Re: [PATCH 2/3] dmapool: improve scalability of dma_pool_free

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 07/26/2018 08:07 PM, Matthew Wilcox wrote:
> On Thu, Jul 26, 2018 at 04:06:05PM -0400, Tony Battersby wrote:
>> On 07/26/2018 03:42 PM, Matthew Wilcox wrote:
>>> On Thu, Jul 26, 2018 at 02:54:56PM -0400, Tony Battersby wrote:
>>>> dma_pool_free() scales poorly when the pool contains many pages because
>>>> pool_find_page() does a linear scan of all allocated pages.  Improve its
>>>> scalability by replacing the linear scan with a red-black tree lookup. 
>>>> In big O notation, this improves the algorithm from O(n^2) to O(n * log n).
>>> This is a lot of code to get us to O(n * log(n)) when we can use less
>>> code to go to O(n).  dma_pool_free() is passed the virtual address.
>>> We can go from virtual address to struct page with virt_to_page().
>>> In struct page, we have 5 words available (20/40 bytes), and it's trivial
>>> to use one of them to point to the struct dma_page.
>>>
>> Thanks for the tip.  I will give that a try.
> If you're up for more major surgery, then I think we can put all the
> information currently stored in dma_page into struct page.  Something
> like this:
>
> +++ b/include/linux/mm_types.h
> @@ -152,6 +152,12 @@ struct page {
>                         unsigned long hmm_data;
>                         unsigned long _zd_pad_1;        /* uses mapping */
>                 };
> +               struct {        /* dma_pool pages */
> +                       struct list_head dma_list;
> +                       unsigned short in_use;
> +                       unsigned short offset;
> +                       dma_addr_t dma;
> +               };
>  
>                 /** @rcu_head: You can use this to free a page by RCU. */
>                 struct rcu_head rcu_head;
>
> page_list -> dma_list
> vaddr goes away (page_to_virt() exists)
> dma -> dma
> in_use and offset shrink from 4 bytes to 2.
>
> Some 32-bit systems have a 64-bit dma_addr_t, and on those systems,
> this will be 8 + 2 + 2 + 8 = 20 bytes.  On 64-bit systems, it'll be
> 16 + 2 + 2 + 4 bytes of padding + 8 = 32 bytes (we have 40 available).
>
>
offset at least needs more bits, since allocations can be multi-page. 
See the following from mpt3sas:

cat /sys/devices/pci0000:80/0000:80:07.0/0000:85:00.0/pools
(manually cleaned up column alignment)
poolinfo - 0.1
reply_post_free_array pool  1      21     192     1
reply_free pool             1      1      41728   1
reply pool                  1      1      1335296 1
sense pool                  1      1      970272  1
chain pool                  373959 386048 128     12064
reply_post_free pool        12     12     166528  12
                                          ^size^

In my initial implementation I also added a pointer from struct page to
the dma_pool so that dma_pool_free() could sanity-check that the page
really belongs to the pool, as in:

pg = virt_to_page(vaddr);
if (unlikely(pg->dma_pool != pool)) {
	handle error...
}
page = pg->dma_page;

The linked-list search previously implemented that check as a
by-product, and I didn't want to lose it.  It all seems to be working so
far.




[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[Index of Archives]     [SCSI Target Devel]     [Linux SCSI Target Infrastructure]     [Kernel Newbies]     [IDE]     [Security]     [Git]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux ATA RAID]     [Linux IIO]     [Samba]     [Device Mapper]

  Powered by Linux