Re: [PATCH 2/2] selftests/hmm-tests: Add test for dirty bits

[Alistair Popple <apopple@nvidia.com>]

Mika Penttilä <mpenttil@redhat.com> writes:

> On 15.8.2022 5.35, Alistair Popple wrote:
>> Mika Penttilä <mpenttil@redhat.com> writes:
>>
>>> Hi Alistair!
>>>
>>> On 12.8.2022 8.22, Alistair Popple wrote:
>> [...]
>>
>>>> +	buffer->ptr = mmap(NULL, size,
>>>> +			   PROT_READ | PROT_WRITE,
>>>> +			   MAP_PRIVATE | MAP_ANONYMOUS,
>>>> +			   buffer->fd, 0);
>>>> +	ASSERT_NE(buffer->ptr, MAP_FAILED);
>>>> +
>>>> +	/* Initialize buffer in system memory. */
>>>> +	for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i)
>>>> +		ptr[i] = 0;
>>>> +
>>>> +	ASSERT_FALSE(write_cgroup_param(cgroup, "memory.reclaim", 1UL<<30));
>>>> +
>>>> +	/* Fault pages back in from swap as clean pages */
>>>> +	for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i)
>>>> +		tmp += ptr[i];
>>>> +
>>>> +	/* Dirty the pte */
>>>> +	for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i)
>>>> +		ptr[i] = i;
>>>> +
>>>
>>> The anon pages are quite likely in memory at this point, and dirty in pte.
>> Why would the pte be dirty? I just confirmed using some modified pagemap
>> code that on my system at least this isn't the case.
>>
>>>> +	/*
>>>> +	 * Attempt to migrate memory to device, which should fail because
>>>> +	 * hopefully some pages are backed by swap storage.
>>>> +	 */
>>>> +	ASSERT_TRUE(hmm_migrate_sys_to_dev(self->fd, buffer, npages));
>>>
>>> And pages marked dirty also now. But could you elaborate how and where the above
>>> fails in more detail, couldn't immediately see it...
>> Not if you don't have patch 1 of this series applied. If the
>> trylock_page() in migrate_vma_collect_pmd() succeeds (which it almost
>> always does) it will have cleared the pte without setting PageDirty.
>>
>
> Ah yes but I meant with the patch 1 applied, the comment "Attempt to migrate
> memory to device, which should fail because hopefully some pages are backed by
> swap storage" indicates that hmm_migrate_sys_to_dev() would fail..and there's
> that ASSERT_TRUE which means fail here.
>
> So I understand the data loss but where is the hmm_migrate_sys_to_dev() failing,
> with or wihtout patch 1 applied?

Oh right. hmm_migrate_sys_to_dev() will fail because the page is in the
swap cache, and migrate_vma_*() doesn't currently support migrating
pages with a mapping.

>> So now we have a dirty page without PageDirty set and without a dirty
>> pte. If this page gets swapped back to disk and is still in the swap
>> cache data will be lost because reclaim will see a clean page and won't
>> write it out again.
>> At least that's my understanding - please let me know if you see
>> something that doesn't make sense.
>>
>>>> +
>>>> +	ASSERT_FALSE(write_cgroup_param(cgroup, "memory.reclaim", 1UL<<30));
>>>> +
>>>> +	/* Check we still see the updated data after restoring from swap. */
>>>> +	for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i)
>>>> +		ASSERT_EQ(ptr[i], i);
>>>> +
>>>> +	hmm_buffer_free(buffer);
>>>> +	destroy_cgroup();
>>>> +}
>>>> +
>>>>    /*
>>>>     * Read anonymous memory multiple times.
>>>>     */
>>>
>>>
>>> --Mika
>>