Date: Thu, 27 Jun 2024 14:05:29 +0300
From: Usama Arif <usamaarif642@...il.com>
To: Yosry Ahmed <yosryahmed@...gle.com>, Hugh Dickins <hughd@...gle.com>,
 Yury Norov <yury.norov@...il.com>,
 Rasmus Villemoes <linux@...musvillemoes.dk>
Cc: kernel test robot <oliver.sang@...el.com>, oe-lkp@...ts.linux.dev,
 lkp@...el.com, Linux Memory Management List <linux-mm@...ck.org>,
 Andrew Morton <akpm@...ux-foundation.org>,
 Chengming Zhou <chengming.zhou@...ux.dev>, Nhat Pham <nphamcs@...il.com>,
 David Hildenbrand <david@...hat.com>, "Huang, Ying" <ying.huang@...el.com>,
 Johannes Weiner <hannes@...xchg.org>, Matthew Wilcox <willy@...radead.org>,
 Shakeel Butt <shakeel.butt@...ux.dev>, Andi Kleen <ak@...ux.intel.com>,
 linux-kernel@...r.kernel.org, "Vlastimil Babka (SUSE)" <vbabka@...nel.org>
Subject: Re: [linux-next:master] [mm] 0fa2857d23:
 WARNING:at_mm/page_alloc.c:#__alloc_pages_noprof


On 24/06/2024 21:26, Usama Arif wrote:
>
> On 24/06/2024 20:31, Yosry Ahmed wrote:
>> On Mon, Jun 24, 2024 at 10:26 AM Usama Arif <usamaarif642@...il.com> 
>> wrote:
>>>
>>> On 24/06/2024 19:56, Yosry Ahmed wrote:
>>>> [..]
>>>>>>> -       p->zeromap = bitmap_zalloc(maxpages, GFP_KERNEL);
>>>>>>> +       p->zeromap = kvzalloc(DIV_ROUND_UP(maxpages, 8), 
>>>>>>> GFP_KERNEL);
>>>>>> No, 8 is not right for 32-bit kernels. I think you want
>>>>>>         p->zeromap = kvzalloc(BITS_TO_LONGS(maxpages), GFP_KERNEL);
>>>>>> but please check it carefully, I'm easily confused by such 
>>>>>> conversions.
>>>>>>
>>>>>> Hugh
>>>>> Ah yes, didnt take into account 32-bit kernel. I think its 
>>>>> supposed to be
>>>>>
>>>>>     p->zeromap = kvzalloc(BITS_TO_LONGS(maxpages) * 
>>>>> sizeof(unsigned long),
>>>>> GFP_KERNEL);
>>>> You can do something similar to bitmap_zalloc() and use:
>>>>
>>>> kvmalloc_array(BITS_TO_LONGS(nbits), sizeof(unsigned long), GFP_KERNEL
>>>> | __GFP_ZERO)
>>>>
>>>> I don't see a kvzalloc_array() variant to use directly, but it should
>>>> be trivial to add it. I can see other users of kvmalloc_array() that
>>>> pass in __GFP_ZERO (e.g. fs/ntfs3/bitmap.c).
>>>>
>>>> , or you could take it a step further and add bitmap_kvzalloc(),
>>>> assuming the maintainers are open to that.
>>> Thanks! bitmap_kvzalloc makes most sense to me. It doesnt make sense
>>> that bitmap should only be limited to MAX_PAGE_ORDER size. I can add
>>> this patch below at the start of the series and use it in the patch for
>>> zeropage swap optimization.
>>>
>>>
>>>       bitmap: add support for virtually contiguous bitmap
>>>
>>>       The current bitmap_zalloc API limits the allocation to 
>>> MAX_PAGE_ORDER,
>>>       which prevents larger order bitmap allocations. Introduce
>>>       bitmap_kvzalloc that will allow larger allocations of bitmap.
>>>       kvmalloc_array still attempts to allocate physically 
>>> contiguous memory,
>>>       but upon failure, falls back to non-contiguous (vmalloc) 
>>> allocation.
>>>
>>>       Suggested-by: Yosry Ahmed <yosryahmed@...gle.com>
>>>       Signed-off-by: Usama Arif <usamaarif642@...il.com>
>>>
>> LGTM with a small fix below.
>>
>>> diff --git a/include/linux/bitmap.h b/include/linux/bitmap.h
>>> index 8c4768c44a01..881c2ff2e834 100644
>>> --- a/include/linux/bitmap.h
>>> +++ b/include/linux/bitmap.h
>>> @@ -131,9 +131,11 @@ struct device;
>>>     */
>>>    unsigned long *bitmap_alloc(unsigned int nbits, gfp_t flags);
>>>    unsigned long *bitmap_zalloc(unsigned int nbits, gfp_t flags);
>>> +unsigned long *bitmap_kvzalloc(unsigned int nbits, gfp_t flags);
>>>    unsigned long *bitmap_alloc_node(unsigned int nbits, gfp_t flags, 
>>> int
>>> node);
>>>    unsigned long *bitmap_zalloc_node(unsigned int nbits, gfp_t 
>>> flags, int
>>> node);
>>>    void bitmap_free(const unsigned long *bitmap);
>>> +void bitmap_kvfree(const unsigned long *bitmap);
>>>
>>>    DEFINE_FREE(bitmap, unsigned long *, if (_T) bitmap_free(_T))
>>>
>>> diff --git a/lib/bitmap.c b/lib/bitmap.c
>>> index b97692854966..eabbfb85fb45 100644
>>> --- a/lib/bitmap.c
>>> +++ b/lib/bitmap.c
>>> @@ -727,6 +727,13 @@ unsigned long *bitmap_zalloc(unsigned int nbits,
>>> gfp_t flags)
>>>    }
>>>    EXPORT_SYMBOL(bitmap_zalloc);
>>>
>>> +unsigned long *bitmap_kvzalloc(unsigned int nbits, gfp_t flags)
>>> +{
>>> +       return kvmalloc_array(BITS_TO_LONGS(nbits), sizeof(unsigned 
>>> long),
>>> +                             flags | __GFP_ZERO);
>>> +}
>>> +EXPORT_SYMBOL(bitmap_zalloc);
>> EXPORT_SYMBOL(bitmap_kvzalloc)*
>
>
> Actually, does it make more sense to change the behaviour of the 
> current APIs like below instead of above patch? Or is there an 
> expectation that the current bitmap API is supposed to work only on 
> physically contiguous bits?
>
> I believe in the kernel if the allocation/free starts with 'k' its 
> physically contiguous and with "kv" its physically contiguous if 
> possible, otherwise virtually contiguous. The bitmap functions dont 
> have either, so we could change the current implementation. I believe 
> it would not impact the current users of the functions as the first 
> attempt is physically contiguous which is how it works currently, and 
> only upon failure it would be virtual and it would increase the use of 
> current bitmap API to greater than MAX_PAGE_ORDER size allocations.
>
> Yury Norov and Rasmus Villemoes, any views on this?
>
> Thanks
>
> diff --git a/include/linux/slab.h b/include/linux/slab.h
> index 7247e217e21b..ad771dc81afa 100644
> --- a/include/linux/slab.h
> +++ b/include/linux/slab.h
> @@ -804,6 +804,7 @@ kvmalloc_array_node_noprof(size_t n, size_t size, 
> gfp_t flags, int node)
>  #define kvcalloc_node_noprof(_n,_s,_f,_node) 
> kvmalloc_array_node_noprof(_n,_s,(_f)|__GFP_ZERO,_node)
>  #define kvcalloc_noprof(...) kvcalloc_node_noprof(__VA_ARGS__, 
> NUMA_NO_NODE)
>
> +#define kvmalloc_array_node(...) 
> alloc_hooks(kvmalloc_array_node_noprof(__VA_ARGS__))
>  #define kvmalloc_array(...) 
> alloc_hooks(kvmalloc_array_noprof(__VA_ARGS__))
>  #define kvcalloc_node(...) 
> alloc_hooks(kvcalloc_node_noprof(__VA_ARGS__))
>  #define kvcalloc(...) alloc_hooks(kvcalloc_noprof(__VA_ARGS__))
> diff --git a/lib/bitmap.c b/lib/bitmap.c
> index b97692854966..272164dcbef1 100644
> --- a/lib/bitmap.c
> +++ b/lib/bitmap.c
> @@ -716,7 +716,7 @@ void bitmap_fold(unsigned long *dst, const 
> unsigned long *orig,
>
>  unsigned long *bitmap_alloc(unsigned int nbits, gfp_t flags)
>  {
> -       return kmalloc_array(BITS_TO_LONGS(nbits), sizeof(unsigned long),
> +       return kvmalloc_array(BITS_TO_LONGS(nbits), sizeof(unsigned 
> long),
>                              flags);
>  }
>  EXPORT_SYMBOL(bitmap_alloc);
> @@ -729,7 +729,7 @@ EXPORT_SYMBOL(bitmap_zalloc);
>
>  unsigned long *bitmap_alloc_node(unsigned int nbits, gfp_t flags, int 
> node)
>  {
> -       return kmalloc_array_node(BITS_TO_LONGS(nbits), 
> sizeof(unsigned long),
> +       return kvmalloc_array_node(BITS_TO_LONGS(nbits), 
> sizeof(unsigned long),
>                                   flags, node);
>  }
>  EXPORT_SYMBOL(bitmap_alloc_node);
> @@ -742,7 +742,7 @@ EXPORT_SYMBOL(bitmap_zalloc_node);
>
>  void bitmap_free(const unsigned long *bitmap)
>  {
> -       kfree(bitmap);
> +       kvfree(bitmap);
>  }
>  EXPORT_SYMBOL(bitmap_free);
>

I decided to go with just using simple kvmalloc_array for v7 [1] with 
__GFP_ZERO instead of adding a new API to bitmap or changing the 
existing API to kvmalloc/kvfree as I didnt want to make this series 
dependent of bitmap API changes and there are other places where its 
done using kvmalloc_array like ceph and ntfs3. I am happy to send a 
follow up patch after this series that changes the existing API to be kv 
if thats something the bitmap maintainers think makes sense.

[1] 
https://lore.kernel.org/all/20240627105730.3110705-1-usamaarif642@gmail.com/

Powered by blists - more mailing lists