[PATCH v2 03/11] mm/gup: migrate PIN_LONGTERM dev coherent pages to system

[Felix Kuehling]

Am 2021-12-08 um 6:31 a.m. schrieb Alistair Popple:
> On Tuesday, 7 December 2021 5:52:43 AM AEDT Alex Sierra wrote:
>> Avoid long term pinning for Coherent device type pages. This could
>> interfere with their own device memory manager.
>> If caller tries to get user device coherent pages with PIN_LONGTERM flag
>> set, those pages will be migrated back to system memory.
>>
>> Signed-off-by: Alex Sierra <alex.sierra at amd.com>
>> ---
>>  mm/gup.c | 32 ++++++++++++++++++++++++++++++--
>>  1 file changed, 30 insertions(+), 2 deletions(-)
>>
>> diff --git a/mm/gup.c b/mm/gup.c
>> index 886d6148d3d0..1572eacf07f4 100644
>> --- a/mm/gup.c
>> +++ b/mm/gup.c
>> @@ -1689,17 +1689,37 @@ struct page *get_dump_page(unsigned long addr)
>>  #endif /* CONFIG_ELF_CORE */
>>  
>>  #ifdef CONFIG_MIGRATION
>> +static int migrate_device_page(unsigned long address,
>> +				struct page *page)
>> +{
>> +	struct vm_area_struct *vma = find_vma(current->mm, address);
>> +	struct vm_fault vmf = {
>> +		.vma = vma,
>> +		.address = address & PAGE_MASK,
>> +		.flags = FAULT_FLAG_USER,
>> +		.pgoff = linear_page_index(vma, address),
>> +		.gfp_mask = GFP_KERNEL,
>> +		.page = page,
>> +	};
>> +	if (page->pgmap && page->pgmap->ops->migrate_to_ram)
>> +		return page->pgmap->ops->migrate_to_ram(&vmf);
> How does this synchronise against pgmap being released? As I understand things
> at this point we're not holding a reference on either the page or pgmap, so
> the page and therefore the pgmap may have been freed.
>
> I think a similar problem exists for device private fault handling as well and
> it has been on my list of things to fix for a while. I think the solution is to
> call try_get_page(), except it doesn't work with device pages due to the whole
> refcount thing. That issue is blocking a fair bit of work now so I've started
> looking into it.

At least the page should have been pinned by the __get_user_pages_locked
call in __gup_longterm_locked. That refcount is dropped in
check_and_migrate_movable_pages when it returns 0 or an error.


>
>> +
>> +	return -EBUSY;
>> +}
>> +
>>  /*
>>   * Check whether all pages are pinnable, if so return number of pages.  If some
>>   * pages are not pinnable, migrate them, and unpin all pages. Return zero if
>>   * pages were migrated, or if some pages were not successfully isolated.
>>   * Return negative error if migration fails.
>>   */
>> -static long check_and_migrate_movable_pages(unsigned long nr_pages,
>> +static long check_and_migrate_movable_pages(unsigned long start,
>> +					    unsigned long nr_pages,
>>  					    struct page **pages,
>>  					    unsigned int gup_flags)
>>  {
>>  	unsigned long i;
>> +	unsigned long page_index;
>>  	unsigned long isolation_error_count = 0;
>>  	bool drain_allow = true;
>>  	LIST_HEAD(movable_page_list);
>> @@ -1720,6 +1740,10 @@ static long check_and_migrate_movable_pages(unsigned long nr_pages,
>>  		 * If we get a movable page, since we are going to be pinning
>>  		 * these entries, try to move them out if possible.
>>  		 */
>> +		if (is_device_page(head)) {
>> +			page_index = i;
>> +			goto unpin_pages;
>> +		}
>>  		if (!is_pinnable_page(head)) {
>>  			if (PageHuge(head)) {
>>  				if (!isolate_huge_page(head, &movable_page_list))
>> @@ -1750,12 +1774,16 @@ static long check_and_migrate_movable_pages(unsigned long nr_pages,
>>  	if (list_empty(&movable_page_list) && !isolation_error_count)
>>  		return nr_pages;
>>  
>> +unpin_pages:
>>  	if (gup_flags & FOLL_PIN) {
>>  		unpin_user_pages(pages, nr_pages);
>>  	} else {
>>  		for (i = 0; i < nr_pages; i++)
>>  			put_page(pages[i]);
>>  	}
>> +	if (is_device_page(head))
>> +		return migrate_device_page(start + page_index * PAGE_SIZE, head);
> This isn't very optimal - if a range contains more than one device page (which
> seems likely) we will have to go around the whole gup/check_and_migrate loop
> once for each device page which seems unnecessary. You should be able to either
> build a list or migrate them as you go through the loop. I'm also currently
> looking into how to extend migrate_pages() to support device pages which might
> be useful here too.

We have to do it this way because page->pgmap->ops->migrate_to_ram can
migrate multiple pages per "CPU page fault" to amortize the cost of
migration. The AMD driver typically migrates 2MB at a time. Calling
page->pgmap->ops->migrate_to_ram for each page would probably be even
less optimal.

Regards,
  Felix


>
>> +
>>  	if (!list_empty(&movable_page_list)) {
>>  		ret = migrate_pages(&movable_page_list, alloc_migration_target,
>>  				    NULL, (unsigned long)&mtc, MIGRATE_SYNC,
>> @@ -1798,7 +1826,7 @@ static long __gup_longterm_locked(struct mm_struct *mm,
>>  					     NULL, gup_flags);
>>  		if (rc <= 0)
>>  			break;
>> -		rc = check_and_migrate_movable_pages(rc, pages, gup_flags);
>> +		rc = check_and_migrate_movable_pages(start, rc, pages, gup_flags);
>>  	} while (!rc);
>>  	memalloc_pin_restore(flags);
>>  
>>
>
>