[Devel] [PATCH vz9 01/20] core: Add glob_kstat, percpu kstat and account mm stat
Nikita Yushchenko
nikita.yushchenko at virtuozzo.com
Wed Oct 13 18:26:12 MSK 2021
From: Kirill Tkhai <ktkhai at virtuozzo.com>
Adds latency calculation for:
kstat_glob.swap_in
kstat_glob.page_in
kstat_glob.alloc_lat
And fail count in:
kstat_glob.alloc_fails
Also incorporates fixups patches:
kstat: Make kstat_glob::swap_in percpu - core part
ve/mm/kstat: Port diff-ve-kstat-disable-interrupts-around-seqcount-write-lock
Related buglinks:
https://jira.sw.ru/browse/PCLIN-31259
https://jira.sw.ru/browse/PSBM-33650
Signed-off-by: Kirill Tkhai <ktkhai at virtuozzo.com>
Signed-off-by: Konstantin Khlebnikov <khlebnikov at openvz.org>
Signed-off-by: Vladimir Davydov <vdavydov at parallels.com>
Rebase to vz8:
The commit [1] is trying to reimplement swap_in part of this patch, but
loses "goto out" hunk added in vz7.150.1 on rebase, bring the hunk back.
Note: On rebase I would prefere merging [1] to this patch instead of
merging this to [1].
Add vzstat.h where needed and replace __GFP_WAIT with it's successor
__GFP_RECLAIM, skip kstat_init as it is already there.
https://jira.sw.ru/browse/PSBM-127780
(cherry-picked from vz7 commit 9caa91f6a857 ("core: Add glob_kstat, percpu kstat
and account mm stat"))
Signed-off-by: Pavel Tikhomirov <ptikhomirov at virtuozzo.com>
+++
vzstat: account "page_in" and "swap_in" in nanoseconds
Up to now "page_in" and "swap_in" in /proc/vz/latency has been provided
in cpu cycles while other latencies are in nanoseconds there.
Let's make a single measure unit for all latencies, so provide swap_in
and page_in in nanoseconds as well.
Note: we left time accounting using direct rdtsc() with converting to ns
afterwards. We understand there are some issues possible with
correctness and using ktime_to_ns(ktime_get()) would be better (as it's
done for other latencies), but switching to ktime_get() results in 2%
performance loss on first memory access (pagefault + memory read),
so decided not to slowdown fastpath and be aware of possible stats
incorrectness.
https://pmc.acronis.com/browse/VSTOR-16659
Signed-off-by: Konstantin Khorenko <khorenko at virtuozzo.com>
(cherry-picked from vz7 commit aedfe36c7fc5 ("vzstat: account "page_in" and
"swap_in" in nanoseconds"))
Signed-off-by: Andrey Zhadchenko <andrey.zhadchenko at virtuozzo.com>
+++
kstat: Make kstat_glob::swap_in percpu
Patchset description:
Make kstat_glob::swap_in percpu and cleanup
This patchset continues escaping of kstat_glb_lock
and makes swap_in percpu. Also, newly unused primitives
are dropped and reduced memory usage by using percpu
seqcount (instead of separate percpu seqcount for every
kstat percpu variable).
Kirill Tkhai (4):
kstat: Make kstat_glob::swap_in percpu
kstat: Drop global kstat_lat_struct
kstat: Drop cpu argument in KSTAT_LAT_PCPU_ADD()
kstat: Make global percpu kstat_pcpu_seq instead of percpu seq for every
variable
==========================================
This patch description:
Using of global local is not good for scalability.
Better we make swap_in percpu, and it will be updated
lockless like other statistics (e.g., page_in).
Signed-off-by: Kirill Tkhai <ktkhai at virtuozzo.com>
Ported to vz8:
- Dropped all patchset but this patch, since it is already partially included
- Introduced start in do_swap_page to use it for kstat_glob.swap_in
(cherry picked from vz7 commit ed033a381e01 ("kstat: Make kstat_glob::swap_in
percpu"))
Signed-off-by: Andrey Zhadchenko <andrey.zhadchenko at virtuozzo.com>
Reviewed-by: Kirill Tkhai <ktkhai at virtuozzo.com>
+++
vzstat: account "page_in" and "swap_in" in nanoseconds
Up to now "page_in" and "swap_in" in /proc/vz/latency has been provided
in cpu cycles while other latencies are in nanoseconds there.
Let's make a single measure unit for all latencies, so provide swap_in
and page_in in nanoseconds as well.
Note: we left time accounting using direct rdtsc() with converting to ns
afterwards. We understand there are some issues possible with
correctness and using ktime_to_ns(ktime_get()) would be better (as it's
done for other latencies), but switching to ktime_get() results in 2%
performance loss on first memory access (pagefault + memory read),
so decided not to slowdown fastpath and be aware of possible stats
incorrectness.
https://pmc.acronis.com/browse/VSTOR-16659
Signed-off-by: Konstantin Khorenko <khorenko at virtuozzo.com>
(cherry-picked from vz7 commit aedfe36c7fc5 ("vzstat: account "page_in" and
"swap_in" in nanoseconds"))
Signed-off-by: Andrey Zhadchenko <andrey.zhadchenko at virtuozzo.com>
+++
mm/page_alloc: use sched_clock() instead of jiffies to measure latency
sched_clock() (which is rdtsc() on x86) gives us more precise result
than jiffies.
Q: Why do we need greater accuracy?
A: Because if we target to, say, 10000 IOPS (per cpu) then
1 ms memory allocation latency is too much and we need
to achieve less alloc latency and thus measure it.
https://pmc.acronis.com/browse/VSTOR-19040
Signed-off-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
(cherry-picked from vz7 commit 99407f6d6f50 ("mm/page_alloc: use
sched_clock() instead of jiffies to measure latency"))
Signed-off-by: Andrey Zhadchenko <andrey.zhadchenko at virtuozzo.com>
+++
ve/kstat/alloc_lat: Don't separate GFP_HIGHMEM and !GFP_HIGHMEM allocation latencies
We use mostly 64-bit systems this days. Since they don't have highmem
it's better to not segregate GFP_HIGHMEM and !GFP_HIGHMEM latencies.
For backward compatibility we still output alochigh/alochighmp fields
in /proc/vz/latency but show only zeroes.
https://jira.sw.ru/browse/PSBM-81395
Signed-off-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
https://jira.sw.ru/browse/PSBM-127780
(cherry-picked from vz7 commit 1fcbaf6d1fb2 ("ve/kstat/alloc_lat: Don't separate
GFP_HIGHMEM and !GFP_HIGHMEM allocation latencies"))
Signed-off-by: Pavel Tikhomirov <ptikhomirov at virtuozzo.com>
(cherry-picked from vz8 commit ad75d76f5a08 ("core: Add glob_kstat,
percpu kstat and account mm stat"))
Signed-off-by: Nikita Yushchenko <nikita.yushchenko at virtuozzo.com>
---
mm/memory.c | 22 +++++++++++++++++++---
mm/page_alloc.c | 32 ++++++++++++++++++++++++++++++++
2 files changed, 51 insertions(+), 3 deletions(-)
diff --git a/mm/memory.c b/mm/memory.c
index 2d0bc5ab5884..2511db99634e 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -83,6 +83,7 @@
#include <linux/uaccess.h>
#include <asm/tlb.h>
#include <asm/tlbflush.h>
+#include <asm/tsc.h>
#include "pgalloc-track.h"
#include "internal.h"
@@ -3470,6 +3471,8 @@ static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
return 0;
}
+#define CLKS2NSEC(c) ((c) * 1000000 / tsc_khz)
+
/*
* We enter with non-exclusive mmap_lock (to exclude vma changes,
* but allow concurrent faults), and pte mapped but not yet locked.
@@ -3489,7 +3492,9 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
int exclusive = 0;
vm_fault_t ret = 0;
void *shadow = NULL;
+ cycles_t start;
+ start = get_cycles();
if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
goto out;
@@ -3693,6 +3698,12 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
out:
if (si)
put_swap_device(si);
+
+ local_irq_disable();
+ KSTAT_LAT_PCPU_ADD(&kstat_glob.swap_in,
+ CLKS2NSEC(get_cycles() - start));
+ local_irq_enable();
+
return ret;
out_nomap:
pte_unmap_unlock(vmf->pte, vmf->ptl);
@@ -3704,9 +3715,7 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
unlock_page(swapcache);
put_page(swapcache);
}
- if (si)
- put_swap_device(si);
- return ret;
+ goto out;
}
/*
@@ -3834,6 +3843,7 @@ static vm_fault_t __do_fault(struct vm_fault *vmf)
{
struct vm_area_struct *vma = vmf->vma;
vm_fault_t ret;
+ cycles_t start;
/*
* Preallocate pte before we take page_lock because this might lead to
@@ -3857,6 +3867,7 @@ static vm_fault_t __do_fault(struct vm_fault *vmf)
smp_wmb(); /* See comment in __pte_alloc() */
}
+ start = get_cycles();
ret = vma->vm_ops->fault(vmf);
if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
VM_FAULT_DONE_COW)))
@@ -3875,6 +3886,11 @@ static vm_fault_t __do_fault(struct vm_fault *vmf)
else
VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
+ local_irq_disable();
+ KSTAT_LAT_PCPU_ADD(&kstat_glob.page_in,
+ CLKS2NSEC(get_cycles() - start));
+ local_irq_enable();
+
return ret;
}
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 0e511e1e45c5..ec4f4301a134 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -72,6 +72,7 @@
#include <linux/padata.h>
#include <linux/khugepaged.h>
#include <linux/buffer_head.h>
+#include <linux/vzstat.h>
#include <asm/sections.h>
#include <asm/tlbflush.h>
#include <asm/div64.h>
@@ -5404,6 +5405,34 @@ static __always_inline void warn_high_order(int order, gfp_t gfp_mask)
}
}
+static void __alloc_collect_stats(gfp_t gfp_mask, unsigned int order,
+ struct page *page, u64 time)
+{
+#ifdef CONFIG_VE
+ unsigned long flags;
+ u64 current_clock, delta;
+ int ind, cpu;
+
+ current_clock = sched_clock();
+ delta = current_clock - time;
+ if (!(gfp_mask & __GFP_RECLAIM))
+ ind = KSTAT_ALLOCSTAT_ATOMIC;
+ else
+ if (order > 0)
+ ind = KSTAT_ALLOCSTAT_LOW_MP;
+ else
+ ind = KSTAT_ALLOCSTAT_LOW;
+
+ local_irq_save(flags);
+ cpu = smp_processor_id();
+ KSTAT_LAT_PCPU_ADD(&kstat_glob.alloc_lat[ind], delta);
+
+ if (!page)
+ kstat_glob.alloc_fails[cpu][ind]++;
+ local_irq_restore(flags);
+#endif
+}
+
/*
* This is the 'heart' of the zoned buddy allocator.
*/
@@ -5414,6 +5443,7 @@ struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
unsigned int alloc_flags = ALLOC_WMARK_LOW;
gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
struct alloc_context ac = { };
+ u64 start;
/*
* There are several places where we assume that the order value is sane
@@ -5447,6 +5477,7 @@ struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
*/
alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
+ start = sched_clock();
/* First allocation attempt */
page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
if (likely(page))
@@ -5470,6 +5501,7 @@ struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
page = NULL;
}
+ __alloc_collect_stats(alloc_gfp, order, page, start);
trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
return page;
--
2.30.2
More information about the Devel
mailing list