On 12/26/2018 11:40 AM, Michal Hocko wrote:
> Appart from general comments as a reply to the cover (btw. this all
> should be in the changelog because this is the _why_ part of the
> justification which should be _always_ part of the changelog).
Thank you, will add in the next version of the patch alltogether
with other changes if any.
> On Tue 25-12-18 18:39:27, Konstantin Khorenko wrote:
> [...]
>> +config WARN_HIGH_ORDER
>> + bool "Enable complains about high order memory allocations"
>> + depends on !LOCKDEP
>
> Why?
LOCKDEP makes structures big, so if we see a high order allocation warning
on a debug kernel with lockdep, it does not give us a lot - lockdep enabled
kernel performance is not our target.
i can remove !LOCKDEP dependence here, but then need to adjust default
warning level i think, or logs will be spammed.
>> + default n
>> + help
>> + Enables warnings on high order memory allocations. This allows to
>> + determine users of large memory chunks and rework them to decrease
>> + allocation latency. Note, some debug options make kernel structures
>> + fat.
>> +
>> +config WARN_HIGH_ORDER_LEVEL
>> + int "Define page order level considered as too high"
>> + depends on WARN_HIGH_ORDER
>> + default 3
>> + help
>> + Defines page order starting which the system to complain about.
>> + Default is current PAGE_ALLOC_COSTLY_ORDER.
>> +
>> config HWPOISON_INJECT
>> tristate "HWPoison pages injector"
>> depends on MEMORY_FAILURE && DEBUG_KERNEL && PROC_FS
>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>> index e95b5b7c9c3d..258892adb861 100644
>> --- a/mm/page_alloc.c
>> +++ b/mm/page_alloc.c
>> @@ -4341,6 +4341,30 @@ static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
>> ac->high_zoneidx, ac->nodemask);
>> }
>>
>> +#ifdef CONFIG_WARN_HIGH_ORDER
>> +int warn_order = CONFIG_WARN_HIGH_ORDER_LEVEL;
>> +
>> +/*
>> + * Complain if we allocate a high order page unless there is a __GFP_NOWARN
>> + * flag provided.
>> + *
>> + * Shuts up after 32 complains.
>> + */
>> +static __always_inline void warn_high_order(int order, gfp_t gfp_mask)
>> +{
>> + static atomic_t warn_count = ATOMIC_INIT(32);
>> +
>> + if (order >= warn_order && !(gfp_mask & __GFP_NOWARN))
>> + WARN(atomic_dec_if_positive(&warn_count) >= 0,
>> + "order %d >= %d, gfp 0x%x\n",
>> + order, warn_order, gfp_mask);
>> +}
>
> We do have ratelimit functionality, so why cannot you use it?
Well, my idea was to really shut up the warning after some number of messages
(if a node is in production and its uptime, say, a year, i don't want to see
many warnings in logs, first several is enough - let's fix them first).
If i use printk_ratelimited() i could get 24 days delay at most AFAIK,
but okay, i can switch to printk_ratelimited() if you prefer.
struct ratelimit_state {
int interval;
(gdb) p ((1LL<<31) -1)/1000/60/60/24
$11 = 24
>> +#else
>> +static __always_inline void warn_high_order(int order, gfp_t gfp_mask)
>> +{
>> +}
>> +#endif
>> +
>> /*
>> * This is the 'heart' of the zoned buddy allocator.
>> */
>> @@ -4361,6 +4385,7 @@ __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
>> WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
>> return NULL;
>> }
>> + warn_high_order(order, gfp_mask);
>>
>> gfp_mask &= gfp_allowed_mask;
>> alloc_mask = gfp_mask;
>
> Why do you warn about all allocations in the hot path? I thought you
> want to catch expensive allocations so I would assume that you would
> stick that into a slow path after we are not able to allocate anything
> after the first round of compaction.
The idea is to catch big allocations soon and preferably during testing,
not on production nodes under high load, that's why i've chosen hot path.
And if we switch to the slow path, we'll have to run all tests under
additional serious load - to generate memory fragmentation.
Not so convenient.
> Also do you want to warn about opportunistic GFP_NOWAIT allocations that
> have a reasonable fallback?
Yes, i would like to. Sometimes allocation flags come from upper level and
it's not always evident if there is GFP_NOWAIT flag or not, so let's we
are noticed about such a case and verify if it's legal and not called often.
If yes - we'll just mark it with NO_WARN.
> And forgot to mention other opportunistic allocations like THP of
> course.
hugepages are allocated with NOWARN already, AFAIK.
alloc_fresh_huge_page_node(), __hugetlb_alloc_buddy_huge_page()
Thank you once again, Michal.
--
Best regards,
Konstantin Khorenko,
Virtuozzo Linux Kernel Team
