Dear David,
During my work on NOMMU system (mm/nommu.c), I saw definition and
usage of kenter/kleave/kdebug macros. These macros are compiled as
empty because of "#if 0" construction.
45 #if 0
46 #define kenter(FMT, ...) \
47 printk(KERN_DEBUG "==> %s("FMT")\n", __func__, ##__VA_ARGS__)
48 #define kleave(FMT, ...) \
49 printk(KERN_DEBUG "<== %s()"FMT"\n", __func__, ##__VA_ARGS__)
50 #define kdebug(FMT, ...) \
51 printk(KERN_DEBUG "xxx" FMT"yyy\n", ##__VA_ARGS__)
52 #else
53 #define kenter(FMT, ...) \
54 no_printk(KERN_DEBUG "==> %s("FMT")\n", __func__, ##__VA_ARGS__)
55 #define kleave(FMT, ...) \
56 no_printk(KERN_DEBUG "<== %s()"FMT"\n", __func__, ##__VA_ARGS__)
57 #define kdebug(FMT, ...) \
58 no_printk(KERN_DEBUG FMT"\n", ##__VA_ARGS__)
59 #endif
This code was changed in 2009 [1] and similar definitions can be found
in 9 other files [2]. The protection of these definitions is slightly
different. There are places with "#if 0" protection and others with
"#if defined(__KDEBUG)" protection. __KDEBUG is supposed to be
inserted by GCC.
My question is how we should handle such duplicated debug print code?
As possible solutions, I see five options:
1. Leave it as is.
2. Move it to general include file (for example linux/printk.h) and
commonize the output to be consistent between different kdebug users.
3. Add CONFIG_*_DEBUG definition for every kdebug user.
4. Move everything to "#if 0" construction.
5. Move everything to "#if defined(__KDEBUG)" construction.
What do you think?
[1] commit 8feae13110d60cc6287afabc2887366b0eb226c2
Author: David Howells <[email protected]>
Date: Thu Jan 8 12:04:47 2009 +0000
[2] List of all files there kdebug was defined:
* arch/mn10300/kernel/mn10300-serial.c
* arch/mn10300/mm/misalignment.c
* fs/cachefiles/internal.h
* fs/afs/internal.h
* fs/fscache/internal.h
* fs/binfmt_elf_fdpic.c
* kernel/cred.c
* mm/nommu.c
* net/rxrpc/ar-internal.h
* security/keys/internal.h
Thank you.
--
Leon Romanovsky | Independent Linux Consultant
http://www.leon.nu | [email protected]
On Sat, 2015-05-16 at 20:01 +0300, Leon Romanovsky wrote:
> Dear David,
>
> During my work on NOMMU system (mm/nommu.c), I saw definition and
> usage of kenter/kleave/kdebug macros. These macros are compiled as
> empty because of "#if 0" construction.
> 45 #if 0
> 46 #define kenter(FMT, ...) \
> 47 printk(KERN_DEBUG "==> %s("FMT")\n", __func__, ##__VA_ARGS__)
> 48 #define kleave(FMT, ...) \
> 49 printk(KERN_DEBUG "<== %s()"FMT"\n", __func__, ##__VA_ARGS__)
> 50 #define kdebug(FMT, ...) \
> 51 printk(KERN_DEBUG "xxx" FMT"yyy\n", ##__VA_ARGS__)
> 52 #else
> 53 #define kenter(FMT, ...) \
> 54 no_printk(KERN_DEBUG "==> %s("FMT")\n", __func__, ##__VA_ARGS__)
> 55 #define kleave(FMT, ...) \
> 56 no_printk(KERN_DEBUG "<== %s()"FMT"\n", __func__, ##__VA_ARGS__)
> 57 #define kdebug(FMT, ...) \
> 58 no_printk(KERN_DEBUG FMT"\n", ##__VA_ARGS__)
> 59 #endif
[]
> My question is how we should handle such duplicated debug print code?
> As possible solutions, I see five options:
> 1. Leave it as is.
> 2. Move it to general include file (for example linux/printk.h) and
> commonize the output to be consistent between different kdebug users.
> 3. Add CONFIG_*_DEBUG definition for every kdebug user.
> 4. Move everything to "#if 0" construction.
> 5. Move everything to "#if defined(__KDEBUG)" construction.
6: delete the macros and uses
On Sat, May 16, 2015 at 8:09 PM, Joe Perches <[email protected]> wrote:
> On Sat, 2015-05-16 at 20:01 +0300, Leon Romanovsky wrote:
[]
>> My question is how we should handle such duplicated debug print code?
>> As possible solutions, I see five options:
>> 1. Leave it as is.
>> 2. Move it to general include file (for example linux/printk.h) and
>> commonize the output to be consistent between different kdebug users.
>> 3. Add CONFIG_*_DEBUG definition for every kdebug user.
>> 4. Move everything to "#if 0" construction.
>> 5. Move everything to "#if defined(__KDEBUG)" construction.
>
> 6: delete the macros and uses
Thank you, It is indeed possible option, since in last six years there
were no attempts to "open" this code.
--
Leon Romanovsky | Independent Linux Consultant
http://www.leon.nu | [email protected]
Leon Romanovsky <[email protected]> wrote:
> During my work on NOMMU system (mm/nommu.c), I saw definition and
> usage of kenter/kleave/kdebug macros. These macros are compiled as
> empty because of "#if 0" construction.
Because you only need them if you're debugging. They shouldn't, generally, be
turned on upstream.
> This code was changed in 2009 [1] and similar definitions can be found
> in 9 other files [2]. The protection of these definitions is slightly
> different. There are places with "#if 0" protection and others with
> "#if defined(__KDEBUG)" protection. __KDEBUG is supposed to be
> inserted by GCC.
I can turn on all the macros in a file just be #defining __KDEBUG at the top.
When I first did this, pr_xxx() didn't exist.
Note that the macros in afs, cachefiles, fscache and rxrpc are more complex
than a grep tells you. There are _enter(), _leave() and _debug() macros which
are conditional via a module parameter. These are trivially individually
enableable during debugging by changing the initial underscore to a 'k'. They
are otherwise enableable by module parameter (macros are individually
selectable) or enableably by file __KDEBUG. These are well used. Note that
just turning them all into pr_devel() would represent a loss of useful
function.
The ones in the keys directory are also very well used, though they aren't
externally selectable. I've added functionality to the debugging, but haven't
necessarily needed to backport it to earlier variants.
For the mn10300 macros, I would just recommend leaving them as is.
For the nommu macros, you could convert them to pr_devel() - but putting all
the information in the kenter/kleave/kdebug macro into each pr_devel macro
would be more intrusive in the code since you'd have to move the stuff out of
there macro definition into each caller. You could also reexpress the macros
in terms of pr_devel and get rid of the conditional. OTOH, there's not that
much in the nommu code, so you could probably slim down a lot of what's
printed.
For the cred macro, just convert to pr_devel() or pr_debug() and make pr_fmt
insert current->comm and current->pid.
> 2. Move it to general include file (for example linux/printk.h) and
> commonize the output to be consistent between different kdebug users.
I would quite like to see kenter() and kleave() be moved to printk.h,
expressed in a similar way to pr_devel() or pr_debug() (and perhaps renamed
pr_enter() and pr_leave()) but separately so they can be enabled separately.
OTOH, possibly they should be enableable by compilation block rather than by
macro set.
The main thing I like out of the ones in afs, cachefiles, fscache and rxrpc is
the ability to just turn on a few across a bunch of files so as not to get
overwhelmed by data.
David
Sorry for reposting.
On Mon, May 18, 2015 at 1:31 PM, David Howells <[email protected]> wrote:
>
> I can turn on all the macros in a file just be #defining __KDEBUG at the
> top.
> When I first did this, pr_xxx() didn't exist.
>
> Note that the macros in afs, cachefiles, fscache and rxrpc are more
> complex
> than a grep tells you. There are _enter(), _leave() and _debug() macros
> which
> are conditional via a module parameter. These are trivially individually
> enableable during debugging by changing the initial underscore to a 'k'.
> They
> are otherwise enableable by module parameter (macros are individually
> selectable) or enableably by file __KDEBUG. These are well used. Note
> that
> just turning them all into pr_devel() would represent a loss of useful
> function.
>
> The ones in the keys directory are also very well used, though they aren't
> externally selectable. I've added functionality to the debugging, but
> haven't
> necessarily needed to backport it to earlier variants.
>
> For the mn10300 macros, I would just recommend leaving them as is.
>
> For the nommu macros, you could convert them to pr_devel() - but putting
> all
> the information in the kenter/kleave/kdebug macro into each pr_devel macro
> would be more intrusive in the code since you'd have to move the stuff out
> of
> there macro definition into each caller. You could also reexpress the
> macros
> in terms of pr_devel and get rid of the conditional. OTOH, there's not
> that
> much in the nommu code, so you could probably slim down a lot of what's
> printed.
>
> For the cred macro, just convert to pr_devel() or pr_debug() and make
> pr_fmt
> insert current->comm and current->pid.
>
> > 2. Move it to general include file (for example linux/printk.h) and
> > commonize the output to be consistent between different kdebug users.
>
> I would quite like to see kenter() and kleave() be moved to printk.h,
> expressed in a similar way to pr_devel() or pr_debug() (and perhaps
> renamed
> pr_enter() and pr_leave()) but separately so they can be enabled
> separately.
> OTOH, possibly they should be enableable by compilation block rather than
> by
> macro set.
>
> The main thing I like out of the ones in afs, cachefiles, fscache and
> rxrpc is
> the ability to just turn on a few across a bunch of files so as not to get
> overwhelmed by data.
Blind conversion to pr_debug will blow the code because it will be always
compiled in. In current implementation, it replaced by empty functions which
is thrown by compiler.
Additionally, It looks like the output of these macros can be viewed by
ftrace mechanism.
Maybe we should delete them from mm/nommu.c as was pointed by Joe?
>
>
> David
--
Leon Romanovsky | Independent Linux Consultant
http://www.leon.nu | [email protected]
--
Leon Romanovsky | Independent Linux Consultant
http://www.leon.nu | [email protected]
Leon Romanovsky <[email protected]> wrote:
> Blind conversion to pr_debug will blow the code because it will be always
> compiled in.
No, it won't.
> Additionally, It looks like the output of these macros can be viewed by ftrace
> mechanism.
*blink* It can?
> Maybe we should delete them from mm/nommu.c as was pointed by Joe?
Why?
David
On Mon, May 18, 2015 at 4:29 PM, David Howells <[email protected]> wrote:
> Leon Romanovsky <[email protected]> wrote:
>
>> Blind conversion to pr_debug will blow the code because it will be always
>> compiled in.
>
> No, it won't.
Sorry, you are right.
>
>> Additionally, It looks like the output of these macros can be viewed by ftrace
>> mechanism.
>
> *blink* It can?
I was under strong impression that "function" and "function_graph"
tracers will give similar kenter/kleave information. Do I miss
anything important, except the difference in output format?
>
>> Maybe we should delete them from mm/nommu.c as was pointed by Joe?
>
> Why?
If ftrace is sufficient to get the debug information, there will no
need to duplicate it.
>
> David
Leon Romanovsky <[email protected]> wrote:
> >> Additionally, It looks like the output of these macros can be viewed by
> >> ftrace mechanism.
> >
> > *blink* It can?
> I was under strong impression that "function" and "function_graph"
> tracers will give similar kenter/kleave information. Do I miss
> anything important, except the difference in output format?
>
> >
> >> Maybe we should delete them from mm/nommu.c as was pointed by Joe?
> >
> > Why?
> If ftrace is sufficient to get the debug information, there will no
> need to duplicate it.
It isn't sufficient. It doesn't store the parameters or the return value, it
doesn't distinguish the return path in a function when there's more than one,
eg.:
kleave(" = %d [val]", ret);
vs:
kleave(" = %lx", result);
in do_mmap_pgoff() and it doesn't permit you to retrieve data from where the
argument pointers that you don't have pointed to, eg.:
kenter("%p{%d}", region, region->vm_usage);
David
On Mon, May 18, 2015 at 6:20 PM, David Howells <[email protected]> wrote:
> Leon Romanovsky <[email protected]> wrote:
>
>> >> Additionally, It looks like the output of these macros can be viewed by
>> >> ftrace mechanism.
>> >
>> > *blink* It can?
>> I was under strong impression that "function" and "function_graph"
>> tracers will give similar kenter/kleave information. Do I miss
>> anything important, except the difference in output format?
>>
>> >
>> >> Maybe we should delete them from mm/nommu.c as was pointed by Joe?
>> >
>> > Why?
>> If ftrace is sufficient to get the debug information, there will no
>> need to duplicate it.
>
> It isn't sufficient. It doesn't store the parameters or the return value, it
> doesn't distinguish the return path in a function when there's more than one,
> eg.:
>
> kleave(" = %d [val]", ret);
>
> vs:
>
> kleave(" = %lx", result);
>
> in do_mmap_pgoff() and it doesn't permit you to retrieve data from where the
> argument pointers that you don't have pointed to, eg.:
>
> kenter("%p{%d}", region, region->vm_usage);
>
> David
Thanks you for explanation, I'll send the patch in near future.
--
Leon Romanovsky | Independent Linux Consultant
http://www.leon.nu | [email protected]