[PATCH v6 05/42] powerpc/powernv: Track IO/M32/M64 segments from PE

[Gavin Shan]

On Wed, Aug 12, 2015 at 10:57:33PM +1000, Alexey Kardashevskiy wrote:
>On 08/12/2015 09:20 PM, Gavin Shan wrote:
>>On Wed, Aug 12, 2015 at 09:05:09PM +1000, Alexey Kardashevskiy wrote:
>>>On 08/12/2015 08:45 PM, Gavin Shan wrote:
>>>>On Tue, Aug 11, 2015 at 12:23:42PM +1000, Alexey Kardashevskiy wrote:
>>>>>On 08/11/2015 10:03 AM, Gavin Shan wrote:
>>>>>>On Mon, Aug 10, 2015 at 05:16:40PM +1000, Alexey Kardashevskiy wrote:
>>>>>>>On 08/06/2015 02:11 PM, Gavin Shan wrote:
>>>>>>>>The patch is adding 6 bitmaps, three to PE and three to PHB, to track
>>>>>>>
>>>>>>>The patch is also removing 2 arrays (io_segmap and m32_segmap), what is that
>>>>>>>all about? Also, there was no m64_segmap, now there is, needs an explanation
>>>>>>>may be.
>>>>>>>
>>>>>>
>>>>>>Originally, the bitmaps (io_segmap and m32_segmap) are allocated dynamically.
>>>>>>Now, they have fixed sizes - 512 bits.
>>>>>>
>>>>>>The subject "powerpc/powernv: Track IO/M32/M64 segments from PE" indicates
>>>>>>why m64_segmap is added.
>>>>>
>>>>>
>>>>>But before this patch, you somehow managed to keep it working without a map
>>>>>for M64, by the same time you needed map for IO and M32. It seems you are
>>>>>making things consistent in this patch but it also feels like you do not have
>>>>>to do so as M64 did not need a map before and I cannot see why it needs one
>>>>>now.
>>>>>
>>>>
>>>>The M64 map is used by [PATCH v6 23/42] powerpc/powernv: Release PEs dynamically
>>>>where the M64 segments consumed by one particular PE will be released.
>>>
>>>
>>>Then add it where it is really started being used. It is really hard to
>>>review a patch which is actually spread between patches. Do not count that
>>>reviewers will just trust you.
>>>
>>
>>Ok. I'll try.
>>
>>>
>>>>>>>
>>>>>>>>the consumed by one particular PE, which can be released once the PE
>>>>>>>>is destroyed during PCI unplugging time. Also, we're using fixed
>>>>>>>>quantity of bits to trace the used IO, M32 and M64 segments by PEs
>>>>>>>>in one particular PHB.
>>>>>>>>
>>>>>>>
>>>>>>>Out of curiosity - have you considered having just 3 arrays, in PHB, storing
>>>>>>>PE numbers, and ditching PE's arrays? Does PE itself need to know what PEs it
>>>>>>>is using? Not sure about this master/slave PEs though.
>>>>>>>
>>>>>>
>>>>>>I don't follow your suggestion. Can you rephrase and explain it a bit more?
>>>>>
>>>>>
>>>>>Please explains in what situations you need same map in both PHB and PE and
>>>>>how you are going to use them. For example, pe::m64_segmap and
>>>>>phb::m64_segmap.
>>>>>
>>>>>I believe you need to know what segment is used by what PE and that's it and
>>>>>having 2 bitmaps is overcomplicated hard to follow. Is there anything else
>>>>>what I am missing?
>>>>>
>>>>
>>>>The situation is same to all (IO, M32 and M64) segment maps. Taking m64_segmap
>>>>as an example, it will be used when creating or destroying the PE who consumes
>>>>M64 segments. phb::m64_segmap is recording the M64 segment usage in PHB's domain.
>>>>It's used to check same M64 segment won't be used for towice. pe::m64_segmap tracks
>>>>the M64 segments consumed by the PE.
>>>
>>>
>>>You could have a single map in PHB, key would be a segment number and value
>>>would be PE number. No need to have a map in PE. At all. No need to
>>>initialize bitmaps, etc.
>>>
>>
>>So it would be arrays for various segmant maps if I understood your suggestion
>>as below. Please confirm:
>>
>>#define PNV_IODA_MAX_SEG_NUM	512
>>
>>	int struct pnv_phb::io_segmap[PNV_IODA_MAX_SEG_NUM];
>>			    m32_segmap[PNV_IODA_MAX_SEG_NUM];
>>			    m64_segmap[PNV_IODA_MAX_SEG_NUM];
>>- Initially, they are initialize to IODA_INVALID_PE;
>>- When one segment is assigned to one PE, the corresponding entry
>>   of the array is set to PE number.
>>- When one segment is relased, the corresponding entry of the array
>>   is set to IODA_INVALID_PE;
>
>
>No, not arrays, I meant DEFINE_HASHTABLE(), hash_add(), etc from
>include/linux/hashtable.h.
>
>http://kernelnewbies.org/FAQ/Hashtables is a good place to start :)
>

Are you sure it needs hashtable to represent the simple data struct?
I really don't understand the benefits, could you provide more details
about the benefits?

With hashtable, every bucket will include multiple items with conflicting
hash key, each of which would be represented by data struct as below. The
data struct uses 24 bytes memory and not efficient enough from this aspect.
When one more segment consued, instance of "struct pnv_ioda_segment" is
allocated and put into the conflicting list of the target bucket. At later
point, the instance is removed from the list and released when the segment
is detached from the PE. It's more complex than it should be.

struct pnv_ioda_segment {
	int               pe_number;
	int               seg_number;
	struct hlist_node node;
};

Thanks,
Gavin