Re:Re: [PATCH v2] mm/page_alloc: fix alloc_pages_bulk/set_page_owner panic on irq disabled

[杨欢]

>> BUG: sleeping function called from invalid context at mm/page_alloc.c:5179
>> in_atomic(): 0, irqs_disabled(): 1, non_block: 0, pid: 1, name: swapper/0
>> 
>> __dump_stack lib/dump_stack.c:79 [inline]
>> dump_stack_lvl+0xcd/0x134 lib/dump_stack.c:96
>> ___might_sleep.cold+0x1f1/0x237 kernel/sched/core.c:9153
>> prepare_alloc_pages+0x3da/0x580 mm/page_alloc.c:5179
>> __alloc_pages+0x12f/0x500 mm/page_alloc.c:5375
>> alloc_pages+0x18c/0x2a0 mm/mempolicy.c:2272
>> stack_depot_save+0x39d/0x4e0 lib/stackdepot.c:303
>> save_stack+0x15e/0x1e0 mm/page_owner.c:120
>> __set_page_owner+0x50/0x290 mm/page_owner.c:181
>> prep_new_page mm/page_alloc.c:2445 [inline]
>> __alloc_pages_bulk+0x8b9/0x1870 mm/page_alloc.c:5313
>> 
>> The problem is caused by set_page_owner alloc memory to save stack with
>> GFP_KERNEL in local_riq disabled.
>> So, we just can't assume that alloc flags should be same with new page,
>> prep_new_page should prep/trace the page gfp, but shouldn't use the same
>> gfp to get memory, let's depend on caller.
>> So, here is two gfp flags, alloc_gfp used to alloc memory, depend on
>> caller, page_gfp_mask is page's gfp, used to trace/prep itself
>> But in most situation, same is ok, in alloc_pages_bulk, use GFP_ATOMIC
>> is ok.(even if set_page_owner save backtrace failed, limited impact)
>> 
>> v2:
>> - add more description.
>> 
>> Fixes: 0f87d9d30f21 ("mm/page_alloc: add an array-based interface to the bulk page allocator")
>> Reported-by: syzbot+b07d8440edb5f8988eea@xxxxxxxxxxxxxxxxxxxxxxxxx
>> Suggested-by: Wang Qing <wangqing@xxxxxxxx>
>> Signed-off-by: Yang Huan <link@xxxxxxxx>
>
>https://lore.kernel.org/lkml/20210713152100.10381-2-mgorman@xxxxxxxxxxxxxxxxxxx/
>is now part of a series that has being sent to Linus. Hence, the Fixes
>part is no longer applicable and the patch will no longer be addresing
>an atomic sleep bug.  This patch should be treated as an enhancement
Hi Mel Gorman, thanks for your reply.
I see the fix patch, it fix this bug by abandon alloc bulk feature when page_owner is set. 
But in my opinion, it can't really fix this bug, it's a circumvention plan.
>to allow bulk allocations when PAGE_OWNER is set. For that, it should
>include a note on the performance if PAGE_OWNER is used with either NFS
>or high-speed networking to justify the additional complexity.
My patch just split the prep_new_page page_gfp into alloc_gfp(for alloc bulk is GFP_ATOMIC,
for other's no change) and trace page gfp.  So, we will not use the error way to get memory. 
So, I think this will not affect alloc bulk performance when page_owner is on(compare with origin patch) but
can really fix this  bug rather than evade.
And this patch can let alloc bulk feature and page_owner feature work togher
So, I will send patch again based on the fix patch.
Thank you
Yang Huan
>
>-- 
>Mel Gorman
>SUSE Labs