| 1 | |
| 2 | |
| 3 | * memory tiering: migrate cold pages in node 0 and hot pages in node 1 to node
|
| 4 | * 1 and node 0, respectively. Adjust the hotness/coldness threshold aiming
|
| 5 | * resulting 99.6 % node 0 utilization ratio.
|
| 6 |
|
| 7 |
|
| 8 | #define pr_fmt(fmt) "damon_sample_mtier: " fmt
|
| 9 |
|
| 10 | #include <linux/damon.h>
|
| 11 | #include <linux/init.h>
|
| 12 | #include <linux/kernel.h>
|
| 13 | #include <linux/module.h>
|
| 14 |
|
| 15 | #ifdef MODULE_PARAM_PREFIX
|
| 16 | #undef MODULE_PARAM_PREFIX
|
| 17 | #endif
|
| 18 | #define MODULE_PARAM_PREFIX "damon_sample_mtier."
|
| 19 |
|
| 20 | static unsigned long node0_start_addr __read_mostly;
|
| 21 | module_param(node0_start_addr, ulong, 0600);
|
| 22 |
|
| 23 | static unsigned long node0_end_addr __read_mostly;
|
| 24 | module_param(node0_end_addr, ulong, 0600);
|
| 25 |
|
| 26 | static unsigned long node1_start_addr __read_mostly;
|
| 27 | module_param(node1_start_addr, ulong, 0600);
|
| 28 |
|
| 29 | static unsigned long node1_end_addr __read_mostly;
|
| 30 | module_param(node1_end_addr, ulong, 0600);
|
| 31 |
|
| 32 | static unsigned long node0_mem_used_bp __read_mostly = 9970;
|
| 33 | module_param(node0_mem_used_bp, ulong, 0600);
|
| 34 |
|
| 35 | static unsigned long node0_mem_free_bp __read_mostly = 50;
|
| 36 | module_param(node0_mem_free_bp, ulong, 0600);
|
| 37 |
|
| 38 | static int damon_sample_mtier_enable_store(
|
| 39 | const char *val, const struct kernel_param *kp);
|
| 40 |
|
| 41 | static const struct kernel_param_ops enabled_param_ops = {
|
| 42 | .set = damon_sample_mtier_enable_store,
|
| 43 | .get = param_get_bool,
|
| 44 | };
|
| 45 |
|
| 46 | static bool enabled __read_mostly;
|
| 47 | module_param_cb(enabled, &enabled_param_ops, &enabled, 0600);
|
| 48 | MODULE_PARM_DESC(enabled, "Enable or disable DAMON_SAMPLE_MTIER");
|
| 49 |
|
| 50 | static bool detect_node_addresses __read_mostly;
|
| 51 | module_param(detect_node_addresses, bool, 0600);
|
| 52 |
|
| 53 | static struct damon_ctx *ctxs[2];
|
| 54 |
|
| 55 | struct region_range {
|
| 56 | phys_addr_t start;
|
| 57 | phys_addr_t end;
|
| 58 | };
|
| 59 |
|
| 60 | static int nid_to_phys(int target_node, struct region_range *range)
|
| 61 | {
|
| 62 | if (!node_online(target_node)) {
|
| 63 | pr_err("NUMA node %d is not online\n", target_node);
|
| 64 | return -EINVAL;
|
| 65 | }
|
| 66 |
|
| 67 | range->start = PFN_PHYS(node_start_pfn(target_node));
|
| 68 | range->end = PFN_PHYS(node_end_pfn(target_node));
|
| 69 |
|
| 70 | return 0;
|
| 71 | }
|
| 72 |
|
| 73 | static struct damon_ctx *damon_sample_mtier_build_ctx(bool promote)
|
| 74 | {
|
| 75 | struct damon_ctx *ctx;
|
| 76 | struct damon_attrs attrs;
|
| 77 | struct damon_target *target;
|
| 78 | struct damos *scheme;
|
| 79 | struct damos_quota_goal *quota_goal;
|
| 80 | struct damos_filter *filter;
|
| 81 | struct region_range addr;
|
| 82 | struct damon_addr_range range;
|
| 83 | int ret;
|
| 84 |
|
| 85 | ctx = damon_new_ctx();
|
| 86 | if (!ctx)
|
| 87 | return NULL;
|
| 88 | attrs = (struct damon_attrs) {
|
| 89 | .sample_interval = 5 * USEC_PER_MSEC,
|
| 90 | .aggr_interval = 100 * USEC_PER_MSEC,
|
| 91 | .ops_update_interval = 60 * USEC_PER_MSEC * MSEC_PER_SEC,
|
| 92 | .min_nr_regions = 10,
|
| 93 | .max_nr_regions = 1000,
|
| 94 | };
|
| 95 |
|
| 96 | |
| 97 | * auto-tune sampling and aggregation interval aiming 4% DAMON-observed
|
| 98 | * accesses ratio, keeping sampling interval in [5ms, 10s] range.
|
| 99 |
|
| 100 | attrs.intervals_goal = (struct damon_intervals_goal) {
|
| 101 | .access_bp = 400, .aggrs = 3,
|
| 102 | .min_sample_us = 5000, .max_sample_us = 10000000,
|
| 103 | };
|
| 104 | if (damon_set_attrs(ctx, &attrs))
|
| 105 | goto free_out;
|
| 106 | if (damon_select_ops(ctx, DAMON_OPS_PADDR))
|
| 107 | goto free_out;
|
| 108 |
|
| 109 | target = damon_new_target();
|
| 110 | if (!target)
|
| 111 | goto free_out;
|
| 112 | damon_add_target(ctx, target);
|
| 113 |
|
| 114 | if (detect_node_addresses) {
|
| 115 | ret = promote ? nid_to_phys(1, &addr) : nid_to_phys(0, &addr);
|
| 116 | if (ret)
|
| 117 | goto free_out;
|
| 118 | } else {
|
| 119 | addr.start = promote ? node1_start_addr : node0_start_addr;
|
| 120 | addr.end = promote ? node1_end_addr : node0_end_addr;
|
| 121 | }
|
| 122 |
|
| 123 | if (addr.start >= addr.end)
|
| 124 | goto free_out;
|
| 125 |
|
| 126 | range.start = addr.start;
|
| 127 | range.end = addr.end;
|
| 128 |
|
| 129 | ret = damon_set_regions(target, &range, 1, DAMON_MIN_REGION_SZ);
|
| 130 | if (ret)
|
| 131 | goto free_out;
|
| 132 |
|
| 133 | scheme = damon_new_scheme(
|
| 134 |
|
| 135 | &(struct damos_access_pattern) {
|
| 136 | .min_sz_region = PAGE_SIZE,
|
| 137 | .max_sz_region = ULONG_MAX,
|
| 138 | .min_nr_accesses = promote ? 1 : 0,
|
| 139 | .max_nr_accesses = promote ? UINT_MAX : 0,
|
| 140 | .min_age_region = 0,
|
| 141 | .max_age_region = UINT_MAX},
|
| 142 |
|
| 143 | promote ? DAMOS_MIGRATE_HOT : DAMOS_MIGRATE_COLD,
|
| 144 | 1000000,
|
| 145 | &(struct damos_quota){
|
| 146 |
|
| 147 | .reset_interval = 1000,
|
| 148 | .sz = 200 * 1024 * 1024,
|
| 149 |
|
| 150 | .weight_sz = 0,
|
| 151 | .weight_nr_accesses = 100,
|
| 152 | .weight_age = 100,
|
| 153 | },
|
| 154 | &(struct damos_watermarks){},
|
| 155 | promote ? 0 : 1);
|
| 156 | if (!scheme)
|
| 157 | goto free_out;
|
| 158 | damon_set_schemes(ctx, &scheme, 1);
|
| 159 |
|
| 160 | if (!node0_mem_used_bp || !node0_mem_free_bp)
|
| 161 | goto free_out;
|
| 162 | quota_goal = damos_new_quota_goal(
|
| 163 | promote ? DAMOS_QUOTA_NODE_MEM_USED_BP :
|
| 164 | DAMOS_QUOTA_NODE_MEM_FREE_BP,
|
| 165 | promote ? node0_mem_used_bp : node0_mem_free_bp);
|
| 166 | if (!quota_goal)
|
| 167 | goto free_out;
|
| 168 | quota_goal->nid = 0;
|
| 169 | damos_add_quota_goal(&scheme->quota, quota_goal);
|
| 170 | filter = damos_new_filter(DAMOS_FILTER_TYPE_YOUNG, true, promote);
|
| 171 | if (!filter)
|
| 172 | goto free_out;
|
| 173 | damos_add_filter(scheme, filter);
|
| 174 | return ctx;
|
| 175 | free_out:
|
| 176 | damon_destroy_ctx(ctx);
|
| 177 | return NULL;
|
| 178 | }
|
| 179 |
|
| 180 | static int damon_sample_mtier_start(void)
|
| 181 | {
|
| 182 | struct damon_ctx *ctx;
|
| 183 | int err;
|
| 184 |
|
| 185 | ctx = damon_sample_mtier_build_ctx(true);
|
| 186 | if (!ctx)
|
| 187 | return -ENOMEM;
|
| 188 | ctxs[0] = ctx;
|
| 189 | ctx = damon_sample_mtier_build_ctx(false);
|
| 190 | if (!ctx) {
|
| 191 | damon_destroy_ctx(ctxs[0]);
|
| 192 | return -ENOMEM;
|
| 193 | }
|
| 194 | ctxs[1] = ctx;
|
| 195 | err = damon_start(ctxs, 2, true);
|
| 196 | if (!err)
|
| 197 | return 0;
|
| 198 |
|
| 199 | damon_destroy_ctx(ctxs[0]);
|
| 200 | damon_destroy_ctx(ctxs[1]);
|
| 201 | return err;
|
| 202 | }
|
| 203 |
|
| 204 | static void damon_sample_mtier_stop(void)
|
| 205 | {
|
| 206 | damon_stop(ctxs, 2);
|
| 207 | damon_destroy_ctx(ctxs[0]);
|
| 208 | damon_destroy_ctx(ctxs[1]);
|
| 209 | }
|
| 210 |
|
| 211 | static int damon_sample_mtier_enable_store(
|
| 212 | const char *val, const struct kernel_param *kp)
|
| 213 | {
|
| 214 | bool is_enabled = enabled;
|
| 215 | int err;
|
| 216 |
|
| 217 | err = kstrtobool(val, &enabled);
|
| 218 | if (err)
|
| 219 | return err;
|
| 220 |
|
| 221 | if (enabled == is_enabled)
|
| 222 | return 0;
|
| 223 |
|
| 224 | if (!damon_initialized())
|
| 225 | return 0;
|
| 226 |
|
| 227 | if (enabled) {
|
| 228 | err = damon_sample_mtier_start();
|
| 229 | if (err)
|
| 230 | enabled = false;
|
| 231 | return err;
|
| 232 | }
|
| 233 | damon_sample_mtier_stop();
|
| 234 | return 0;
|
| 235 | }
|
| 236 |
|
| 237 | static int __init damon_sample_mtier_init(void)
|
| 238 | {
|
| 239 | int err = 0;
|
| 240 |
|
| 241 | if (!damon_initialized()) {
|
| 242 | if (enabled)
|
| 243 | enabled = false;
|
| 244 | return -ENOMEM;
|
| 245 | }
|
| 246 |
|
| 247 | if (enabled) {
|
| 248 | err = damon_sample_mtier_start();
|
| 249 | if (err)
|
| 250 | enabled = false;
|
| 251 | }
|
| 252 | return 0;
|
| 253 | }
|
| 254 |
|
| 255 | module_init(damon_sample_mtier_init);
|