Re: [PATCH v3 21/26] mm/page_alloc: implement FREETYPE_UNMAPPED allocations

On Wed Aug 5, 2026 at 12:53 AM BST, Yosry Ahmed wrote:
>> @@ -3400,6 +3426,127 @@ static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
>>  #endif
>>  }
>>  
>> +#ifdef CONFIG_PAGE_ALLOC_UNMAPPED
>> +/* Try to allocate a page by mapping/unmapping a block from the direct map. */
>> +static inline struct page *
>> +__rmqueue_direct_map(struct zone *zone, unsigned int request_order,
>> +		     unsigned int alloc_flags, freetype_t freetype)
>> +{
>> +	unsigned int ft_flags_other = freetype_flags(freetype) ^ FREETYPE_UNMAPPED;
>> +	freetype_t ft_other = migrate_to_freetype(free_to_migratetype(freetype),
>> +						  ft_flags_other);
>> +	bool want_mapped = !(freetype_flags(freetype) & FREETYPE_UNMAPPED);
>> +	enum rmqueue_mode rmqm = RMQUEUE_NORMAL;
>> +	unsigned long irq_flags;
>> +	int nr_pageblocks, nr_freed;
>> +	struct page *page;
>> +	int alloc_order;
>> +	int err;
>> +
>> +	if (freetype_idx(ft_other) < 0)
>> +		return NULL;
>> +
>> +	/*
>> +	 * Might need a TLB shootdown. Even if IRQs are on this isn't
>> +	 * safe if the caller holds a lock (in case the other CPUs need that
>> +	 * lock to handle the shootdown IPI).
>> +	 */
>> +	if (alloc_flags & ALLOC_NOBLOCK)
>> +		return NULL;
>> +
>> +	if (!can_set_direct_map() || alloc_flags & ALLOC_NOLOCK)
>> +		return NULL;
>> +
>> +	lockdep_assert(!irqs_disabled() || unlikely(early_boot_irqs_disabled));
>> +
>> +	/*
>> +	 * Need to [un]map a whole pageblock (otherwise it might require
>> +	 * allocating pagetables). First allocate it.
>> +	 */
>> +	alloc_order = max(request_order, pageblock_order);
>> +	nr_pageblocks = 1 << (alloc_order - pageblock_order);
>> +	spin_lock_irqsave(&zone->lock, irq_flags);
>> +	/* First try a block that already has the right migratetype. */
>> +	page = __rmqueue(zone, alloc_order, ft_other, alloc_flags, &rmqm);
>> +	if (!page) {
>> +		/* Fallback to changing a block's migratetype. */
>> +		rmqm = RMQUEUE_CLAIM;
>> +		page = __rmqueue(zone, alloc_order, ft_other, alloc_flags, &rmqm);
>> +	}
>> +	spin_unlock_irqrestore(&zone->lock, irq_flags);
>> +	if (!page)
>> +		return NULL;
>
> IIUC we only try to change an entire pageblock here, but what if we
> can't? If memory is fragmented enough that many pageblocks have few
> unmapped pages in them, how do we serve a mapped allocation (e.g. a slab
> allocation)?
>
> We'll go into reclaim/compaction, but there's a chance we'll end up with
> unexpected allocation failures or OOM kills even though we have free
> memory, because unmapped memory is not movable or reclaimable (as of
> now, at least).
>
> The same could happen if many pageblocks have few mapped but unmovable
> pages in them, and we make an unmapped allocation.

Yeah, although just to be clear: those unmovable/unreclaimable pages
are _already_ un{mov,reclaim}able, ALLOC_UNMAPPED doesn't change that.
So in the situation where they prevent compaction, we were already
somewhat in trouble re memory pressure. It's just that, with
ALLOC_UNMAPPED, the allocation failures can start hitting us for
lower-order allocations whereas before they only mattered for
higher-order ones.

How much worse is that? I'm not sure - my understanding from working
with GCE people was that if fragmentation prevents allocating hugetlb
pages for guest memory, yes in theory the system can fall back to base
pages, but in practice the perf hit from doing that is intolerable, i.e.
successful higher-order allocations are basically a hard requirement in
the real world already. Is that unique to GCE or are other workloads
like that too? I don't really know :/

The OTHER thing, trying to see this as zoomed-out as possible, is that
the "unmapped pages are un{mov,reclaim}able" thing is hopefully very
temporary? I'm pretty sure we want at least compaction for guest_memfd
[0]. If we extend ALLOC_UNMAPPED into any other areas I also expect that
to be the same. Secretmem is... well, whatever. So I think the "real"
fix is to support compaction.

[0]: https://lore.kernel.org/all/[email protected]/

> I wonder if we still need a fallback case where a pageblock contains a
> mix of mapped and unmapped pages. We need to carefully handle such
> pageblocks:
> - For unmapped allocations, we need to unmap the relevant PTEs and
>   potentially do a TLB shootdown (if they were previously mapped). Maybe
>   we should always flush the TLB for simplicity for now.
> - For mapped allocations, we need to map the relevant PTEs. No TLB
>   shootdown should be needed.
>
> Assuming unmapped allocations are always zeroed by the users on alloc
> and free, we don't need to worry about zeroing pages either way.
>
> We may want to track the number of unmapped pages in such page blocks to
> now when it's fully mapped or fully unmapped and change its type, but
> maybe this can be a followup if needed.

... However, yes this might be unavoidable despite what I said above.
This was also DavidH's feeling when I chatted to him a few months back.
The hard parts of it are a) the tracking as you hinted at and b) in the
general case that means allocating pagetables. 

So yeah I think it's probably doable but I _really_ want to avoid doing
it in this series if we possibly can.