kernel/linux.git/include/linux/hugetlb.h, branch v6.12.80

mm/hugetlb: fix excessive IPI broadcasts when unsharing PMD tables using mmu_gather

2026-02-19T15:29:55+00:00

commit 8ce720d5bd91e9dc16db3604aa4b1bf76770a9a1 upstream. As reported, ever since commit 1013af4f585f ("mm/hugetlb: fix huge_pmd_unshare() vs GUP-fast race") we can end up in some situations where we perform so many IPI broadcasts when unsharing hugetlb PMD page tables that it severely regresses some workloads. In particular, when we fork()+exit(), or when we munmap() a large area backed by many shared PMD tables, we perform one IPI broadcast per unshared PMD table. There are two optimizations to be had: (1) When we process (unshare) multiple such PMD tables, such as during exit(), it is sufficient to send a single IPI broadcast (as long as we respect locking rules) instead of one per PMD table. Locking prevents that any of these PMD tables could get reused before we drop the lock. (2) When we are not the last sharer (> 2 users including us), there is no need to send the IPI broadcast. The shared PMD tables cannot become exclusive (fully unshared) before an IPI will be broadcasted by the last sharer. Concurrent GUP-fast could walk into a PMD table just before we unshared it. It could then succeed in grabbing a page from the shared page table even after munmap() etc succeeded (and supressed an IPI). But there is not difference compared to GUP-fast just sleeping for a while after grabbing the page and re-enabling IRQs. Most importantly, GUP-fast will never walk into page tables that are no-longer shared, because the last sharer will issue an IPI broadcast. (if ever required, checking whether the PUD changed in GUP-fast after grabbing the page like we do in the PTE case could handle this) So let's rework PMD sharing TLB flushing + IPI sync to use the mmu_gather infrastructure so we can implement these optimizations and demystify the code at least a bit. Extend the mmu_gather infrastructure to be able to deal with our special hugetlb PMD table sharing implementation. To make initialization of the mmu_gather easier when working on a single VMA (in particular, when dealing with hugetlb), provide tlb_gather_mmu_vma(). We'll consolidate the handling for (full) unsharing of PMD tables in tlb_unshare_pmd_ptdesc() and tlb_flush_unshared_tables(), and track in "struct mmu_gather" whether we had (full) unsharing of PMD tables. Because locking is very special (concurrent unsharing+reuse must be prevented), we disallow deferring flushing to tlb_finish_mmu() and instead require an explicit earlier call to tlb_flush_unshared_tables(). From hugetlb code, we call huge_pmd_unshare_flush() where we make sure that the expected lock protecting us from concurrent unsharing+reuse is still held. Check with a VM_WARN_ON_ONCE() in tlb_finish_mmu() that tlb_flush_unshared_tables() was properly called earlier. Document it all properly. Notes about tlb_remove_table_sync_one() interaction with unsharing: There are two fairly tricky things: (1) tlb_remove_table_sync_one() is a NOP on architectures without CONFIG_MMU_GATHER_RCU_TABLE_FREE. Here, the assumption is that the previous TLB flush would send an IPI to all relevant CPUs. Careful: some architectures like x86 only send IPIs to all relevant CPUs when tlb->freed_tables is set. The relevant architectures should be selecting MMU_GATHER_RCU_TABLE_FREE, but x86 might not do that in stable kernels and it might have been problematic before this patch. Also, the arch flushing behavior (independent of IPIs) is different when tlb->freed_tables is set. Do we have to enlighten them to also take care of tlb->unshared_tables? So far we didn't care, so hopefully we are fine. Of course, we could be setting tlb->freed_tables as well, but that might then unnecessarily flush too much, because the semantics of tlb->freed_tables are a bit fuzzy. This patch changes nothing in this regard. (2) tlb_remove_table_sync_one() is not a NOP on architectures with CONFIG_MMU_GATHER_RCU_TABLE_FREE that actually don't need a sync. Take x86 as an example: in the common case (!pv, !X86_FEATURE_INVLPGB) we still issue IPIs during TLB flushes and don't actually need the second tlb_remove_table_sync_one(). This optimized can be implemented on top of this, by checking e.g., in tlb_remove_table_sync_one() whether we really need IPIs. But as described in (1), it really must honor tlb->freed_tables then to send IPIs to all relevant CPUs. Notes on TLB flushing changes: (1) Flushing for non-shared PMD tables We're converting from flush_hugetlb_tlb_range() to tlb_remove_huge_tlb_entry(). Given that we properly initialize the MMU gather in tlb_gather_mmu_vma() to be hugetlb aware, similar to __unmap_hugepage_range(), that should be fine. (2) Flushing for shared PMD tables We're converting from various things (flush_hugetlb_tlb_range(), tlb_flush_pmd_range(), flush_tlb_range()) to tlb_flush_pmd_range(). tlb_flush_pmd_range() achieves the same that tlb_remove_huge_tlb_entry() would achieve in these scenarios. Note that tlb_remove_huge_tlb_entry() also calls __tlb_remove_tlb_entry(), however that is only implemented on powerpc, which does not support PMD table sharing. Similar to (1), tlb_gather_mmu_vma() should make sure that TLB flushing keeps on working as expected. Further, note that the ptdesc_pmd_pts_dec() in huge_pmd_share() is not a concern, as we are holding the i_mmap_lock the whole time, preventing concurrent unsharing. That ptdesc_pmd_pts_dec() usage will be removed separately as a cleanup later. There are plenty more cleanups to be had, but they have to wait until this is fixed. [david@kernel.org: fix kerneldoc] Link: https://lkml.kernel.org/r/f223dd74-331c-412d-93fc-69e360a5006c@kernel.org Link: https://lkml.kernel.org/r/20251223214037.580860-5-david@kernel.org Fixes: 1013af4f585f ("mm/hugetlb: fix huge_pmd_unshare() vs GUP-fast race") Signed-off-by: David Hildenbrand (Red Hat) Reported-by: "Uschakow, Stanislav" Closes: https://lore.kernel.org/all/4d3878531c76479d9f8ca9789dc6485d@amazon.de/ Tested-by: Laurence Oberman Acked-by: Harry Yoo Reviewed-by: Lorenzo Stoakes Cc: Lance Yang Cc: Liu Shixin Cc: Oscar Salvador Cc: Rik van Riel Cc: Signed-off-by: Andrew Morton Signed-off-by: David Hildenbrand (Arm) Signed-off-by: Greg Kroah-Hartman

mm/hugetlb: fix hugetlb_pmd_shared()

2026-02-19T15:29:54+00:00

commit ca1a47cd3f5f4c46ca188b1c9a27af87d1ab2216 upstream. Patch series "mm/hugetlb: fixes for PMD table sharing (incl. using mmu_gather)", v3. One functional fix, one performance regression fix, and two related comment fixes. I cleaned up my prototype I recently shared [1] for the performance fix, deferring most of the cleanups I had in the prototype to a later point. While doing that I identified the other things. The goal of this patch set is to be backported to stable trees "fairly" easily. At least patch #1 and #4. Patch #1 fixes hugetlb_pmd_shared() not detecting any sharing Patch #2 + #3 are simple comment fixes that patch #4 interacts with. Patch #4 is a fix for the reported performance regression due to excessive IPI broadcasts during fork()+exit(). The last patch is all about TLB flushes, IPIs and mmu_gather. Read: complicated There are plenty of cleanups in the future to be had + one reasonable optimization on x86. But that's all out of scope for this series. Runtime tested, with a focus on fixing the performance regression using the original reproducer [2] on x86. This patch (of 4): We switched from (wrongly) using the page count to an independent shared count. Now, shared page tables have a refcount of 1 (excluding speculative references) and instead use ptdesc->pt_share_count to identify sharing. We didn't convert hugetlb_pmd_shared(), so right now, we would never detect a shared PMD table as such, because sharing/unsharing no longer touches the refcount of a PMD table. Page migration, like mbind() or migrate_pages() would allow for migrating folios mapped into such shared PMD tables, even though the folios are not exclusive. In smaps we would account them as "private" although they are "shared", and we would be wrongly setting the PM_MMAP_EXCLUSIVE in the pagemap interface. Fix it by properly using ptdesc_pmd_is_shared() in hugetlb_pmd_shared(). Link: https://lkml.kernel.org/r/20251223214037.580860-1-david@kernel.org Link: https://lkml.kernel.org/r/20251223214037.580860-2-david@kernel.org Link: https://lore.kernel.org/all/8cab934d-4a56-44aa-b641-bfd7e23bd673@kernel.org/ [1] Link: https://lore.kernel.org/all/8cab934d-4a56-44aa-b641-bfd7e23bd673@kernel.org/ [2] Fixes: 59d9094df3d7 ("mm: hugetlb: independent PMD page table shared count") Signed-off-by: David Hildenbrand (Red Hat) Reviewed-by: Rik van Riel Reviewed-by: Lance Yang Tested-by: Lance Yang Reviewed-by: Harry Yoo Tested-by: Laurence Oberman Reviewed-by: Lorenzo Stoakes Acked-by: Oscar Salvador Cc: Liu Shixin Cc: "Uschakow, Stanislav" Cc: Signed-off-by: Andrew Morton Signed-off-by: David Hildenbrand (Arm) Signed-off-by: Greg Kroah-Hartman

mm/hugetlb: unshare page tables during VMA split, not before

2025-06-27T10:11:40+00:00

commit 081056dc00a27bccb55ccc3c6f230a3d5fd3f7e0 upstream. Currently, __split_vma() triggers hugetlb page table unsharing through vm_ops->may_split(). This happens before the VMA lock and rmap locks are taken - which is too early, it allows racing VMA-locked page faults in our process and racing rmap walks from other processes to cause page tables to be shared again before we actually perform the split. Fix it by explicitly calling into the hugetlb unshare logic from __split_vma() in the same place where THP splitting also happens. At that point, both the VMA and the rmap(s) are write-locked. An annoying detail is that we can now call into the helper hugetlb_unshare_pmds() from two different locking contexts: 1. from hugetlb_split(), holding: - mmap lock (exclusively) - VMA lock - file rmap lock (exclusively) 2. hugetlb_unshare_all_pmds(), which I think is designed to be able to call us with only the mmap lock held (in shared mode), but currently only runs while holding mmap lock (exclusively) and VMA lock Backporting note: This commit fixes a racy protection that was introduced in commit b30c14cd6102 ("hugetlb: unshare some PMDs when splitting VMAs"); that commit claimed to fix an issue introduced in 5.13, but it should actually also go all the way back. [jannh@google.com: v2] Link: https://lkml.kernel.org/r/20250528-hugetlb-fixes-splitrace-v2-1-1329349bad1a@google.com Link: https://lkml.kernel.org/r/20250528-hugetlb-fixes-splitrace-v2-0-1329349bad1a@google.com Link: https://lkml.kernel.org/r/20250527-hugetlb-fixes-splitrace-v1-1-f4136f5ec58a@google.com Fixes: 39dde65c9940 ("[PATCH] shared page table for hugetlb page") Signed-off-by: Jann Horn Cc: Liam Howlett Reviewed-by: Lorenzo Stoakes Reviewed-by: Oscar Salvador Cc: Lorenzo Stoakes Cc: Vlastimil Babka Cc: [b30c14cd6102: hugetlb: unshare some PMDs when splitting VMAs] Cc: Signed-off-by: Andrew Morton [stable backport: added missing include] Signed-off-by: Jann Horn Signed-off-by: Greg Kroah-Hartman

mm/hugetlb: wait for hugetlb folios to be freed

2025-03-22T19:54:28+00:00

[ Upstream commit 67bab13307c83fb742c2556b06cdc39dbad27f07 ] Since the introduction of commit c77c0a8ac4c52 ("mm/hugetlb: defer freeing of huge pages if in non-task context"), which supports deferring the freeing of hugetlb pages, the allocation of contiguous memory through cma_alloc() may fail probabilistically. In the CMA allocation process, if it is found that the CMA area is occupied by in-use hugetlb folios, these in-use hugetlb folios need to be migrated to another location. When there are no available hugetlb folios in the free hugetlb pool during the migration of in-use hugetlb folios, new folios are allocated from the buddy system. A temporary state is set on the newly allocated folio. Upon completion of the hugetlb folio migration, the temporary state is transferred from the new folios to the old folios. Normally, when the old folios with the temporary state are freed, it is directly released back to the buddy system. However, due to the deferred freeing of hugetlb pages, the PageBuddy() check fails, ultimately leading to the failure of cma_alloc(). Here is a simplified call trace illustrating the process: cma_alloc() ->__alloc_contig_migrate_range() // Migrate in-use hugetlb folios ->unmap_and_move_huge_page() ->folio_putback_hugetlb() // Free old folios ->test_pages_isolated() ->__test_page_isolated_in_pageblock() ->PageBuddy(page) // Check if the page is in buddy To resolve this issue, we have implemented a function named wait_for_freed_hugetlb_folios(). This function ensures that the hugetlb folios are properly released back to the buddy system after their migration is completed. By invoking wait_for_freed_hugetlb_folios() before calling PageBuddy(), we ensure that PageBuddy() will succeed. Link: https://lkml.kernel.org/r/1739936804-18199-1-git-send-email-yangge1116@126.com Fixes: c77c0a8ac4c5 ("mm/hugetlb: defer freeing of huge pages if in non-task context") Signed-off-by: Ge Yang Reviewed-by: Muchun Song Acked-by: David Hildenbrand Cc: Baolin Wang Cc: Barry Song <21cnbao@gmail.com> Cc: Oscar Salvador Cc: Signed-off-by: Andrew Morton Signed-off-by: Sasha Levin

mm: hugetlb: Add huge page size param to huge_ptep_get_and_clear()

2025-03-13T12:02:17+00:00

commit 02410ac72ac3707936c07ede66e94360d0d65319 upstream. In order to fix a bug, arm64 needs to be told the size of the huge page for which the huge_pte is being cleared in huge_ptep_get_and_clear(). Provide for this by adding an `unsigned long sz` parameter to the function. This follows the same pattern as huge_pte_clear() and set_huge_pte_at(). This commit makes the required interface modifications to the core mm as well as all arches that implement this function (arm64, loongarch, mips, parisc, powerpc, riscv, s390, sparc). The actual arm64 bug will be fixed in a separate commit. Cc: stable@vger.kernel.org Fixes: 66b3923a1a0f ("arm64: hugetlb: add support for PTE contiguous bit") Acked-by: David Hildenbrand Reviewed-by: Alexandre Ghiti # riscv Reviewed-by: Christophe Leroy Reviewed-by: Catalin Marinas Reviewed-by: Anshuman Khandual Signed-off-by: Ryan Roberts Acked-by: Alexander Gordeev # s390 Link: https://lore.kernel.org/r/20250226120656.2400136-2-ryan.roberts@arm.com Signed-off-by: Will Deacon Signed-off-by: Greg Kroah-Hartman

mm/hugetlb: fix memfd_pin_folios resv_huge_pages leak

2024-09-26T21:01:43+00:00

memfd_pin_folios followed by unpin_folios leaves resv_huge_pages elevated if the pages were not already faulted in. During a normal page fault, resv_huge_pages is consumed here: hugetlb_fault() alloc_hugetlb_folio() dequeue_hugetlb_folio_vma() dequeue_hugetlb_folio_nodemask() dequeue_hugetlb_folio_node_exact() free_huge_pages-- resv_huge_pages-- During memfd_pin_folios, the page is created by calling alloc_hugetlb_folio_nodemask instead of alloc_hugetlb_folio, and resv_huge_pages is not modified: memfd_alloc_folio() alloc_hugetlb_folio_nodemask() dequeue_hugetlb_folio_nodemask() dequeue_hugetlb_folio_node_exact() free_huge_pages-- alloc_hugetlb_folio_nodemask has other callers that must not modify resv_huge_pages. Therefore, to fix, define an alternate version of alloc_hugetlb_folio_nodemask for this call site that adjusts resv_huge_pages. Link: https://lkml.kernel.org/r/1725373521-451395-4-git-send-email-steven.sistare@oracle.com Fixes: 89c1905d9c14 ("mm/gup: introduce memfd_pin_folios() for pinning memfd folios") Signed-off-by: Steve Sistare Acked-by: Vivek Kasireddy Cc: David Hildenbrand Cc: Jason Gunthorpe Cc: Matthew Wilcox Cc: Muchun Song Cc: Peter Xu Cc: Signed-off-by: Andrew Morton

mm/hugetlb: use __GFP_COMP for gigantic folios

2024-09-04T04:15:36+00:00

Use __GFP_COMP for gigantic folios to greatly reduce not only the amount of code but also the allocation and free time. LOC (approximately): +60, -240 Allocate and free 500 1GB hugeTLB memory without HVO by: time echo 500 >/sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages time echo 0 >/sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages Before After Alloc ~13s ~10s Free ~15s <1s The above magnitude generally holds for multiple x86 and arm64 CPU models. Link: https://lkml.kernel.org/r/20240814035451.773331-4-yuzhao@google.com Signed-off-by: Yu Zhao Reported-by: Frank van der Linden Acked-by: Zi Yan Cc: Matthew Wilcox (Oracle) Cc: Muchun Song Signed-off-by: Andrew Morton

mm/hugetlb: remove hugetlb_follow_page_mask() leftover

2024-09-02T03:25:57+00:00

We removed hugetlb_follow_page_mask() in commit 9cb28da54643 ("mm/gup: handle hugetlb in the generic follow_page_mask code") but forgot to cleanup some leftovers. While at it, simplify the hugetlb comment, it's overly detailed and rather confusing. Stating that we may end up in there during coredumping is sufficient to explain the PF_DUMPCORE usage. Link: https://lkml.kernel.org/r/20240731142000.625044-1-david@redhat.com Signed-off-by: David Hildenbrand Reviewed-by: Peter Xu Cc: Muchun Song Cc: Alexander Viro Cc: Christian Brauner Cc: Jan Kara Signed-off-by: Andrew Morton

mm/hugetlb: enforce that PMD PT sharing has split PMD PT locks

2024-09-02T03:25:51+00:00

Sharing page tables between processes but falling back to per-MM page table locks cannot possibly work. So, let's make sure that we do have split PMD locks by adding a new Kconfig option and letting that depend on CONFIG_SPLIT_PMD_PTLOCKS. Link: https://lkml.kernel.org/r/20240726150728.3159964-3-david@redhat.com Signed-off-by: David Hildenbrand Acked-by: Mike Rapoport (Microsoft) Cc: Alexander Viro Cc: Borislav Petkov Cc: Boris Ostrovsky Cc: Christian Brauner Cc: Christophe Leroy Cc: Dave Hansen Cc: "H. Peter Anvin" Cc: Ingo Molnar Cc: Juergen Gross Cc: Michael Ellerman Cc: Muchun Song Cc: "Naveen N. Rao" Cc: Nicholas Piggin Cc: Oscar Salvador Cc: Peter Xu Cc: Russell King Cc: Thomas Gleixner Signed-off-by: Andrew Morton

mm/hugetlb: fix hugetlb vs. core-mm PT locking

2024-08-16T05:16:13+00:00

We recently made GUP's common page table walking code to also walk hugetlb VMAs without most hugetlb special-casing, preparing for the future of having less hugetlb-specific page table walking code in the codebase. Turns out that we missed one page table locking detail: page table locking for hugetlb folios that are not mapped using a single PMD/PUD. Assume we have hugetlb folio that spans multiple PTEs (e.g., 64 KiB hugetlb folios on arm64 with 4 KiB base page size). GUP, as it walks the page tables, will perform a pte_offset_map_lock() to grab the PTE table lock. However, hugetlb that concurrently modifies these page tables would actually grab the mm->page_table_lock: with USE_SPLIT_PTE_PTLOCKS, the locks would differ. Something similar can happen right now with hugetlb folios that span multiple PMDs when USE_SPLIT_PMD_PTLOCKS. This issue can be reproduced [1], for example triggering: [ 3105.936100] ------------[ cut here ]------------ [ 3105.939323] WARNING: CPU: 31 PID: 2732 at mm/gup.c:142 try_grab_folio+0x11c/0x188 [ 3105.944634] Modules linked in: [...] [ 3105.974841] CPU: 31 PID: 2732 Comm: reproducer Not tainted 6.10.0-64.eln141.aarch64 #1 [ 3105.980406] Hardware name: QEMU KVM Virtual Machine, BIOS edk2-20240524-4.fc40 05/24/2024 [ 3105.986185] pstate: 60000005 (nZCv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--) [ 3105.991108] pc : try_grab_folio+0x11c/0x188 [ 3105.994013] lr : follow_page_pte+0xd8/0x430 [ 3105.996986] sp : ffff80008eafb8f0 [ 3105.999346] x29: ffff80008eafb900 x28: ffffffe8d481f380 x27: 00f80001207cff43 [ 3106.004414] x26: 0000000000000001 x25: 0000000000000000 x24: ffff80008eafba48 [ 3106.009520] x23: 0000ffff9372f000 x22: ffff7a54459e2000 x21: ffff7a546c1aa978 [ 3106.014529] x20: ffffffe8d481f3c0 x19: 0000000000610041 x18: 0000000000000001 [ 3106.019506] x17: 0000000000000001 x16: ffffffffffffffff x15: 0000000000000000 [ 3106.024494] x14: ffffb85477fdfe08 x13: 0000ffff9372ffff x12: 0000000000000000 [ 3106.029469] x11: 1fffef4a88a96be1 x10: ffff7a54454b5f0c x9 : ffffb854771b12f0 [ 3106.034324] x8 : 0008000000000000 x7 : ffff7a546c1aa980 x6 : 0008000000000080 [ 3106.038902] x5 : 00000000001207cf x4 : 0000ffff9372f000 x3 : ffffffe8d481f000 [ 3106.043420] x2 : 0000000000610041 x1 : 0000000000000001 x0 : 0000000000000000 [ 3106.047957] Call trace: [ 3106.049522] try_grab_folio+0x11c/0x188 [ 3106.051996] follow_pmd_mask.constprop.0.isra.0+0x150/0x2e0 [ 3106.055527] follow_page_mask+0x1a0/0x2b8 [ 3106.058118] __get_user_pages+0xf0/0x348 [ 3106.060647] faultin_page_range+0xb0/0x360 [ 3106.063651] do_madvise+0x340/0x598 Let's make huge_pte_lockptr() effectively use the same PT locks as any core-mm page table walker would. Add ptep_lockptr() to obtain the PTE page table lock using a pte pointer -- unfortunately we cannot convert pte_lockptr() because virt_to_page() doesn't work with kmap'ed page tables we can have with CONFIG_HIGHPTE. Handle CONFIG_PGTABLE_LEVELS correctly by checking in reverse order, such that when e.g., CONFIG_PGTABLE_LEVELS==2 with PGDIR_SIZE==P4D_SIZE==PUD_SIZE==PMD_SIZE will work as expected. Document why that works. There is one ugly case: powerpc 8xx, whereby we have an 8 MiB hugetlb folio being mapped using two PTE page tables. While hugetlb wants to take the PMD table lock, core-mm would grab the PTE table lock of one of both PTE page tables. In such corner cases, we have to make sure that both locks match, which is (fortunately!) currently guaranteed for 8xx as it does not support SMP and consequently doesn't use split PT locks. [1] https://lore.kernel.org/all/1bbfcc7f-f222-45a5-ac44-c5a1381c596d@redhat.com/ Link: https://lkml.kernel.org/r/20240801204748.99107-1-david@redhat.com Fixes: 9cb28da54643 ("mm/gup: handle hugetlb in the generic follow_page_mask code") Signed-off-by: David Hildenbrand Acked-by: Peter Xu Reviewed-by: Baolin Wang Tested-by: Baolin Wang Cc: Peter Xu Cc: Oscar Salvador Cc: Muchun Song Cc: Signed-off-by: Andrew Morton