1#!/usr/bin/env drgn
2#
3# Copyright (C) 2020 Roman Gushchin <guro@fb.com>
4# Copyright (C) 2020 Facebook
5
6from os import stat
7import argparse
8import sys
9
10from drgn.helpers.linux import list_for_each_entry, list_empty
11from drgn.helpers.linux import for_each_page
12from drgn.helpers.linux.cpumask import for_each_online_cpu
13from drgn.helpers.linux.percpu import per_cpu_ptr
14from drgn import container_of, FaultError, Object
15
16
17DESC = """
18This is a drgn script to provide slab statistics for memory cgroups.
19It supports cgroup v2 and v1 and can emulate memory.kmem.slabinfo
20interface of cgroup v1.
21For drgn, visit https://github.com/osandov/drgn.
22"""
23
24
25MEMCGS = {}
26
27OO_SHIFT = 16
28OO_MASK = ((1 << OO_SHIFT) - 1)
29
30
31def err(s):
32    print('slabinfo.py: error: %s' % s, file=sys.stderr, flush=True)
33    sys.exit(1)
34
35
36def find_memcg_ids(css=prog['root_mem_cgroup'].css, prefix=''):
37    if not list_empty(css.children.address_of_()):
38        for css in list_for_each_entry('struct cgroup_subsys_state',
39                                       css.children.address_of_(),
40                                       'sibling'):
41            name = prefix + '/' + css.cgroup.kn.name.string_().decode('utf-8')
42            memcg = container_of(css, 'struct mem_cgroup', 'css')
43            MEMCGS[css.cgroup.kn.id.value_()] = memcg
44            find_memcg_ids(css, name)
45
46
47def is_root_cache(s):
48    try:
49        return False if s.memcg_params.root_cache else True
50    except AttributeError:
51        return True
52
53
54def cache_name(s):
55    if is_root_cache(s):
56        return s.name.string_().decode('utf-8')
57    else:
58        return s.memcg_params.root_cache.name.string_().decode('utf-8')
59
60
61# SLUB
62
63def oo_order(s):
64    return s.oo.x >> OO_SHIFT
65
66
67def oo_objects(s):
68    return s.oo.x & OO_MASK
69
70
71def count_partial(n, fn):
72    nr_pages = 0
73    for page in list_for_each_entry('struct page', n.partial.address_of_(),
74                                    'lru'):
75         nr_pages += fn(page)
76    return nr_pages
77
78
79def count_free(page):
80    return page.objects - page.inuse
81
82
83def slub_get_slabinfo(s, cfg):
84    nr_slabs = 0
85    nr_objs = 0
86    nr_free = 0
87
88    for node in range(cfg['nr_nodes']):
89        n = s.node[node]
90        nr_slabs += n.nr_slabs.counter.value_()
91        nr_objs += n.total_objects.counter.value_()
92        nr_free += count_partial(n, count_free)
93
94    return {'active_objs': nr_objs - nr_free,
95            'num_objs': nr_objs,
96            'active_slabs': nr_slabs,
97            'num_slabs': nr_slabs,
98            'objects_per_slab': oo_objects(s),
99            'cache_order': oo_order(s),
100            'limit': 0,
101            'batchcount': 0,
102            'shared': 0,
103            'shared_avail': 0}
104
105
106def cache_show(s, cfg, objs):
107    if cfg['allocator'] == 'SLUB':
108        sinfo = slub_get_slabinfo(s, cfg)
109    else:
110        err('SLAB isn\'t supported yet')
111
112    if cfg['shared_slab_pages']:
113        sinfo['active_objs'] = objs
114        sinfo['num_objs'] = objs
115
116    print('%-17s %6lu %6lu %6u %4u %4d'
117          ' : tunables %4u %4u %4u'
118          ' : slabdata %6lu %6lu %6lu' % (
119              cache_name(s), sinfo['active_objs'], sinfo['num_objs'],
120              s.size, sinfo['objects_per_slab'], 1 << sinfo['cache_order'],
121              sinfo['limit'], sinfo['batchcount'], sinfo['shared'],
122              sinfo['active_slabs'], sinfo['num_slabs'],
123              sinfo['shared_avail']))
124
125
126def detect_kernel_config():
127    cfg = {}
128
129    cfg['nr_nodes'] = prog['nr_online_nodes'].value_()
130
131    if prog.type('struct kmem_cache').members[1][1] == 'flags':
132        cfg['allocator'] = 'SLUB'
133    elif prog.type('struct kmem_cache').members[1][1] == 'batchcount':
134        cfg['allocator'] = 'SLAB'
135    else:
136        err('Can\'t determine the slab allocator')
137
138    cfg['shared_slab_pages'] = False
139    try:
140        if prog.type('struct obj_cgroup'):
141            cfg['shared_slab_pages'] = True
142    except:
143        pass
144
145    return cfg
146
147
148def for_each_slab_page(prog):
149    PGSlab = 1 << prog.constant('PG_slab')
150    PGHead = 1 << prog.constant('PG_head')
151
152    for page in for_each_page(prog):
153        try:
154            if page.flags.value_() & PGSlab:
155                yield page
156        except FaultError:
157            pass
158
159
160def main():
161    parser = argparse.ArgumentParser(description=DESC,
162                                     formatter_class=
163                                     argparse.RawTextHelpFormatter)
164    parser.add_argument('cgroup', metavar='CGROUP',
165                        help='Target memory cgroup')
166    args = parser.parse_args()
167
168    try:
169        cgroup_id = stat(args.cgroup).st_ino
170        find_memcg_ids()
171        memcg = MEMCGS[cgroup_id]
172    except KeyError:
173        err('Can\'t find the memory cgroup')
174
175    cfg = detect_kernel_config()
176
177    print('# name            <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>'
178          ' : tunables <limit> <batchcount> <sharedfactor>'
179          ' : slabdata <active_slabs> <num_slabs> <sharedavail>')
180
181    if cfg['shared_slab_pages']:
182        obj_cgroups = set()
183        stats = {}
184        caches = {}
185
186        # find memcg pointers belonging to the specified cgroup
187        obj_cgroups.add(memcg.objcg.value_())
188        for ptr in list_for_each_entry('struct obj_cgroup',
189                                       memcg.objcg_list.address_of_(),
190                                       'list'):
191            obj_cgroups.add(ptr.value_())
192
193        # look over all slab pages, belonging to non-root memcgs
194        # and look for objects belonging to the given memory cgroup
195        for page in for_each_slab_page(prog):
196            objcg_vec_raw = page.obj_cgroups.value_()
197            if objcg_vec_raw == 0:
198                continue
199            cache = page.slab_cache
200            if not cache:
201                continue
202            addr = cache.value_()
203            caches[addr] = cache
204            # clear the lowest bit to get the true obj_cgroups
205            objcg_vec = Object(prog, page.obj_cgroups.type_,
206                               value=objcg_vec_raw & ~1)
207
208            if addr not in stats:
209                stats[addr] = 0
210
211            for i in range(oo_objects(cache)):
212                if objcg_vec[i].value_() in obj_cgroups:
213                    stats[addr] += 1
214
215        for addr in caches:
216            if stats[addr] > 0:
217                cache_show(caches[addr], cfg, stats[addr])
218
219    else:
220        for s in list_for_each_entry('struct kmem_cache',
221                                     memcg.kmem_caches.address_of_(),
222                                     'memcg_params.kmem_caches_node'):
223            cache_show(s, cfg, None)
224
225
226main()
227