On 17/12/24 4:02 pm, David Hildenbrand wrote:
On 17.12.24 11:07, Dev Jain wrote:On 16/12/24 10:36 pm, David Hildenbrand wrote:On 16.12.24 17:51, Dev Jain wrote:In contrast to PMD-collapse, we do not need to operate on two levels of pagetable simultaneously. Therefore, downgrade the mmap lock from write to read mode. Still take the anon_vma lock in exclusive mode so as to not waste time in the rmap path, which is anyways going to fail since the PTEs are going to be changed. Under the PTL, copy page contents, clear the PTEs, remove folio pins, and (try to) unmap the old folios. Set the PTEs to the new folio using the set_ptes() API. Signed-off-by: Dev Jain <dev.jain@xxxxxxx> --- Note: I have been trying hard to get rid of the locks in here: we still are taking the PTL around the page copying; dropping the PTL and taking it after the copying should lead to a deadlock, for example: khugepaged madvise(MADV_COLD) folio_lock() lock(ptl) lock(ptl) folio_lock() We can create a locked folio list, altogether drop both the locks, take the PTL, do everything which __collapse_huge_page_isolate() does *except* the isolation and again try locking folios, but then it will reduce efficiency of khugepaged and almost looks like a forced solution :) Please note the following discussion if anyone is interested:https://lore.kernel.org/all/66bb7496-a445-4ad7-8e56-4f2863465c54@xxxxxxx/(Apologies for not CCing the mailing list from the start)mm/khugepaged.c | 108 ++++++++++++++++++++++++++++++++++++++----------1 file changed, 87 insertions(+), 21 deletions(-) diff --git a/mm/khugepaged.c b/mm/khugepaged.c index 88beebef773e..8040b130e677 100644 --- a/mm/khugepaged.c +++ b/mm/khugepaged.c @@ -714,24 +714,28 @@ static void __collapse_huge_page_copy_succeeded(pte_t *pte, struct vm_area_struct *vma, unsigned long address, spinlock_t *ptl, - struct list_head *compound_pagelist)+ struct list_head *compound_pagelist, int order){ struct folio *src, *tmp; pte_t *_pte; pte_t pteval; - for (_pte = pte; _pte < pte + HPAGE_PMD_NR; + for (_pte = pte; _pte < pte + (1UL << order); _pte++, address += PAGE_SIZE) { pteval = ptep_get(_pte); if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); if (is_zero_pfn(pte_pfn(pteval))) { - /* - * ptl mostly unnecessary. - */ - spin_lock(ptl); - ptep_clear(vma->vm_mm, address, _pte); - spin_unlock(ptl); + if (order == HPAGE_PMD_ORDER) { + /* + * ptl mostly unnecessary. + */ + spin_lock(ptl); + ptep_clear(vma->vm_mm, address, _pte); + spin_unlock(ptl); + } else { + ptep_clear(vma->vm_mm, address, _pte); + } ksm_might_unmap_zero_page(vma->vm_mm, pteval); } } else { @@ -740,15 +744,20 @@ static void __collapse_huge_page_copy_succeeded(pte_t *pte, src = page_folio(src_page); if (!folio_test_large(src)) release_pte_folio(src); - /* - * ptl mostly unnecessary, but preempt has to - * be disabled to update the per-cpu stats - * inside folio_remove_rmap_pte(). - */ - spin_lock(ptl); - ptep_clear(vma->vm_mm, address, _pte); - folio_remove_rmap_pte(src, src_page, vma); - spin_unlock(ptl); + if (order == HPAGE_PMD_ORDER) { + /* + * ptl mostly unnecessary, but preempt has to + * be disabled to update the per-cpu stats + * inside folio_remove_rmap_pte(). + */ + spin_lock(ptl); + ptep_clear(vma->vm_mm, address, _pte);+ folio_remove_rmap_pte(src, src_page, vma); + spin_unlock(ptl); + } else { + ptep_clear(vma->vm_mm, address, _pte); + folio_remove_rmap_pte(src, src_page, vma); + }As I've talked to Nico about this code recently ... :) Are you clearing the PTE after the copy succeeded? If so, where is the TLB flush? How do you sync against concurrent write acess + GUP-fast? The sequence really must be: (1) clear PTE/PMD + flush TLB (2) check if there are unexpected page references (e.g., GUP) if so back off (3) copy page content (4) set updated PTE/PMD.Thanks...we need to ensure GUP-fast does not write when we are copying contents, so (2) will ensure that GUP-fast will see the cleared PTE and back-off.Yes, and of course, that also the CPU cannot concurrently still modify the page content while/after you copy the page content, but before you unmap+flush.To Nico, I suggested doing it simple initially, and still clear the high-level PMD entry + flush under mmap write lock, then re-map the PTE table after modifying the page table. It's not as efficient, but "harder to get wrong". Maybe that's already happening, but I stumbled over this clearing logic in __collapse_huge_page_copy_succeeded(), so I'm curious.No, I am not even touching the PMD. I guess the sequence you described should work? I just need to reverse the copying and PTE clearing order to implement this sequence.That would work, but you really have to hold the PTL for the whole period: from when you temporarily clear PTEs +_ flush the TLB, when you copy, until you re-insert the updated ones.
Ignoring the implementation and code churn part :) Is the following algorithm theoretically correct: (1) Take PTL, scan PTEs, isolate and lock the folios, set the PTEs to migration entries, check folio references. This will solve concurrent write access races. Now, we can drop the PTL...no one can write to the old folios because (1) rmap cannot run (2) folio from PTE cannot be derived. Note that migration_entry_wait_on_locked() path can be scheduled out, so this is not the same as the fault handlers spinning on the PTL. We can now safely copy old folios to new folio, then take the PTL: The PTL is available because every pagetable walker will see a migration entry and back off. We batch set the PTEs now, and release the folio locks, making the fault handlers getting out of migration_entry_wait_on_locked(). As compared to the old code, the point of failure we need to handle is when copying fails, or at some point folio isolation fails...therefore, we need to maintain a list of old PTEs corresponding to the PTEs set to migration entries.
Note that, I had suggested this "setting the PTEs to a global invalid state" thingy in our previous discussion too, but I guess simultaneously working on the PMD and PTE was the main problem there, since the walkers do not take a lock on the PMD to check if someone is changing it, when what they really are interested in is to make change at the PTE level. In fact, leaving all specifics like racing with a specific pagetable walker etc aside, I do not see why the following claim isn't true:
Claim: The (anon-private) mTHP khugepaged collapse problem is mathematically equivalent to the (anon-private) page migration problem.
The difference being, in khugepaged we need the VMA to be stable, hence have to take the mmap_read_lock(), and have to "migrate"
to a large folio instead of individual pages.If at all my theory is correct, I'll leave it to the community to decide if it's worth it to go through my brain-rot :)
So, I understand the following: Some CPU spinning on the PTL for a long time is worse than taking the mmap_write_lock(). The latter blocks this process from doing mmap()s, which, in my limited knowledge, is bad for memory-intensive processes (aligned with the fact that the maple tree was introduced to optimize VMA operations), and the former literally nukes one unit of computation from the system for a long time.When having to back-off (restore original PTEs), or for copying, you'll likely need access to the original PTEs, which were already cleared. So likely you need a temporary copy of the original PTEs somehow.That's why temporarily clearing the PMD und mmap write lock is easier to implement, at the cost of requiring the mmap lock in write mode like PMD collapse.
