, does that means we should addNO_MCOUNT_FILESfor these single uages as well?yes, cc-ing Björn to make sure it's valid use case for dispatcher
Hmm, could you expand a bit on how this would work?
Quoted 8 lines
the goal here is to remove bpf_dispatcher_<FUNC>_func functions from ftrace, because it's updated by dispatcher code with
bpf_arch_text_poke, but it's also visible and attachable to ftrace.. and will cause problems when these 2 updates will racequestion was if
DEFINE_BPF_DISPATCHERcan be used in kernel module, which would bring another realm of problems ;-)
Oh, now I follow. AFAIK there is only one flavor of BPF dispatcher in use, and that's the XDP dispatcher, which does not reside in module code, but is typically *called* by module code.
A module could define a BPF dispatcher, but it wouldn't be able to update it, since bpf_arch_text_poke() does not support poking in modules.
Björn
Quoted 33 lines · Björn Töpel
Quoted 31 lines · Jiri Olsa
Quoted 15 lines · Björn Töpel
On Mon, 15 Aug 2022 at 10:04, Jiri Olsa <[email protected]> wrote:
[...]Quoted 13 lines
Quoted 13 lines
Quoted 5 lines
Quoted 3 lines
Today, objtool has also got involved, and added an "--mcount" option that will create the section too.
I overlooked that objtool is involved as well,
will check on thatobjtool --mcount option only involves
mcount_locgeneration (seeannotate_call_site) and other validation check call destination directly (seeis_fentry_call).Some simply removing --mcount option dose work for this.
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines
Another question, it seems we can export and use
DEFINE_BPF_DISPATCHERout of kernel, does that means we should addNO_MCOUNT_FILESfor these single uages as well?yes, cc-ing Björn to make sure it's valid use case for dispatcher
Hmm, could you expand a bit on how this would work?
the goal here is to remove bpf_dispatcher_<FUNC>_func functions from ftrace, because it's updated by dispatcher code with
bpf_arch_text_poke, but it's also visible and attachable to ftrace.. and will cause problems when these 2 updates will race
Quoted 8 lines
Quoted 3 lines
question was if
DEFINE_BPF_DISPATCHERcan be used in kernel module, which would bring another realm of problems ;-)Oh, now I follow. AFAIK there is only one flavor of BPF dispatcher in use, and that's the XDP dispatcher, which does not reside in module code, but is typically *called* by module code.
Some history why the EXPORT is required:
https://lore.kernel.org/bpf/CAADnVQ+eD-=FZrg8L+YcdCyAS+E30W=Z-ShtEXAXVFjmxV4usg@mail.gmail.com/
Quoted 6 lines · Steven Rostedt
Quoted 3 lines · On Fri, 12 Aug 2022 23:18:15 +0200 Jiri Olsa <[email protected]> wrote:
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
Why ?!?
Quoted 9 lines · Peter Zijlstra
Quoted 6 lines · Steven Rostedt
Quoted 3 lines · On Fri, 12 Aug 2022 23:18:15 +0200 Jiri Olsa <[email protected]> wrote:
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
Why ?!?
there's bpf dispatcher code that updates bpf_dispatcher_xdp_func function with bpf_arch_text_poke and that can race with ftrace update if the function is traced
the idea to solve it is to 'mark' the function independent of ftrace, and add a way to make the function invissible to ftrace but with the profile code fentry call generated
jirka
Quoted 9 lines · Jiri Olsa
Quoted 8 lines · Peter Zijlstra
Quoted 6 lines · Steven Rostedt
Quoted 3 lines · On Fri, 12 Aug 2022 23:18:15 +0200 Jiri Olsa <[email protected]> wrote:
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
Quoted 5 lines
Quoted 1 line
Why ?!?
there's bpf dispatcher code that updates
bpf_dispatcher_xdp_funcfunction withbpf_arch_text_pokeand that can race with ftrace update if the function is traced
I thought bpf_arch_text_poke() wasn't allowed to touch kernel code and ftrace is in full control of it ?
Quoted 11 lines · Peter Zijlstra
Quoted 9 lines · Jiri Olsa
Quoted 8 lines · Peter Zijlstra
Quoted 6 lines · Steven Rostedt
Quoted 3 lines · On Fri, 12 Aug 2022 23:18:15 +0200 Jiri Olsa <[email protected]> wrote:
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
Why ?!?
there's bpf dispatcher code that updates
bpf_dispatcher_xdp_funcfunction withbpf_arch_text_pokeand that can race with ftrace update if the function is tracedI thought
bpf_arch_text_poke()wasn't allowed to touch kernel code and ftrace is in full control of it ?
ftrace is not in "full control" of nop5 and must not be. Soon we will have nop5 in the middle of the function. ftrace must not touch it.
Quoted 14 lines · Alexei Starovoitov
Quoted 13 lines · Peter Zijlstra
Quoted 11 lines · Jiri Olsa
Quoted 9 lines · Peter Zijlstra
Quoted 6 lines · Steven Rostedt
Quoted 3 lines · On Fri, 12 Aug 2022 23:18:15 +0200 Jiri Olsa <[email protected]> wrote:
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
Why ?!?
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines
there's bpf dispatcher code that updates
bpf_dispatcher_xdp_funcfunction withbpf_arch_text_pokeand that can race with ftrace update if the function is tracedI thought
bpf_arch_text_poke()wasn't allowed to touch kernel code and ftrace is in full control of it ?ftrace is not in "full control" of nop5 and must not be.
It is in full control of the 'call __fentry__'. Absolute full NAK on you trying to make it otherwise.
Quoted 2 lines
Soon we will have nop5 in the middle of the function. ftrace must not touch it.
How are you generating that NOP and what for?
Quoted 20 lines · Peter Zijlstra
Quoted 18 lines · Alexei Starovoitov
Quoted 17 lines · Peter Zijlstra
Quoted 14 lines · Jiri Olsa
Quoted 9 lines · Peter Zijlstra
Quoted 6 lines · Steven Rostedt
Quoted 3 lines · On Fri, 12 Aug 2022 23:18:15 +0200 Jiri Olsa <[email protected]> wrote:
the patch below moves the bpf function into sepatate object and switches off the -mrecord-mcount for it.. so the function gets profile call generated but it's not visible to ftrace
Why ?!?
there's bpf dispatcher code that updates
bpf_dispatcher_xdp_funcfunction withbpf_arch_text_pokeand that can race with ftrace update if the function is traced
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
I thought
bpf_arch_text_poke()wasn't allowed to touch kernel code and ftrace is in full control of it ?ftrace is not in "full control" of nop5 and must not be.
It is in full control of the 'call
__fentry__'. Absolute full NAK on you trying to make it otherwise.
Don't mix 'call fentry' generated by the compiler with nop5 inserted by macroses or JITs.
Quoted 4 lines
Quoted 2 lines
Soon we will have nop5 in the middle of the function. ftrace must not touch it.
How are you generating that NOP and what for?
We're generating nop5-s in JITed code to further
attach to. For example when one bpf prog is being replaced by another. Currently it's in the func prologue only.
In the future it will be anywhere in the body.
Quoted 7 lines · Alexei Starovoitov
Quoted 2 lines · Peter Zijlstra
It is in full control of the 'call
__fentry__'. Absolute full NAK on you trying to make it otherwise.Don't mix 'call fentry' generated by the compiler with nop5 inserted by macroses or JITs.
Looking at:
https://lore.kernel.org/all/[email protected]/
this seems to want to prod at the __fentry__ sites.
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
Soon we will have nop5 in the middle of the function. ftrace must not touch it.
How are you generating that NOP and what for?
We're generating nop5-s in JITed code to further
attach to.
Ftrace doesn't know about those; so how can it break that?
Likewise it doesn't know about the static_branch/static_call NOPs and nothing is broken.
Ftrace only knows about the __fentry__ sites, and it *does* own them. Are you saying ftrace is writing to a code location not a __fentry__ site?
Quoted 24 lines · Peter Zijlstra
Quoted 7 lines · Alexei Starovoitov
Quoted 2 lines · Peter Zijlstra
It is in full control of the 'call
__fentry__'. Absolute full NAK on you trying to make it otherwise.Don't mix 'call fentry' generated by the compiler with nop5 inserted by macroses or JITs.
Looking at:
https://lore.kernel.org/all/[email protected]/
this seems to want to prod at the
__fentry__sites.Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
Soon we will have nop5 in the middle of the function. ftrace must not touch it.
How are you generating that NOP and what for?
We're generating nop5-s in JITed code to further
attach to.
Quoted 8 lines
Ftrace doesn't know about those; so how can it break that?
Likewise it doesn't know about the static_branch/static_call NOPs and nothing is broken.
Ftrace only knows about the
__fentry__sites, and it *does* own them. Are you saying ftrace is writing to a code location not a__fentry__site?
Let's keep it in one thread:
Quoted 6 lines
It wasn't long before. Yes it landed a few months prior to the
static_callwork, but the wholestatic_callthing was in progress for a long long time.Anyway, yes it is different. But it's still very much broken. You simply cannot step on
__fentry__sites like that.
Ask yourself: should static_call patching logic go through ftrace infra ? No. Right?static_call has nothing to do with ftrace (function tracing). Same thing here. bpf dispatching logic is nothing to do with function tracing.
In this case bpf_dispatcher_xdp_func is a placeholder written C. If it was written in asm, fentry recording wouldn't have known about it. And that's more or less what Jiri patch is doing.
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.
Quoted 3 lines · Alexei Starovoitov
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.
Then make it a notrace function with a nop5 in it. That isn't hard.
The whole problem is that it isn't a notrace function and you're abusing a __fentry__ site.
Quoted 7 lines · Peter Zijlstra
Quoted 3 lines · Alexei Starovoitov
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.Then make it a notrace function with a nop5 in it. That isn't hard.
That's exactly what we're trying to do.
Jiri's patch is one way to achieve that.
What is your suggestion?
Move it from C to asm ?
Make it naked function with explicit inline asm?
What else?
Steven Rostedt
maintainerQuoted 8 lines · On Mon, 15 Aug 2022 08:35:53 -0700 Alexei Starovoitov…
Quoted 1 line
Then make it a notrace function with a nop5 in it. That isn't hard.
That's exactly what we're trying to do.
Jiri's patch is one way to achieve that.
What is your suggestion?
Move it from C to asm ?
Make it naked function with explicit inline asm?
What else?
The dispatcher is already in the kernel so it's too late to complain about it. Jiri's patch (with my extensions) will hopefully fix the breakage BPF did to ftrace.
My ask now is to be more inclusive when doing anything that deals with modification of text, or other infrastructures. This "go it alone" approach really needs to stop. Linux is an open source project and collaboration is key. I know you don't care about others use cases (as you told me in that BPF meeting last year), but any maintainer in the Linux kernel must care about the use case of others or this will all fail.
-- Steve
Quoted 17 lines · Steven Rostedt
Quoted 8 lines · On Mon, 15 Aug 2022 08:35:53 -0700 Alexei Starovoitov…
Quoted 1 line
Then make it a notrace function with a nop5 in it. That isn't hard.
That's exactly what we're trying to do.
Jiri's patch is one way to achieve that.
What is your suggestion?
Move it from C to asm ?
Make it naked function with explicit inline asm?
What else?The dispatcher is already in the kernel so it's too late to complain about it. Jiri's patch (with my extensions) will hopefully fix the breakage BPF did to ftrace.
Quoted 6 lines
My ask now is to be more inclusive when doing anything that deals with modification of text, or other infrastructures. This "go it alone" approach really needs to stop. Linux is an open source project and collaboration is key. I know you don't care about others use cases (as you told me in that BPF meeting last year), but any maintainer in the Linux kernel must care about the use case of others or this will all fail.
Please don't misrepresent. Not cool.
Steven Rostedt
maintainerQuoted 8 lines · On Mon, 15 Aug 2022 08:53:05 -0700 Alexei Starovoitov…
Quoted 6 lines
My ask now is to be more inclusive when doing anything that deals with modification of text, or other infrastructures. This "go it alone" approach really needs to stop. Linux is an open source project and collaboration is key. I know you don't care about others use cases (as you told me in that BPF meeting last year), but any maintainer in the Linux kernel must care about the use case of others or this will all fail.
Please don't misrepresent. Not cool.
Sorry. To quote exactly what you told me when you cut me off in that meeting, "I don't care about your use cases". I guess you care about others use cases, just not mine.
-- Steve
Quoted 10 lines · Alexei Starovoitov
Quoted 7 lines · Peter Zijlstra
Quoted 3 lines · Alexei Starovoitov
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.Then make it a notrace function with a nop5 in it. That isn't hard.
That's exactly what we're trying to do.
All the while claiming ftrace is broken while it is not.
Quoted 1 line
Jiri's patch is one way to achieve that.
Fairly horrible way.
Quoted 1 line
What is your suggestion?
Mailed it already.
Quoted 1 line
Move it from C to asm ?
Would be much better than proposed IMO.
Quoted 1 line
Make it naked function with explicit inline asm?
Can be made to work but is iffy because the compiler can do horrible things with placing the asm().
Quoted 18 lines · Peter Zijlstra
Quoted 10 lines · Alexei Starovoitov
Quoted 7 lines · Peter Zijlstra
Quoted 3 lines · Alexei Starovoitov
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.Then make it a notrace function with a nop5 in it. That isn't hard.
That's exactly what we're trying to do.
All the while claiming ftrace is broken while it is not.
Quoted 1 line
Jiri's patch is one way to achieve that.
Fairly horrible way.
Quoted 7 lines
Quoted 1 line
What is your suggestion?
Mailed it already.
Quoted 1 line
Move it from C to asm ?
Would be much better than proposed IMO.
nice, that would be independent of the compiler atributes and config checking.. will check on this one ;-)
thanks,
jirka
Quoted 5 lines
Quoted 2 lines
Make it naked function with explicit inline asm?
Can be made to work but is iffy because the compiler can do horrible things with placing the
asm().
Quoted 21 lines · Jiri Olsa
Quoted 20 lines · Peter Zijlstra
Quoted 10 lines · Alexei Starovoitov
Quoted 7 lines · Peter Zijlstra
Quoted 3 lines · Alexei Starovoitov
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.Then make it a notrace function with a nop5 in it. That isn't hard.
That's exactly what we're trying to do.
All the while claiming ftrace is broken while it is not.
Quoted 1 line
Jiri's patch is one way to achieve that.
Fairly horrible way.
Quoted 1 line
What is your suggestion?
Quoted 8 lines
Quoted 5 lines
Mailed it already.
Quoted 1 line
Move it from C to asm ?
Would be much better than proposed IMO.
nice, that would be independent of the compiler atributes and config checking.. will check on this one ;-)
how about something like below?
dispatcher code is generated only for x86_64, so that will be covered by the assembly version (free of ftrace table) other archs stay same
jirka
----
arch/x86/net/Makefile +1 -0
@@ -7,4 +7,5 @@ ifeq ($(CONFIG_X86_32),y) obj-$(CONFIG_BPF_JIT) += bpf_jit_comp32.o else obj-$(CONFIG_BPF_JIT) += bpf_jit_comp.o + obj-$(CONFIG_BPF_JIT) += bpf_dispatcher.o endif
arch/x86/net/bpf_dispatcher.S new file +10 -0
@@ -0,0 +1,10 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#include <linux/linkage.h> +#include <asm/nops.h> +#include <asm/nospec-branch.h> + + .text +SYM_FUNC_START(bpf_dispatcher_xdp_func) + ASM_NOP5 + JMP_NOSPEC rdx +SYM_FUNC_END(bpf_dispatcher_xdp_func)
include/linux/bpf.h +1 -1
@@ -924,7 +924,7 @@ int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs); }#define DEFINE_BPF_DISPATCHER (name) \ - noinline __nocfi unsigned int bpf_dispatcher_##name##_func( \ + noinline __nocfi unsigned int __weak bpf_dispatcher_##name##_func(\ const void *ctx, \ const struct bpf_insn *insnsi, \ bpf_func_t bpf_func) \
Quoted 57 lines · Jiri Olsa
Quoted 21 lines · Jiri Olsa
Quoted 20 lines · Peter Zijlstra
Quoted 10 lines · Alexei Starovoitov
Quoted 7 lines · Peter Zijlstra
Quoted 3 lines · Alexei Starovoitov
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.Then make it a notrace function with a nop5 in it. That isn't hard.
That's exactly what we're trying to do.
All the while claiming ftrace is broken while it is not.
Quoted 1 line
Jiri's patch is one way to achieve that.
Fairly horrible way.
Quoted 1 line
What is your suggestion?
Quoted 8 lines
Quoted 5 lines
Mailed it already.
Quoted 1 line
Move it from C to asm ?
Would be much better than proposed IMO.
nice, that would be independent of the compiler atributes and config checking.. will check on this one ;-)
how about something like below?
dispatcher code is generated only for
x86_64, so that will be covered by the assembly version (free of ftrace table) other archs stay samejirka
----
arch/x86/net/Makefile +1
-0@@ -7,4 +7,5 @@ ifeq ($(CONFIG_X86_32),y) obj-$(CONFIG_BPF_JIT) += bpf_jit_comp32.o else obj-$(CONFIG_BPF_JIT) += bpf_jit_comp.o + obj-$(CONFIG_BPF_JIT) += bpf_dispatcher.o endifarch/x86/net/bpf_dispatcher.S new file +0
-0
arch/x86/net/bpf_dispatcher.S line 10
Quoted 10 lines
arch/x86/net/bpf_dispatcher.S new file +10
-0@@ -0,0 +1,10 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#include <linux/linkage.h> +#include <asm/nops.h> +#include <asm/nospec-branch.h> + + .text +SYM_FUNC_START(bpf_dispatcher_xdp_func) + ASM_NOP5 + JMP_NOSPEC rdx +SYM_FUNC_END(bpf_dispatcher_xdp_func)
Wait. Why asm ? Did you try Peter's suggestion:
__attribute__((__no_instrument_function__))
__attribute__((patchable_function_entry(5)))
?
Quoted 64 lines · Alexei Starovoitov
Quoted 63 lines · Jiri Olsa
Quoted 21 lines · Jiri Olsa
Quoted 20 lines · Peter Zijlstra
Quoted 10 lines · Alexei Starovoitov
Quoted 7 lines · Peter Zijlstra
Quoted 3 lines · Alexei Starovoitov
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.Then make it a notrace function with a nop5 in it. That isn't hard.
That's exactly what we're trying to do.
All the while claiming ftrace is broken while it is not.
Quoted 1 line
Jiri's patch is one way to achieve that.
Fairly horrible way.
Quoted 1 line
What is your suggestion?
Quoted 8 lines
Quoted 5 lines
Mailed it already.
Quoted 1 line
Move it from C to asm ?
Would be much better than proposed IMO.
nice, that would be independent of the compiler atributes and config checking.. will check on this one ;-)
how about something like below?
dispatcher code is generated only for
x86_64, so that will be covered by the assembly version (free of ftrace table) other archs stay samejirka
----
arch/x86/net/Makefile +1
-0@@ -7,4 +7,5 @@ ifeq ($(CONFIG_X86_32),y) obj-$(CONFIG_BPF_JIT) += bpf_jit_comp32.o else obj-$(CONFIG_BPF_JIT) += bpf_jit_comp.o + obj-$(CONFIG_BPF_JIT) += bpf_dispatcher.o endifarch/x86/net/bpf_dispatcher.S new file +6
-0@@ -0,0 +1,10 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#include <linux/linkage.h> +#include <asm/nops.h> +#include <asm/nospec-branch.h> + + .text
Quoted 8 lines
arch/x86/net/bpf_dispatcher.S line 10
Quoted 4 lines
arch/x86/net/bpf_dispatcher.S new file +4
-0+SYM_FUNC_START(bpf_dispatcher_xdp_func) + ASM_NOP5 + JMP_NOSPEC rdx +SYM_FUNC_END(bpf_dispatcher_xdp_func)Wait. Why asm ? Did you try Peter's suggestion:
__attribute__((__no_instrument_function__))
__attribute__((patchable_function_entry(5)))
ah so this suggestion came in the other thread after the asm one.. ok, will check
jirka
Quoted 9 lines · Peter Zijlstra
Quoted 3 lines · Alexei Starovoitov
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.Then make it a notrace function with a nop5 in it. That isn't hard.
The whole problem is that it isn't a notrace function and you're abusing a
__fentry__site.
foo.c:
__attribute__((__no_instrument_function__))
__attribute__((patchable_function_entry(5)))
void my_func(void)
{
}
void my_foo(void)
{
}
gcc -c foo.c -pg -mfentry -mcmodel=kernel -fno-PIE -O2
foo.o: file format elf64-x86-64
Disassembly of section .text:
0000000000000000 <my_func>:
0: f3 0f 1e fa endbr64 4: 90 nop 5: 90 nop 6: 90 nop 7: 90 nop 8: 90 nop 9: c3 ret a: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1)
0000000000000010 <my_foo>:
10: f3 0f 1e fa endbr64 14: e8 00 00 00 00 call 19 <my_foo+0x9> 15: R_X86_64_PLT32 __fentry__-0x4 19: c3 ret
Quoted 7 lines · Peter Zijlstra
Quoted 5 lines · Peter Zijlstra
Quoted 3 lines · Alexei Starovoitov
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.
Quoted 6 lines
Quoted 4 lines
Then make it a notrace function with a nop5 in it. That isn't hard.
The whole problem is that it isn't a notrace function and you're abusing a
__fentry__site.
Brand new stuff.
Awesome. That should fit perfectly.
Quoted 4 lines
foo.c:__attribute__((__no_instrument_function__))
__attribute__((patchable_function_entry(5)))
Interesting. Didn't know about this attribute.
Quoted 7 lines
void
my_func(void)
{
}void
my_foo(void)
{
}
Great.
Jiri, could you please revise your patch with this approach?
Steven Rostedt
maintainerQuoted 19 lines · On Mon, 15 Aug 2022 08:49:11 -0700 Alexei Starovoitov…
Quoted 4 lines
Quoted 2 lines
The whole problem is that it isn't a notrace function and you're abusing a
__fentry__site.Brand new stuff.
Awesome. That should fit perfectly.Quoted 4 lines
foo.c:__attribute__((__no_instrument_function__))
__attribute__((patchable_function_entry(5)))Interesting. Didn't know about this attribute.
Quoted 4 lines
void
my_func(void)
{
}
Quoted 6 lines
Quoted 3 lines
void
my_foo(void)
{
}Great.
Jiri, could you please revise your patch with this approach?
This is the exact result I was looking for. Something we can all agree to.
The point being, include others when developing code that is similar to what other subsystems do. On the code modification front, please Cc the ftrace, kprobe, static_call and jump_label maintainers, as we like to work together. The BPF dispatcher modifications should be no different. There's a lot of experience in this field throughout the kernel. Please utilize it. If it wasn't for this thread, we would never had found out about this easy solution.
-- Steve
Quoted 42 lines · Peter Zijlstra
Quoted 9 lines · Peter Zijlstra
Quoted 3 lines · Alexei Starovoitov
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.Then make it a notrace function with a nop5 in it. That isn't hard.
The whole problem is that it isn't a notrace function and you're abusing a
__fentry__site.
foo.c:__attribute__((__no_instrument_function__))
__attribute__((patchable_function_entry(5)))
voidmy_func(void)
{
}void
my_foo(void)
{
}gcc -c
foo.c-pg -mfentry -mcmodel=kernel -fno-PIE -O2foo.o: file format elf64-x86-64
Disassembly of section .text:
0000000000000000 <
my_func>:0: f3 0f 1e fa endbr64 4: 90 nop 5: 90 nop 6: 90 nop 7: 90 nop 8: 90 nop 9: c3 ret a: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1)
Quoted 5 lines
0000000000000010 <
my_foo>:10: f3 0f 1e fa endbr64 14: e8 00 00 00 00 call 19 <my_foo+0x9> 15: R_X86_64_PLT32 __fentry__-0x4 19: c3 ret
ok, so the problem with __attribute__((patchable_function_entry(5))) is that it puts function address into __patchable_function_entries section, which is one of ftrace locations source:
#define MCOUNT_REC() . = ALIGN(8); \
__start_mcount_loc = .; \
KEEP(*(__mcount_loc)) \
KEEP(*(__patchable_function_entries)) \
__stop_mcount_loc = .; \...
it looks like __patchable_function_entries is used for other than x86 archs, so we perhaps we could have x86 specific MCOUNT_REC macro just with __mcount_loc section?
jirka
Steven Rostedt
maintainerQuoted 9 lines · On Thu, 18 Aug 2022 22:27:07 +0200 Jiri Olsa <[email protected]> wrote:
ok, so the problem with __attribute__((patchable_function_entry(5))) is that it puts function address into
__patchable_function_entriessection, which is one of ftrace locations source:#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ KEEP(*(__patchable_function_entries)) \ __stop_mcount_loc = .; \
Quoted 6 lines
...
it looks like
__patchable_function_entriesis used for other than x86 archs, so we perhaps we could have x86 specificMCOUNT_RECmacro just with__mcount_locsection?
So something like this:
#ifdef CONFIG_X86
# define NON_MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) # define MCOUNT_PATCHABLE #else # define NON_MCOUNT_PATCHABLE # define MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) #endif
#define MCOUNT_REC() . = ALIGN(8); \
__start_mcount_loc = .; \
KEEP(*(__mcount_loc)) \
MCOUNT_PATCHABLE \
__stop_mcount_loc = .; \
NON_MCOUNT_PATCHABLE \...
??
-- Steve
Quoted 36 lines · Steven Rostedt
Quoted 9 lines · On Thu, 18 Aug 2022 22:27:07 +0200 Jiri Olsa <[email protected]> wrote:
ok, so the problem with __attribute__((patchable_function_entry(5))) is that it puts function address into
__patchable_function_entriessection, which is one of ftrace locations source:#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ KEEP(*(__patchable_function_entries)) \ __stop_mcount_loc = .; \Quoted 6 lines
...
it looks like
__patchable_function_entriesis used for other than x86 archs, so we perhaps we could have x86 specificMCOUNT_RECmacro just with__mcount_locsection?So something like this:
#ifdef
CONFIG_X86# define NON_MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) # define MCOUNT_PATCHABLE #else # define NON_MCOUNT_PATCHABLE # define MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) #endif#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ MCOUNT_PATCHABLE \ __stop_mcount_loc = .; \ NON_MCOUNT_PATCHABLE \
Quoted 3 lines
...
??
That's what more or less Peter's patch is doing:
Here it is again for reference:
https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/commit/?id=8d075bdf11193f1d276bf19fa56b4b8dfe24df9e
Steven Rostedt
maintainerQuoted 16 lines · On Thu, 18 Aug 2022 14:00:21 -0700 Alexei Starovoitov…
Quoted 16 lines
#ifdef
CONFIG_X86# define NON_MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) # define MCOUNT_PATCHABLE #else # define NON_MCOUNT_PATCHABLE # define MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) #endif#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ MCOUNT_PATCHABLE \ __stop_mcount_loc = .; \ NON_MCOUNT_PATCHABLE \...
Quoted 3 lines
Quoted 1 line
??
That's what more or less Peter's patch is doing:
Heh.
Quoted 2 lines
Here it is again for reference:
https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/commit/?id=8d075bdf11193f1d276bf19fa56b4b8dfe24df9e
Thanks, I missed seeing this.
-- Steve
Quoted 39 lines · Alexei Starovoitov
Quoted 38 lines · Steven Rostedt
Quoted 9 lines · On Thu, 18 Aug 2022 22:27:07 +0200 Jiri Olsa <[email protected]> wrote:
ok, so the problem with __attribute__((patchable_function_entry(5))) is that it puts function address into
__patchable_function_entriessection, which is one of ftrace locations source:#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ KEEP(*(__patchable_function_entries)) \ __stop_mcount_loc = .; \Quoted 6 lines
...
it looks like
__patchable_function_entriesis used for other than x86 archs, so we perhaps we could have x86 specificMCOUNT_RECmacro just with__mcount_locsection?So something like this:
#ifdef
CONFIG_X86# define NON_MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) # define MCOUNT_PATCHABLE #else # define NON_MCOUNT_PATCHABLE # define MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) #endif#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ MCOUNT_PATCHABLE \ __stop_mcount_loc = .; \ NON_MCOUNT_PATCHABLE \...
Quoted 5 lines
Quoted 1 line
??
That's what more or less Peter's patch is doing:
Here it is again for reference:
https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/commit/?id=8d075bdf11193f1d276bf19fa56b4b8dfe24df9e
ah nice, and discards the __patchable_function_entries section, great
jirka
Quoted 40 lines · Jiri Olsa
Quoted 39 lines · Alexei Starovoitov
Quoted 38 lines · Steven Rostedt
Quoted 9 lines · On Thu, 18 Aug 2022 22:27:07 +0200 Jiri Olsa <[email protected]> wrote:
ok, so the problem with __attribute__((patchable_function_entry(5))) is that it puts function address into
__patchable_function_entriessection, which is one of ftrace locations source:#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ KEEP(*(__patchable_function_entries)) \ __stop_mcount_loc = .; \Quoted 6 lines
...
it looks like
__patchable_function_entriesis used for other than x86 archs, so we perhaps we could have x86 specificMCOUNT_RECmacro just with__mcount_locsection?So something like this:
#ifdef
CONFIG_X86# define NON_MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) # define MCOUNT_PATCHABLE #else # define NON_MCOUNT_PATCHABLE # define MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) #endif#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ MCOUNT_PATCHABLE \ __stop_mcount_loc = .; \ NON_MCOUNT_PATCHABLE \...
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
??
That's what more or less Peter's patch is doing:
Here it is again for reference:
https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/commit/?id=8d075bdf11193f1d276bf19fa56b4b8dfe24df9eah nice, and discards the
__patchable_function_entriessection, great
tested change below with Peter's change above and it seems to work, once it get merged I'll send full patch
thanks,
jirka
---
include/linux/bpf.h +2 -0
@@ -924,6 +924,8 @@ int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs); }#define DEFINE_BPF_DISPATCHER (name) \ + __attribute__((__no_instrument_function__)) \ + __attribute__((patchable_function_entry(5))) \ noinline __nocfi unsigned int bpf_dispatcher_##name##_func( \ const void *ctx, \ const struct bpf_insn *insnsi, \
Quoted 48 lines · Jiri Olsa
Quoted 46 lines · Jiri Olsa
Quoted 44 lines · Alexei Starovoitov
Quoted 36 lines · Steven Rostedt
Quoted 9 lines · On Thu, 18 Aug 2022 22:27:07 +0200 Jiri Olsa <[email protected]> wrote:
ok, so the problem with __attribute__((patchable_function_entry(5))) is that it puts function address into
__patchable_function_entriessection, which is one of ftrace locations source:#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ KEEP(*(__patchable_function_entries)) \ __stop_mcount_loc = .; \Quoted 6 lines
...
it looks like
__patchable_function_entriesis used for other than x86 archs, so we perhaps we could have x86 specificMCOUNT_RECmacro just with__mcount_locsection?So something like this:
#ifdef
CONFIG_X86# define NON_MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) # define MCOUNT_PATCHABLE #else # define NON_MCOUNT_PATCHABLE # define MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) #endif#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ MCOUNT_PATCHABLE \ __stop_mcount_loc = .; \ NON_MCOUNT_PATCHABLE \Quoted 3 lines
...
??
That's what more or less Peter's patch is doing:
Here it is again for reference:
https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/commit/?id=8d075bdf11193f1d276bf19fa56b4b8dfe24df9e
Quoted 5 lines
Quoted 2 lines
ah nice, and discards the
__patchable_function_entriessection, greattested change below with Peter's change above and it seems to work, once it get merged I'll send full patch
Peter,
what is the ETA to land your changes?
That particular commit is detached in your git tree. Did you move it to a different branch?
Just trying to figure out the logistics to land Jiri's fix below. We can take it into bpf-next, since it's harmless as-is, but it won't have an effect until your change lands. Sounds like they both will get in during the next merge window?
Quoted 18 lines
thanks,
jirka---
include/linux/bpf.h +2
-0@@ -924,6 +924,8 @@ int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs); }#define DEFINE_BPF_DISPATCHER (name) \ + __attribute__((__no_instrument_function__)) \ + __attribute__((patchable_function_entry(5))) \ noinline __nocfi unsigned int bpf_dispatcher_##name##_func( \ const void *ctx, \ const struct bpf_insn *insnsi, \
Quoted 60 lines · Alexei Starovoitov
Quoted 48 lines · Jiri Olsa
Quoted 46 lines · Jiri Olsa
Quoted 44 lines · Alexei Starovoitov
Quoted 36 lines · Steven Rostedt
Quoted 9 lines · On Thu, 18 Aug 2022 22:27:07 +0200 Jiri Olsa <[email protected]> wrote:
ok, so the problem with __attribute__((patchable_function_entry(5))) is that it puts function address into
__patchable_function_entriessection, which is one of ftrace locations source:#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ KEEP(*(__patchable_function_entries)) \ __stop_mcount_loc = .; \Quoted 6 lines
...
it looks like
__patchable_function_entriesis used for other than x86 archs, so we perhaps we could have x86 specificMCOUNT_RECmacro just with__mcount_locsection?So something like this:
#ifdef
CONFIG_X86# define NON_MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) # define MCOUNT_PATCHABLE #else # define NON_MCOUNT_PATCHABLE # define MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) #endif#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ MCOUNT_PATCHABLE \ __stop_mcount_loc = .; \ NON_MCOUNT_PATCHABLE \Quoted 3 lines
...
??
That's what more or less Peter's patch is doing:
Here it is again for reference:
https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/commit/?id=8d075bdf11193f1d276bf19fa56b4b8dfe24df9eQuoted 5 lines
Quoted 2 lines
ah nice, and discards the
__patchable_function_entriessection, greattested change below with Peter's change above and it seems to work, once it get merged I'll send full patch
Peter,
what is the ETA to land your changes?
That particular commit is detached in your git tree. Did you move it to a different branch?
Quoted 4 lines
Just trying to figure out the logistics to land Jiri's fix below. We can take it into bpf-next, since it's harmless as-is, but it won't have an effect until your change lands. Sounds like they both will get in during the next merge window?
I discussed with Peter and I'll send his change together with my fix
jirka
Quoted 29 lines · Steven Rostedt
Quoted 9 lines · On Thu, 18 Aug 2022 22:27:07 +0200 Jiri Olsa <[email protected]> wrote:
ok, so the problem with __attribute__((patchable_function_entry(5))) is that it puts function address into
__patchable_function_entriessection, which is one of ftrace locations source:#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ KEEP(*(__patchable_function_entries)) \ __stop_mcount_loc = .; \Quoted 6 lines
...
it looks like
__patchable_function_entriesis used for other than x86 archs, so we perhaps we could have x86 specificMCOUNT_RECmacro just with__mcount_locsection?So something like this:
#ifdef
CONFIG_X86# define NON_MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) # define MCOUNT_PATCHABLE #else # define NON_MCOUNT_PATCHABLE # define MCOUNT_PATCHABLE KEEP(*(__patchable_function_entries)) #endif
Quoted 8 lines
#define MCOUNT_REC() . = ALIGN(8); \ __start_mcount_loc = .; \ KEEP(*(__mcount_loc)) \ MCOUNT_PATCHABLE \ __stop_mcount_loc = .; \ NON_MCOUNT_PATCHABLE \...
is there a reason to keep NON_MCOUNT_PATCHABLE section for x86? otherwise LGTM
jirka
Steven Rostedt
maintainerQuoted 2 lines · On Mon, 15 Aug 2022 08:17:42 -0700 Alexei Starovoitov…
Ask yourself: should
static_callpatching logic go through ftrace infra ? No. Right?
I agree that static_call (and jump_labels) are not part of the ftrace infrastructure (but ftrace was a strong motivator for those).
Quoted 1 line
static_callhas nothing to do with ftrace (function tracing).
Besides the motivation, I agree.
Quoted 2 lines
Same thing here. bpf dispatching logic is nothing to do with function tracing.
But it used fentry, which is part of function tracing. Which is what I'm against. And why it broke ftrace.
Quoted 2 lines
In this case
bpf_dispatcher_xdp_funcis a placeholder written C. If it was written in asm, fentry recording wouldn't have known about it.
And I would not have had an issue with that approach (for ftrace that is). But that brings up other concerns (see below).
Quoted 4 lines
And that's more or less what Jiri patch is doing.
It's hiding a fake function from ftrace, since it's not a function and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.
On the ftrace side, I'm perfectly happy with Jiri's approach (the one I help extend).
But dynamic code modification is something we need to take very seriously. It's very similar to writing your own locking primitives (which Linus always says "Don't do"). It's complex and easy to get wrong. The more dynamic code modifications we have, the less secure the kernel is.
Here's the list of dynamic code modification infrastructures:
ftrace
kprobes
jump_labels
static_calls
We now have the bpf dispatcher.
The ftrace, kprobes, jump_labels and static_calls developers work together to make sure that we are all in line, not breaking anything, and try to consolidate when possible. We also review each others code.
The issue I have is that BPF is largely doing it alone, and not communicating with the others. This gives me cause for concern on both a robustness and security point of view.
-- Steve
Quoted 11 lines · Steven Rostedt
Quoted 6 lines · On Fri, 22 Jul 2022 09:04:29 -0700 Alexei Starovoitov…
ftrace must not peek into bpf specific functions. Currently ftrace is causing the kernel to crash. What Jiri is proposing is to fix ftrace bug. And you're saying nack? let ftrace be broken ?
If you don't like Jiri's approach please propose something else.
So, why not mark it as notrace? That will prevent ftrace from looking at it.
there's still needs to be the instrument jump generated in order to use bpf_arch_text_poke on that
jirka
Steven Rostedt
maintainerQuoted 4 lines · On Fri, 22 Jul 2022 18:26:09 +0200 Jiri Olsa <[email protected]> wrote:
Quoted 1 line
So, why not mark it as notrace? That will prevent ftrace from looking at it.
there's still needs to be the instrument jump generated in order to use
bpf_arch_text_pokeon that
Can you explain the background on this. Why is bpf doing text_poke on function entries? What was the direct use case for?
-- Steve
Quoted 11 lines · Steven Rostedt
Quoted 4 lines · On Fri, 22 Jul 2022 18:26:09 +0200 Jiri Olsa <[email protected]> wrote:
Quoted 1 line
So, why not mark it as notrace? That will prevent ftrace from looking at it.
there's still needs to be the instrument jump generated in order to use
bpf_arch_text_pokeon thatCan you explain the background on this. Why is bpf doing
text_pokeon function entries? What was the direct use case for?
It's a bpf specific dispatcher.
Like static_call but more dynamic and for bpf progs.