Re: [PATCH v3] Tracepoint Tapset for Memory Subsystem

[Rajasekhar Duddu <rajduddu at linux dot vnet dot ibm dot com>]

Hi Frank, 
	thanks for your comments. I am answering for your
questions here and in my next patch I will apply all the possible
changes. 

On Tue, Oct 06, 2009 at 03:01:09PM -0400, Frank Ch. Eigler wrote:
> Rajasekhar Duddu <rajduddu@linux.vnet.ibm.com> writes:
> 
> > [...]
> > +/* Function to convert the GFP_FLAGS . */
> > +
> > +function gfp_flag_str:string (gfp_flag:long)
> > +%{
> > +int flags = (int)THIS->gfp_flag;
> > +THIS->__retvalue[0] = '\0';
> > +
> > +#ifdef __GFP_HIGH
> > +        if (flags & __GFP_HIGH)
> > +                strlcat (THIS->__retvalue, "GFP_HIGH",MAXSTRINGLEN);
> > +#endif
> > +
> > +#ifdef __GFP_WAIT
> > +        if (flags & __GFP_WAIT)
> > +                strlcat (THIS->__retvalue, "GFP_WAIT",MAXSTRINGLEN);
> > +#endif
> > +
> > +#ifdef __GFP_IO
> > +        if (flags & __GFP_IO)
> > +                strlcat (THIS->__retvalue, "|GFP_IO",MAXSTRINGLEN);
> > +#endif
> 
> Why no "|" before GFP_HIGH/GFP_WAIT?

All the other Flags (except GFP_WAIT/GFP_HIGH) are composition of 
one or the other Flags, and the composition starts with either
GFP_HIGH or GFP_WAIT, so there is no "|" before GFP_HIGH/GFP_WAIT.

> Also, why no "__" before the stringified version?
This I will apply in the next patch.
> 
> 
> > +#ifdef __GFP_FS
> > +        if (flags & __GFP_FS)
> > +                strlcat (THIS->__retvalue, "|GFP_FS",MAXSTRINGLEN);
> > +#endif
> 
> (How about a macro to generate all these near-identical branches?)
> 
> 
Sure I will have macro something like this:
%{

#define FLAG_TEST(TYPE)  if (flags & TYPE) { strlcat \
(THIS->__retvalue, #TYPE,MAXSTRINGLEN); }

%}

and I will use that in my gfp_flag_str as fallowing.

#ifdef __GFP_HIGH
        FLAG_TEST(__GFP_HIGH);
#endif

> > +%}
> 
> 
> > +/**
> > + * probe vm.kmalloc - Fires when <command>kmalloc</command> is requested.
> > + * @call_site: Address of the caller function.
> > + * @caller_function: Name of the caller function.
> > + * @bytes_req: Requested Bytes
> > + * @bytes_alloc: Allocated Bytes
> > + * @gfp_flags: type of kmemory to allocate
> > + * @ptr: Pointer to the kmemory allocated
> > + */
> > +
> > +probe vm.kmalloc = kernel.trace("kmalloc") {
> > +	name = "kmalloc"
> > +	call_site = $call_site
> > +	caller_function = symname(call_site)
> > +	bytes_req = $bytes_req
> > +	bytes_alloc = $bytes_alloc
> > +	gfp_flags = gfp_flag_str($gfp_flags)
> > +	ptr = $ptr
> > +}
> 
> Nice.  I thought that the raison d'etre for these aliases was to
> abstract the presence or absence of tracepoints, so is there no
> fallback kprobe available?  Something like this:
> 
Fallback kprobe is not available for other memory functions because
the variables exported by them are will be modified.
> 
> > +probe __vm.kfree.kp = kernel.function("kfree") {
> > +	name = "kfree"
> > +	call_site = "0"
> 
> (Note though that this will fail type checking on a non-tracepoint
> kernel -- have you tried it? -- it should be just 0 instead of "0".)
> 
This also I will apply in my next patch.
> > +	caller_function = "unknown"
> > +	ptr = $x
> > +}
> > +
> > +probe __vm.kfree.tp = kernel.trace("kfree") {
> > +	name = "kfree"
> > +	call_site = $call_site
> > +	caller_function = symname(call_site)
> > +	ptr = $ptr
> > +}
> > +
> > +/**
> > + * probe vm.kfree - Fires when <command>kfree</comand> is requested.
> > + * @call_site: Address of the caller function (displayed if available)
> > + * @caller_function - Name of the caller function (displayed if available)
> > + * @ptr: Pointer to the kmemory allocated which is returned by kmalloc
> > + */
> > +probe vm.kfree = __vm.kfree.tp !,
> > +		   __vm.kfree.kp
> > +{}
> 
> Right.
> 
> 
> > +/**
> > + * probe vm.kmalloc_node - Fires when <command>kmalloc_node</command> is requested.
> > + * @call_site: Address of the caller function.
> > + * @caller_function: Name of the caller function.
> > + * @bytes_req: Requested Bytes
> > + * @bytes_alloc: Allocated Bytes
> > + * @gfp_flags: Type of kmemory to allocate
> > + * @ptr: Pointer to the kmemory allocated
> > + */
> 
> Please, no "<command>" markup in there, it is not valid.
> 
This I will cange in my next patch.
> 
> > +probe vm.kmalloc_node = kernel.trace("kmalloc_node")? {
> > [...]
> 
> Why is this marked with "?"?

Kmalloc_node will be called when "CONFIG_NUMA" is defined, so I have
used "?" to make it an optional.
> 
> 
> > --- a/testsuite/buildok/vm.tracepoints.stp	1969-12-31 19:00:00.000000000 -0500
> > +++ b/testsuite/buildok/vm.tracepoints.stp	2009-10-02 10:59:20.000000000 -0400
> > @@ -0,0 +1,32 @@
> > +#!/usr/bin/stp -up4
> 
> Other similar test cases just use
> 
> #! stap -up4
> 
This also I will change in my next patch.
> 

Thanks
-- 
Rajasekhar Duddu (rajduddu@linux.vnet.ibm.com),
Linux on System z - CSVT, IBM LTC, Bangalore.