| Message ID | 20250217092907.3474806-1-quic_zhenhuah@quicinc.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [v7] arm64: mm: Populate vmemmap at the page level if not section aligned | expand |
On 17.02.25 10:29, 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. > > The first problem is that 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. > > The Next problem is if we hotplug something that spans part of 128 MiB > section (subsections, let's call it memblock1), and then hotplug something > that spans another part of a 128 MiB section(subsections, let's call it > memblock2), and subsequently unplug memblock1, vmemmap_free() will clear > the entire PMD entry which also supports memblock2 even though memblock2 > is still active. > > Assuming hotplug/unplug sizes are guaranteed to be symmetric. Do the > fix similar to x86-64: populate to pages levels if start/end is not aligned > with section boundary. > > Signed-off-by: Zhenhua Huang <quic_zhenhuah@quicinc.com> > --- > arch/arm64/mm/mmu.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c > index b4df5bc5b1b8..eec1666da368 100644 > --- a/arch/arm64/mm/mmu.c > +++ b/arch/arm64/mm/mmu.c > @@ -1178,7 +1178,8 @@ 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)) > + if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || > + (end - start < PAGES_PER_SECTION * sizeof(struct page))) > return vmemmap_populate_basepages(start, end, node, altmap); > else > return vmemmap_populate_hugepages(start, end, node, altmap); Yes, this does mimic what x86 does. That handling does look weird, because it doesn't care about any address alignments, only about the size, which is odd. I wonder if we could do better and move this handling into vmemmap_populate_hugepages(), where we already have a fallback to vmemmap_populate_basepages(). Something like: One thing that confuses me is the "altmap" handling in x86-64 code: in particular why it is ignored in some cases. So that might need a bit of thought / double-checking. diff --git a/arch/x86/mm/init_64.c b/arch/x86/mm/init_64.c index 01ea7c6df3036..57542313c0000 100644 --- a/arch/x86/mm/init_64.c +++ b/arch/x86/mm/init_64.c @@ -1546,10 +1546,10 @@ int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, VM_BUG_ON(!PAGE_ALIGNED(start)); VM_BUG_ON(!PAGE_ALIGNED(end)); - 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)) + if (boot_cpu_has(X86_FEATURE_PSE)) err = vmemmap_populate_hugepages(start, end, node, altmap); + else + err = vmemmap_populate_basepages(start, end, node, NULL); else if (altmap) { pr_err_once("%s: no cpu support for altmap allocations\n", __func__); diff --git a/mm/sparse-vmemmap.c b/mm/sparse-vmemmap.c index 3287ebadd167d..8b217265b25b1 100644 --- a/mm/sparse-vmemmap.c +++ b/mm/sparse-vmemmap.c @@ -300,6 +300,10 @@ int __weak __meminit vmemmap_check_pmd(pmd_t *pmd, int node, return 0; } +/* + * Try to populate PMDs, but fallback to populating base pages when ranges + * would only partially cover a PMD. + */ int __meminit vmemmap_populate_hugepages(unsigned long start, unsigned long end, int node, struct vmem_altmap *altmap) { @@ -313,6 +317,9 @@ int __meminit vmemmap_populate_hugepages(unsigned long start, unsigned long end, for (addr = start; addr < end; addr = next) { next = pmd_addr_end(addr, end); + if (!IS_ALIGNED(addr, PMD_SIZE) || !IS_ALIGNED(next, PMD_SIZE)) + goto fallback; + pgd = vmemmap_pgd_populate(addr, node); if (!pgd) return -ENOMEM; @@ -346,6 +353,7 @@ int __meminit vmemmap_populate_hugepages(unsigned long start, unsigned long end, } } else if (vmemmap_check_pmd(pmd, node, addr, next)) continue; +fallback: if (vmemmap_populate_basepages(addr, next, node, altmap)) return -ENOMEM; }
On 2025/2/17 17:44, David Hildenbrand wrote: > On 17.02.25 10:29, 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. >> >> The first problem is that 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. >> >> The Next problem is if we hotplug something that spans part of 128 MiB >> section (subsections, let's call it memblock1), and then hotplug >> something >> that spans another part of a 128 MiB section(subsections, let's call it >> memblock2), and subsequently unplug memblock1, vmemmap_free() will clear >> the entire PMD entry which also supports memblock2 even though memblock2 >> is still active. >> >> Assuming hotplug/unplug sizes are guaranteed to be symmetric. Do the >> fix similar to x86-64: populate to pages levels if start/end is not >> aligned >> with section boundary. >> >> Signed-off-by: Zhenhua Huang <quic_zhenhuah@quicinc.com> >> --- >> arch/arm64/mm/mmu.c | 3 ++- >> 1 file changed, 2 insertions(+), 1 deletion(-) >> >> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c >> index b4df5bc5b1b8..eec1666da368 100644 >> --- a/arch/arm64/mm/mmu.c >> +++ b/arch/arm64/mm/mmu.c >> @@ -1178,7 +1178,8 @@ 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)) >> + if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || >> + (end - start < PAGES_PER_SECTION * sizeof(struct page))) >> return vmemmap_populate_basepages(start, end, node, altmap); >> else >> return vmemmap_populate_hugepages(start, end, node, altmap); > > Yes, this does mimic what x86 does. That handling does look weird, > because it > doesn't care about any address alignments, only about the size, which is > odd. > > I wonder if we could do better and move this handling > into vmemmap_populate_hugepages(), where we already have a fallback > to vmemmap_populate_basepages(). Hi David, I had the same doubt initially. After going through the codes, I noticed for vmemmap_populate(), the arguments "start" and "end" passed down should already be within one section. early path: for_each_present_section_nr __populate_section_memmap .. vmemmap_populate() hotplug path: __add_pages section_activate vmemmap_populate() Therefore.. focusing only on the size seems OK to me, and fall back solution below appears unnecessary? BTW, I have few more doubt about the original codes below, but they're not bugs, so I have not raised them. Please correct me if it's incorrect. > > Something like: > > One thing that confuses me is the "altmap" handling in x86-64 code: in > particular > why it is ignored in some cases. So that might need a bit of thought / > double-checking. > > > diff --git a/arch/x86/mm/init_64.c b/arch/x86/mm/init_64.c > index 01ea7c6df3036..57542313c0000 100644 > --- a/arch/x86/mm/init_64.c > +++ b/arch/x86/mm/init_64.c > @@ -1546,10 +1546,10 @@ int __meminit vmemmap_populate(unsigned long > start, unsigned long end, int node, > VM_BUG_ON(!PAGE_ALIGNED(start)); > VM_BUG_ON(!PAGE_ALIGNED(end)); > > - 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)) > + if (boot_cpu_has(X86_FEATURE_PSE)) > err = vmemmap_populate_hugepages(start, end, node, > altmap); > + else > + err = vmemmap_populate_basepages(start, end, node, NULL); > else if (altmap) { > pr_err_once("%s: no cpu support for altmap allocations\n", > __func__); > diff --git a/mm/sparse-vmemmap.c b/mm/sparse-vmemmap.c > index 3287ebadd167d..8b217265b25b1 100644 > --- a/mm/sparse-vmemmap.c > +++ b/mm/sparse-vmemmap.c > @@ -300,6 +300,10 @@ int __weak __meminit vmemmap_check_pmd(pmd_t *pmd, > int node, > return 0; > } > > +/* > + * Try to populate PMDs, but fallback to populating base pages when ranges > + * would only partially cover a PMD. > + */ > int __meminit vmemmap_populate_hugepages(unsigned long start, unsigned > long end, > int node, struct vmem_altmap > *altmap) > { > @@ -313,6 +317,9 @@ int __meminit vmemmap_populate_hugepages(unsigned > long start, unsigned long end, > for (addr = start; addr < end; addr = next) { This for loop appears to be redundant for arm64 as well, as above mentioned, a single call to pmd_addr_end() should suffice. > next = pmd_addr_end(addr, end); > > + if (!IS_ALIGNED(addr, PMD_SIZE) || !IS_ALIGNED(next, > PMD_SIZE)) > + goto fallback; > + > pgd = vmemmap_pgd_populate(addr, node); > if (!pgd) > return -ENOMEM; > @@ -346,6 +353,7 @@ int __meminit vmemmap_populate_hugepages(unsigned > long start, unsigned long end, > } > } else if (vmemmap_check_pmd(pmd, node, addr, next)) > continue; > +fallback: > if (vmemmap_populate_basepages(addr, next, node, altmap)) > return -ENOMEM; It seems we have no chance to call populate_basepages here? > } > >
On 17.02.25 11:34, Zhenhua Huang wrote: > > > On 2025/2/17 17:44, David Hildenbrand wrote: >> On 17.02.25 10:29, 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. >>> >>> The first problem is that 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. >>> >>> The Next problem is if we hotplug something that spans part of 128 MiB >>> section (subsections, let's call it memblock1), and then hotplug >>> something >>> that spans another part of a 128 MiB section(subsections, let's call it >>> memblock2), and subsequently unplug memblock1, vmemmap_free() will clear >>> the entire PMD entry which also supports memblock2 even though memblock2 >>> is still active. >>> >>> Assuming hotplug/unplug sizes are guaranteed to be symmetric. Do the >>> fix similar to x86-64: populate to pages levels if start/end is not >>> aligned >>> with section boundary. >>> >>> Signed-off-by: Zhenhua Huang <quic_zhenhuah@quicinc.com> >>> --- >>> arch/arm64/mm/mmu.c | 3 ++- >>> 1 file changed, 2 insertions(+), 1 deletion(-) >>> >>> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c >>> index b4df5bc5b1b8..eec1666da368 100644 >>> --- a/arch/arm64/mm/mmu.c >>> +++ b/arch/arm64/mm/mmu.c >>> @@ -1178,7 +1178,8 @@ 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)) >>> + if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || >>> + (end - start < PAGES_PER_SECTION * sizeof(struct page))) >>> return vmemmap_populate_basepages(start, end, node, altmap); >>> else >>> return vmemmap_populate_hugepages(start, end, node, altmap); >> >> Yes, this does mimic what x86 does. That handling does look weird, >> because it >> doesn't care about any address alignments, only about the size, which is >> odd. >> >> I wonder if we could do better and move this handling >> into vmemmap_populate_hugepages(), where we already have a fallback >> to vmemmap_populate_basepages(). > > Hi David, > > I had the same doubt initially. > After going through the codes, I noticed for vmemmap_populate(), the > arguments "start" and "end" passed down should already be within one > section. > early path: > for_each_present_section_nr > __populate_section_memmap > .. > vmemmap_populate() > > hotplug path: > __add_pages > section_activate > vmemmap_populate() > > Therefore.. focusing only on the size seems OK to me, and fall back > solution below appears unnecessary? Ah, in that case it is fine. Might make sense to document/enforce that somehow for the time being ... >> +/* >> + * Try to populate PMDs, but fallback to populating base pages when ranges >> + * would only partially cover a PMD. >> + */ >> int __meminit vmemmap_populate_hugepages(unsigned long start, unsigned >> long end, >> int node, struct vmem_altmap >> *altmap) >> { >> @@ -313,6 +317,9 @@ int __meminit vmemmap_populate_hugepages(unsigned >> long start, unsigned long end, >> for (addr = start; addr < end; addr = next) { > > This for loop appears to be redundant for arm64 as well, as above > mentioned, a single call to pmd_addr_end() should suffice. Right, that was what was confusing me in the first place. > >> next = pmd_addr_end(addr, end); >> >> + if (!IS_ALIGNED(addr, PMD_SIZE) || !IS_ALIGNED(next, >> PMD_SIZE)) >> + goto fallback; >> + >> pgd = vmemmap_pgd_populate(addr, node); >> if (!pgd) >> return -ENOMEM; >> @@ -346,6 +353,7 @@ int __meminit vmemmap_populate_hugepages(unsigned >> long start, unsigned long end, >> } >> } else if (vmemmap_check_pmd(pmd, node, addr, next)) >> continue; >> +fallback: >> if (vmemmap_populate_basepages(addr, next, node, altmap)) >> return -ENOMEM; > > It seems we have no chance to call populate_basepages here? Can you elaborate?
On 2025/2/17 22:30, David Hildenbrand wrote: > On 17.02.25 11:34, Zhenhua Huang wrote: >> >> >> On 2025/2/17 17:44, David Hildenbrand wrote: >>> On 17.02.25 10:29, 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. >>>> >>>> The first problem is that 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. >>>> >>>> The Next problem is if we hotplug something that spans part of 128 MiB >>>> section (subsections, let's call it memblock1), and then hotplug >>>> something >>>> that spans another part of a 128 MiB section(subsections, let's call it >>>> memblock2), and subsequently unplug memblock1, vmemmap_free() will >>>> clear >>>> the entire PMD entry which also supports memblock2 even though >>>> memblock2 >>>> is still active. >>>> >>>> Assuming hotplug/unplug sizes are guaranteed to be symmetric. Do the >>>> fix similar to x86-64: populate to pages levels if start/end is not >>>> aligned >>>> with section boundary. >>>> >>>> Signed-off-by: Zhenhua Huang <quic_zhenhuah@quicinc.com> >>>> --- >>>> arch/arm64/mm/mmu.c | 3 ++- >>>> 1 file changed, 2 insertions(+), 1 deletion(-) >>>> >>>> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c >>>> index b4df5bc5b1b8..eec1666da368 100644 >>>> --- a/arch/arm64/mm/mmu.c >>>> +++ b/arch/arm64/mm/mmu.c >>>> @@ -1178,7 +1178,8 @@ 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)) >>>> + if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || >>>> + (end - start < PAGES_PER_SECTION * sizeof(struct page))) >>>> return vmemmap_populate_basepages(start, end, node, altmap); >>>> else >>>> return vmemmap_populate_hugepages(start, end, node, altmap); >>> >>> Yes, this does mimic what x86 does. That handling does look weird, >>> because it >>> doesn't care about any address alignments, only about the size, which is >>> odd. >>> >>> I wonder if we could do better and move this handling >>> into vmemmap_populate_hugepages(), where we already have a fallback >>> to vmemmap_populate_basepages(). >> >> Hi David, >> >> I had the same doubt initially. >> After going through the codes, I noticed for vmemmap_populate(), the >> arguments "start" and "end" passed down should already be within one >> section. >> early path: >> for_each_present_section_nr >> __populate_section_memmap >> .. >> vmemmap_populate() >> >> hotplug path: >> __add_pages >> section_activate >> vmemmap_populate() >> >> Therefore.. focusing only on the size seems OK to me, and fall back >> solution below appears unnecessary? > > Ah, in that case it is fine. Might make sense to document/enforce that > somehow for the time being ... Shall I document and WARN_ON if the size exceeds? like: WARN_ON(end - start > PAGES_PER_SECTION * sizeof(struct page)) Since vmemmap_populate() is implemented per architecture, the change should apply for other architectures as well. However I have no setup to test it on... Therefore, May I implement it only for arm64 now ? Additionally, from previous discussion, the change is worth backporting(apologies for missing to CC stable kernel in this version). Keeping it for arm64 should simplify for backporting. WDYT? > > >>> +/* >>> + * Try to populate PMDs, but fallback to populating base pages when >>> ranges >>> + * would only partially cover a PMD. >>> + */ >>> int __meminit vmemmap_populate_hugepages(unsigned long start, >>> unsigned >>> long end, >>> int node, struct vmem_altmap >>> *altmap) >>> { >>> @@ -313,6 +317,9 @@ int __meminit vmemmap_populate_hugepages(unsigned >>> long start, unsigned long end, >>> for (addr = start; addr < end; addr = next) { >> >> This for loop appears to be redundant for arm64 as well, as above >> mentioned, a single call to pmd_addr_end() should suffice. > > Right, that was what was confusing me in the first place. > >> >>> next = pmd_addr_end(addr, end); >>> >>> + if (!IS_ALIGNED(addr, PMD_SIZE) || !IS_ALIGNED(next, >>> PMD_SIZE)) >>> + goto fallback; >>> + >>> pgd = vmemmap_pgd_populate(addr, node); >>> if (!pgd) >>> return -ENOMEM; >>> @@ -346,6 +353,7 @@ int __meminit vmemmap_populate_hugepages(unsigned >>> long start, unsigned long end, >>> } >>> } else if (vmemmap_check_pmd(pmd, node, addr, next)) >>> continue; >>> +fallback: >>> if (vmemmap_populate_basepages(addr, next, node, >>> altmap)) >>> return -ENOMEM; >> >> It seems we have no chance to call populate_basepages here? > > > Can you elaborate? It's invoked within vmemmap_populate_hugepages(), which is called by vmemmap_populate(). This implies that we are always performing a whole section hotplug? However, since it's common code used by other architectures like x86, RISC-V and LoongArch, it is still necessary to review the code for these architectures as well. At the very least, it's not a BUG :) > >
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c index b4df5bc5b1b8..eec1666da368 100644 --- a/arch/arm64/mm/mmu.c +++ b/arch/arm64/mm/mmu.c @@ -1178,7 +1178,8 @@ 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)) + if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || + (end - start < PAGES_PER_SECTION * sizeof(struct page))) return vmemmap_populate_basepages(start, end, node, altmap); else return vmemmap_populate_hugepages(start, end, node, altmap);
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. The first problem is that 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. The Next problem is if we hotplug something that spans part of 128 MiB section (subsections, let's call it memblock1), and then hotplug something that spans another part of a 128 MiB section(subsections, let's call it memblock2), and subsequently unplug memblock1, vmemmap_free() will clear the entire PMD entry which also supports memblock2 even though memblock2 is still active. Assuming hotplug/unplug sizes are guaranteed to be symmetric. Do the fix similar to x86-64: populate to pages levels if start/end is not aligned with section boundary. Signed-off-by: Zhenhua Huang <quic_zhenhuah@quicinc.com> --- arch/arm64/mm/mmu.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)