On 09/20/2016 02:54 PM, Nicholas Piggin wrote:
> On Tue, 20 Sep 2016 14:02:26 +0800
> zijun_hu <zijun...@zoho.com> wrote:
>> From: zijun_hu <zijun...@htc.com>
>> correct a few logic error in __insert_vmap_area() since the else if
>> condition is always true and meaningless
>> avoid endless loop under [un]mapping improper ranges whose boundary
>> are not aligned to page
>> correct lazy_max_pages() return value if the number of online cpus
>> is power of 2
>> improve performance for pcpu_get_vm_areas() via optimizing vmap_areas
>> overlay checking algorithm and finding near vmap_areas by list_head
>> other than rbtree
>> simplify /proc/vmallocinfo implementation via seq_file helpers
>> for list_head
>> Signed-off-by: zijun_hu <zijun...@htc.com>
>> Signed-off-by: zijun_hu <zijun...@zoho.com>
> Could you submit each of these changes as a separate patch? Would you
> consider using capitalisation and punctuation in the changelog?
thanks for your advisement
i will follow it and split this patch to smaller patches finally

> Did you measure any performance improvements, or do you have a workload
> where vmalloc shows up in profiles?
don't have measurement in practice, but i am sure there are
performance improvements for pcpu_get_vm_areas() theoretically
due to below reasons:
1) the counter of vmap_area overlay checkup loop is reduced to half
2) the previous and next vmap_area of one on list_head are just the
   nearest ones due to address sorted vmap_areas on list_head, so no
   walk and compare is needed
>> @@ -108,6 +108,9 @@ static void vunmap_page_range(unsigned long addr, 
>> unsigned long end)
>>      unsigned long next;
>>      BUG_ON(addr >= end);
>> +    WARN_ON(!PAGE_ALIGNED(addr | end));
> I prefer to avoid mixing bitwise and arithmetic operations unless it's
> necessary. Gcc should be able to optimise
i agree with you, i will apply your suggestion finally
>> +    addr = round_down(addr, PAGE_SIZE);
> I don't know if it's really necessary to relax the API like this for
> internal vmalloc.c functions. If garbage is detected here, it's likely
> due to a bug, and I'm not sure that rounding it would solve the problem.
> For API functions perhaps it's reasonable -- in such cases you should
> consider using WARN_ON_ONCE() or similar.
actually, another patch for API function within /lib/ioremap.c used the 
way as pointed by you as below, i am not sure which is better, perhaps i
will exchange each other

Subject: [PATCH 2/3] lib/ioremap.c: avoid endless loop under ioremapping
improper ranges

for ioremap_page_range(), endless loop maybe happen if either of parameter
addr and end is not page aligned, in order to fix this issue and hint range
parameter requirements BUG_ON() checkup are performed firstly

for ioremap_pte_range(), loop end condition is optimized due to lack of
relevant macro pte_addr_end()

Signed-off-by: zijun_hu <zijun...@htc.com>
 lib/ioremap.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/lib/ioremap.c b/lib/ioremap.c
index 86c8911..0058cc8 100644
--- a/lib/ioremap.c
+++ b/lib/ioremap.c
@@ -64,7 +64,7 @@ static int ioremap_pte_range(pmd_t *pmd, unsigned long addr,
                set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
-       } while (pte++, addr += PAGE_SIZE, addr != end);
+       } while (pte++, addr += PAGE_SIZE, addr < end);
        return 0;
@@ -129,6 +129,7 @@ int ioremap_page_range(unsigned long addr,
        int err;
        BUG_ON(addr >= end);
+       BUG_ON(!PAGE_ALIGNED(addr | end));
        start = addr;
        phys_addr -= addr;

> Thanks,
> Nick

Reply via email to