On the arm64 platform with 4K base page config, SECTION_SIZE_BITS is set to 27, making one section 128M. The related page struct which vmemmap points to is 2M then. Commit c1cc1552616d ("arm64: MMU initialisation") optimizes the vmemmap to populate at the PMD section level which was suitable initially since hot plug granule is always one section(128M). However, commit ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") introduced a 2M(SUBSECTION_SIZE) hot plug granule, which disrupted the existing arm64 assumptions.
Considering the vmemmap_free -> unmap_hotplug_pmd_range path, when pmd_sect() is true, the entire PMD section is cleared, even if there is other effective subsection. For example page_struct_map1 and page_strcut_map2 are part of a single PMD entry and they are hot-added sequentially. Then page_struct_map1 is removed, vmemmap_free() will clear the entire PMD entry freeing the struct page map for the whole section, even though page_struct_map2 is still active. Similar problem exists with linear mapping as well, for 16K base page(PMD size = 32M) or 64K base page(PMD = 512M), their block mappings exceed SUBSECTION_SIZE. Tearing down the entire PMD mapping too will leave other subsections unmapped in the linear mapping.
To address the issue, we need to prevent PMD/PUD/CONT mappings for both linear and vmemmap for non-boot sections if corresponding size on the given base page exceeds SUBSECTION_SIZE(2MB now).
Cc: stable@vger.kernel.org # v5.4+ Fixes: ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") Reviewed-by: Catalin Marinas catalin.marinas@arm.com Signed-off-by: Zhenhua Huang quic_zhenhuah@quicinc.com --- arch/arm64/mm/mmu.c | 46 ++++++++++++++++++++++++++++++++++++--------- 1 file changed, 37 insertions(+), 9 deletions(-)
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c index b4df5bc5b1b8..b1089365f3e7 100644 --- a/arch/arm64/mm/mmu.c +++ b/arch/arm64/mm/mmu.c @@ -42,9 +42,13 @@ #include <asm/pgalloc.h> #include <asm/kfence.h>
-#define NO_BLOCK_MAPPINGS BIT(0) -#define NO_CONT_MAPPINGS BIT(1) -#define NO_EXEC_MAPPINGS BIT(2) /* assumes FEAT_HPDS is not used */ +#define NO_PMD_BLOCK_MAPPINGS BIT(0) +#define NO_PUD_BLOCK_MAPPINGS BIT(1) /* Hotplug case: do not want block mapping for PUD */ +#define NO_BLOCK_MAPPINGS (NO_PMD_BLOCK_MAPPINGS | NO_PUD_BLOCK_MAPPINGS) +#define NO_PTE_CONT_MAPPINGS BIT(2) +#define NO_PMD_CONT_MAPPINGS BIT(3) /* Hotplug case: do not want cont mapping for PMD */ +#define NO_CONT_MAPPINGS (NO_PTE_CONT_MAPPINGS | NO_PMD_CONT_MAPPINGS) +#define NO_EXEC_MAPPINGS BIT(4) /* assumes FEAT_HPDS is not used */
u64 kimage_voffset __ro_after_init; EXPORT_SYMBOL(kimage_voffset); @@ -224,7 +228,7 @@ static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
/* use a contiguous mapping if the range is suitably aligned */ if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) && - (flags & NO_CONT_MAPPINGS) == 0) + (flags & NO_PTE_CONT_MAPPINGS) == 0) __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
init_pte(ptep, addr, next, phys, __prot); @@ -254,7 +258,7 @@ static void init_pmd(pmd_t *pmdp, unsigned long addr, unsigned long end,
/* try section mapping first */ if (((addr | next | phys) & ~PMD_MASK) == 0 && - (flags & NO_BLOCK_MAPPINGS) == 0) { + (flags & NO_PMD_BLOCK_MAPPINGS) == 0) { pmd_set_huge(pmdp, phys, prot);
/* @@ -311,7 +315,7 @@ static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
/* use a contiguous mapping if the range is suitably aligned */ if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) && - (flags & NO_CONT_MAPPINGS) == 0) + (flags & NO_PMD_CONT_MAPPINGS) == 0) __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
init_pmd(pmdp, addr, next, phys, __prot, pgtable_alloc, flags); @@ -358,8 +362,8 @@ static void alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end, * For 4K granule only, attempt to put down a 1GB block */ if (pud_sect_supported() && - ((addr | next | phys) & ~PUD_MASK) == 0 && - (flags & NO_BLOCK_MAPPINGS) == 0) { + ((addr | next | phys) & ~PUD_MASK) == 0 && + (flags & NO_PUD_BLOCK_MAPPINGS) == 0) { pud_set_huge(pudp, phys, prot);
/* @@ -1178,7 +1182,13 @@ int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, { WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
- if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES)) + /* + * Hotplugged section does not support hugepages as + * PMD_SIZE (hence PUD_SIZE) section mapping covers + * struct page range that exceeds a SUBSECTION_SIZE + * i.e 2MB - for all available base page sizes. + */ + if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || system_state != SYSTEM_BOOTING) return vmemmap_populate_basepages(start, end, node, altmap); else return vmemmap_populate_hugepages(start, end, node, altmap); @@ -1340,9 +1350,27 @@ int arch_add_memory(int nid, u64 start, u64 size, struct mhp_params *params) { int ret, flags = NO_EXEC_MAPPINGS; + unsigned long start_pfn = PFN_DOWN(start); + struct mem_section *ms = __pfn_to_section(start_pfn);
VM_BUG_ON(!mhp_range_allowed(start, size, true));
+ /* should not be invoked by early section */ + WARN_ON(early_section(ms)); + + /* + * Disallow BLOCK/CONT mappings if the corresponding size exceeds + * SUBSECTION_SIZE which now is 2MB. + * + * PUD_BLOCK or PMD_CONT should consistently exceed SUBSECTION_SIZE + * across all variable page size configurations, so add them directly + */ + flags |= NO_PUD_BLOCK_MAPPINGS | NO_PMD_CONT_MAPPINGS; + if (SUBSECTION_SHIFT < PMD_SHIFT) + flags |= NO_PMD_BLOCK_MAPPINGS; + if (SUBSECTION_SHIFT < CONT_PTE_SHIFT) + flags |= NO_PTE_CONT_MAPPINGS; + if (can_set_direct_map()) flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
On 13.02.25 08:57, Zhenhua Huang wrote:
On the arm64 platform with 4K base page config, SECTION_SIZE_BITS is set to 27, making one section 128M. The related page struct which vmemmap points to is 2M then. Commit c1cc1552616d ("arm64: MMU initialisation") optimizes the vmemmap to populate at the PMD section level which was suitable initially since hot plug granule is always one section(128M). However, commit ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") introduced a 2M(SUBSECTION_SIZE) hot plug granule, which disrupted the existing arm64 assumptions.
Considering the vmemmap_free -> unmap_hotplug_pmd_range path, when pmd_sect() is true, the entire PMD section is cleared, even if there is other effective subsection. For example page_struct_map1 and page_strcut_map2 are part of a single PMD entry and they are hot-added sequentially. Then page_struct_map1 is removed, vmemmap_free() will clear the entire PMD entry freeing the struct page map for the whole section, even though page_struct_map2 is still active. Similar problem exists with linear mapping as well, for 16K base page(PMD size = 32M) or 64K base page(PMD = 512M), their block mappings exceed SUBSECTION_SIZE. Tearing down the entire PMD mapping too will leave other subsections unmapped in the linear mapping.
To address the issue, we need to prevent PMD/PUD/CONT mappings for both linear and vmemmap for non-boot sections if corresponding size on the given base page exceeds SUBSECTION_SIZE(2MB now).
Cc: stable@vger.kernel.org # v5.4+ Fixes: ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") Reviewed-by: Catalin Marinas catalin.marinas@arm.com Signed-off-by: Zhenhua Huang quic_zhenhuah@quicinc.com
Just so I understand correctly: for ordinary memory-sections-size hotplug (NVDIMM, virtio-mem), we still get a large mapping where possible?
On Thu, Feb 13, 2025 at 01:59:25PM +0100, David Hildenbrand wrote:
On 13.02.25 08:57, Zhenhua Huang wrote:
On the arm64 platform with 4K base page config, SECTION_SIZE_BITS is set to 27, making one section 128M. The related page struct which vmemmap points to is 2M then. Commit c1cc1552616d ("arm64: MMU initialisation") optimizes the vmemmap to populate at the PMD section level which was suitable initially since hot plug granule is always one section(128M). However, commit ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") introduced a 2M(SUBSECTION_SIZE) hot plug granule, which disrupted the existing arm64 assumptions.
Considering the vmemmap_free -> unmap_hotplug_pmd_range path, when pmd_sect() is true, the entire PMD section is cleared, even if there is other effective subsection. For example page_struct_map1 and page_strcut_map2 are part of a single PMD entry and they are hot-added sequentially. Then page_struct_map1 is removed, vmemmap_free() will clear the entire PMD entry freeing the struct page map for the whole section, even though page_struct_map2 is still active. Similar problem exists with linear mapping as well, for 16K base page(PMD size = 32M) or 64K base page(PMD = 512M), their block mappings exceed SUBSECTION_SIZE. Tearing down the entire PMD mapping too will leave other subsections unmapped in the linear mapping.
To address the issue, we need to prevent PMD/PUD/CONT mappings for both linear and vmemmap for non-boot sections if corresponding size on the given base page exceeds SUBSECTION_SIZE(2MB now).
Cc: stable@vger.kernel.org # v5.4+ Fixes: ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") Reviewed-by: Catalin Marinas catalin.marinas@arm.com Signed-off-by: Zhenhua Huang quic_zhenhuah@quicinc.com
Just so I understand correctly: for ordinary memory-sections-size hotplug (NVDIMM, virtio-mem), we still get a large mapping where possible?
Up to 2MB blocks only since that's the SUBSECTION_SIZE value. The vmemmap mapping is also limited to PAGE_SIZE mappings (we could use contiguous mappings for vmemmap but it's not wired up; I don't think it's worth the hassle).
On 13.02.25 16:49, Catalin Marinas wrote:
On Thu, Feb 13, 2025 at 01:59:25PM +0100, David Hildenbrand wrote:
On 13.02.25 08:57, Zhenhua Huang wrote:
On the arm64 platform with 4K base page config, SECTION_SIZE_BITS is set to 27, making one section 128M. The related page struct which vmemmap points to is 2M then. Commit c1cc1552616d ("arm64: MMU initialisation") optimizes the vmemmap to populate at the PMD section level which was suitable initially since hot plug granule is always one section(128M). However, commit ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") introduced a 2M(SUBSECTION_SIZE) hot plug granule, which disrupted the existing arm64 assumptions.
Considering the vmemmap_free -> unmap_hotplug_pmd_range path, when pmd_sect() is true, the entire PMD section is cleared, even if there is other effective subsection. For example page_struct_map1 and page_strcut_map2 are part of a single PMD entry and they are hot-added sequentially. Then page_struct_map1 is removed, vmemmap_free() will clear the entire PMD entry freeing the struct page map for the whole section, even though page_struct_map2 is still active. Similar problem exists with linear mapping as well, for 16K base page(PMD size = 32M) or 64K base page(PMD = 512M), their block mappings exceed SUBSECTION_SIZE. Tearing down the entire PMD mapping too will leave other subsections unmapped in the linear mapping.
To address the issue, we need to prevent PMD/PUD/CONT mappings for both linear and vmemmap for non-boot sections if corresponding size on the given base page exceeds SUBSECTION_SIZE(2MB now).
Cc: stable@vger.kernel.org # v5.4+ Fixes: ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") Reviewed-by: Catalin Marinas catalin.marinas@arm.com Signed-off-by: Zhenhua Huang quic_zhenhuah@quicinc.com
Just so I understand correctly: for ordinary memory-sections-size hotplug (NVDIMM, virtio-mem), we still get a large mapping where possible?
Up to 2MB blocks only since that's the SUBSECTION_SIZE value. The vmemmap mapping is also limited to PAGE_SIZE mappings (we could use contiguous mappings for vmemmap but it's not wired up; I don't think it's worth the hassle).
But that's messed up, no?
If someone hotplugs a memory section, they have to hotunplug a memory section, not parts of it.
That's why x86 does in vmemmap_populate():
if (end - start < PAGES_PER_SECTION * sizeof(struct page)) err = vmemmap_populate_basepages(start, end, node, NULL); else if (boot_cpu_has(X86_FEATURE_PSE)) err = vmemmap_populate_hugepages(start, end, node, altmap); ...
Maybe I'm missing something. Most importantly, why the weird subsection stuff is supposed to degrade ordinary hotplug of dimms/virtio-mem etc.
On Thu, Feb 13, 2025 at 05:16:37PM +0100, David Hildenbrand wrote:
On 13.02.25 16:49, Catalin Marinas wrote:
On Thu, Feb 13, 2025 at 01:59:25PM +0100, David Hildenbrand wrote:
On 13.02.25 08:57, Zhenhua Huang wrote:
On the arm64 platform with 4K base page config, SECTION_SIZE_BITS is set to 27, making one section 128M. The related page struct which vmemmap points to is 2M then. Commit c1cc1552616d ("arm64: MMU initialisation") optimizes the vmemmap to populate at the PMD section level which was suitable initially since hot plug granule is always one section(128M). However, commit ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") introduced a 2M(SUBSECTION_SIZE) hot plug granule, which disrupted the existing arm64 assumptions.
Considering the vmemmap_free -> unmap_hotplug_pmd_range path, when pmd_sect() is true, the entire PMD section is cleared, even if there is other effective subsection. For example page_struct_map1 and page_strcut_map2 are part of a single PMD entry and they are hot-added sequentially. Then page_struct_map1 is removed, vmemmap_free() will clear the entire PMD entry freeing the struct page map for the whole section, even though page_struct_map2 is still active. Similar problem exists with linear mapping as well, for 16K base page(PMD size = 32M) or 64K base page(PMD = 512M), their block mappings exceed SUBSECTION_SIZE. Tearing down the entire PMD mapping too will leave other subsections unmapped in the linear mapping.
To address the issue, we need to prevent PMD/PUD/CONT mappings for both linear and vmemmap for non-boot sections if corresponding size on the given base page exceeds SUBSECTION_SIZE(2MB now).
Cc: stable@vger.kernel.org # v5.4+ Fixes: ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") Reviewed-by: Catalin Marinas catalin.marinas@arm.com Signed-off-by: Zhenhua Huang quic_zhenhuah@quicinc.com
Just so I understand correctly: for ordinary memory-sections-size hotplug (NVDIMM, virtio-mem), we still get a large mapping where possible?
Up to 2MB blocks only since that's the SUBSECTION_SIZE value. The vmemmap mapping is also limited to PAGE_SIZE mappings (we could use contiguous mappings for vmemmap but it's not wired up; I don't think it's worth the hassle).
But that's messed up, no?
If someone hotplugs a memory section, they have to hotunplug a memory section, not parts of it.
That's why x86 does in vmemmap_populate():
if (end - start < PAGES_PER_SECTION * sizeof(struct page)) err = vmemmap_populate_basepages(start, end, node, NULL); else if (boot_cpu_has(X86_FEATURE_PSE)) err = vmemmap_populate_hugepages(start, end, node, altmap); ...
Maybe I'm missing something. Most importantly, why the weird subsection stuff is supposed to degrade ordinary hotplug of dimms/virtio-mem etc.
I think that's based on the discussion for a previous version assuming that the hotplug/unplug sizes are not guaranteed to be symmetric:
https://lore.kernel.org/lkml/a720aaa5-a75e-481e-b396-a5f2b50ed362@quicinc.co...
If that's not the case, we can indeed ignore the SUBSECTION_SIZE altogether and just rely on the start/end of the hotplugged region.
On 13.02.25 18:56, Catalin Marinas wrote:
On Thu, Feb 13, 2025 at 05:16:37PM +0100, David Hildenbrand wrote:
On 13.02.25 16:49, Catalin Marinas wrote:
On Thu, Feb 13, 2025 at 01:59:25PM +0100, David Hildenbrand wrote:
On 13.02.25 08:57, Zhenhua Huang wrote:
On the arm64 platform with 4K base page config, SECTION_SIZE_BITS is set to 27, making one section 128M. The related page struct which vmemmap points to is 2M then. Commit c1cc1552616d ("arm64: MMU initialisation") optimizes the vmemmap to populate at the PMD section level which was suitable initially since hot plug granule is always one section(128M). However, commit ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") introduced a 2M(SUBSECTION_SIZE) hot plug granule, which disrupted the existing arm64 assumptions.
Considering the vmemmap_free -> unmap_hotplug_pmd_range path, when pmd_sect() is true, the entire PMD section is cleared, even if there is other effective subsection. For example page_struct_map1 and page_strcut_map2 are part of a single PMD entry and they are hot-added sequentially. Then page_struct_map1 is removed, vmemmap_free() will clear the entire PMD entry freeing the struct page map for the whole section, even though page_struct_map2 is still active. Similar problem exists with linear mapping as well, for 16K base page(PMD size = 32M) or 64K base page(PMD = 512M), their block mappings exceed SUBSECTION_SIZE. Tearing down the entire PMD mapping too will leave other subsections unmapped in the linear mapping.
To address the issue, we need to prevent PMD/PUD/CONT mappings for both linear and vmemmap for non-boot sections if corresponding size on the given base page exceeds SUBSECTION_SIZE(2MB now).
Cc: stable@vger.kernel.org # v5.4+ Fixes: ba72b4c8cf60 ("mm/sparsemem: support sub-section hotplug") Reviewed-by: Catalin Marinas catalin.marinas@arm.com Signed-off-by: Zhenhua Huang quic_zhenhuah@quicinc.com
Just so I understand correctly: for ordinary memory-sections-size hotplug (NVDIMM, virtio-mem), we still get a large mapping where possible?
Up to 2MB blocks only since that's the SUBSECTION_SIZE value. The vmemmap mapping is also limited to PAGE_SIZE mappings (we could use contiguous mappings for vmemmap but it's not wired up; I don't think it's worth the hassle).
But that's messed up, no?
If someone hotplugs a memory section, they have to hotunplug a memory section, not parts of it.
That's why x86 does in vmemmap_populate():
if (end - start < PAGES_PER_SECTION * sizeof(struct page)) err = vmemmap_populate_basepages(start, end, node, NULL); else if (boot_cpu_has(X86_FEATURE_PSE)) err = vmemmap_populate_hugepages(start, end, node, altmap); ...
Maybe I'm missing something. Most importantly, why the weird subsection stuff is supposed to degrade ordinary hotplug of dimms/virtio-mem etc.
I think that's based on the discussion for a previous version assuming that the hotplug/unplug sizes are not guaranteed to be symmetric:
https://lore.kernel.org/lkml/a720aaa5-a75e-481e-b396-a5f2b50ed362@quicinc.co...
If that's not the case, we can indeed ignore the SUBSECTION_SIZE>
altogether and just rely on the start/end of the hotplugged region.
All cases I know about hotunplug system RAM in the same granularity they hotplugged (virtio-mem, dax/kmem, dimm, dlpar), and if they wouldn't, they wouldn't operate on sub-section sizes either way.
Regarding dax/pmem, I also recall that it happens always in the same granularity. If not, it should be fixed: this weird subsection hotplug should not make all other hotplug users suffer (e.g., no vmemmap PMD).
What can likely happen (dax/pmem) is that we hotplug something that spans part of 128 MiB section (subsections), to then hotplug something that spans another part of a 128 MiB section (subsections). Hotunplugging either should not hotplug something part of the other device (e.g., rip out the vmemmap PMD).
I think this was expressed with:
"However, if start or end is not aligned to a section boundary, such as when a subsection is hot added, populating the entire section is wasteful." -- which is what we should focus on.
I thought x86-64 would handle that case; it would surprise me if handling between both archs would have to differ in that regard: with 4k arm64 we have the same section/subsection sizes as on x86-64.
linux-stable-mirror@lists.linaro.org
-
Catalin Marinas -
David Hildenbrand -
Zhenhua Huang