On Fri, Apr 12, 2024 at 10:25 PM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
>
> On 12/04/2024 11:17, Barry Song wrote:
> > On Fri, Apr 12, 2024 at 9:56 PM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
> >>
> >> On 12/04/2024 10:43, Barry Song wrote:
> >>> On Fri, Apr 12, 2024 at 9:27 PM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
> >>>>
> >>>> Hi Barry,
> >>>>
> >>>> 2 remaining comments - otherwise looks good. (same comments I just made in the
> >>>> v4 conversation).
> >>>>
> >>>> On 12/04/2024 08:37, Barry Song wrote:
> >>>>> From: Barry Song <v-songbaohua@xxxxxxxx>
> >>>>>
> >>>>> Profiling a system blindly with mTHP has become challenging due to the
> >>>>> lack of visibility into its operations. Presenting the success rate of
> >>>>> mTHP allocations appears to be pressing need.
> >>>>>
> >>>>> Recently, I've been experiencing significant difficulty debugging
> >>>>> performance improvements and regressions without these figures. It's
> >>>>> crucial for us to understand the true effectiveness of mTHP in real-world
> >>>>> scenarios, especially in systems with fragmented memory.
> >>>>>
> >>>>> This patch establishes the framework for per-order mTHP
> >>>>> counters. It begins by introducing the anon_fault_alloc and
> >>>>> anon_fault_fallback counters. Additionally, to maintain consistency
> >>>>> with thp_fault_fallback_charge in /proc/vmstat, this patch also tracks
> >>>>> anon_fault_fallback_charge when mem_cgroup_charge fails for mTHP.
> >>>>> Incorporating additional counters should now be straightforward as well.
> >>>>>
> >>>>> Signed-off-by: Barry Song <v-songbaohua@xxxxxxxx>
> >>>>> Cc: Chris Li <chrisl@xxxxxxxxxx>
> >>>>> Cc: David Hildenbrand <david@xxxxxxxxxx>
> >>>>> Cc: Domenico Cerasuolo <cerasuolodomenico@xxxxxxxxx>
> >>>>> Cc: Kairui Song <kasong@xxxxxxxxxxx>
> >>>>> Cc: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> >>>>> Cc: Peter Xu <peterx@xxxxxxxxxx>
> >>>>> Cc: Ryan Roberts <ryan.roberts@xxxxxxx>
> >>>>> Cc: Suren Baghdasaryan <surenb@xxxxxxxxxx>
> >>>>> Cc: Yosry Ahmed <yosryahmed@xxxxxxxxxx>
> >>>>> Cc: Yu Zhao <yuzhao@xxxxxxxxxx>
> >>>>> ---
> >>>>> include/linux/huge_mm.h | 51 ++++++++++++++++++++++++++++++++++
> >>>>> mm/huge_memory.c | 61 +++++++++++++++++++++++++++++++++++++++++
> >>>>> mm/memory.c | 3 ++
> >>>>> mm/page_alloc.c | 4 +++
> >>>>> 4 files changed, 119 insertions(+)
> >>>>>
> >>>>> diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
> >>>>> index e896ca4760f6..c5beb54b97cb 100644
> >>>>> --- a/include/linux/huge_mm.h
> >>>>> +++ b/include/linux/huge_mm.h
> >>>>> @@ -264,6 +264,57 @@ unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma,
> >>>>> enforce_sysfs, orders);
> >>>>> }
> >>>>>
> >>>>> +enum mthp_stat_item {
> >>>>> + MTHP_STAT_ANON_FAULT_ALLOC,
> >>>>> + MTHP_STAT_ANON_FAULT_FALLBACK,
> >>>>> + MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE,
> >>>>> + __MTHP_STAT_COUNT
> >>>>> +};
> >>>>> +
> >>>>> +struct mthp_stat {
> >>>>> + unsigned long stats[0][__MTHP_STAT_COUNT];
> >>>>> +};
> >>>>> +
> >>>>> +extern struct mthp_stat __percpu *mthp_stats;
> >>>>> +
> >>>>> +static inline void count_mthp_stat(int order, enum mthp_stat_item item)
> >>>>> +{
> >>>>> + if (order <= 0 || order > PMD_ORDER || !mthp_stats)
> >>>>> + return;
> >>>>> +
> >>>>> + this_cpu_inc(mthp_stats->stats[order][item]);
> >>>>> +}
> >>>>> +
> >>>>> +static inline void count_mthp_stats(int order, enum mthp_stat_item item, long delta)
> >>>>> +{
> >>>>> + if (order <= 0 || order > PMD_ORDER || !mthp_stats)
> >>>>> + return;
> >>>>> +
> >>>>> + this_cpu_add(mthp_stats->stats[order][item], delta);
> >>>>> +}
> >>>>> +
> >>>>> +/*
> >>>>> + * Fold the foreign cpu mthp stats into our own.
> >>>>> + *
> >>>>> + * This is adding to the stats on one processor
> >>>>> + * but keeps the global counts constant.
> >>>>> + */
> >>>>> +static inline void mthp_stats_fold_cpu(int cpu)
> >>>>> +{
> >>>>> + struct mthp_stat *fold_stat;
> >>>>> + int i, j;
> >>>>> +
> >>>>> + if (!mthp_stats)
> >>>>> + return;
> >>>>> + fold_stat = per_cpu_ptr(mthp_stats, cpu);
> >>>>> + for (i = 1; i <= PMD_ORDER; i++) {
> >>>>> + for (j = 0; j < __MTHP_STAT_COUNT; j++) {
> >>>>> + count_mthp_stats(i, j, fold_stat->stats[i][j]);
> >>>>> + fold_stat->stats[i][j] = 0;
> >>>>> + }
> >>>>> + }
> >>>>> +}
> >>>>
> >>>> This is a pretty horrible hack; I'm pretty sure just summing for all *possible*
> >>>> cpus should work.
> >>>>
> >>>>> +
> >>>>> #define transparent_hugepage_use_zero_page() \
> >>>>> (transparent_hugepage_flags & \
> >>>>> (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
> >>>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> >>>>> index dc30139590e6..21c4ac74b484 100644
> >>>>> --- a/mm/huge_memory.c
> >>>>> +++ b/mm/huge_memory.c
> >>>>> @@ -526,6 +526,50 @@ static const struct kobj_type thpsize_ktype = {
> >>>>> .sysfs_ops = &kobj_sysfs_ops,
> >>>>> };
> >>>>>
> >>>>> +struct mthp_stat __percpu *mthp_stats;
> >>>>> +
> >>>>> +static unsigned long sum_mthp_stat(int order, enum mthp_stat_item item)
> >>>>> +{
> >>>>> + unsigned long sum = 0;
> >>>>> + int cpu;
> >>>>> +
> >>>>> + cpus_read_lock();
> >>>>> + for_each_online_cpu(cpu) {
> >>>>> + struct mthp_stat *this = per_cpu_ptr(mthp_stats, cpu);
> >>>>> +
> >>>>> + sum += this->stats[order][item];
> >>>>> + }
> >>>>> + cpus_read_unlock();
> >>>>> +
> >>>>> + return sum;
> >>>>> +}
> >>>>> +
> >>>>> +#define DEFINE_MTHP_STAT_ATTR(_name, _index) \
> >>>>> +static ssize_t _name##_show(struct kobject *kobj, \
> >>>>> + struct kobj_attribute *attr, char *buf) \
> >>>>> +{ \
> >>>>> + int order = to_thpsize(kobj)->order; \
> >>>>> + \
> >>>>> + return sysfs_emit(buf, "%lu\n", sum_mthp_stat(order, _index)); \
> >>>>> +} \
> >>>>> +static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
> >>>>> +
> >>>>> +DEFINE_MTHP_STAT_ATTR(anon_fault_alloc, MTHP_STAT_ANON_FAULT_ALLOC);
> >>>>> +DEFINE_MTHP_STAT_ATTR(anon_fault_fallback, MTHP_STAT_ANON_FAULT_FALLBACK);
> >>>>> +DEFINE_MTHP_STAT_ATTR(anon_fault_fallback_charge, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
> >>>>> +
> >>>>> +static struct attribute *stats_attrs[] = {
> >>>>> + &anon_fault_alloc_attr.attr,
> >>>>> + &anon_fault_fallback_attr.attr,
> >>>>> + &anon_fault_fallback_charge_attr.attr,
> >>>>> + NULL,
> >>>>> +};
> >>>>> +
> >>>>> +static struct attribute_group stats_attr_group = {
> >>>>> + .name = "stats",
> >>>>> + .attrs = stats_attrs,
> >>>>> +};
> >>>>> +
> >>>>> static struct thpsize *thpsize_create(int order, struct kobject *parent)
> >>>>> {
> >>>>> unsigned long size = (PAGE_SIZE << order) / SZ_1K;
> >>>>> @@ -549,6 +593,12 @@ static struct thpsize *thpsize_create(int order, struct kobject *parent)
> >>>>> return ERR_PTR(ret);
> >>>>> }
> >>>>>
> >>>>> + ret = sysfs_create_group(&thpsize->kobj, &stats_attr_group);
> >>>>> + if (ret) {
> >>>>> + kobject_put(&thpsize->kobj);
> >>>>> + return ERR_PTR(ret);
> >>>>> + }
> >>>>> +
> >>>>> thpsize->order = order;
> >>>>> return thpsize;
> >>>>> }
> >>>>> @@ -691,6 +741,11 @@ static int __init hugepage_init(void)
> >>>>> */
> >>>>> MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
> >>>>>
> >>>>> + mthp_stats = __alloc_percpu((PMD_ORDER + 1) * sizeof(mthp_stats->stats[0]),
> >>>>> + sizeof(unsigned long));
> >>>>
> >>>> Personally I think it would be cleaner to allocate statically using
> >>>> ilog2(MAX_PTRS_PER_PTE) instead of PMD_ORDER.
> >>>
> >>> Hi Ryan,
> >>>
> >>> I don't understand why MAX_PTRS_PER_PTE is the correct size. For ARM64,
> >>>
> >>> #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT)
> >>>
> >>> #define MAX_PTRS_PER_PTE PTRS_PER_PTE
> >>>
> >>> #define PTRS_PER_PTE (1 << (PAGE_SHIFT - 3))
> >>>
> >>> while PAGE_SIZE is 16KiB or 64KiB, PTRS_PER_PTE can be a huge number?
> >>>
> >>>
> >>> Am I missing something?
> >>
> >> PTRS_PER_PTE is the number of PTE entries in a PTE table. On arm64 its as follows:
> >>
> >> PAGE_SIZE PAGE_SHIFT PTRS_PER_PTE
> >> 4K 12 512
> >> 16K 14 2048
> >> 64K 16 8192
> >>
> >> So (PTRS_PER_PTE * PAGE_SIZE) = PMD_SIZE
> >>
> >> PMD_ORDER is ilog2(PMD_SIZE / PAGE_SIZE) = ilog2(PTRS_PER_PTE)
> >>
> >> MAX_PTRS_PER_PTE is just the maximum value that PTRS_PER_PTE will ever have,
> >> (and its equal to PTRS_PER_PTE except for powerpc).
> >>
> >> Pretty sure the math is correct?
> >
> > I am not convinced the math is correct :-)
> >
> > while page size is 64KiB, the page table is as below,
> > PMD_ORDER = L2 index bits = [41:29] = 13 != ilog2(8192)
>
> 1 << 13 = 8192
>
> Right? So:
>
> ilog2(8192) = 13
>
> What's wrong with that?
>
> I even checked in Python to make sure I'm not going mad:
>
> >>> import math
> >>> math.log2(8192)
> 13.0
You're correct. My mind fixated on the '16' in the line '64K 16 8192'.
I mistakenly thought ilog2(8192) equals 16. Apologies for the confusion.
>
> >
> >
> > +--------+--------+--------+--------+--------+--------+--------+--------+
> > |63 56|55 48|47 40|39 32|31 24|23 16|15 8|7 0|
> > +--------+--------+--------+--------+--------+--------+--------+--------+
> > | | | | |
> > | | | | v
> > | | | | [15:0] in-page offset
> > | | | +----------> [28:16] L3 index
> > | | +--------------------------> [41:29] L2 index
> > | +-------------------------------> [47:42] L1 index (48-bit)
> > | [51:42] L1 index (52-bit)
> > +-------------------------------------------------> [63] TTBR0/1
> >
> > while page size is 4KiB, the page table is as below,
> >
> > +--------+--------+--------+--------+--------+--------+--------+--------+
> > |63 56|55 48|47 40|39 32|31 24|23 16|15 8|7 0|
> > +--------+--------+--------+--------+--------+--------+--------+--------+
> > | | | | | |
> > | | | | | v
> > | | | | | [11:0] in-page offset
> > | | | | +-> [20:12] L3 index
> > | | | +-----------> [29:21] L2 index
> > | | +---------------------> [38:30] L1 index
> > | +-------------------------------> [47:39] L0 index
> > +-------------------------------------------------> [63] TTBR0/1
> >
> > PMD_ORDER = L2 index bits = [29:21] = 9 = ilog2(512).
> >
> > You are only correct while page size = 4KiB.
> >
> >
> >
> >
>
