Re : sparsemem usage

Re : sparsemem usage

[moreau francis]

KAMEZAWA Hiroyuki wrote:
> On Thu, 10 Aug 2006 14:40:52 +0200 (CEST)
> moreau francis <francis_moreau2000@yahoo.fr> wrote: 
>>> BTW, ioresouce information (see kernel/resouce.c)
>>>
>>> [kamezawa@aworks Development]$ cat /proc/iomem | grep RAM
>>> 00000000-0009fbff : System RAM
>>> 000a0000-000bffff : Video RAM area
>>> 00100000-2dfeffff : System RAM
>>>
>>> is not enough ?
>>>
>> well actually you show that to get a really simple information, ie does
>> a page exist ?, we need to parse some kernel data structures like 
>> ioresource (which is, IMHO, hackish) or duplicate in each architecture
>> some data to keep track of existing pages.
>>
> 
> becasue memory map from e820(x86) or efi(ia64) are registered to iomem_resource,
> we should avoid duplicates that information. kdump and memory hotplug uses
> this information. (memory hotplug updates this iomem_resource.)
> 
> Implementing "page_is_exist" function based on ioresouce is one of generic
> and rubust way to go, I think.
> (if performance of list walking is problem, enhancing ioresouce code is
>  better.)
>  

Why not implementing page_exist() by simply using mem_map[] ? When
allocating mem_map[], we can just fill it with a special value. And
then when registering memory area, we clear this special value with
the "reserved" value. Hence for flatmem model, we can have:

#define page_exist(pfn)        (mem_map[pfn] != SPECIAL_VALUE)

and it should work for sparsemem too and other models that will use
mem_map[].

Another point, is page_exist() going to replace page_valid() ?
I mean page_exist() is going to be something more accurate than
page_valid(). All tests on page_valid() _only_ will be fine to test
page_exist(). But all tests such:

    if (page_valid(x) && page_is_ram(x))

can be replaced by

    if (page_exist(x))

So, again, why not simply improving page_valid() definition rather
than introduce a new service ?

Francis

Re: Re : sparsemem usage

[KAMEZAWA Hiroyuki]

On Thu, 10 Aug 2006 14:46:01 +0000 (GMT)
moreau francis <francis_moreau2000@yahoo.fr> wrote:
> 
> Why not implementing page_exist() by simply using mem_map[] ? When
> allocating mem_map[], we can just fill it with a special value. And
> then when registering memory area, we clear this special value with
> the "reserved" value. Hence for flatmem model, we can have:
> 
> #define page_exist(pfn)        (mem_map[pfn] != SPECIAL_VALUE)
>  
putting a special value to a page struct at mem_map + pfn ?

> and it should work for sparsemem too and other models that will use
> mem_map[].
> 
> Another point, is page_exist() going to replace page_valid() ?
what is page_valid() here ? pfn_valid() (in current kernel) ?

> I mean page_exist() is going to be something more accurate than
> page_valid(). All tests on page_valid() _only_ will be fine to test
> page_exist(). But all tests such:
> 
>     if (page_valid(x) && page_is_ram(x))
> 
> can be replaced by
> 
>     if (page_exist(x))
> 
> So, again, why not simply improving page_valid() definition rather
> than introduce a new service ?
> 
I welcome to do that if implementation is sane.
pfn_valid() --- check there is a page struct
page_exist() --- check there is a physical memory.

but discussing without patch is not very good. please post your patch.
Then we can discuss more concrete things.

-Kame

Re : Re : sparsemem usage

[moreau francis]

KAMEZAWA Hiroyuki wrote:
> On Thu, 10 Aug 2006 14:46:01 +0000 (GMT)
> moreau francis <francis_moreau2000@yahoo.fr> wrote:
>> Why not implementing page_exist() by simply using mem_map[] ? When
>> allocating mem_map[], we can just fill it with a special value. And
>> then when registering memory area, we clear this special value with
>> the "reserved" value. Hence for flatmem model, we can have:
>>
>> #define page_exist(pfn)        (mem_map[pfn] != SPECIAL_VALUE)
>>  
> putting a special value to a page struct at mem_map + pfn ?

yes

> 
>> and it should work for sparsemem too and other models that will use
>> mem_map[].
>>
>> Another point, is page_exist() going to replace page_valid() ?
> what is page_valid() here ? pfn_valid() (in current kernel) ?

sorry I was meaning pfn_valid() instead of page_valid() in the
whole email.

> 
>> I mean page_exist() is going to be something more accurate than
>> page_valid(). All tests on page_valid() _only_ will be fine to test
>> page_exist(). But all tests such:
>>
>>     if (page_valid(x) && page_is_ram(x))
>>
>> can be replaced by
>>
>>     if (page_exist(x))
>>
>> So, again, why not simply improving page_valid() definition rather
>> than introduce a new service ?
>>

s/page_valid/pfn_valid

> I welcome to do that if implementation is sane.
> pfn_valid() --- check there is a page struct
> page_exist() --- check there is a physical memory.
> 

new definition of pfn_valid() would be "a physical page exists". And
this definition imply the old one "it's safe to read the page struct *"

> but discussing without patch is not very good. please post your patch.
> Then we can discuss more concrete things.
> 

Since I'm not kernel hacker, or rather a newbie one, I try to make sure
that it worth to dig in that direction before working hard to write a
patch.

thanks

Francis

Re: Re : sparsemem usage

[Andy Whitcroft]

moreau francis wrote:
> KAMEZAWA Hiroyuki wrote:
>> On Thu, 10 Aug 2006 14:40:52 +0200 (CEST)
>> moreau francis <francis_moreau2000@yahoo.fr> wrote: 
>>>> BTW, ioresouce information (see kernel/resouce.c)
>>>>
>>>> [kamezawa@aworks Development]$ cat /proc/iomem | grep RAM
>>>> 00000000-0009fbff : System RAM
>>>> 000a0000-000bffff : Video RAM area
>>>> 00100000-2dfeffff : System RAM
>>>>
>>>> is not enough ?
>>>>
>>> well actually you show that to get a really simple information, ie does
>>> a page exist ?, we need to parse some kernel data structures like 
>>> ioresource (which is, IMHO, hackish) or duplicate in each architecture
>>> some data to keep track of existing pages.
>>>
>> becasue memory map from e820(x86) or efi(ia64) are registered to iomem_resource,
>> we should avoid duplicates that information. kdump and memory hotplug uses
>> this information. (memory hotplug updates this iomem_resource.)
>>
>> Implementing "page_is_exist" function based on ioresouce is one of generic
>> and rubust way to go, I think.
>> (if performance of list walking is problem, enhancing ioresouce code is
>>  better.)
>>  
> 
> Why not implementing page_exist() by simply using mem_map[] ? When
> allocating mem_map[], we can just fill it with a special value. And
> then when registering memory area, we clear this special value with
> the "reserved" value. Hence for flatmem model, we can have:
> 
> #define page_exist(pfn)        (mem_map[pfn] != SPECIAL_VALUE)
> 
> and it should work for sparsemem too and other models that will use
> mem_map[].

The mem_map isn't a pointer, its a physical structure.  We have a 
special value to tell you if the page is usable within that, thats 
called PG_reserved.  If this page is reserved the kernel can't touch it, 
can't look at it.

> Another point, is page_exist() going to replace page_valid() ?
> I mean page_exist() is going to be something more accurate than
> page_valid(). All tests on page_valid() _only_ will be fine to test
> page_exist(). But all tests such:
> 
>     if (page_valid(x) && page_is_ram(x))
> 
> can be replaced by
> 
>     if (page_exist(x))
> 
> So, again, why not simply improving page_valid() definition rather
> than introduce a new service ?

Whilst I can understand that not knowing if a page is real or not is 
perhaps unappealing, I've yet to see any case where we need or care. 
Changing things to make things 'nicer' interlectually is sometimes 
worthwhile.  But what is the user here.

The only consumer you have shown is show_mem() which is a debug 
function, and that only dumps out the current memory counts.  Its not 
clear it cares to really know if a page is real or not.

-ap

Re : Re : sparsemem usage

[moreau francis]

Andy Whitcroft wrote:
> moreau francis wrote:
>> KAMEZAWA Hiroyuki wrote:
>>> On Thu, 10 Aug 2006 14:40:52 +0200 (CEST)
>>> moreau francis <francis_moreau2000@yahoo.fr> wrote:
>>>>> BTW, ioresouce information (see kernel/resouce.c)
>>>>>
>>>>> [kamezawa@aworks Development]$ cat /proc/iomem | grep RAM
>>>>> 00000000-0009fbff : System RAM
>>>>> 000a0000-000bffff : Video RAM area
>>>>> 00100000-2dfeffff : System RAM
>>>>>
>>>>> is not enough ?
>>>>>
>>>> well actually you show that to get a really simple information, ie does
>>>> a page exist ?, we need to parse some kernel data structures like
>>>> ioresource (which is, IMHO, hackish) or duplicate in each architecture
>>>> some data to keep track of existing pages.
>>>>
>>> becasue memory map from e820(x86) or efi(ia64) are registered to
>>> iomem_resource,
>>> we should avoid duplicates that information. kdump and memory hotplug
>>> uses
>>> this information. (memory hotplug updates this iomem_resource.)
>>>
>>> Implementing "page_is_exist" function based on ioresouce is one of
>>> generic
>>> and rubust way to go, I think.
>>> (if performance of list walking is problem, enhancing ioresouce code is
>>>  better.)
>>>  
>>
>> Why not implementing page_exist() by simply using mem_map[] ? When
>> allocating mem_map[], we can just fill it with a special value. And
>> then when registering memory area, we clear this special value with
>> the "reserved" value. Hence for flatmem model, we can have:
>>
>> #define page_exist(pfn)        (mem_map[pfn] != SPECIAL_VALUE)
>>
>> and it should work for sparsemem too and other models that will use
>> mem_map[].
> 
> The mem_map isn't a pointer, its a physical structure.  We have a

ok

> special value to tell you if the page is usable within that, thats
> called PG_reserved.  If this page is reserved the kernel can't touch it,
> can't look at it.

can't we introduce a new special value, such as "PG_real" ?

> 
>> Another point, is page_exist() going to replace page_valid() ?
>> I mean page_exist() is going to be something more accurate than
>> page_valid(). All tests on page_valid() _only_ will be fine to test
>> page_exist(). But all tests such:
>>
>>     if (page_valid(x) && page_is_ram(x))
>>
>> can be replaced by
>>
>>     if (page_exist(x))
>>
>> So, again, why not simply improving page_valid() definition rather
>> than introduce a new service ?
> 
> Whilst I can understand that not knowing if a page is real or not is
> perhaps unappealing, I've yet to see any case where we need or care.
> Changing things to make things 'nicer' interlectually is sometimes
> worthwhile.  But what is the user here.
> 
> The only consumer you have shown is show_mem() which is a debug
> function, and that only dumps out the current memory counts.  Its not
> clear it cares to really know if a page is real or not.
> 

I understand your point of view, but even if it's a debug function,
it must exist and report correct information. And my point is that
I think it should be really easy to implement :) that by using
a new "special value". Can you confirm that it's really easy to
implement that ?

thanks

Francis

Re: Re : Re : sparsemem usage

[Andy Whitcroft]

moreau francis wrote:
> Andy Whitcroft wrote:
>> moreau francis wrote:
>>> KAMEZAWA Hiroyuki wrote:
>>>> On Thu, 10 Aug 2006 14:40:52 +0200 (CEST)
>>>> moreau francis <francis_moreau2000@yahoo.fr> wrote:
>>>>>> BTW, ioresouce information (see kernel/resouce.c)
>>>>>>
>>>>>> [kamezawa@aworks Development]$ cat /proc/iomem | grep RAM
>>>>>> 00000000-0009fbff : System RAM
>>>>>> 000a0000-000bffff : Video RAM area
>>>>>> 00100000-2dfeffff : System RAM
>>>>>>
>>>>>> is not enough ?
>>>>>>
>>>>> well actually you show that to get a really simple information, ie does
>>>>> a page exist ?, we need to parse some kernel data structures like
>>>>> ioresource (which is, IMHO, hackish) or duplicate in each architecture
>>>>> some data to keep track of existing pages.
>>>>>
>>>> becasue memory map from e820(x86) or efi(ia64) are registered to
>>>> iomem_resource,
>>>> we should avoid duplicates that information. kdump and memory hotplug
>>>> uses
>>>> this information. (memory hotplug updates this iomem_resource.)
>>>>
>>>> Implementing "page_is_exist" function based on ioresouce is one of
>>>> generic
>>>> and rubust way to go, I think.
>>>> (if performance of list walking is problem, enhancing ioresouce code is
>>>>  better.)
>>>>  
>>> Why not implementing page_exist() by simply using mem_map[] ? When
>>> allocating mem_map[], we can just fill it with a special value. And
>>> then when registering memory area, we clear this special value with
>>> the "reserved" value. Hence for flatmem model, we can have:
>>>
>>> #define page_exist(pfn)        (mem_map[pfn] != SPECIAL_VALUE)
>>>
>>> and it should work for sparsemem too and other models that will use
>>> mem_map[].
>> The mem_map isn't a pointer, its a physical structure.  We have a
> 
> ok
> 
>> special value to tell you if the page is usable within that, thats
>> called PG_reserved.  If this page is reserved the kernel can't touch it,
>> can't look at it.
> 
> can't we introduce a new special value, such as "PG_real" ?
> 
>>> Another point, is page_exist() going to replace page_valid() ?
>>> I mean page_exist() is going to be something more accurate than
>>> page_valid(). All tests on page_valid() _only_ will be fine to test
>>> page_exist(). But all tests such:
>>>
>>>     if (page_valid(x) && page_is_ram(x))
>>>
>>> can be replaced by
>>>
>>>     if (page_exist(x))
>>>
>>> So, again, why not simply improving page_valid() definition rather
>>> than introduce a new service ?
>> Whilst I can understand that not knowing if a page is real or not is
>> perhaps unappealing, I've yet to see any case where we need or care.
>> Changing things to make things 'nicer' interlectually is sometimes
>> worthwhile.  But what is the user here.
>>
>> The only consumer you have shown is show_mem() which is a debug
>> function, and that only dumps out the current memory counts.  Its not
>> clear it cares to really know if a page is real or not.
>>
> 
> I understand your point of view, but even if it's a debug function,
> it must exist and report correct information. And my point is that
> I think it should be really easy to implement :) that by using
> a new "special value". Can you confirm that it's really easy to
> implement that ?

It does produce real numbers, it tells you how many reserved pages you 
have.  The places that this is triggered we are interested in why we 
have no memory left.  We are not interested in how many pages are known 
but reserved as against how many pages are backed by page*'s but are 
really holes; they are mearly pages we can't use out of the total we are 
tracking.  We care about how many are not reserved, and how many of 
those are available.

It would be 'as simple' as adding a PG_real page bit except for two things:

1) page flags bits are seriously short supply; there are some 24 
available of which 22 are in use.  Any new user of a bit would have to 
be an extremely valuable change with major benefit to the kernel, and

2) if you were to try and populate a PG_real flag it would need to be 
populated for _all_ architectures (and there are a lot) for it to be of 
any use.  As you have already noted there is no consistent way to find 
out whether a page is ram so it would be major exercise to get these 
bits setup during boot.

I think we should take (2) as a hint here.  If we don't have a 
consistent interface for finding whether a page is real or not, we 
obviously have no general need of that information in the kernel.

Yes we obviously care if we can use a page, but we do not care if the 
page is unusable because it contains an ACPI table or the video driver 
BIOS or there is a memory hole.  Its either usable (!PG_reserved) or its 
not (PG_reserved).

-apw

Re : Re : Re : sparsemem usage

[moreau francis]

Andy Whitcroft wrote:
> It does produce real numbers, it tells you how many reserved pages you
> have.  The places that this is triggered we are interested in why we
> have no memory left.  We are not interested in how many pages are known
> but reserved as against how many pages are backed by page*'s but are
> really holes; they are mearly pages we can't use out of the total we are
> tracking.  We care about how many are not reserved, and how many of
> those are available.
> 
> It would be 'as simple' as adding a PG_real page bit except for two things:
> 
> 1) page flags bits are seriously short supply; there are some 24
> available of which 22 are in use.  Any new user of a bit would have to
> be an extremely valuable change with major benefit to the kernel, and
> 

It's indeed an issue. Could we instead use a combination of flags
that can't happen together. For example PG_Free|PG_Reserved ?

> 2) if you were to try and populate a PG_real flag it would need to be
> populated for _all_ architectures (and there are a lot) for it to be of
> any use.  As you have already noted there is no consistent way to find
> out whether a page is ram so it would be major exercise to get these
> bits setup during boot.
> 
> I think we should take (2) as a hint here.  If we don't have a
> consistent interface for finding whether a page is real or not, we
> obviously have no general need of that information in the kernel.
> 

or maybe _because_ we don't have a consistent interface for finding 
whether a page is real or not, we end up with a strange thing called
page_is_ram() which could be the same for all arch and be implemented
very simply.

BTW, can you try in a linux tree:

$ grep -r page_is_ram arch/

and see how it's implemented...

thanks

Francis

Re: Re : Re : Re : sparsemem usage

[Andy Whitcroft]

moreau francis wrote:
> Andy Whitcroft wrote:
>> It does produce real numbers, it tells you how many reserved pages you
>> have.  The places that this is triggered we are interested in why we
>> have no memory left.  We are not interested in how many pages are known
>> but reserved as against how many pages are backed by page*'s but are
>> really holes; they are mearly pages we can't use out of the total we are
>> tracking.  We care about how many are not reserved, and how many of
>> those are available.
>>
>> It would be 'as simple' as adding a PG_real page bit except for two things:
>>
>> 1) page flags bits are seriously short supply; there are some 24
>> available of which 22 are in use.  Any new user of a bit would have to
>> be an extremely valuable change with major benefit to the kernel, and
>>
> 
> It's indeed an issue. Could we instead use a combination of flags
> that can't happen together. For example PG_Free|PG_Reserved ?
> 

You'd need to audit all other users of the bits you wanted to borrow to 
check they would understand.  Like if you used PG_buddy (which I assume 
is what you are referring to above) then you'd get non-real memory 
getting merged into your buddies.  Badness.

>> 2) if you were to try and populate a PG_real flag it would need to be
>> populated for _all_ architectures (and there are a lot) for it to be of
>> any use.  As you have already noted there is no consistent way to find
>> out whether a page is ram so it would be major exercise to get these
>> bits setup during boot.
>>
>> I think we should take (2) as a hint here.  If we don't have a
>> consistent interface for finding whether a page is real or not, we
>> obviously have no general need of that information in the kernel.
>>
> 
> or maybe _because_ we don't have a consistent interface for finding 
> whether a page is real or not, we end up with a strange thing called
> page_is_ram() which could be the same for all arch and be implemented
> very simply.
> 
> BTW, can you try in a linux tree:
> 
> $ grep -r page_is_ram arch/
> 
> and see how it's implemented...

Well it depends how you look at it.  You are going to need to know which 
pages are ram in each architecture to set the bits in the page*'s to 
tell us later.  So the problem is the same problem.  We don't 
necessarily have the information for all architectures.  As we don't use 
this for anything its questionable whether we need it.

Feel free to try and figure it out for all these architectures :).

-apw

Re : Re : Re : Re : sparsemem usage

[moreau francis]

Andy Whitcroft wrote:
> moreau francis wrote:
>>
>> It's indeed an issue. Could we instead use a combination of flags
>> that can't happen together. For example PG_Free|PG_Reserved ?
>>
> 
> You'd need to audit all other users of the bits you wanted to borrow to
> check they would understand.  Like if you used PG_buddy (which I assume
> is what you are referring to above) then you'd get non-real memory
> getting merged into your buddies.  Badness.
> 

It would be great if we could define:

#define page_is_real(p) (p->_count > 0 || p->flags != 0)

Hence mem_map[] would be automatically initialized as full of page without
any real memory instead of initializing it with a magic value.
>>
>> or maybe _because_ we don't have a consistent interface for finding
>> whether a page is real or not, we end up with a strange thing called
>> page_is_ram() which could be the same for all arch and be implemented
>> very simply.
>>
>> BTW, can you try in a linux tree:
>>
>> $ grep -r page_is_ram arch/
>>
>> and see how it's implemented...
> 
> Well it depends how you look at it.  You are going to need to know which
> pages are ram in each architecture to set the bits in the page*'s to

I don't see the problem there. You can init mem_map[] each time it is
allocated with the magic value (if the above definition can't be used).
Then, as usual, archs free all zone area by initializing all mem_map
entries with something different from the magic value. After that all
entries of mem_map[] with the magic value can be fastly discarded
because they don't have real memory.

Francis

Re : Re : sparsemem usage

[moreau francis]

Alan Cox wrote:
>
> Mapping out parts of a section is quite normal - think about the 640K to
> 1Mb hole in PC memory space.

OK. But I'm still worry. Please consider the following code

       for (...; ...; ...) {
                [...]
                if (pfn_valid(i))
                       num_physpages++;
                [...]
        }

In that case num_physpages won't store an accurate value. Still it will be
used by the kernel to make some statistic assumptions on other kernel
data structure sizes.

Francis
        

Re: Re : Re : sparsemem usage

[KAMEZAWA Hiroyuki]

On Thu, 3 Aug 2006 09:07:06 +0000 (GMT)
moreau francis <francis_moreau2000@yahoo.fr> wrote:

> Alan Cox wrote:
> >
> > Mapping out parts of a section is quite normal - think about the 640K to
> > 1Mb hole in PC memory space.
> 
> OK. But I'm still worry. Please consider the following code
> 
>        for (...; ...; ...) {
>                 [...]
>                 if (pfn_valid(i))
>                        num_physpages++;
>                 [...]
>         }
> 
> In that case num_physpages won't store an accurate value. Still it will be
> used by the kernel to make some statistic assumptions on other kernel
> data structure sizes.
> 
In my understanding, pfn_valid() just returns "the page has page struct or not".
So, don't use pfn_valid() to count physical pages..


-Kame

Re: Re : Re : sparsemem usage

[Andy Whitcroft]

moreau francis wrote:
> Alan Cox wrote:
>> Mapping out parts of a section is quite normal - think about the 640K to
>> 1Mb hole in PC memory space.
> 
> OK. But I'm still worry. Please consider the following code
> 
>        for (...; ...; ...) {
>                 [...]
>                 if (pfn_valid(i))
>                        num_physpages++;
>                 [...]
>         }
> 
> In that case num_physpages won't store an accurate value. Still it will be
> used by the kernel to make some statistic assumptions on other kernel
> data structure sizes.

That would be incorrect usage.  pfn_valid() simply doesn't tell you if 
you have memory backing a pfn, it mearly means you can interrogate the 
page* for it.  A good example of code which counts pages in a region is 
in count_highmem_pages() which has a form as below:

			for (pfn = start; pfn < end; pfn++) {
  				if (!pfn_valid(pfn))
                                         continue;
                                 page = pfn_to_page(pfn);
                                 if (PageReserved(page))
                                         continue;
				num_physpages++;
			}

-apw

Re: Re : sparsemem usage

[Andy Whitcroft]

moreau francis wrote:
> Andy Whitcroft wrote:
>> The memory allocator buddy location algorithm has an implicit assumption 
>> that the memory map will be contigious and valid out to MAX_ORDER.  ie 
>> that we can do relative arithmetic on a page* for a page to find its 
>> buddy at all times.  The allocator never looks outside a MAX_ORDER 
>> block, aligned to MAX_ORDER in physical pages.  SPARSEMEM's 
>> implementation by it nature breaks up the mem_map at the section size. 
>> Thus for the buddy to work a section must be >= MAX_ORDER in size to 
>> maintain the contiguity constraint.
> 
> thanks for the explanation. But still something I'm missing, how can a
> MAX_ORDER block be allocated in a memory whose size is only 128Ko ?
> Can't it be detected by the buddy allocatorvery early without doing any 
> relative arithmetic on a page* ?

When allocating we do not have a problem as we simply pull a free page 
off the appropriately sizes free list.  Its when freeing we have an 
issue, all the allocator has to work with is the page you are freeing. 
As MAX_ORDER is >128K we can get to the situation where all but one page 
is free.  When we free that page we then need to merge this 128Kb page 
with its buddy if its free.   To tell if that one is free it has to look 
at the page* for it, so that page* must also exist for this check to work.

>> However, just because you have a small memory block in your memory map 
>> doesn't mean that the sparsemem section size needs to be that small to 
>> match.  If there is any valid memory in any section that section will be 
>> instantiated and the valid memory marked within it, any invalid memory 
>> is marked reserved.  
> 
> ah ok but that means that pfn_valid() will still returns ok for invalid page which
> are in a invalid memory marked as reserved. Is it not risky ?

pfn_valid() will indeed say 'ok'.  But that is defined only to mean that 
it is safe to look at the page* for that page.  It says nothing else 
about the page itself.  Pages which are reserved never get freed into 
the allocator so they are not there to be allocated so we should not be 
refering to them.

>> The section size bounds the amount of internal 
>> fragmentation we can have in the mem_map.  SPARSEMEM as its name 
>> suggests wins biggest when memory is very sparsly populate. 
> 
> sorry but I don't understand. I would say that sparsemem section size should
> be chosen to make mem_map[] and mem_section[] sizes as small as possible.

There are tradeoffs here.  The smaller the section size the better the 
internal fragmentation will be.  However also the more of them there 
will be, the more space that will be used tracking them, the more 
cachelines touched with them.  Also as we have seen we can't have things 
in the allocator bigger than the section size.  This can constrain the 
lower bound on the section size.  Finally, on 32 bit systems the overall 
number of sections is bounded by the available space in the fields 
section of the page* flags field.

If your system has 256 1Gb sections and 1 128Kb section then it could 
well make sense to have a 1GB section size or perhaps a 256Mb section 
size as you are only wasting space in the last section.

> 
>> If I am 
>> reading correctly your memory is actually contigious.
> 
> well there're big holes in address space.
> 

I read that as saying there was a major gap to 3Gb and then it was 
contigious from there; but then I was guessing at the units :).

-apw

Re : Re : sparsemem usage

[moreau francis]

Andy Whitcroft wrote:
> When allocating we do not have a problem as we simply pull a free page 
> off the appropriately sizes free list.  Its when freeing we have an 
> issue, all the allocator has to work with is the page you are freeing. 
> As MAX_ORDER is >128K we can get to the situation where all but one page 
> is free.  When we free that page we then need to merge this 128Kb page 
> with its buddy if its free.   To tell if that one is free it has to look 
> at the page* for it, so that page* must also exist for this check to work.

Maybe in sparsemem code, we could mark a well chosen page as reserved if
the size of mem region is < MAX_ORDER. That way the buddy allocator
will never have to free a block of 128 KO...

> pfn_valid() will indeed say 'ok'.  But that is defined only to mean that 
> it is safe to look at the page* for that page.  It says nothing else 
> about the page itself.  Pages which are reserved never get freed into 
> the allocator so they are not there to be allocated so we should not be 
> refering to them.

wouldn't it be safer to mark these pages as "invalid" instead of "reserved"
with a special value stored in mem_map[] ?

> There are tradeoffs here.  The smaller the section size the better the 
> internal fragmentation will be.  However also the more of them there 
> will be, the more space that will be used tracking them, the more 
> cachelines touched with them.  Also as we have seen we can't have things 
> in the allocator bigger than the section size.  This can constrain the 
> lower bound on the section size.  Finally, on 32 bit systems the overall 
> number of sections is bounded by the available space in the fields 
> section of the page* flags field.

thanks for that.

> If your system has 256 1Gb sections and 1 128Kb section then it could 
> well make sense to have a 1GB section size or perhaps a 256Mb section 
> size as you are only wasting space in the last section.

> I read that as saying there was a major gap to 3Gb and then it was 
> contigious from there; but then I was guessing at the units :).

here is a updated version of my mapping, it should be clear now:

HOLE0: 0 - 3 Go
MEM1: 0xc000 0000 - 32 Mo
HOLE1: 0xc200 0000 - 224 Mo
MEM2: 0xd000 0000 - 8 Mo
HOLE2: 0xd080 0000 - 120 Mo
MEM3: 0xd800 0000 - 128 Ko
HOLE3: rest of mem

Francis