diff mbox series

[2/3] bpf: Optimize get_modules_for_addrs()

Message ID 20221230112729.351-3-thunder.leizhen@huawei.com (mailing list archive)
State Changes Requested
Delegated to: BPF
Headers show
Series kallsyms: Optimize the search for module symbols by livepatch and bpf | expand

Checks

Context Check Description
netdev/tree_selection success Guessed tree name to be net-next
netdev/fixes_present success Fixes tag not required for -next series
netdev/subject_prefix success Link
netdev/cover_letter success Series has a cover letter
netdev/patch_count success Link
netdev/header_inline success No static functions without inline keyword in header files
netdev/build_32bit success Errors and warnings before: 15 this patch: 15
netdev/cc_maintainers success CCed 15 of 15 maintainers
netdev/build_clang success Errors and warnings before: 1 this patch: 1
netdev/module_param success Was 0 now: 0
netdev/verify_signedoff success Signed-off-by tag matches author and committer
netdev/check_selftest success No net selftest shell script
netdev/verify_fixes success No Fixes tag
netdev/build_allmodconfig_warn success Errors and warnings before: 15 this patch: 15
netdev/checkpatch warning WARNING: line length of 93 exceeds 80 columns WARNING: line length of 96 exceeds 80 columns
netdev/kdoc success Errors and warnings before: 0 this patch: 0
netdev/source_inline success Was 0 now: 0
bpf/vmtest-bpf-next-PR success PR summary
bpf/vmtest-bpf-next-VM_Test-2 success Logs for build for aarch64 with gcc
bpf/vmtest-bpf-next-VM_Test-3 success Logs for build for aarch64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-4 success Logs for build for s390x with gcc
bpf/vmtest-bpf-next-VM_Test-5 success Logs for build for x86_64 with gcc
bpf/vmtest-bpf-next-VM_Test-6 success Logs for build for x86_64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-9 success Logs for test_maps on aarch64 with gcc
bpf/vmtest-bpf-next-VM_Test-10 success Logs for test_maps on aarch64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-11 success Logs for test_maps on s390x with gcc
bpf/vmtest-bpf-next-VM_Test-12 success Logs for test_maps on x86_64 with gcc
bpf/vmtest-bpf-next-VM_Test-13 success Logs for test_maps on x86_64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-15 success Logs for test_progs on aarch64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-16 success Logs for test_progs on s390x with gcc
bpf/vmtest-bpf-next-VM_Test-17 success Logs for test_progs on x86_64 with gcc
bpf/vmtest-bpf-next-VM_Test-18 success Logs for test_progs on x86_64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-20 success Logs for test_progs_no_alu32 on aarch64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-21 success Logs for test_progs_no_alu32 on s390x with gcc
bpf/vmtest-bpf-next-VM_Test-22 success Logs for test_progs_no_alu32 on x86_64 with gcc
bpf/vmtest-bpf-next-VM_Test-23 success Logs for test_progs_no_alu32 on x86_64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-24 success Logs for test_progs_no_alu32_parallel on aarch64 with gcc
bpf/vmtest-bpf-next-VM_Test-25 success Logs for test_progs_no_alu32_parallel on aarch64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-26 success Logs for test_progs_no_alu32_parallel on s390x with gcc
bpf/vmtest-bpf-next-VM_Test-27 success Logs for test_progs_no_alu32_parallel on x86_64 with gcc
bpf/vmtest-bpf-next-VM_Test-28 success Logs for test_progs_no_alu32_parallel on x86_64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-29 success Logs for test_progs_parallel on aarch64 with gcc
bpf/vmtest-bpf-next-VM_Test-30 success Logs for test_progs_parallel on aarch64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-31 success Logs for test_progs_parallel on s390x with gcc
bpf/vmtest-bpf-next-VM_Test-32 success Logs for test_progs_parallel on x86_64 with gcc
bpf/vmtest-bpf-next-VM_Test-33 success Logs for test_progs_parallel on x86_64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-34 success Logs for test_verifier on aarch64 with gcc
bpf/vmtest-bpf-next-VM_Test-35 success Logs for test_verifier on aarch64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-36 success Logs for test_verifier on s390x with gcc
bpf/vmtest-bpf-next-VM_Test-37 success Logs for test_verifier on x86_64 with gcc
bpf/vmtest-bpf-next-VM_Test-38 success Logs for test_verifier on x86_64 with llvm-16
bpf/vmtest-bpf-next-VM_Test-14 success Logs for test_progs on aarch64 with gcc
bpf/vmtest-bpf-next-VM_Test-19 success Logs for test_progs_no_alu32 on aarch64 with gcc
bpf/vmtest-bpf-next-VM_Test-1 success Logs for ShellCheck
bpf/vmtest-bpf-next-VM_Test-7 success Logs for llvm-toolchain
bpf/vmtest-bpf-next-VM_Test-8 success Logs for set-matrix

Commit Message

Zhen Lei Dec. 30, 2022, 11:27 a.m. UTC
Function __module_address() can quickly return the pointer of the module
to which an address belongs. We do not need to traverse the symbols of all
modules to check whether each address in addrs[] is the start address of
the corresponding symbol, because register_fprobe_ips() will do this check
later.

Assuming that there are m modules, each module has n symbols on average,
and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
and the time complexity of current method is O(K * (log(m) + M)), M <= m.
(m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
the ratio is still greater than 1. Therefore, the new method will
generally have better performance.

Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
---
 kernel/trace/bpf_trace.c | 101 ++++++++++++++++-----------------------
 1 file changed, 40 insertions(+), 61 deletions(-)

Comments

Petr Mladek Jan. 4, 2023, 4:25 p.m. UTC | #1
On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
> Function __module_address() can quickly return the pointer of the module
> to which an address belongs. We do not need to traverse the symbols of all
> modules to check whether each address in addrs[] is the start address of
> the corresponding symbol, because register_fprobe_ips() will do this check
> later.
> 
> Assuming that there are m modules, each module has n symbols on average,
> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
> the ratio is still greater than 1. Therefore, the new method will
> generally have better performance.
> 
> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
> ---
>  kernel/trace/bpf_trace.c | 101 ++++++++++++++++-----------------------
>  1 file changed, 40 insertions(+), 61 deletions(-)
> 
> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> index 5f3be4bc16403a5..0ff9037098bd241 100644
> --- a/kernel/trace/bpf_trace.c
> +++ b/kernel/trace/bpf_trace.c
> @@ -2684,69 +2684,55 @@ static void symbols_swap_r(void *a, void *b, int size, const void *priv)
>  	}
>  }
>  
> -struct module_addr_args {
> -	unsigned long *addrs;
> -	u32 addrs_cnt;
> -	struct module **mods;
> -	int mods_cnt;
> -	int mods_cap;
> -};
> -
> -static int module_callback(void *data, const char *name,
> -			   struct module *mod, unsigned long addr)
> +static int get_modules_for_addrs(struct module ***out_mods, unsigned long *addrs, u32 addrs_cnt)
>  {
> -	struct module_addr_args *args = data;
> -	struct module **mods;
> -
> -	/* We iterate all modules symbols and for each we:
> -	 * - search for it in provided addresses array
> -	 * - if found we check if we already have the module pointer stored
> -	 *   (we iterate modules sequentially, so we can check just the last
> -	 *   module pointer)
> -	 * - take module reference and store it
> -	 */
> -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
> -		       bpf_kprobe_multi_addrs_cmp))
> -		return 0;
> +	int i, j, err;
> +	int mods_cnt = 0;
> +	int mods_cap = 0;
> +	struct module *mod;
> +	struct module **mods = NULL;
>  
> -	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
> -		return 0;
> +	for (i = 0; i < addrs_cnt; i++) {
> +		mod = __module_address(addrs[i]);

This must be called under module_mutex to make sure that the module
would not disappear.

> +		if (!mod)
> +			continue;
>  
> -	if (args->mods_cnt == args->mods_cap) {
> -		args->mods_cap = max(16, args->mods_cap * 3 / 2);
> -		mods = krealloc_array(args->mods, args->mods_cap, sizeof(*mods), GFP_KERNEL);
> -		if (!mods)
> -			return -ENOMEM;
> -		args->mods = mods;
> -	}
> +		/* check if we already have the module pointer stored */
> +		for (j = 0; j < mods_cnt; j++) {
> +			if (mods[j] == mod)
> +				break;
> +		}

This might get optimized like the original code.

My understanding is that the addresses are sorted in "addrs" array.
So, the address is either part of the last found module or it belongs
to a completely new module.

	for (i = 0; i < addrs_cnt; i++) {
		/*
		 * The adresses are sorted. The adress either belongs
		 * to the last found module or a new one.
		 *
		 * This is safe because we already have reference
		 * on the found modules.
		 */
		 if (mods_cnt && within_module(addrs[i], mods[mods_cnt - 1]))
			continue;

		mutex_lock(&module_mutex);
		mod = __module_address(addrs[i]);
		if (mod && !try_module_get(mod)) {
			mutex_unlock(&module_mutex);
			goto failed;
		}
		mutex_unlock(&module_mutex);

		/*
		 * Nope when the address was not from a module.
		 *
		 * Is this correct? What if the module has gone in
		 * the meantime? Anyway, the original code
		 * worked this way.
		 *
		 * FIXME: I would personally make sure that it is part
		 * of vmlinux or so.
		 */
		if (!mod)
			continue;

		/* store the module into mods array */
		...




> +		if (j < mods_cnt)
> +			continue;
>  
> -	if (!try_module_get(mod))
> -		return -EINVAL;
> +		if (mods_cnt == mods_cap) {
> +			struct module **new_mods;
>  
> -	args->mods[args->mods_cnt] = mod;
> -	args->mods_cnt++;
> -	return 0;
> -}
> +			mods_cap = max(16, mods_cap * 3 / 2);
> +			new_mods = krealloc_array(mods, mods_cap, sizeof(*mods), GFP_KERNEL);
> +			if (!new_mods) {
> +				err = -ENOMEM;
> +				goto failed;
> +			}
> +			mods = new_mods;
> +		}
>  
> -static int get_modules_for_addrs(struct module ***mods, unsigned long *addrs, u32 addrs_cnt)
> -{
> -	struct module_addr_args args = {
> -		.addrs     = addrs,
> -		.addrs_cnt = addrs_cnt,
> -	};
> -	int err;
> +		if (!try_module_get(mod)) {
> +			err = -EINVAL;
> +			goto failed;
> +		}
>  
> -	/* We return either err < 0 in case of error, ... */
> -	err = module_kallsyms_on_each_symbol(NULL, module_callback, &args);
> -	if (err) {
> -		kprobe_multi_put_modules(args.mods, args.mods_cnt);
> -		kfree(args.mods);
> -		return err;
> +		mods[mods_cnt] = mod;
> +		mods_cnt++;
>  	}
>  
> -	/* or number of modules found if everything is ok. */
> -	*mods = args.mods;
> -	return args.mods_cnt;
> +	*out_mods = mods;
> +	return mods_cnt;
> +
> +failed:
> +	kprobe_multi_put_modules(mods, mods_cnt);
> +	kfree(mods);
> +	return err;
>  }
>  
>  int bpf_kprobe_multi_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)

Otherwise, it looks good. IMHO, the new code looks more straightforward
than the original one.

Best Regards,
Petr
Song Liu Jan. 4, 2023, 5:07 p.m. UTC | #2
On Wed, Jan 4, 2023 at 8:25 AM Petr Mladek <pmladek@suse.com> wrote:
>
> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
> > Function __module_address() can quickly return the pointer of the module
> > to which an address belongs. We do not need to traverse the symbols of all
> > modules to check whether each address in addrs[] is the start address of
> > the corresponding symbol, because register_fprobe_ips() will do this check
> > later.
> >
> > Assuming that there are m modules, each module has n symbols on average,
> > and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
> > complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
> > and the time complexity of current method is O(K * (log(m) + M)), M <= m.
> > (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
> > the ratio is still greater than 1. Therefore, the new method will
> > generally have better performance.
> >
> > Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
> > ---
> >  kernel/trace/bpf_trace.c | 101 ++++++++++++++++-----------------------
> >  1 file changed, 40 insertions(+), 61 deletions(-)
> >
> > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > index 5f3be4bc16403a5..0ff9037098bd241 100644
> > --- a/kernel/trace/bpf_trace.c
> > +++ b/kernel/trace/bpf_trace.c
> > @@ -2684,69 +2684,55 @@ static void symbols_swap_r(void *a, void *b, int size, const void *priv)
> >       }
> >  }
> >
> > -struct module_addr_args {
> > -     unsigned long *addrs;
> > -     u32 addrs_cnt;
> > -     struct module **mods;
> > -     int mods_cnt;
> > -     int mods_cap;
> > -};
> > -
> > -static int module_callback(void *data, const char *name,
> > -                        struct module *mod, unsigned long addr)
> > +static int get_modules_for_addrs(struct module ***out_mods, unsigned long *addrs, u32 addrs_cnt)
> >  {
> > -     struct module_addr_args *args = data;
> > -     struct module **mods;
> > -
> > -     /* We iterate all modules symbols and for each we:
> > -      * - search for it in provided addresses array
> > -      * - if found we check if we already have the module pointer stored
> > -      *   (we iterate modules sequentially, so we can check just the last
> > -      *   module pointer)
> > -      * - take module reference and store it
> > -      */
> > -     if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
> > -                    bpf_kprobe_multi_addrs_cmp))
> > -             return 0;
> > +     int i, j, err;
> > +     int mods_cnt = 0;
> > +     int mods_cap = 0;
> > +     struct module *mod;
> > +     struct module **mods = NULL;
> >
> > -     if (args->mods && args->mods[args->mods_cnt - 1] == mod)
> > -             return 0;
> > +     for (i = 0; i < addrs_cnt; i++) {
> > +             mod = __module_address(addrs[i]);
>
> This must be called under module_mutex to make sure that the module
> would not disappear.

module_mutex is not available outside kernel/module/. The common
practice is to disable preempt before calling __module_address().
CONFIG_LOCKDEP should catch this.

Thanks,
Song

[...]
Zhen Lei Jan. 5, 2023, 7:31 a.m. UTC | #3
On 2023/1/5 1:07, Song Liu wrote:
> On Wed, Jan 4, 2023 at 8:25 AM Petr Mladek <pmladek@suse.com> wrote:
>>
>> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
>>> Function __module_address() can quickly return the pointer of the module
>>> to which an address belongs. We do not need to traverse the symbols of all
>>> modules to check whether each address in addrs[] is the start address of
>>> the corresponding symbol, because register_fprobe_ips() will do this check
>>> later.
>>>
>>> Assuming that there are m modules, each module has n symbols on average,
>>> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
>>> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
>>> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
>>> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
>>> the ratio is still greater than 1. Therefore, the new method will
>>> generally have better performance.
>>>
>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>>> ---
>>>  kernel/trace/bpf_trace.c | 101 ++++++++++++++++-----------------------
>>>  1 file changed, 40 insertions(+), 61 deletions(-)
>>>
>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>> index 5f3be4bc16403a5..0ff9037098bd241 100644
>>> --- a/kernel/trace/bpf_trace.c
>>> +++ b/kernel/trace/bpf_trace.c
>>> @@ -2684,69 +2684,55 @@ static void symbols_swap_r(void *a, void *b, int size, const void *priv)
>>>       }
>>>  }
>>>
>>> -struct module_addr_args {
>>> -     unsigned long *addrs;
>>> -     u32 addrs_cnt;
>>> -     struct module **mods;
>>> -     int mods_cnt;
>>> -     int mods_cap;
>>> -};
>>> -
>>> -static int module_callback(void *data, const char *name,
>>> -                        struct module *mod, unsigned long addr)
>>> +static int get_modules_for_addrs(struct module ***out_mods, unsigned long *addrs, u32 addrs_cnt)
>>>  {
>>> -     struct module_addr_args *args = data;
>>> -     struct module **mods;
>>> -
>>> -     /* We iterate all modules symbols and for each we:
>>> -      * - search for it in provided addresses array
>>> -      * - if found we check if we already have the module pointer stored
>>> -      *   (we iterate modules sequentially, so we can check just the last
>>> -      *   module pointer)
>>> -      * - take module reference and store it
>>> -      */
>>> -     if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
>>> -                    bpf_kprobe_multi_addrs_cmp))
>>> -             return 0;
>>> +     int i, j, err;
>>> +     int mods_cnt = 0;
>>> +     int mods_cap = 0;
>>> +     struct module *mod;
>>> +     struct module **mods = NULL;
>>>
>>> -     if (args->mods && args->mods[args->mods_cnt - 1] == mod)
>>> -             return 0;
>>> +     for (i = 0; i < addrs_cnt; i++) {
>>> +             mod = __module_address(addrs[i]);
>>
>> This must be called under module_mutex to make sure that the module
>> would not disappear.

Yes, mod needs to be protected, thanks.

> 
> module_mutex is not available outside kernel/module/. The common
> practice is to disable preempt before calling __module_address().

Yes, I've looked elsewhere, and all calling preempt_disable() for
RCU read protection. I will fix it.

> CONFIG_LOCKDEP should catch this.
> 
> Thanks,
> Song
> 
> [...]
> .
>
Zhen Lei Jan. 5, 2023, 7:48 a.m. UTC | #4
On 2023/1/5 0:25, Petr Mladek wrote:
> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
>> Function __module_address() can quickly return the pointer of the module
>> to which an address belongs. We do not need to traverse the symbols of all
>> modules to check whether each address in addrs[] is the start address of
>> the corresponding symbol, because register_fprobe_ips() will do this check
>> later.
>>
>> Assuming that there are m modules, each module has n symbols on average,
>> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
>> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
>> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
>> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
>> the ratio is still greater than 1. Therefore, the new method will
>> generally have better performance.
>>
>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>> ---
>>  kernel/trace/bpf_trace.c | 101 ++++++++++++++++-----------------------
>>  1 file changed, 40 insertions(+), 61 deletions(-)
>>
>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>> index 5f3be4bc16403a5..0ff9037098bd241 100644
>> --- a/kernel/trace/bpf_trace.c
>> +++ b/kernel/trace/bpf_trace.c
>> @@ -2684,69 +2684,55 @@ static void symbols_swap_r(void *a, void *b, int size, const void *priv)
>>  	}
>>  }
>>  
>> -struct module_addr_args {
>> -	unsigned long *addrs;
>> -	u32 addrs_cnt;
>> -	struct module **mods;
>> -	int mods_cnt;
>> -	int mods_cap;
>> -};
>> -
>> -static int module_callback(void *data, const char *name,
>> -			   struct module *mod, unsigned long addr)
>> +static int get_modules_for_addrs(struct module ***out_mods, unsigned long *addrs, u32 addrs_cnt)
>>  {
>> -	struct module_addr_args *args = data;
>> -	struct module **mods;
>> -
>> -	/* We iterate all modules symbols and for each we:
>> -	 * - search for it in provided addresses array
>> -	 * - if found we check if we already have the module pointer stored
>> -	 *   (we iterate modules sequentially, so we can check just the last
>> -	 *   module pointer)
>> -	 * - take module reference and store it
>> -	 */
>> -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
>> -		       bpf_kprobe_multi_addrs_cmp))
>> -		return 0;
>> +	int i, j, err;
>> +	int mods_cnt = 0;
>> +	int mods_cap = 0;
>> +	struct module *mod;
>> +	struct module **mods = NULL;
>>  
>> -	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
>> -		return 0;
>> +	for (i = 0; i < addrs_cnt; i++) {
>> +		mod = __module_address(addrs[i]);
> 
> This must be called under module_mutex to make sure that the module
> would not disappear.
> 
>> +		if (!mod)
>> +			continue;
>>  
>> -	if (args->mods_cnt == args->mods_cap) {
>> -		args->mods_cap = max(16, args->mods_cap * 3 / 2);
>> -		mods = krealloc_array(args->mods, args->mods_cap, sizeof(*mods), GFP_KERNEL);
>> -		if (!mods)
>> -			return -ENOMEM;
>> -		args->mods = mods;
>> -	}
>> +		/* check if we already have the module pointer stored */
>> +		for (j = 0; j < mods_cnt; j++) {
>> +			if (mods[j] == mod)
>> +				break;
>> +		}
> 
> This might get optimized like the original code.
> 
> My understanding is that the addresses are sorted in "addrs" array.
> So, the address is either part of the last found module or it belongs
> to a completely new module.

I'm in a hurry to get to the airport now. I will reply next week.
move_module() shows that a module has three layouts, and the memory
area is discontinuous. I originally wanted to implement what you
suggested below. I'll analyze it in depth next week. Maybe it'll work.


> 
> 	for (i = 0; i < addrs_cnt; i++) {
> 		/*
> 		 * The adresses are sorted. The adress either belongs
> 		 * to the last found module or a new one.
> 		 *
> 		 * This is safe because we already have reference
> 		 * on the found modules.
> 		 */
> 		 if (mods_cnt && within_module(addrs[i], mods[mods_cnt - 1]))
> 			continue;
> 
> 		mutex_lock(&module_mutex);
> 		mod = __module_address(addrs[i]);
> 		if (mod && !try_module_get(mod)) {
> 			mutex_unlock(&module_mutex);
> 			goto failed;
> 		}
> 		mutex_unlock(&module_mutex);
> 
> 		/*
> 		 * Nope when the address was not from a module.
> 		 *
> 		 * Is this correct? What if the module has gone in
> 		 * the meantime? Anyway, the original code
> 		 * worked this way.
> 		 *
> 		 * FIXME: I would personally make sure that it is part
> 		 * of vmlinux or so.
> 		 */
> 		if (!mod)
> 			continue;
> 
> 		/* store the module into mods array */
> 		...
> 
> 
> 
> 
>> +		if (j < mods_cnt)
>> +			continue;
>>  
>> -	if (!try_module_get(mod))
>> -		return -EINVAL;
>> +		if (mods_cnt == mods_cap) {
>> +			struct module **new_mods;
>>  
>> -	args->mods[args->mods_cnt] = mod;
>> -	args->mods_cnt++;
>> -	return 0;
>> -}
>> +			mods_cap = max(16, mods_cap * 3 / 2);
>> +			new_mods = krealloc_array(mods, mods_cap, sizeof(*mods), GFP_KERNEL);
>> +			if (!new_mods) {
>> +				err = -ENOMEM;
>> +				goto failed;
>> +			}
>> +			mods = new_mods;
>> +		}
>>  
>> -static int get_modules_for_addrs(struct module ***mods, unsigned long *addrs, u32 addrs_cnt)
>> -{
>> -	struct module_addr_args args = {
>> -		.addrs     = addrs,
>> -		.addrs_cnt = addrs_cnt,
>> -	};
>> -	int err;
>> +		if (!try_module_get(mod)) {
>> +			err = -EINVAL;
>> +			goto failed;
>> +		}
>>  
>> -	/* We return either err < 0 in case of error, ... */
>> -	err = module_kallsyms_on_each_symbol(NULL, module_callback, &args);
>> -	if (err) {
>> -		kprobe_multi_put_modules(args.mods, args.mods_cnt);
>> -		kfree(args.mods);
>> -		return err;
>> +		mods[mods_cnt] = mod;
>> +		mods_cnt++;
>>  	}
>>  
>> -	/* or number of modules found if everything is ok. */
>> -	*mods = args.mods;
>> -	return args.mods_cnt;
>> +	*out_mods = mods;
>> +	return mods_cnt;
>> +
>> +failed:
>> +	kprobe_multi_put_modules(mods, mods_cnt);
>> +	kfree(mods);
>> +	return err;
>>  }
>>  
>>  int bpf_kprobe_multi_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)
> 
> Otherwise, it looks good. IMHO, the new code looks more straightforward
> than the original one.
> 
> Best Regards,
> Petr
> .
>
Petr Mladek Jan. 5, 2023, 9:05 a.m. UTC | #5
On Wed 2023-01-04 09:07:02, Song Liu wrote:
> On Wed, Jan 4, 2023 at 8:25 AM Petr Mladek <pmladek@suse.com> wrote:
> >
> > On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
> > > Function __module_address() can quickly return the pointer of the module
> > > to which an address belongs. We do not need to traverse the symbols of all
> > > modules to check whether each address in addrs[] is the start address of
> > > the corresponding symbol, because register_fprobe_ips() will do this check
> > > later.
> > >
> > > Assuming that there are m modules, each module has n symbols on average,
> > > and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
> > > complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
> > > and the time complexity of current method is O(K * (log(m) + M)), M <= m.
> > > (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
> > > the ratio is still greater than 1. Therefore, the new method will
> > > generally have better performance.
> > >
> > > Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
> > > ---
> > >  kernel/trace/bpf_trace.c | 101 ++++++++++++++++-----------------------
> > >  1 file changed, 40 insertions(+), 61 deletions(-)
> > >
> > > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > > index 5f3be4bc16403a5..0ff9037098bd241 100644
> > > --- a/kernel/trace/bpf_trace.c
> > > +++ b/kernel/trace/bpf_trace.c
> > > @@ -2684,69 +2684,55 @@ static void symbols_swap_r(void *a, void *b, int size, const void *priv)
> > >       }
> > >  }
> > >
> > > -struct module_addr_args {
> > > -     unsigned long *addrs;
> > > -     u32 addrs_cnt;
> > > -     struct module **mods;
> > > -     int mods_cnt;
> > > -     int mods_cap;
> > > -};
> > > -
> > > -static int module_callback(void *data, const char *name,
> > > -                        struct module *mod, unsigned long addr)
> > > +static int get_modules_for_addrs(struct module ***out_mods, unsigned long *addrs, u32 addrs_cnt)
> > >  {
> > > -     struct module_addr_args *args = data;
> > > -     struct module **mods;
> > > -
> > > -     /* We iterate all modules symbols and for each we:
> > > -      * - search for it in provided addresses array
> > > -      * - if found we check if we already have the module pointer stored
> > > -      *   (we iterate modules sequentially, so we can check just the last
> > > -      *   module pointer)
> > > -      * - take module reference and store it
> > > -      */
> > > -     if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
> > > -                    bpf_kprobe_multi_addrs_cmp))
> > > -             return 0;
> > > +     int i, j, err;
> > > +     int mods_cnt = 0;
> > > +     int mods_cap = 0;
> > > +     struct module *mod;
> > > +     struct module **mods = NULL;
> > >
> > > -     if (args->mods && args->mods[args->mods_cnt - 1] == mod)
> > > -             return 0;
> > > +     for (i = 0; i < addrs_cnt; i++) {
> > > +             mod = __module_address(addrs[i]);
> >
> > This must be called under module_mutex to make sure that the module
> > would not disappear.
> 
> module_mutex is not available outside kernel/module/. The common
> practice is to disable preempt before calling __module_address().
> CONFIG_LOCKDEP should catch this.

I see. Sigh, it is always better to take mutex than disable
preemption. But it might be acceptable in this case. We just need
to be careful.

First, the preemption must stay disabled all the time until
try_module_get(). Otherwise the returned struct module could
disappear in the meantime.

Second, krealloc_array() has to be called with preemption
enabled. It is perfectly fine to do it after try_module_get().

Best Regards,
Petr
Petr Mladek Jan. 5, 2023, 9:32 a.m. UTC | #6
On Wed 2023-01-04 17:25:08, Petr Mladek wrote:
> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
> > Function __module_address() can quickly return the pointer of the module
> > to which an address belongs. We do not need to traverse the symbols of all
> > modules to check whether each address in addrs[] is the start address of
> > the corresponding symbol, because register_fprobe_ips() will do this check
> > later.
> > 
> > Assuming that there are m modules, each module has n symbols on average,
> > and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
> > complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
> > and the time complexity of current method is O(K * (log(m) + M)), M <= m.
> > (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
> > the ratio is still greater than 1. Therefore, the new method will
> > generally have better performance.
> > 
> > Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
> > ---
> >  kernel/trace/bpf_trace.c | 101 ++++++++++++++++-----------------------
> >  1 file changed, 40 insertions(+), 61 deletions(-)
> > 
> > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > index 5f3be4bc16403a5..0ff9037098bd241 100644
> > --- a/kernel/trace/bpf_trace.c
> > +++ b/kernel/trace/bpf_trace.c
> > @@ -2684,69 +2684,55 @@ static void symbols_swap_r(void *a, void *b, int size, const void *priv)
> >  	}
> >  }
> >  
> > -struct module_addr_args {
> > -	unsigned long *addrs;
> > -	u32 addrs_cnt;
> > -	struct module **mods;
> > -	int mods_cnt;
> > -	int mods_cap;
> > -};
> > -
> > -static int module_callback(void *data, const char *name,
> > -			   struct module *mod, unsigned long addr)
> > +static int get_modules_for_addrs(struct module ***out_mods, unsigned long *addrs, u32 addrs_cnt)
> >  {
> > -	struct module_addr_args *args = data;
> > -	struct module **mods;
> > -
> > -	/* We iterate all modules symbols and for each we:
> > -	 * - search for it in provided addresses array
> > -	 * - if found we check if we already have the module pointer stored
> > -	 *   (we iterate modules sequentially, so we can check just the last
> > -	 *   module pointer)
> > -	 * - take module reference and store it
> > -	 */
> > -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
> > -		       bpf_kprobe_multi_addrs_cmp))
> > -		return 0;
> > +	int i, j, err;
> > +	int mods_cnt = 0;
> > +	int mods_cap = 0;
> > +	struct module *mod;
> > +	struct module **mods = NULL;
> >  
> > -	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
> > -		return 0;
> > +	for (i = 0; i < addrs_cnt; i++) {
> > +		mod = __module_address(addrs[i]);
> 
> This must be called under module_mutex to make sure that the module
> would not disappear.
> 
> > +		if (!mod)
> > +			continue;
> >  
> > -	if (args->mods_cnt == args->mods_cap) {
> > -		args->mods_cap = max(16, args->mods_cap * 3 / 2);
> > -		mods = krealloc_array(args->mods, args->mods_cap, sizeof(*mods), GFP_KERNEL);
> > -		if (!mods)
> > -			return -ENOMEM;
> > -		args->mods = mods;
> > -	}
> > +		/* check if we already have the module pointer stored */
> > +		for (j = 0; j < mods_cnt; j++) {
> > +			if (mods[j] == mod)
> > +				break;
> > +		}
> 
> This might get optimized like the original code.
> 
> My understanding is that the addresses are sorted in "addrs" array.
> So, the address is either part of the last found module or it belongs
> to a completely new module.

I thought more about it and I think that I was wrong, see below.

> 	for (i = 0; i < addrs_cnt; i++) {
> 		/*
> 		 * The adresses are sorted. The adress either belongs
> 		 * to the last found module or a new one.
> 		 *
> 		 * This is safe because we already have reference
> 		 * on the found modules.
> 		 */
> 		 if (mods_cnt && within_module(addrs[i], mods[mods_cnt - 1]))
> 			continue;

within_module() checks two sections (init and core). They are
allocated separately, see module_alloc() called in move_module().

There might be a section from another modules between the init
and core section of a module.

The optimization worked in the original code because
module_kallsyms_on_each_symbol() always iterated over all
symbols from a module.

That said, I am not sure if bpf trace might be added for
symbols in the module init section. But it might be
better to stay on the safe side.

Best Regards,
Petr
Jiri Olsa Jan. 5, 2023, 9:31 p.m. UTC | #7
On Wed, Jan 04, 2023 at 05:25:08PM +0100, Petr Mladek wrote:
> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
> > Function __module_address() can quickly return the pointer of the module
> > to which an address belongs. We do not need to traverse the symbols of all
> > modules to check whether each address in addrs[] is the start address of
> > the corresponding symbol, because register_fprobe_ips() will do this check
> > later.

hum, for some reason I can see only replies to this patch and
not the actual patch.. I'll dig it out of the lore I guess

> > 
> > Assuming that there are m modules, each module has n symbols on average,
> > and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
> > complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
> > and the time complexity of current method is O(K * (log(m) + M)), M <= m.
> > (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
> > the ratio is still greater than 1. Therefore, the new method will
> > generally have better performance.

could you try to benchmark that? I tried something similar but was not
able to get better performance

I'll review and run my benchmark test tomorrow

thanks,
jirka

> > 
> > Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
> > ---
> >  kernel/trace/bpf_trace.c | 101 ++++++++++++++++-----------------------
> >  1 file changed, 40 insertions(+), 61 deletions(-)
> > 
> > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > index 5f3be4bc16403a5..0ff9037098bd241 100644
> > --- a/kernel/trace/bpf_trace.c
> > +++ b/kernel/trace/bpf_trace.c
> > @@ -2684,69 +2684,55 @@ static void symbols_swap_r(void *a, void *b, int size, const void *priv)
> >  	}
> >  }
> >  
> > -struct module_addr_args {
> > -	unsigned long *addrs;
> > -	u32 addrs_cnt;
> > -	struct module **mods;
> > -	int mods_cnt;
> > -	int mods_cap;
> > -};
> > -
> > -static int module_callback(void *data, const char *name,
> > -			   struct module *mod, unsigned long addr)
> > +static int get_modules_for_addrs(struct module ***out_mods, unsigned long *addrs, u32 addrs_cnt)
> >  {
> > -	struct module_addr_args *args = data;
> > -	struct module **mods;
> > -
> > -	/* We iterate all modules symbols and for each we:
> > -	 * - search for it in provided addresses array
> > -	 * - if found we check if we already have the module pointer stored
> > -	 *   (we iterate modules sequentially, so we can check just the last
> > -	 *   module pointer)
> > -	 * - take module reference and store it
> > -	 */
> > -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
> > -		       bpf_kprobe_multi_addrs_cmp))
> > -		return 0;
> > +	int i, j, err;
> > +	int mods_cnt = 0;
> > +	int mods_cap = 0;
> > +	struct module *mod;
> > +	struct module **mods = NULL;
> >  
> > -	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
> > -		return 0;
> > +	for (i = 0; i < addrs_cnt; i++) {
> > +		mod = __module_address(addrs[i]);
> 
> This must be called under module_mutex to make sure that the module
> would not disappear.
> 
> > +		if (!mod)
> > +			continue;
> >  
> > -	if (args->mods_cnt == args->mods_cap) {
> > -		args->mods_cap = max(16, args->mods_cap * 3 / 2);
> > -		mods = krealloc_array(args->mods, args->mods_cap, sizeof(*mods), GFP_KERNEL);
> > -		if (!mods)
> > -			return -ENOMEM;
> > -		args->mods = mods;
> > -	}
> > +		/* check if we already have the module pointer stored */
> > +		for (j = 0; j < mods_cnt; j++) {
> > +			if (mods[j] == mod)
> > +				break;
> > +		}
> 
> This might get optimized like the original code.
> 
> My understanding is that the addresses are sorted in "addrs" array.
> So, the address is either part of the last found module or it belongs
> to a completely new module.
> 
> 	for (i = 0; i < addrs_cnt; i++) {
> 		/*
> 		 * The adresses are sorted. The adress either belongs
> 		 * to the last found module or a new one.
> 		 *
> 		 * This is safe because we already have reference
> 		 * on the found modules.
> 		 */
> 		 if (mods_cnt && within_module(addrs[i], mods[mods_cnt - 1]))
> 			continue;
> 
> 		mutex_lock(&module_mutex);
> 		mod = __module_address(addrs[i]);
> 		if (mod && !try_module_get(mod)) {
> 			mutex_unlock(&module_mutex);
> 			goto failed;
> 		}
> 		mutex_unlock(&module_mutex);
> 
> 		/*
> 		 * Nope when the address was not from a module.
> 		 *
> 		 * Is this correct? What if the module has gone in
> 		 * the meantime? Anyway, the original code
> 		 * worked this way.
> 		 *
> 		 * FIXME: I would personally make sure that it is part
> 		 * of vmlinux or so.
> 		 */
> 		if (!mod)
> 			continue;
> 
> 		/* store the module into mods array */
> 		...
> 
> 
> 
> 
> > +		if (j < mods_cnt)
> > +			continue;
> >  
> > -	if (!try_module_get(mod))
> > -		return -EINVAL;
> > +		if (mods_cnt == mods_cap) {
> > +			struct module **new_mods;
> >  
> > -	args->mods[args->mods_cnt] = mod;
> > -	args->mods_cnt++;
> > -	return 0;
> > -}
> > +			mods_cap = max(16, mods_cap * 3 / 2);
> > +			new_mods = krealloc_array(mods, mods_cap, sizeof(*mods), GFP_KERNEL);
> > +			if (!new_mods) {
> > +				err = -ENOMEM;
> > +				goto failed;
> > +			}
> > +			mods = new_mods;
> > +		}
> >  
> > -static int get_modules_for_addrs(struct module ***mods, unsigned long *addrs, u32 addrs_cnt)
> > -{
> > -	struct module_addr_args args = {
> > -		.addrs     = addrs,
> > -		.addrs_cnt = addrs_cnt,
> > -	};
> > -	int err;
> > +		if (!try_module_get(mod)) {
> > +			err = -EINVAL;
> > +			goto failed;
> > +		}
> >  
> > -	/* We return either err < 0 in case of error, ... */
> > -	err = module_kallsyms_on_each_symbol(NULL, module_callback, &args);
> > -	if (err) {
> > -		kprobe_multi_put_modules(args.mods, args.mods_cnt);
> > -		kfree(args.mods);
> > -		return err;
> > +		mods[mods_cnt] = mod;
> > +		mods_cnt++;
> >  	}
> >  
> > -	/* or number of modules found if everything is ok. */
> > -	*mods = args.mods;
> > -	return args.mods_cnt;
> > +	*out_mods = mods;
> > +	return mods_cnt;
> > +
> > +failed:
> > +	kprobe_multi_put_modules(mods, mods_cnt);
> > +	kfree(mods);
> > +	return err;
> >  }
> >  
> >  int bpf_kprobe_multi_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)
> 
> Otherwise, it looks good. IMHO, the new code looks more straightforward
> than the original one.
> 
> Best Regards,
> Petr
Jiri Olsa Jan. 6, 2023, 9:45 a.m. UTC | #8
On Thu, Jan 05, 2023 at 10:31:12PM +0100, Jiri Olsa wrote:
> On Wed, Jan 04, 2023 at 05:25:08PM +0100, Petr Mladek wrote:
> > On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
> > > Function __module_address() can quickly return the pointer of the module
> > > to which an address belongs. We do not need to traverse the symbols of all
> > > modules to check whether each address in addrs[] is the start address of
> > > the corresponding symbol, because register_fprobe_ips() will do this check
> > > later.
> 
> hum, for some reason I can see only replies to this patch and
> not the actual patch.. I'll dig it out of the lore I guess
> 
> > > 
> > > Assuming that there are m modules, each module has n symbols on average,
> > > and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
> > > complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
> > > and the time complexity of current method is O(K * (log(m) + M)), M <= m.
> > > (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
> > > the ratio is still greater than 1. Therefore, the new method will
> > > generally have better performance.
> 
> could you try to benchmark that? I tried something similar but was not
> able to get better performance

hm looks like I tried the smilar thing (below) like you did,
but wasn't able to get better performace

I guess your goal is to get rid of the module arg in
module_kallsyms_on_each_symbol callback that we use?
I'm ok with the change if the performace is not worse

jirka


---
diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
index 5b9008bc597b..3280c22009f1 100644
--- a/kernel/trace/bpf_trace.c
+++ b/kernel/trace/bpf_trace.c
@@ -2692,23 +2692,16 @@ struct module_addr_args {
 	int mods_cap;
 };
 
-static int module_callback(void *data, const char *name,
-			   struct module *mod, unsigned long addr)
+static int add_module(struct module_addr_args *args, struct module *mod)
 {
-	struct module_addr_args *args = data;
 	struct module **mods;
 
-	/* We iterate all modules symbols and for each we:
-	 * - search for it in provided addresses array
-	 * - if found we check if we already have the module pointer stored
-	 *   (we iterate modules sequentially, so we can check just the last
-	 *   module pointer)
+	/* We iterate sorted addresses and for each within module we:
+	 * - check if we already have the module pointer stored for it
+	 *   (we iterate sorted addresses sequentially, so we can check
+	 *   just the last module pointer)
 	 * - take module reference and store it
 	 */
-	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
-		       bpf_kprobe_multi_addrs_cmp))
-		return 0;
-
 	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
 		return 0;
 
@@ -2734,10 +2727,24 @@ static int get_modules_for_addrs(struct module ***mods, unsigned long *addrs, u3
 		.addrs     = addrs,
 		.addrs_cnt = addrs_cnt,
 	};
-	int err;
+	u32 i, err = 0;
+
+	for (i = 0; !err && i < addrs_cnt; i++) {
+		struct module *mod;
+		bool found = false;
+
+		preempt_disable();
+		mod = __module_text_address(addrs[i]);
+		found = mod && try_module_get(mod);
+		preempt_enable();
+
+		if (found) {
+			err = add_module(&args, mod);
+			module_put(mod);
+		}
+	}
 
 	/* We return either err < 0 in case of error, ... */
-	err = module_kallsyms_on_each_symbol(module_callback, &args);
 	if (err) {
 		kprobe_multi_put_modules(args.mods, args.mods_cnt);
 		kfree(args.mods);
@@ -2862,7 +2869,8 @@ int bpf_kprobe_multi_link_attach(const union bpf_attr *attr, struct bpf_prog *pr
 	} else {
 		/*
 		 * We need to sort addrs array even if there are no cookies
-		 * provided, to allow bsearch in get_modules_for_addrs.
+		 * provided, to allow sequential address walk in
+		 * get_modules_for_addrs.
 		 */
 		sort(addrs, cnt, sizeof(*addrs),
 		       bpf_kprobe_multi_addrs_cmp, NULL);
Zhen Lei Jan. 9, 2023, 4:02 a.m. UTC | #9
On 2023/1/5 17:05, Petr Mladek wrote:
> On Wed 2023-01-04 09:07:02, Song Liu wrote:
>> On Wed, Jan 4, 2023 at 8:25 AM Petr Mladek <pmladek@suse.com> wrote:
>>>
>>> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
>>>> Function __module_address() can quickly return the pointer of the module
>>>> to which an address belongs. We do not need to traverse the symbols of all
>>>> modules to check whether each address in addrs[] is the start address of
>>>> the corresponding symbol, because register_fprobe_ips() will do this check
>>>> later.
>>>>
>>>> Assuming that there are m modules, each module has n symbols on average,
>>>> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
>>>> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
>>>> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
>>>> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
>>>> the ratio is still greater than 1. Therefore, the new method will
>>>> generally have better performance.
>>>>
>>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>>>> ---
>>>>  kernel/trace/bpf_trace.c | 101 ++++++++++++++++-----------------------
>>>>  1 file changed, 40 insertions(+), 61 deletions(-)
>>>>
>>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>>> index 5f3be4bc16403a5..0ff9037098bd241 100644
>>>> --- a/kernel/trace/bpf_trace.c
>>>> +++ b/kernel/trace/bpf_trace.c
>>>> @@ -2684,69 +2684,55 @@ static void symbols_swap_r(void *a, void *b, int size, const void *priv)
>>>>       }
>>>>  }
>>>>
>>>> -struct module_addr_args {
>>>> -     unsigned long *addrs;
>>>> -     u32 addrs_cnt;
>>>> -     struct module **mods;
>>>> -     int mods_cnt;
>>>> -     int mods_cap;
>>>> -};
>>>> -
>>>> -static int module_callback(void *data, const char *name,
>>>> -                        struct module *mod, unsigned long addr)
>>>> +static int get_modules_for_addrs(struct module ***out_mods, unsigned long *addrs, u32 addrs_cnt)
>>>>  {
>>>> -     struct module_addr_args *args = data;
>>>> -     struct module **mods;
>>>> -
>>>> -     /* We iterate all modules symbols and for each we:
>>>> -      * - search for it in provided addresses array
>>>> -      * - if found we check if we already have the module pointer stored
>>>> -      *   (we iterate modules sequentially, so we can check just the last
>>>> -      *   module pointer)
>>>> -      * - take module reference and store it
>>>> -      */
>>>> -     if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
>>>> -                    bpf_kprobe_multi_addrs_cmp))
>>>> -             return 0;
>>>> +     int i, j, err;
>>>> +     int mods_cnt = 0;
>>>> +     int mods_cap = 0;
>>>> +     struct module *mod;
>>>> +     struct module **mods = NULL;
>>>>
>>>> -     if (args->mods && args->mods[args->mods_cnt - 1] == mod)
>>>> -             return 0;
>>>> +     for (i = 0; i < addrs_cnt; i++) {
>>>> +             mod = __module_address(addrs[i]);
>>>
>>> This must be called under module_mutex to make sure that the module
>>> would not disappear.
>>
>> module_mutex is not available outside kernel/module/. The common
>> practice is to disable preempt before calling __module_address().
>> CONFIG_LOCKDEP should catch this.
> 
> I see. Sigh, it is always better to take mutex than disable
> preemption. But it might be acceptable in this case. We just need
> to be careful.
> 
> First, the preemption must stay disabled all the time until
> try_module_get(). Otherwise the returned struct module could
> disappear in the meantime.
> 
> Second, krealloc_array() has to be called with preemption
> enabled. It is perfectly fine to do it after try_module_get().

Okay, thanks for the heads-up.

> 
> Best Regards,
> Petr
> .
>
Zhen Lei Jan. 9, 2023, 4:10 a.m. UTC | #10
On 2023/1/5 17:32, Petr Mladek wrote:
> On Wed 2023-01-04 17:25:08, Petr Mladek wrote:
>> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
>>> Function __module_address() can quickly return the pointer of the module
>>> to which an address belongs. We do not need to traverse the symbols of all
>>> modules to check whether each address in addrs[] is the start address of
>>> the corresponding symbol, because register_fprobe_ips() will do this check
>>> later.
>>>
>>> Assuming that there are m modules, each module has n symbols on average,
>>> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
>>> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
>>> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
>>> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
>>> the ratio is still greater than 1. Therefore, the new method will
>>> generally have better performance.
>>>
>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>>> ---
>>>  kernel/trace/bpf_trace.c | 101 ++++++++++++++++-----------------------
>>>  1 file changed, 40 insertions(+), 61 deletions(-)
>>>
>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>> index 5f3be4bc16403a5..0ff9037098bd241 100644
>>> --- a/kernel/trace/bpf_trace.c
>>> +++ b/kernel/trace/bpf_trace.c
>>> @@ -2684,69 +2684,55 @@ static void symbols_swap_r(void *a, void *b, int size, const void *priv)
>>>  	}
>>>  }
>>>  
>>> -struct module_addr_args {
>>> -	unsigned long *addrs;
>>> -	u32 addrs_cnt;
>>> -	struct module **mods;
>>> -	int mods_cnt;
>>> -	int mods_cap;
>>> -};
>>> -
>>> -static int module_callback(void *data, const char *name,
>>> -			   struct module *mod, unsigned long addr)
>>> +static int get_modules_for_addrs(struct module ***out_mods, unsigned long *addrs, u32 addrs_cnt)
>>>  {
>>> -	struct module_addr_args *args = data;
>>> -	struct module **mods;
>>> -
>>> -	/* We iterate all modules symbols and for each we:
>>> -	 * - search for it in provided addresses array
>>> -	 * - if found we check if we already have the module pointer stored
>>> -	 *   (we iterate modules sequentially, so we can check just the last
>>> -	 *   module pointer)
>>> -	 * - take module reference and store it
>>> -	 */
>>> -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
>>> -		       bpf_kprobe_multi_addrs_cmp))
>>> -		return 0;
>>> +	int i, j, err;
>>> +	int mods_cnt = 0;
>>> +	int mods_cap = 0;
>>> +	struct module *mod;
>>> +	struct module **mods = NULL;
>>>  
>>> -	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
>>> -		return 0;
>>> +	for (i = 0; i < addrs_cnt; i++) {
>>> +		mod = __module_address(addrs[i]);
>>
>> This must be called under module_mutex to make sure that the module
>> would not disappear.
>>
>>> +		if (!mod)
>>> +			continue;
>>>  
>>> -	if (args->mods_cnt == args->mods_cap) {
>>> -		args->mods_cap = max(16, args->mods_cap * 3 / 2);
>>> -		mods = krealloc_array(args->mods, args->mods_cap, sizeof(*mods), GFP_KERNEL);
>>> -		if (!mods)
>>> -			return -ENOMEM;
>>> -		args->mods = mods;
>>> -	}
>>> +		/* check if we already have the module pointer stored */
>>> +		for (j = 0; j < mods_cnt; j++) {
>>> +			if (mods[j] == mod)
>>> +				break;
>>> +		}
>>
>> This might get optimized like the original code.
>>
>> My understanding is that the addresses are sorted in "addrs" array.
>> So, the address is either part of the last found module or it belongs
>> to a completely new module.
> 
> I thought more about it and I think that I was wrong, see below.
> 
>> 	for (i = 0; i < addrs_cnt; i++) {
>> 		/*
>> 		 * The adresses are sorted. The adress either belongs
>> 		 * to the last found module or a new one.
>> 		 *
>> 		 * This is safe because we already have reference
>> 		 * on the found modules.
>> 		 */
>> 		 if (mods_cnt && within_module(addrs[i], mods[mods_cnt - 1]))
>> 			continue;
> 
> within_module() checks two sections (init and core). They are
> allocated separately, see module_alloc() called in move_module().
> 
> There might be a section from another modules between the init
> and core section of a module.
> 
> The optimization worked in the original code because
> module_kallsyms_on_each_symbol() always iterated over all
> symbols from a module.
> 
> That said, I am not sure if bpf trace might be added for
> symbols in the module init section. But it might be
> better to stay on the safe side.

Yes.

> 
> Best Regards,
> Petr
> .
>
Zhen Lei Jan. 9, 2023, 7:03 a.m. UTC | #11
On 2023/1/6 5:31, Jiri Olsa wrote:
> On Wed, Jan 04, 2023 at 05:25:08PM +0100, Petr Mladek wrote:
>> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
>>> Function __module_address() can quickly return the pointer of the module
>>> to which an address belongs. We do not need to traverse the symbols of all
>>> modules to check whether each address in addrs[] is the start address of
>>> the corresponding symbol, because register_fprobe_ips() will do this check
>>> later.
> 
> hum, for some reason I can see only replies to this patch and
> not the actual patch.. I'll dig it out of the lore I guess

https://lkml.org/lkml/2022/12/30/195

> 
>>>
>>> Assuming that there are m modules, each module has n symbols on average,
>>> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
>>> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
>>> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
>>> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
>>> the ratio is still greater than 1. Therefore, the new method will
>>> generally have better performance.
> 
> could you try to benchmark that? I tried something similar but was not
> able to get better performance

I'm just theoretically analyzing, at least the performance won't get worse.

> 
> I'll review and run my benchmark test tomorrow
> 
> thanks,
> jirka
> 
>>>
>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>>> ---
>>>  kernel/trace/bpf_trace.c | 101 ++++++++++++++++-----------------------
>>>  1 file changed, 40 insertions(+), 61 deletions(-)
>>>
>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>> index 5f3be4bc16403a5..0ff9037098bd241 100644
>>> --- a/kernel/trace/bpf_trace.c
>>> +++ b/kernel/trace/bpf_trace.c
>>> @@ -2684,69 +2684,55 @@ static void symbols_swap_r(void *a, void *b, int size, const void *priv)
>>>  	}
>>>  }
>>>  
>>> -struct module_addr_args {
>>> -	unsigned long *addrs;
>>> -	u32 addrs_cnt;
>>> -	struct module **mods;
>>> -	int mods_cnt;
>>> -	int mods_cap;
>>> -};
>>> -
>>> -static int module_callback(void *data, const char *name,
>>> -			   struct module *mod, unsigned long addr)
>>> +static int get_modules_for_addrs(struct module ***out_mods, unsigned long *addrs, u32 addrs_cnt)
>>>  {
>>> -	struct module_addr_args *args = data;
>>> -	struct module **mods;
>>> -
>>> -	/* We iterate all modules symbols and for each we:
>>> -	 * - search for it in provided addresses array
>>> -	 * - if found we check if we already have the module pointer stored
>>> -	 *   (we iterate modules sequentially, so we can check just the last
>>> -	 *   module pointer)
>>> -	 * - take module reference and store it
>>> -	 */
>>> -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
>>> -		       bpf_kprobe_multi_addrs_cmp))
>>> -		return 0;
>>> +	int i, j, err;
>>> +	int mods_cnt = 0;
>>> +	int mods_cap = 0;
>>> +	struct module *mod;
>>> +	struct module **mods = NULL;
>>>  
>>> -	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
>>> -		return 0;
>>> +	for (i = 0; i < addrs_cnt; i++) {
>>> +		mod = __module_address(addrs[i]);
>>
>> This must be called under module_mutex to make sure that the module
>> would not disappear.
>>
>>> +		if (!mod)
>>> +			continue;
>>>  
>>> -	if (args->mods_cnt == args->mods_cap) {
>>> -		args->mods_cap = max(16, args->mods_cap * 3 / 2);
>>> -		mods = krealloc_array(args->mods, args->mods_cap, sizeof(*mods), GFP_KERNEL);
>>> -		if (!mods)
>>> -			return -ENOMEM;
>>> -		args->mods = mods;
>>> -	}
>>> +		/* check if we already have the module pointer stored */
>>> +		for (j = 0; j < mods_cnt; j++) {
>>> +			if (mods[j] == mod)
>>> +				break;
>>> +		}
>>
>> This might get optimized like the original code.
>>
>> My understanding is that the addresses are sorted in "addrs" array.
>> So, the address is either part of the last found module or it belongs
>> to a completely new module.
>>
>> 	for (i = 0; i < addrs_cnt; i++) {
>> 		/*
>> 		 * The adresses are sorted. The adress either belongs
>> 		 * to the last found module or a new one.
>> 		 *
>> 		 * This is safe because we already have reference
>> 		 * on the found modules.
>> 		 */
>> 		 if (mods_cnt && within_module(addrs[i], mods[mods_cnt - 1]))
>> 			continue;
>>
>> 		mutex_lock(&module_mutex);
>> 		mod = __module_address(addrs[i]);
>> 		if (mod && !try_module_get(mod)) {
>> 			mutex_unlock(&module_mutex);
>> 			goto failed;
>> 		}
>> 		mutex_unlock(&module_mutex);
>>
>> 		/*
>> 		 * Nope when the address was not from a module.
>> 		 *
>> 		 * Is this correct? What if the module has gone in
>> 		 * the meantime? Anyway, the original code
>> 		 * worked this way.
>> 		 *
>> 		 * FIXME: I would personally make sure that it is part
>> 		 * of vmlinux or so.
>> 		 */
>> 		if (!mod)
>> 			continue;
>>
>> 		/* store the module into mods array */
>> 		...
>>
>>
>>
>>
>>> +		if (j < mods_cnt)
>>> +			continue;
>>>  
>>> -	if (!try_module_get(mod))
>>> -		return -EINVAL;
>>> +		if (mods_cnt == mods_cap) {
>>> +			struct module **new_mods;
>>>  
>>> -	args->mods[args->mods_cnt] = mod;
>>> -	args->mods_cnt++;
>>> -	return 0;
>>> -}
>>> +			mods_cap = max(16, mods_cap * 3 / 2);
>>> +			new_mods = krealloc_array(mods, mods_cap, sizeof(*mods), GFP_KERNEL);
>>> +			if (!new_mods) {
>>> +				err = -ENOMEM;
>>> +				goto failed;
>>> +			}
>>> +			mods = new_mods;
>>> +		}
>>>  
>>> -static int get_modules_for_addrs(struct module ***mods, unsigned long *addrs, u32 addrs_cnt)
>>> -{
>>> -	struct module_addr_args args = {
>>> -		.addrs     = addrs,
>>> -		.addrs_cnt = addrs_cnt,
>>> -	};
>>> -	int err;
>>> +		if (!try_module_get(mod)) {
>>> +			err = -EINVAL;
>>> +			goto failed;
>>> +		}
>>>  
>>> -	/* We return either err < 0 in case of error, ... */
>>> -	err = module_kallsyms_on_each_symbol(NULL, module_callback, &args);
>>> -	if (err) {
>>> -		kprobe_multi_put_modules(args.mods, args.mods_cnt);
>>> -		kfree(args.mods);
>>> -		return err;
>>> +		mods[mods_cnt] = mod;
>>> +		mods_cnt++;
>>>  	}
>>>  
>>> -	/* or number of modules found if everything is ok. */
>>> -	*mods = args.mods;
>>> -	return args.mods_cnt;
>>> +	*out_mods = mods;
>>> +	return mods_cnt;
>>> +
>>> +failed:
>>> +	kprobe_multi_put_modules(mods, mods_cnt);
>>> +	kfree(mods);
>>> +	return err;
>>>  }
>>>  
>>>  int bpf_kprobe_multi_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)
>>
>> Otherwise, it looks good. IMHO, the new code looks more straightforward
>> than the original one.
>>
>> Best Regards,
>> Petr
> .
>
Zhen Lei Jan. 9, 2023, 8:51 a.m. UTC | #12
On 2023/1/6 17:45, Jiri Olsa wrote:
> On Thu, Jan 05, 2023 at 10:31:12PM +0100, Jiri Olsa wrote:
>> On Wed, Jan 04, 2023 at 05:25:08PM +0100, Petr Mladek wrote:
>>> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
>>>> Function __module_address() can quickly return the pointer of the module
>>>> to which an address belongs. We do not need to traverse the symbols of all
>>>> modules to check whether each address in addrs[] is the start address of
>>>> the corresponding symbol, because register_fprobe_ips() will do this check
>>>> later.
>>
>> hum, for some reason I can see only replies to this patch and
>> not the actual patch.. I'll dig it out of the lore I guess
>>
>>>>
>>>> Assuming that there are m modules, each module has n symbols on average,
>>>> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
>>>> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
>>>> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
>>>> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
>>>> the ratio is still greater than 1. Therefore, the new method will
>>>> generally have better performance.
>>
>> could you try to benchmark that? I tried something similar but was not
>> able to get better performance
> 
> hm looks like I tried the smilar thing (below) like you did,

Yes. I just found out you're working on this improvement, too.

> but wasn't able to get better performace

Your implementation below is already the limit that can be optimized.
If the performance is not improved, it indicates that this place is
not the bottleneck.

> 
> I guess your goal is to get rid of the module arg in
> module_kallsyms_on_each_symbol callback that we use?

It's not a bad thing to keep argument 'mod' for function
module_kallsyms_on_each_symbol(), but for kallsyms_on_each_symbol(),
it's completely redundant. Now these two functions often use the
same hook function. So I carefully analyzed get_modules_for_addrs(),
which is the only place that involves the use of parameter 'mod'.
Looks like there's a possibility of eliminating parameter 'mod'.

> I'm ok with the change if the performace is not worse

OK, thanks.

> 
> jirka
> 
> 
> ---
> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> index 5b9008bc597b..3280c22009f1 100644
> --- a/kernel/trace/bpf_trace.c
> +++ b/kernel/trace/bpf_trace.c
> @@ -2692,23 +2692,16 @@ struct module_addr_args {
>  	int mods_cap;
>  };
>  
> -static int module_callback(void *data, const char *name,
> -			   struct module *mod, unsigned long addr)
> +static int add_module(struct module_addr_args *args, struct module *mod)
>  {
> -	struct module_addr_args *args = data;
>  	struct module **mods;
>  
> -	/* We iterate all modules symbols and for each we:
> -	 * - search for it in provided addresses array
> -	 * - if found we check if we already have the module pointer stored
> -	 *   (we iterate modules sequentially, so we can check just the last
> -	 *   module pointer)
> +	/* We iterate sorted addresses and for each within module we:
> +	 * - check if we already have the module pointer stored for it
> +	 *   (we iterate sorted addresses sequentially, so we can check
> +	 *   just the last module pointer)
>  	 * - take module reference and store it
>  	 */
> -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
> -		       bpf_kprobe_multi_addrs_cmp))
> -		return 0;
> -
>  	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
>  		return 0;

There'll be problems Petr mentioned.

https://lkml.org/lkml/2023/1/5/191

>  
> @@ -2734,10 +2727,24 @@ static int get_modules_for_addrs(struct module ***mods, unsigned long *addrs, u3
>  		.addrs     = addrs,
>  		.addrs_cnt = addrs_cnt,
>  	};
> -	int err;
> +	u32 i, err = 0;
> +
> +	for (i = 0; !err && i < addrs_cnt; i++) {
> +		struct module *mod;
> +		bool found = false;
> +
> +		preempt_disable();
> +		mod = __module_text_address(addrs[i]);
> +		found = mod && try_module_get(mod);
> +		preempt_enable();
> +
> +		if (found) {
> +			err = add_module(&args, mod);
> +			module_put(mod);
> +		}
> +	}
>  
>  	/* We return either err < 0 in case of error, ... */
> -	err = module_kallsyms_on_each_symbol(module_callback, &args);
>  	if (err) {
>  		kprobe_multi_put_modules(args.mods, args.mods_cnt);
>  		kfree(args.mods);
> @@ -2862,7 +2869,8 @@ int bpf_kprobe_multi_link_attach(const union bpf_attr *attr, struct bpf_prog *pr
>  	} else {
>  		/*
>  		 * We need to sort addrs array even if there are no cookies
> -		 * provided, to allow bsearch in get_modules_for_addrs.
> +		 * provided, to allow sequential address walk in
> +		 * get_modules_for_addrs.
>  		 */
>  		sort(addrs, cnt, sizeof(*addrs),
>  		       bpf_kprobe_multi_addrs_cmp, NULL);
> .
>
Jiri Olsa Jan. 9, 2023, 1:48 p.m. UTC | #13
On Mon, Jan 09, 2023 at 04:51:37PM +0800, Leizhen (ThunderTown) wrote:
> 
> 
> On 2023/1/6 17:45, Jiri Olsa wrote:
> > On Thu, Jan 05, 2023 at 10:31:12PM +0100, Jiri Olsa wrote:
> >> On Wed, Jan 04, 2023 at 05:25:08PM +0100, Petr Mladek wrote:
> >>> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
> >>>> Function __module_address() can quickly return the pointer of the module
> >>>> to which an address belongs. We do not need to traverse the symbols of all
> >>>> modules to check whether each address in addrs[] is the start address of
> >>>> the corresponding symbol, because register_fprobe_ips() will do this check
> >>>> later.
> >>
> >> hum, for some reason I can see only replies to this patch and
> >> not the actual patch.. I'll dig it out of the lore I guess
> >>
> >>>>
> >>>> Assuming that there are m modules, each module has n symbols on average,
> >>>> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
> >>>> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
> >>>> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
> >>>> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
> >>>> the ratio is still greater than 1. Therefore, the new method will
> >>>> generally have better performance.
> >>
> >> could you try to benchmark that? I tried something similar but was not
> >> able to get better performance
> > 
> > hm looks like I tried the smilar thing (below) like you did,
> 
> Yes. I just found out you're working on this improvement, too.
> 
> > but wasn't able to get better performace
> 
> Your implementation below is already the limit that can be optimized.
> If the performance is not improved, it indicates that this place is
> not the bottleneck.
> 
> > 
> > I guess your goal is to get rid of the module arg in
> > module_kallsyms_on_each_symbol callback that we use?
> 
> It's not a bad thing to keep argument 'mod' for function
> module_kallsyms_on_each_symbol(), but for kallsyms_on_each_symbol(),
> it's completely redundant. Now these two functions often use the
> same hook function. So I carefully analyzed get_modules_for_addrs(),
> which is the only place that involves the use of parameter 'mod'.
> Looks like there's a possibility of eliminating parameter 'mod'.
> 
> > I'm ok with the change if the performace is not worse
> 
> OK, thanks.
> 
> > 
> > jirka
> > 
> > 
> > ---
> > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > index 5b9008bc597b..3280c22009f1 100644
> > --- a/kernel/trace/bpf_trace.c
> > +++ b/kernel/trace/bpf_trace.c
> > @@ -2692,23 +2692,16 @@ struct module_addr_args {
> >  	int mods_cap;
> >  };
> >  
> > -static int module_callback(void *data, const char *name,
> > -			   struct module *mod, unsigned long addr)
> > +static int add_module(struct module_addr_args *args, struct module *mod)
> >  {
> > -	struct module_addr_args *args = data;
> >  	struct module **mods;
> >  
> > -	/* We iterate all modules symbols and for each we:
> > -	 * - search for it in provided addresses array
> > -	 * - if found we check if we already have the module pointer stored
> > -	 *   (we iterate modules sequentially, so we can check just the last
> > -	 *   module pointer)
> > +	/* We iterate sorted addresses and for each within module we:
> > +	 * - check if we already have the module pointer stored for it
> > +	 *   (we iterate sorted addresses sequentially, so we can check
> > +	 *   just the last module pointer)
> >  	 * - take module reference and store it
> >  	 */
> > -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
> > -		       bpf_kprobe_multi_addrs_cmp))
> > -		return 0;
> > -
> >  	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
> >  		return 0;
> 
> There'll be problems Petr mentioned.
> 
> https://lkml.org/lkml/2023/1/5/191

ok, makes sense.. I guess we could just search args->mods in here?
are you going to send new version, or should I update my patch with that?

thanks,
jirka
Zhen Lei Jan. 9, 2023, 3:11 p.m. UTC | #14
On 2023/1/9 21:48, Jiri Olsa wrote:
> On Mon, Jan 09, 2023 at 04:51:37PM +0800, Leizhen (ThunderTown) wrote:
>>
>>
>> On 2023/1/6 17:45, Jiri Olsa wrote:
>>> On Thu, Jan 05, 2023 at 10:31:12PM +0100, Jiri Olsa wrote:
>>>> On Wed, Jan 04, 2023 at 05:25:08PM +0100, Petr Mladek wrote:
>>>>> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
>>>>>> Function __module_address() can quickly return the pointer of the module
>>>>>> to which an address belongs. We do not need to traverse the symbols of all
>>>>>> modules to check whether each address in addrs[] is the start address of
>>>>>> the corresponding symbol, because register_fprobe_ips() will do this check
>>>>>> later.
>>>>
>>>> hum, for some reason I can see only replies to this patch and
>>>> not the actual patch.. I'll dig it out of the lore I guess
>>>>
>>>>>>
>>>>>> Assuming that there are m modules, each module has n symbols on average,
>>>>>> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
>>>>>> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
>>>>>> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
>>>>>> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
>>>>>> the ratio is still greater than 1. Therefore, the new method will
>>>>>> generally have better performance.
>>>>
>>>> could you try to benchmark that? I tried something similar but was not
>>>> able to get better performance
>>>
>>> hm looks like I tried the smilar thing (below) like you did,
>>
>> Yes. I just found out you're working on this improvement, too.
>>
>>> but wasn't able to get better performace
>>
>> Your implementation below is already the limit that can be optimized.
>> If the performance is not improved, it indicates that this place is
>> not the bottleneck.
>>
>>>
>>> I guess your goal is to get rid of the module arg in
>>> module_kallsyms_on_each_symbol callback that we use?
>>
>> It's not a bad thing to keep argument 'mod' for function
>> module_kallsyms_on_each_symbol(), but for kallsyms_on_each_symbol(),
>> it's completely redundant. Now these two functions often use the
>> same hook function. So I carefully analyzed get_modules_for_addrs(),
>> which is the only place that involves the use of parameter 'mod'.
>> Looks like there's a possibility of eliminating parameter 'mod'.
>>
>>> I'm ok with the change if the performace is not worse
>>
>> OK, thanks.
>>
>>>
>>> jirka
>>>
>>>
>>> ---
>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>> index 5b9008bc597b..3280c22009f1 100644
>>> --- a/kernel/trace/bpf_trace.c
>>> +++ b/kernel/trace/bpf_trace.c
>>> @@ -2692,23 +2692,16 @@ struct module_addr_args {
>>>  	int mods_cap;
>>>  };
>>>  
>>> -static int module_callback(void *data, const char *name,
>>> -			   struct module *mod, unsigned long addr)
>>> +static int add_module(struct module_addr_args *args, struct module *mod)
>>>  {
>>> -	struct module_addr_args *args = data;
>>>  	struct module **mods;
>>>  
>>> -	/* We iterate all modules symbols and for each we:
>>> -	 * - search for it in provided addresses array
>>> -	 * - if found we check if we already have the module pointer stored
>>> -	 *   (we iterate modules sequentially, so we can check just the last
>>> -	 *   module pointer)
>>> +	/* We iterate sorted addresses and for each within module we:
>>> +	 * - check if we already have the module pointer stored for it
>>> +	 *   (we iterate sorted addresses sequentially, so we can check
>>> +	 *   just the last module pointer)
>>>  	 * - take module reference and store it
>>>  	 */
>>> -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
>>> -		       bpf_kprobe_multi_addrs_cmp))
>>> -		return 0;
>>> -
>>>  	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
>>>  		return 0;
>>
>> There'll be problems Petr mentioned.
>>
>> https://lkml.org/lkml/2023/1/5/191
> 
> ok, makes sense.. I guess we could just search args->mods in here?
> are you going to send new version, or should I update my patch with that?

It's better for you to update! I'm not familiar with the bpf module.

> 
> thanks,
> jirka
> .
>
Zhen Lei Jan. 11, 2023, 8:41 a.m. UTC | #15
On 2023/1/9 23:11, Leizhen (ThunderTown) wrote:
> 
> 
> On 2023/1/9 21:48, Jiri Olsa wrote:
>> On Mon, Jan 09, 2023 at 04:51:37PM +0800, Leizhen (ThunderTown) wrote:
>>>
>>>
>>> On 2023/1/6 17:45, Jiri Olsa wrote:
>>>> On Thu, Jan 05, 2023 at 10:31:12PM +0100, Jiri Olsa wrote:
>>>>> On Wed, Jan 04, 2023 at 05:25:08PM +0100, Petr Mladek wrote:
>>>>>> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
>>>>>>> Function __module_address() can quickly return the pointer of the module
>>>>>>> to which an address belongs. We do not need to traverse the symbols of all
>>>>>>> modules to check whether each address in addrs[] is the start address of
>>>>>>> the corresponding symbol, because register_fprobe_ips() will do this check
>>>>>>> later.
>>>>>
>>>>> hum, for some reason I can see only replies to this patch and
>>>>> not the actual patch.. I'll dig it out of the lore I guess
>>>>>
>>>>>>>
>>>>>>> Assuming that there are m modules, each module has n symbols on average,
>>>>>>> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
>>>>>>> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
>>>>>>> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
>>>>>>> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
>>>>>>> the ratio is still greater than 1. Therefore, the new method will
>>>>>>> generally have better performance.
>>>>>
>>>>> could you try to benchmark that? I tried something similar but was not
>>>>> able to get better performance
>>>>
>>>> hm looks like I tried the smilar thing (below) like you did,
>>>
>>> Yes. I just found out you're working on this improvement, too.
>>>
>>>> but wasn't able to get better performace
>>>
>>> Your implementation below is already the limit that can be optimized.
>>> If the performance is not improved, it indicates that this place is
>>> not the bottleneck.
>>>
>>>>
>>>> I guess your goal is to get rid of the module arg in
>>>> module_kallsyms_on_each_symbol callback that we use?
>>>
>>> It's not a bad thing to keep argument 'mod' for function
>>> module_kallsyms_on_each_symbol(), but for kallsyms_on_each_symbol(),
>>> it's completely redundant. Now these two functions often use the
>>> same hook function. So I carefully analyzed get_modules_for_addrs(),
>>> which is the only place that involves the use of parameter 'mod'.
>>> Looks like there's a possibility of eliminating parameter 'mod'.
>>>
>>>> I'm ok with the change if the performace is not worse
>>>
>>> OK, thanks.
>>>
>>>>
>>>> jirka
>>>>
>>>>
>>>> ---
>>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>>> index 5b9008bc597b..3280c22009f1 100644
>>>> --- a/kernel/trace/bpf_trace.c
>>>> +++ b/kernel/trace/bpf_trace.c
>>>> @@ -2692,23 +2692,16 @@ struct module_addr_args {
>>>>  	int mods_cap;
>>>>  };
>>>>  
>>>> -static int module_callback(void *data, const char *name,
>>>> -			   struct module *mod, unsigned long addr)
>>>> +static int add_module(struct module_addr_args *args, struct module *mod)
>>>>  {
>>>> -	struct module_addr_args *args = data;
>>>>  	struct module **mods;
>>>>  
>>>> -	/* We iterate all modules symbols and for each we:
>>>> -	 * - search for it in provided addresses array
>>>> -	 * - if found we check if we already have the module pointer stored
>>>> -	 *   (we iterate modules sequentially, so we can check just the last
>>>> -	 *   module pointer)
>>>> +	/* We iterate sorted addresses and for each within module we:
>>>> +	 * - check if we already have the module pointer stored for it
>>>> +	 *   (we iterate sorted addresses sequentially, so we can check
>>>> +	 *   just the last module pointer)
>>>>  	 * - take module reference and store it
>>>>  	 */
>>>> -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
>>>> -		       bpf_kprobe_multi_addrs_cmp))
>>>> -		return 0;
>>>> -
>>>>  	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
>>>>  		return 0;
>>>
>>> There'll be problems Petr mentioned.
>>>
>>> https://lkml.org/lkml/2023/1/5/191
>>
>> ok, makes sense.. I guess we could just search args->mods in here?
>> are you going to send new version, or should I update my patch with that?
> 
> It's better for you to update! I'm not familiar with the bpf module.

Hi Jiri:
  Can you attach patch 1/3 when you send the new patch? There's a little
dependency. Petr has already replied OK to patch 1/3, see [1].
  Patch 3/3 is just a cleanup, I'll delay updating it after v6.3-rc1, it
also has a dependency on another patch [2].

[1] https://lkml.org/lkml/2023/1/4/627
[2] https://lkml.org/lkml/2023/1/10/534



> 
>>
>> thanks,
>> jirka
>> .
>>
>
Jiri Olsa Jan. 11, 2023, 9:53 a.m. UTC | #16
On Wed, Jan 11, 2023 at 04:41:21PM +0800, Leizhen (ThunderTown) wrote:
> 
> 
> On 2023/1/9 23:11, Leizhen (ThunderTown) wrote:
> > 
> > 
> > On 2023/1/9 21:48, Jiri Olsa wrote:
> >> On Mon, Jan 09, 2023 at 04:51:37PM +0800, Leizhen (ThunderTown) wrote:
> >>>
> >>>
> >>> On 2023/1/6 17:45, Jiri Olsa wrote:
> >>>> On Thu, Jan 05, 2023 at 10:31:12PM +0100, Jiri Olsa wrote:
> >>>>> On Wed, Jan 04, 2023 at 05:25:08PM +0100, Petr Mladek wrote:
> >>>>>> On Fri 2022-12-30 19:27:28, Zhen Lei wrote:
> >>>>>>> Function __module_address() can quickly return the pointer of the module
> >>>>>>> to which an address belongs. We do not need to traverse the symbols of all
> >>>>>>> modules to check whether each address in addrs[] is the start address of
> >>>>>>> the corresponding symbol, because register_fprobe_ips() will do this check
> >>>>>>> later.
> >>>>>
> >>>>> hum, for some reason I can see only replies to this patch and
> >>>>> not the actual patch.. I'll dig it out of the lore I guess
> >>>>>
> >>>>>>>
> >>>>>>> Assuming that there are m modules, each module has n symbols on average,
> >>>>>>> and the number of addresses 'addrs_cnt' is abbreviated as K. Then the time
> >>>>>>> complexity of the original method is O(K * log(K)) + O(m * n * log(K)),
> >>>>>>> and the time complexity of current method is O(K * (log(m) + M)), M <= m.
> >>>>>>> (m * n * log(K)) / (K * m) ==> n / log2(K). Even if n is 10 and K is 128,
> >>>>>>> the ratio is still greater than 1. Therefore, the new method will
> >>>>>>> generally have better performance.
> >>>>>
> >>>>> could you try to benchmark that? I tried something similar but was not
> >>>>> able to get better performance
> >>>>
> >>>> hm looks like I tried the smilar thing (below) like you did,
> >>>
> >>> Yes. I just found out you're working on this improvement, too.
> >>>
> >>>> but wasn't able to get better performace
> >>>
> >>> Your implementation below is already the limit that can be optimized.
> >>> If the performance is not improved, it indicates that this place is
> >>> not the bottleneck.
> >>>
> >>>>
> >>>> I guess your goal is to get rid of the module arg in
> >>>> module_kallsyms_on_each_symbol callback that we use?
> >>>
> >>> It's not a bad thing to keep argument 'mod' for function
> >>> module_kallsyms_on_each_symbol(), but for kallsyms_on_each_symbol(),
> >>> it's completely redundant. Now these two functions often use the
> >>> same hook function. So I carefully analyzed get_modules_for_addrs(),
> >>> which is the only place that involves the use of parameter 'mod'.
> >>> Looks like there's a possibility of eliminating parameter 'mod'.
> >>>
> >>>> I'm ok with the change if the performace is not worse
> >>>
> >>> OK, thanks.
> >>>
> >>>>
> >>>> jirka
> >>>>
> >>>>
> >>>> ---
> >>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> >>>> index 5b9008bc597b..3280c22009f1 100644
> >>>> --- a/kernel/trace/bpf_trace.c
> >>>> +++ b/kernel/trace/bpf_trace.c
> >>>> @@ -2692,23 +2692,16 @@ struct module_addr_args {
> >>>>  	int mods_cap;
> >>>>  };
> >>>>  
> >>>> -static int module_callback(void *data, const char *name,
> >>>> -			   struct module *mod, unsigned long addr)
> >>>> +static int add_module(struct module_addr_args *args, struct module *mod)
> >>>>  {
> >>>> -	struct module_addr_args *args = data;
> >>>>  	struct module **mods;
> >>>>  
> >>>> -	/* We iterate all modules symbols and for each we:
> >>>> -	 * - search for it in provided addresses array
> >>>> -	 * - if found we check if we already have the module pointer stored
> >>>> -	 *   (we iterate modules sequentially, so we can check just the last
> >>>> -	 *   module pointer)
> >>>> +	/* We iterate sorted addresses and for each within module we:
> >>>> +	 * - check if we already have the module pointer stored for it
> >>>> +	 *   (we iterate sorted addresses sequentially, so we can check
> >>>> +	 *   just the last module pointer)
> >>>>  	 * - take module reference and store it
> >>>>  	 */
> >>>> -	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
> >>>> -		       bpf_kprobe_multi_addrs_cmp))
> >>>> -		return 0;
> >>>> -
> >>>>  	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
> >>>>  		return 0;
> >>>
> >>> There'll be problems Petr mentioned.
> >>>
> >>> https://lkml.org/lkml/2023/1/5/191
> >>
> >> ok, makes sense.. I guess we could just search args->mods in here?
> >> are you going to send new version, or should I update my patch with that?
> > 
> > It's better for you to update! I'm not familiar with the bpf module.
> 
> Hi Jiri:
>   Can you attach patch 1/3 when you send the new patch? There's a little
> dependency. Petr has already replied OK to patch 1/3, see [1].
>   Patch 3/3 is just a cleanup, I'll delay updating it after v6.3-rc1, it
> also has a dependency on another patch [2].

ok, will do

jirka
diff mbox series

Patch

diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
index 5f3be4bc16403a5..0ff9037098bd241 100644
--- a/kernel/trace/bpf_trace.c
+++ b/kernel/trace/bpf_trace.c
@@ -2684,69 +2684,55 @@  static void symbols_swap_r(void *a, void *b, int size, const void *priv)
 	}
 }
 
-struct module_addr_args {
-	unsigned long *addrs;
-	u32 addrs_cnt;
-	struct module **mods;
-	int mods_cnt;
-	int mods_cap;
-};
-
-static int module_callback(void *data, const char *name,
-			   struct module *mod, unsigned long addr)
+static int get_modules_for_addrs(struct module ***out_mods, unsigned long *addrs, u32 addrs_cnt)
 {
-	struct module_addr_args *args = data;
-	struct module **mods;
-
-	/* We iterate all modules symbols and for each we:
-	 * - search for it in provided addresses array
-	 * - if found we check if we already have the module pointer stored
-	 *   (we iterate modules sequentially, so we can check just the last
-	 *   module pointer)
-	 * - take module reference and store it
-	 */
-	if (!bsearch(&addr, args->addrs, args->addrs_cnt, sizeof(addr),
-		       bpf_kprobe_multi_addrs_cmp))
-		return 0;
+	int i, j, err;
+	int mods_cnt = 0;
+	int mods_cap = 0;
+	struct module *mod;
+	struct module **mods = NULL;
 
-	if (args->mods && args->mods[args->mods_cnt - 1] == mod)
-		return 0;
+	for (i = 0; i < addrs_cnt; i++) {
+		mod = __module_address(addrs[i]);
+		if (!mod)
+			continue;
 
-	if (args->mods_cnt == args->mods_cap) {
-		args->mods_cap = max(16, args->mods_cap * 3 / 2);
-		mods = krealloc_array(args->mods, args->mods_cap, sizeof(*mods), GFP_KERNEL);
-		if (!mods)
-			return -ENOMEM;
-		args->mods = mods;
-	}
+		/* check if we already have the module pointer stored */
+		for (j = 0; j < mods_cnt; j++) {
+			if (mods[j] == mod)
+				break;
+		}
+		if (j < mods_cnt)
+			continue;
 
-	if (!try_module_get(mod))
-		return -EINVAL;
+		if (mods_cnt == mods_cap) {
+			struct module **new_mods;
 
-	args->mods[args->mods_cnt] = mod;
-	args->mods_cnt++;
-	return 0;
-}
+			mods_cap = max(16, mods_cap * 3 / 2);
+			new_mods = krealloc_array(mods, mods_cap, sizeof(*mods), GFP_KERNEL);
+			if (!new_mods) {
+				err = -ENOMEM;
+				goto failed;
+			}
+			mods = new_mods;
+		}
 
-static int get_modules_for_addrs(struct module ***mods, unsigned long *addrs, u32 addrs_cnt)
-{
-	struct module_addr_args args = {
-		.addrs     = addrs,
-		.addrs_cnt = addrs_cnt,
-	};
-	int err;
+		if (!try_module_get(mod)) {
+			err = -EINVAL;
+			goto failed;
+		}
 
-	/* We return either err < 0 in case of error, ... */
-	err = module_kallsyms_on_each_symbol(NULL, module_callback, &args);
-	if (err) {
-		kprobe_multi_put_modules(args.mods, args.mods_cnt);
-		kfree(args.mods);
-		return err;
+		mods[mods_cnt] = mod;
+		mods_cnt++;
 	}
 
-	/* or number of modules found if everything is ok. */
-	*mods = args.mods;
-	return args.mods_cnt;
+	*out_mods = mods;
+	return mods_cnt;
+
+failed:
+	kprobe_multi_put_modules(mods, mods_cnt);
+	kfree(mods);
+	return err;
 }
 
 int bpf_kprobe_multi_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)
@@ -2859,13 +2845,6 @@  int bpf_kprobe_multi_link_attach(const union bpf_attr *attr, struct bpf_prog *pr
 		       bpf_kprobe_multi_cookie_cmp,
 		       bpf_kprobe_multi_cookie_swap,
 		       link);
-	} else {
-		/*
-		 * We need to sort addrs array even if there are no cookies
-		 * provided, to allow bsearch in get_modules_for_addrs.
-		 */
-		sort(addrs, cnt, sizeof(*addrs),
-		       bpf_kprobe_multi_addrs_cmp, NULL);
 	}
 
 	err = get_modules_for_addrs(&link->mods, addrs, cnt);