Question on follow_page_mask
Not able to understand the first code block of follow_page_mask function. follow_huge_addr function is expected to find the page struct for the given address if it turns out to be a HugeTLB page but then when it finds the page we bug on if it had been called with FOLL_GET flag.
page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
if (!IS_ERR(page)) {
BUG_ON(flags & FOLL_GET);
return page;
}do_move_page_to_node_array calls follow_page with FOLL_GET which in turn calls follow_page_mask with FOLL_GET. On POWER, the function follow_huge_addr is defined and does not return -EINVAL like the generic one. It returns the page struct if its a HugeTLB page. Just curious to know what is the purpose behind the BUG_ON. Thank you.
Regards
Anshuman
Quoted 12 lines · Anshuman Khandual
Not able to understand the first code block of
follow_page_maskfunction.follow_huge_addrfunction is expected to find the page struct for the given address if it turns out to be a HugeTLB page but then when it finds the page we bug on if it had been called withFOLL_GETflag.page = follow_huge_addr(mm, address, flags & FOLL_WRITE); if (!IS_ERR(page)) { BUG_ON(flags & FOLL_GET); return page; }
Quoted 5 lines
do_move_page_to_node_arraycallsfollow_pagewithFOLL_GETwhich in turn callsfollow_page_maskwithFOLL_GET. On POWER, the functionfollow_huge_addris defined and does not return-EINVALlike the generic one. It returns the page struct if its a HugeTLB page. Just curious to know what is the purpose behind theBUG_ON.
I would guess requesting pin on non-reclaimable page is considered useless, meaning suspicius behavior. BUG_ON() is overkill, I think. WARN_ON_ONCE() would make it.
Not that this follow_huge_addr() on Power is not reachable via do_move_page_to_node_array(), because the vma is !vma_is_migratable().
Signature · 1 line
Kirill A. Shutemov
Quoted 19 lines · Kirill A. Shutemov
Quoted 12 lines · Anshuman Khandual
Not able to understand the first code block of
follow_page_maskfunction.follow_huge_addrfunction is expected to find the page struct for the given address if it turns out to be a HugeTLB page but then when it finds the page we bug on if it had been called withFOLL_GETflag.page = follow_huge_addr(mm, address, flags & FOLL_WRITE); if (!IS_ERR(page)) { BUG_ON(flags & FOLL_GET); return page; }Quoted 5 lines
do_move_page_to_node_arraycallsfollow_pagewithFOLL_GETwhich in turn callsfollow_page_maskwithFOLL_GET. On POWER, the functionfollow_huge_addris defined and does not return-EINVALlike the generic one. It returns the page struct if its a HugeTLB page. Just curious to know what is the purpose behind theBUG_ON.
Quoted 3 lines
I would guess requesting pin on non-reclaimable page is considered useless, meaning suspicius behavior.
BUG_ON()is overkill, I think.WARN_ON_ONCE()would make it.
No, it's there to guard against abuse, until the correct functionality is implemented: which has not so far been required, I think.
The problem is that a get_page() here is too late: it needs to be done inside each arch's implementation of follow_huge_addr(), while holding whatever is the appropriate lock, dropped by the time it returns here.
If you look through where FOLL_GET is usually implemented, such as in follow_page_pte(), but pud and pmd cases too, I hope you'll still find that they are careful to get the reference on the page while it's safe in the pagetable.
But follow_huge_addr() would need some work to offer the same guarantees:
it's good for those "peep at a page without actually getting a reference" cases, but not good enough for preventing a page for being put to some other use completely, before we've secured it with our reference.
Unless something's changed: the last time I recall the issue coming up, was when Naoya Horiguchi was working on hugetlbfs page migration: see linux-kernel/linux-mm mail thread "BUG at mm/memory.c:1489!" from 28 May 2014; and the resolution there was not to support the follow_huge_addr() case (which IIRC is peculiar to powerpc alone?).
Hugh
Quoted 19 lines · Kirill A. Shutemov
Quoted 12 lines · Anshuman Khandual
Not able to understand the first code block of
follow_page_maskfunction.follow_huge_addrfunction is expected to find the page struct for the given address if it turns out to be a HugeTLB page but then when it finds the page we bug on if it had been called withFOLL_GETflag.page = follow_huge_addr(mm, address, flags & FOLL_WRITE); if (!IS_ERR(page)) { BUG_ON(flags & FOLL_GET); return page; }Quoted 5 lines
do_move_page_to_node_arraycallsfollow_pagewithFOLL_GETwhich in turn callsfollow_page_maskwithFOLL_GET. On POWER, the functionfollow_huge_addris defined and does not return-EINVALlike the generic one. It returns the page struct if its a HugeTLB page. Just curious to know what is the purpose behind theBUG_ON.
Quoted 3 lines
I would guess requesting pin on non-reclaimable page is considered useless, meaning suspicius behavior.
BUG_ON()is overkill, I think.WARN_ON_ONCE()would make it.
pin on non-reclaimable page ? I thought the page reference is obtained for page migration purpose only. I may be missing something here.
Quoted 3 lines
Not that this
follow_huge_addr()on Power is not reachable viado_move_page_to_node_array(), because the vma is !vma_is_migratable().
Was experimenting with that enabled via ARCH_ENABLE_HUGEPAGE_MIGRATION.
Quoted 20 lines · Hugh Dickins via Linuxppc-dev
Quoted 19 lines · Kirill A. Shutemov
Quoted 17 lines · Anshuman Khandual
Not able to understand the first code block of
follow_page_maskfunction.follow_huge_addrfunction is expected to find the page struct for the given address if it turns out to be a HugeTLB page but then when it finds the page we bug on if it had been called withFOLL_GETflag.page = follow_huge_addr(mm, address, flags & FOLL_WRITE); if (!IS_ERR(page)) { BUG_ON(flags & FOLL_GET); return page; }
do_move_page_to_node_arraycallsfollow_pagewithFOLL_GETwhich in turn callsfollow_page_maskwithFOLL_GET. On POWER, the functionfollow_huge_addris defined and does not return-EINVALlike the generic one. It returns the page struct if its a HugeTLB page. Just curious to know what is the purpose behind theBUG_ON.
Quoted 3 lines
Quoted 3 lines
I would guess requesting pin on non-reclaimable page is considered useless, meaning suspicius behavior.
BUG_ON()is overkill, I think.WARN_ON_ONCE()would make it.
Hugh, thanks for such a detailed response.
Quoted 7 lines
No, it's there to guard against abuse, until the correct functionality is implemented: which has not so far been required, I think.
The problem is that a
get_page()here is too late: it needs to be done inside each arch's implementation offollow_huge_addr(), while holding whatever is the appropriate lock, dropped by the time it returns here.
Got it.
Quoted 5 lines
If you look through where
FOLL_GETis usually implemented, such as infollow_page_pte(), but pud and pmd cases too, I hope you'll still find that they are careful to get the reference on the page while it's safe in the pagetable.
yeah, true.
Quoted 5 lines
But
follow_huge_addr()would need some work to offer the same guarantees:
it's good for those "peep at a page without actually getting a reference" cases, but not good enough for preventing a page for being put to some other use completely, before we've secured it with our reference.
Yeah, I understand that now.
Quoted 6 lines
Unless something's changed: the last time I recall the issue coming up, was when Naoya Horiguchi was working on hugetlbfs page migration: see linux-kernel/linux-mm mail thread "BUG at
mm/memory.c:1489!" from 28 May 2014; and the resolution there was not to support thefollow_huge_addr()case (which IIRC is peculiar to powerpc alone?).
Yeah correct, looked into that discussion. Also tried out the discussed small patch where we do a get_page(page) if the returned page is a head of a HugeTLB compound page and the flag contains FOLL_GET.
That made the do_move_page_to_node_array function work before hitting a race condition with the test case in the commit message of e66f17ff7177 ("mm/hugetlb: take page table lock in follow_huge_pmd()").
I believe the test exposes the locking problem which is not being taken care at the follow_huge_addr function level right now on powerpc and forces it to call follow_huge_pmd (which does a BUG_ON() on powerpc) after failing to detect a huge page at the PMD level when it checked the first time around.