1============================
2Subsystem Trace Points: kmem
3============================
4
5The kmem tracing system captures events related to object and page allocation
6within the kernel. Broadly speaking there are five major subheadings.
7
8  - Slab allocation of small objects of unknown type (kmalloc)
9  - Slab allocation of small objects of known type
10  - Page allocation
11  - Per-CPU Allocator Activity
12  - External Fragmentation
13
14This document describes what each of the tracepoints is and why they
15might be useful.
16
171. Slab allocation of small objects of unknown type
18===================================================
19::
20
21  kmalloc		call_site=%lx ptr=%p bytes_req=%zu bytes_alloc=%zu gfp_flags=%s
22  kmalloc_node	call_site=%lx ptr=%p bytes_req=%zu bytes_alloc=%zu gfp_flags=%s node=%d
23  kfree		call_site=%lx ptr=%p
24
25Heavy activity for these events may indicate that a specific cache is
26justified, particularly if kmalloc slab pages are getting significantly
27internal fragmented as a result of the allocation pattern. By correlating
28kmalloc with kfree, it may be possible to identify memory leaks and where
29the allocation sites were.
30
31
322. Slab allocation of small objects of known type
33=================================================
34::
35
36  kmem_cache_alloc	call_site=%lx ptr=%p bytes_req=%zu bytes_alloc=%zu gfp_flags=%s
37  kmem_cache_alloc_node	call_site=%lx ptr=%p bytes_req=%zu bytes_alloc=%zu gfp_flags=%s node=%d
38  kmem_cache_free		call_site=%lx ptr=%p
39
40These events are similar in usage to the kmalloc-related events except that
41it is likely easier to pin the event down to a specific cache. At the time
42of writing, no information is available on what slab is being allocated from,
43but the call_site can usually be used to extrapolate that information.
44
453. Page allocation
46==================
47::
48
49  mm_page_alloc		  page=%p pfn=%lu order=%d migratetype=%d gfp_flags=%s
50  mm_page_alloc_zone_locked page=%p pfn=%lu order=%u migratetype=%d cpu=%d percpu_refill=%d
51  mm_page_free		  page=%p pfn=%lu order=%d
52  mm_page_free_batched	  page=%p pfn=%lu order=%d cold=%d
53
54These four events deal with page allocation and freeing. mm_page_alloc is
55a simple indicator of page allocator activity. Pages may be allocated from
56the per-CPU allocator (high performance) or the buddy allocator.
57
58If pages are allocated directly from the buddy allocator, the
59mm_page_alloc_zone_locked event is triggered. This event is important as high
60amounts of activity imply high activity on the zone->lock. Taking this lock
61impairs performance by disabling interrupts, dirtying cache lines between
62CPUs and serialising many CPUs.
63
64When a page is freed directly by the caller, the only mm_page_free event
65is triggered. Significant amounts of activity here could indicate that the
66callers should be batching their activities.
67
68When pages are freed in batch, the also mm_page_free_batched is triggered.
69Broadly speaking, pages are taken off the LRU lock in bulk and
70freed in batch with a page list. Significant amounts of activity here could
71indicate that the system is under memory pressure and can also indicate
72contention on the zone->lru_lock.
73
744. Per-CPU Allocator Activity
75=============================
76::
77
78  mm_page_alloc_zone_locked	page=%p pfn=%lu order=%u migratetype=%d cpu=%d percpu_refill=%d
79  mm_page_pcpu_drain		page=%p pfn=%lu order=%d cpu=%d migratetype=%d
80
81In front of the page allocator is a per-cpu page allocator. It exists only
82for order-0 pages, reduces contention on the zone->lock and reduces the
83amount of writing on struct page.
84
85When a per-CPU list is empty or pages of the wrong type are allocated,
86the zone->lock will be taken once and the per-CPU list refilled. The event
87triggered is mm_page_alloc_zone_locked for each page allocated with the
88event indicating whether it is for a percpu_refill or not.
89
90When the per-CPU list is too full, a number of pages are freed, each one
91which triggers a mm_page_pcpu_drain event.
92
93The individual nature of the events is so that pages can be tracked
94between allocation and freeing. A number of drain or refill pages that occur
95consecutively imply the zone->lock being taken once. Large amounts of per-CPU
96refills and drains could imply an imbalance between CPUs where too much work
97is being concentrated in one place. It could also indicate that the per-CPU
98lists should be a larger size. Finally, large amounts of refills on one CPU
99and drains on another could be a factor in causing large amounts of cache
100line bounces due to writes between CPUs and worth investigating if pages
101can be allocated and freed on the same CPU through some algorithm change.
102
1035. External Fragmentation
104=========================
105::
106
107  mm_page_alloc_extfrag		page=%p pfn=%lu alloc_order=%d fallback_order=%d pageblock_order=%d alloc_migratetype=%d fallback_migratetype=%d fragmenting=%d change_ownership=%d
108
109External fragmentation affects whether a high-order allocation will be
110successful or not. For some types of hardware, this is important although
111it is avoided where possible. If the system is using huge pages and needs
112to be able to resize the pool over the lifetime of the system, this value
113is important.
114
115Large numbers of this event implies that memory is fragmenting and
116high-order allocations will start failing at some time in the future. One
117means of reducing the occurrence of this event is to increase the size of
118min_free_kbytes in increments of 3*pageblock_size*nr_online_nodes where
119pageblock_size is usually the size of the default hugepage size.
120