As of now, debugobjects framework uses hlist_head and separate spinlock
as a hash table bucket. We have hlist_bl_head APIs which embeds a
bit_spinlock in head pointer and thus no separate spinlock is required.

This patchset first wraps irq variant API for bit_spinlock as well as
hlist_bl_lock and several other APIs and macros. Lastly, It replaces
hlist APIs with hlist_bl counterparts.

Woody Zhang (5):
  bit_spinlock: add irq variant for bit spinlock API
  list_bl: add irq variant for hlist_bl lock API
  list_bl: remove lock check in hlist_bl_set_first
  list_bl: add hlist_bl_move_list and two macros
  debugobjects: use list_bl to save memory for hash slot spinlocks

 include/linux/bit_spinlock.h |  37 +++++
 include/linux/debugobjects.h |   4 +-
 include/linux/list_bl.h      |  40 ++++-
 lib/debugobjects.c           | 288 +++++++++++++++++------------------
 4 files changed, 217 insertions(+), 152 deletions(-)

-- 
2.45.2

On Thu, Aug 22 2024 at 07:05, Woody Zhang wrote:
> As of now, debugobjects framework uses hlist_head and separate spinlock
> as a hash table bucket. We have hlist_bl_head APIs which embeds a
> bit_spinlock in head pointer and thus no separate spinlock is required.
>
> This patchset first wraps irq variant API for bit_spinlock as well as
> hlist_bl_lock and several other APIs and macros. Lastly, It replaces
> hlist APIs with hlist_bl counterparts.

You are telling _what_ your changes are doing, but not _why_ and neither
_what_ they are trying to achieve.

Aside of that you are failing to explain how replacing a spinlock by a
hlist bitlock is equivalent to a lockdep covered locking primitive.

It is NOT.

And you have to come up with a convincing argument why this makes sense
aside of saving an unspecified amount of memory, which you haven't even
bothered to document. Neither in the changelogs nor in the cover letter.

You also completely fail to provide an analysis why converting the debug
object locking from a fair and sensible locking implementation to a
known to be unscalable locking implementation makes sense for a debug
facility which is used in a lot of hotpaths.

Any attempt to substitute a spinlock with a hlist_bl locking scheme
needs to come with a proper analysis to demonstrate that:

   1) this is a completely equivalent locking scheme

   2) the resulting loss of lockdep coverage is justified

   3) there is an actual performance benefit

   4) the actual memory savings

Just handwaving about an unspecified amount of memory savings (probably
in the range of 2 bytes or such) without any of #1 -#3 above is not
cutting it at all.

Try again.

Thanks,

        tglx

On Thu, Aug 22, 2024 at 02:11:56AM +0200, Thomas Gleixner wrote:
>On Thu, Aug 22 2024 at 07:05, Woody Zhang wrote:
>> As of now, debugobjects framework uses hlist_head and separate spinlock
>> as a hash table bucket. We have hlist_bl_head APIs which embeds a
>> bit_spinlock in head pointer and thus no separate spinlock is required.
>>
>> This patchset first wraps irq variant API for bit_spinlock as well as
>> hlist_bl_lock and several other APIs and macros. Lastly, It replaces
>> hlist APIs with hlist_bl counterparts.
>
>You are telling _what_ your changes are doing, but not _why_ and neither
>_what_ they are trying to achieve.
>
>Aside of that you are failing to explain how replacing a spinlock by a
>hlist bitlock is equivalent to a lockdep covered locking primitive.
>
>It is NOT.
>
>And you have to come up with a convincing argument why this makes sense
>aside of saving an unspecified amount of memory, which you haven't even
>bothered to document. Neither in the changelogs nor in the cover letter.
>
>You also completely fail to provide an analysis why converting the debug
>object locking from a fair and sensible locking implementation to a
>known to be unscalable locking implementation makes sense for a debug
>facility which is used in a lot of hotpaths.
>
>Any attempt to substitute a spinlock with a hlist_bl locking scheme
>needs to come with a proper analysis to demonstrate that:
>
>   1) this is a completely equivalent locking scheme
>
>   2) the resulting loss of lockdep coverage is justified
>
>   3) there is an actual performance benefit
>
>   4) the actual memory savings
>
>Just handwaving about an unspecified amount of memory savings (probably
>in the range of 2 bytes or such) without any of #1 -#3 above is not
>cutting it at all.

All right. I will post a v2 to address these issues and try to give a more
detailed explanation.

BR
Woody

>
>Try again.
>
>Thanks,
>
>        tglx