Re: [PATCH] mm/vmstat: Reduce zone lock hold time when reading /proc/pagetypeinfo

[Waiman Long]

On 10/22/19 5:59 PM, Andrew Morton wrote:
> On Tue, 22 Oct 2019 12:21:56 -0400 Waiman Long <longman@xxxxxxxxxx> wrote:
>
>> The pagetypeinfo_showfree_print() function prints out the number of
>> free blocks for each of the page orders and migrate types. The current
>> code just iterates the each of the free lists to get counts.  There are
>> bug reports about hard lockup panics when reading the /proc/pagetyeinfo
>> file just because it look too long to iterate all the free lists within
>> a zone while holing the zone lock with irq disabled.
>>
>> Given the fact that /proc/pagetypeinfo is readable by all, the possiblity
>> of crashing a system by the simple act of reading /proc/pagetypeinfo
>> by any user is a security problem that needs to be addressed.
> Yes.
>
>> There is a free_area structure associated with each page order. There
>> is also a nr_free count within the free_area for all the different
>> migration types combined. Tracking the number of free list entries
>> for each migration type will probably add some overhead to the fast
>> paths like moving pages from one migration type to another which may
>> not be desirable.
>>
>> we can actually skip iterating the list of one of the migration types
>> and used nr_free to compute the missing count. Since MIGRATE_MOVABLE
>> is usually the largest one on large memory systems, this is the one
>> to be skipped. Since the printing order is migration-type => order, we
>> will have to store the counts in an internal 2D array before printing
>> them out.
>>
>> Even by skipping the MIGRATE_MOVABLE pages, we may still be holding the
>> zone lock for too long blocking out other zone lock waiters from being
>> run. This can be problematic for systems with large amount of memory.
>> So a check is added to temporarily release the lock and reschedule if
>> more than 64k of list entries have been iterated for each order. With
>> a MAX_ORDER of 11, the worst case will be iterating about 700k of list
>> entries before releasing the lock.
>>
>> ...
>>
>> --- a/mm/vmstat.c
>> +++ b/mm/vmstat.c
>> @@ -1373,23 +1373,54 @@ static void pagetypeinfo_showfree_print(struct seq_file *m,
>>  					pg_data_t *pgdat, struct zone *zone)
>>  {
>>  	int order, mtype;
>> +	unsigned long nfree[MAX_ORDER][MIGRATE_TYPES];
> 600+ bytes is a bit much.  I guess it's OK in this situation.
>
This function is called by reading /proc/pagetypeinfo. The call stack is
rather shallow:

PID: 58188Â TASK: ffff938a4d4f1fa0Â CPU: 2ÂÂ COMMAND: "sosreport"
Â#0 [ffff9483bf488e48] crash_nmi_callback at ffffffffb8c551d7
Â#1 [ffff9483bf488e58] nmi_handle at ffffffffb931d8cc
Â#2 [ffff9483bf488eb0] do_nmi at ffffffffb931dba8
Â#3 [ffff9483bf488ef0] end_repeat_nmi at ffffffffb931cd69
ÂÂÂ [exception RIP: pagetypeinfo_showfree_print+0x73]
ÂÂÂ RIP: ffffffffb8db7173Â RSP: ffff938b9fcbfda0Â RFLAGS: 00000006
ÂÂÂ RAX: fffff0c9946d7020Â RBX: ffff96073ffd5528Â RCX: 0000000000000000
ÂÂÂ RDX: 00000000001c7764Â RSI: ffffffffb9676ab1Â RDI: 0000000000000000
ÂÂÂ RBP: ffff938b9fcbfdd0ÂÂ R8: 000000000000000aÂÂ R9: 00000000fffffffe
ÂÂÂ R10: 0000000000000000Â R11: ffff938b9fcbfc36Â R12: ffff942b97758240
ÂÂÂ R13: ffffffffb942f730Â R14: ffff96073ffd5000Â R15: ffff96073ffd5180
 ORIG_RAX: ffffffffffffffff CS: 0010 SS: 0018
--- <NMI exception stack> ---
Â#4 [ffff938b9fcbfda0] pagetypeinfo_showfree_print at ffffffffb8db7173
Â#5 [ffff938b9fcbfdd8] walk_zones_in_node at ffffffffb8db74df
Â#6 [ffff938b9fcbfe20] pagetypeinfo_show at ffffffffb8db7a29
Â#7 [ffff938b9fcbfe48] seq_read at ffffffffb8e45c3c
Â#8 [ffff938b9fcbfeb8] proc_reg_read at ffffffffb8e95070
Â#9 [ffff938b9fcbfed8] vfs_read at ffffffffb8e1f2af
#10 [ffff938b9fcbff08] sys_read at ffffffffb8e2017f
#11 [ffff938b9fcbff50] system_call_fastpath at ffffffffb932579b

So we should not be in any risk of overflowing the stack.

>> -	for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
>> -		seq_printf(m, "Node %4d, zone %8s, type %12s ",
>> -					pgdat->node_id,
>> -					zone->name,
>> -					migratetype_names[mtype]);
>> -		for (order = 0; order < MAX_ORDER; ++order) {
>> +	lockdep_assert_held(&zone->lock);
>> +	lockdep_assert_irqs_disabled();
>> +
>> +	/*
>> +	 * MIGRATE_MOVABLE is usually the largest one in large memory
>> +	 * systems. We skip iterating that list. Instead, we compute it by
>> +	 * subtracting the total of the rests from free_area->nr_free.
>> +	 */
>> +	for (order = 0; order < MAX_ORDER; ++order) {
>> +		unsigned long nr_total = 0;
>> +		struct free_area *area = &(zone->free_area[order]);
>> +
>> +		for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
>>  			unsigned long freecount = 0;
>> -			struct free_area *area;
>>  			struct list_head *curr;
>>  
>> -			area = &(zone->free_area[order]);
>> -
>> +			if (mtype == MIGRATE_MOVABLE)
>> +				continue;
>>  			list_for_each(curr, &area->free_list[mtype])
>>  				freecount++;
>> -			seq_printf(m, "%6lu ", freecount);
>> +			nfree[order][mtype] = freecount;
>> +			nr_total += freecount;
>>  		}
>> +		nfree[order][MIGRATE_MOVABLE] = area->nr_free - nr_total;
>> +
>> +		/*
>> +		 * If we have already iterated more than 64k of list
>> +		 * entries, we might have hold the zone lock for too long.
>> +		 * Temporarily release the lock and reschedule before
>> +		 * continuing so that other lock waiters have a chance
>> +		 * to run.
>> +		 */
>> +		if (nr_total > (1 << 16)) {
>> +			spin_unlock_irq(&zone->lock);
>> +			cond_resched();
>> +			spin_lock_irq(&zone->lock);
>> +		}
>> +	}
>> +
>> +	for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
>> +		seq_printf(m, "Node %4d, zone %8s, type %12s ",
>> +					pgdat->node_id,
>> +					zone->name,
>> +					migratetype_names[mtype]);
>> +		for (order = 0; order < MAX_ORDER; ++order)
>> +			seq_printf(m, "%6lu ", nfree[order][mtype]);
>>  		seq_putc(m, '\n');
> This is not exactly a thing of beauty :( Presumably there might still
> be situations where the irq-off times remain excessive.
Yes, that is still possible.
>
> Why are we actually holding zone->lock so much?  Can we get away with
> holding it across the list_for_each() loop and nothing else?  If so,

We can certainly do that with the risk that the counts will be less
reliable for a given order. I can send a v2 patch if you think this is
safer.


> this still isn't a bulletproof fix.  Maybe just terminate the list
> walk if freecount reaches 1024.  Would anyone really care?
>
> Sigh.  I wonder if anyone really uses this thing for anything
> important.  Can we just remove it all?
>
Removing it will be a breakage of kernel API.

Another alternative is to mark the migration type in the page structure
so that we can do per-migration type nr_free tracking. That will be a
major change to the mm code.

I consider this patch lesser of the two evils.Â

Cheers,
Longman