[PATCH 5/8] checkpoint/restart of anonymous hugetlb mappings

[Oren Laadan]

On 09/17/2010 04:23 PM, Nathan Lynch wrote:
> On Thu, 2010-09-16 at 20:44 -0400, Oren Laadan wrote:
>>
>> On 09/14/2010 04:02 PM, Nathan Lynch wrote:
>>> Support checkpoint and restore of both private and shared
>>> hugepage-backed mappings established via mmap(MAP_HUGETLB).  Introduce
>>> APIs for checkpoint and restart of individual huge pages which are to
>>> be used by the sysv SHM_HUGETLB c/r code.
>>>
>>> Signed-off-by: Nathan Lynch <ntl at pobox.com>
>>
>> The code looks clean, but I need to learn more about HUGETLB
>> before I can say much...
>>
>> Do you also have test-suite for this ?
> 
> Included below is a throwaway patch to user-cr's shmem and ipcshm tests
> which will cause them to use huge pages.  You'll need to configure huge
> pages on your system; see Documentation/vm/hugetlbpage.txt in the kernel
> source.

Thanks. I'll add it when I pull these.

>> [...]
>>
>>> +static int hugetlb_dump_contents(struct ckpt_ctx *ctx, struct vm_area_struct *vma)
>>> +{
>>> +	struct ckpt_hdr_hpage hdr;
>>> +	unsigned long pageshift;
>>> +	unsigned long pagesize;
>>> +	unsigned long addr;
>>> +	int ret;
>>> +
>>> +	pageshift = huge_page_shift(hstate_vma(vma));
>>> +	pagesize = vma_kernel_pagesize(vma);
>>> +
>>> +	ckpt_hdr_hpage_init(&hdr, pageshift);
>>> +
>>> +	for (addr = vma->vm_start; addr < vma->vm_end; addr += pagesize) {
>>> +		struct page *page = NULL;
>>> +
>>> +		down_read(&vma->vm_mm->mmap_sem);
>>> +		ret = __get_user_pages(ctx->tsk, vma->vm_mm,
>>> +				       addr, 1, FOLL_DUMP | FOLL_GET,
>>> +				       &page, NULL);
>>> +		/* FOLL_DUMP gives -EFAULT for holes */
>>> +		if (ret == -EFAULT)
>>> +			ret = 0;
>>
>> With regular pages, this didn't always work, especially after they
>> slightly changed the semantics of FOLL_DUMP. So I introduced the
>> FOLL_DIRTY flag to detect dirty (non-zero) pages.  I wonder if
>> something like that may be needed here too ?
> 
> I don't think so - huge pages are never used to map regular files (they
> are always on hugetlbfs), so they can't get out of sync with a backing
> store.

Right.

Oren.