1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
2 #pragma once
3
4 #include <stdbool.h>
5
6 #include "cgroup-util.h"
7 #include "hashmap.h"
8 #include "psi-util.h"
9
10 #define DUMP_ON_KILL_COUNT 10
11 #define GROWING_SIZE_PERCENTILE 80
12
13 extern const struct hash_ops oomd_cgroup_ctx_hash_ops;
14
15 typedef struct OomdCGroupContext OomdCGroupContext;
16 typedef struct OomdSystemContext OomdSystemContext;
17
18 typedef int (oomd_compare_t)(OomdCGroupContext * const *, OomdCGroupContext * const *);
19
20 struct OomdCGroupContext {
21 char *path;
22
23 ResourcePressure memory_pressure;
24
25 uint64_t current_memory_usage;
26
27 uint64_t memory_min;
28 uint64_t memory_low;
29 uint64_t swap_usage;
30
31 uint64_t last_pgscan;
32 uint64_t pgscan;
33
34 ManagedOOMPreference preference;
35
36 /* These are only used for acting on high memory pressure. */
37 loadavg_t mem_pressure_limit;
38 usec_t mem_pressure_limit_hit_start;
39 usec_t last_had_mem_reclaim;
40 };
41
42 struct OomdSystemContext {
43 uint64_t mem_total;
44 uint64_t mem_used;
45 uint64_t swap_total;
46 uint64_t swap_used;
47 };
48
49 OomdCGroupContext *oomd_cgroup_context_free(OomdCGroupContext *ctx);
50 DEFINE_TRIVIAL_CLEANUP_FUNC(OomdCGroupContext*, oomd_cgroup_context_free);
51
52 /* All hashmaps used with these functions are expected to be of the form
53 * key: cgroup paths -> value: OomdCGroupContext. */
54
55 /* Scans all the OomdCGroupContexts in `h` and returns 1 and a set of pointers to those OomdCGroupContexts in `ret`
56 * if any of them have exceeded their supplied memory pressure limits for the `duration` length of time.
57 * `mem_pressure_limit_hit_start` is updated accordingly for the first time the limit is exceeded, and when it returns
58 * below the limit.
59 * Returns 0 and sets `ret` to an empty set if no entries exceeded limits for `duration`.
60 * Returns -ENOMEM for allocation errors. */
61 int oomd_pressure_above(Hashmap *h, usec_t duration, Set **ret);
62
63 /* Returns true if the amount of memory available (see proc(5)) is below the permyriad of memory specified by `threshold_permyriad`. */
64 bool oomd_mem_available_below(const OomdSystemContext *ctx, int threshold_permyriad);
65
66 /* Returns true if the amount of swap free is below the permyriad of swap specified by `threshold_permyriad`. */
67 bool oomd_swap_free_below(const OomdSystemContext *ctx, int threshold_permyriad);
68
69 /* Returns pgscan - last_pgscan, accounting for corner cases. */
70 uint64_t oomd_pgscan_rate(const OomdCGroupContext *c);
71
72 /* The compare functions will sort from largest to smallest, putting all the contexts with "avoid" at the end
73 * (after the smallest values). */
compare_pgscan_rate_and_memory_usage(OomdCGroupContext * const * c1,OomdCGroupContext * const * c2)74 static inline int compare_pgscan_rate_and_memory_usage(OomdCGroupContext * const *c1, OomdCGroupContext * const *c2) {
75 uint64_t diff1, diff2;
76 int r;
77
78 assert(c1);
79 assert(c2);
80
81 r = CMP((*c1)->preference, (*c2)->preference);
82 if (r != 0)
83 return r;
84
85 diff1 = oomd_pgscan_rate(*c1);
86 diff2 = oomd_pgscan_rate(*c2);
87 r = CMP(diff2, diff1);
88 if (r != 0)
89 return r;
90
91 return CMP((*c2)->current_memory_usage, (*c1)->current_memory_usage);
92 }
93
compare_swap_usage(OomdCGroupContext * const * c1,OomdCGroupContext * const * c2)94 static inline int compare_swap_usage(OomdCGroupContext * const *c1, OomdCGroupContext * const *c2) {
95 int r;
96
97 assert(c1);
98 assert(c2);
99
100 r = CMP((*c1)->preference, (*c2)->preference);
101 if (r != 0)
102 return r;
103
104 return CMP((*c2)->swap_usage, (*c1)->swap_usage);
105 }
106
107 /* Get an array of OomdCGroupContexts from `h`, qsorted from largest to smallest values according to `compare_func`.
108 * If `prefix` is not NULL, only include OomdCGroupContexts whose paths start with prefix. Otherwise all paths are sorted.
109 * Returns the number of sorted items; negative on error. */
110 int oomd_sort_cgroup_contexts(Hashmap *h, oomd_compare_t compare_func, const char *prefix, OomdCGroupContext ***ret);
111
112 /* Returns a negative value on error, 0 if no processes were killed, or 1 if processes were killed. */
113 int oomd_cgroup_kill(const char *path, bool recurse, bool dry_run);
114
115 /* The following oomd_kill_by_* functions return 1 if processes were killed, or negative otherwise. */
116 /* If `prefix` is supplied, only cgroups whose paths start with `prefix` are eligible candidates. Otherwise,
117 * everything in `h` is a candidate.
118 * Returns the killed cgroup in ret_selected. */
119 int oomd_kill_by_pgscan_rate(Hashmap *h, const char *prefix, bool dry_run, char **ret_selected);
120 int oomd_kill_by_swap_usage(Hashmap *h, uint64_t threshold_usage, bool dry_run, char **ret_selected);
121
122 int oomd_cgroup_context_acquire(const char *path, OomdCGroupContext **ret);
123 int oomd_system_context_acquire(const char *proc_swaps_path, OomdSystemContext *ret);
124
125 /* Get the OomdCGroupContext of `path` and insert it into `new_h`. The key for the inserted context will be `path`.
126 *
127 * `old_h` is used to get data used to calculate prior interval information. `old_h` can be NULL in which case there
128 * was no prior data to reference. */
129 int oomd_insert_cgroup_context(Hashmap *old_h, Hashmap *new_h, const char *path);
130
131 /* Update each OomdCGroupContext in `curr_h` with prior interval information from `old_h`. */
132 void oomd_update_cgroup_contexts_between_hashmaps(Hashmap *old_h, Hashmap *curr_h);
133
134 void oomd_dump_swap_cgroup_context(const OomdCGroupContext *ctx, FILE *f, const char *prefix);
135 void oomd_dump_memory_pressure_cgroup_context(const OomdCGroupContext *ctx, FILE *f, const char *prefix);
136 void oomd_dump_system_context(const OomdSystemContext *ctx, FILE *f, const char *prefix);
137