|
| 1 | +#include "parse.h" |
| 2 | +#include "builtin-cache.h" |
| 3 | + |
| 4 | +static int restrict_to_data_misses = 0; |
| 5 | + |
| 6 | +static char* source_name[] = { |
| 7 | + "invalid", |
| 8 | + "Local L3", |
| 9 | + "Other core L1/L2", |
| 10 | + "Local DRAM", |
| 11 | + "?", |
| 12 | + "?", |
| 13 | + "?", |
| 14 | + "APIC/PCI", |
| 15 | + |
| 16 | + "invalid", |
| 17 | + "?", |
| 18 | + "Remote L1/L2/L3", |
| 19 | + "Remote DRAM", |
| 20 | + "?", |
| 21 | + "?", |
| 22 | + "?", |
| 23 | + "APIC/PCI", |
| 24 | +}; |
| 25 | +static int global_data_source[16]; |
| 26 | +static int stack_data_source[16]; |
| 27 | +static int kernel_data_source[16]; |
| 28 | +static int softirq_data_source[16]; |
| 29 | +static int user_data_source[16]; |
| 30 | +static int num_local_per_source[16]; |
| 31 | +static int num_distant_per_source[16]; |
| 32 | +static int num_local_per_source_user[16]; |
| 33 | +static int num_distant_per_source_user[16]; |
| 34 | +static int num_local_per_source_kernel[16]; |
| 35 | +static int num_distant_per_source_kernel[16]; |
| 36 | +static int total_samples; |
| 37 | +static int data_miss; |
| 38 | +static int load, store; |
| 39 | +static int lin_valid, phy_valid; |
| 40 | +static int nb_non_null_phys_kern, nb_non_null_phys_user; |
| 41 | + |
| 42 | +void cache_init() { |
| 43 | + memset(global_data_source, 0, sizeof(global_data_source)); |
| 44 | + memset(stack_data_source, 0, sizeof(stack_data_source)); |
| 45 | + memset(kernel_data_source, 0, sizeof(kernel_data_source)); |
| 46 | + memset(softirq_data_source, 0, sizeof(softirq_data_source)); |
| 47 | + memset(user_data_source, 0, sizeof(user_data_source)); |
| 48 | + total_samples = data_miss = load = store = 0; |
| 49 | + lin_valid = phy_valid = 0; |
| 50 | + nb_non_null_phys_user = nb_non_null_phys_kern = 0; |
| 51 | +} |
| 52 | + |
| 53 | +void cache_parse(struct s* s) { |
| 54 | + uint64_t udata3 = (((uint64_t)s->ibs_op_data3_high)<<32) + (uint64_t)s->ibs_op_data3_low; |
| 55 | + ibs_op_data3_t *data3 = (void*)&udata3; |
| 56 | + if(!s->ibs_dc_phys) |
| 57 | + return; |
| 58 | + |
| 59 | + if(!restrict_to_data_misses || (restrict_to_data_misses && data3->ibsdcmiss)) { |
| 60 | + int status = (s->ibs_op_data2_low & 7) + 1*((s->ibs_op_data2_low & 16)>>1); |
| 61 | + /*if(status == 0) |
| 62 | + return;*/ |
| 63 | + |
| 64 | + total_samples++; |
| 65 | + |
| 66 | + global_data_source[status]++; |
| 67 | + if(is_kernel(s)) |
| 68 | + kernel_data_source[status]++; |
| 69 | + else |
| 70 | + user_data_source[status]++; |
| 71 | + if(is_softirq(s)) |
| 72 | + softirq_data_source[status]++; |
| 73 | + |
| 74 | + if(s->stack && s->ibs_dc_linear) { |
| 75 | + if(s->ibs_dc_linear >= s->usersp && s->ibs_dc_linear <= (uint64_t)s->stack) { |
| 76 | + stack_data_source[status]++; |
| 77 | + } |
| 78 | + } |
| 79 | + if(data3->ibsdcmiss) |
| 80 | + data_miss++; |
| 81 | + if(data3->ibsstop) |
| 82 | + store++; |
| 83 | + if(data3->ibsldop) |
| 84 | + load++; |
| 85 | + if(data3->ibsdclinaddrvalid) |
| 86 | + lin_valid++; |
| 87 | + if(data3->ibsdcphyaddrvalid) |
| 88 | + phy_valid++; |
| 89 | + if(s->ibs_dc_phys) { |
| 90 | + /*if(cpu_to_node(s->cpu) != get_addr_node(s) && status == 3) |
| 91 | + printf("#Bug: %d access locally to %d (%lx)\n", cpu_to_node(s->cpu), get_addr_node(s), s->ibs_dc_phys);*/ |
| 92 | + if(is_kernel(s)) { |
| 93 | + nb_non_null_phys_kern++; |
| 94 | + if(cpu_to_node(s->cpu) != get_addr_node(s)) |
| 95 | + num_distant_per_source_kernel[status]++; |
| 96 | + else |
| 97 | + num_local_per_source_kernel[status]++; |
| 98 | + } else { |
| 99 | + nb_non_null_phys_user++; |
| 100 | + if(cpu_to_node(s->cpu) != get_addr_node(s)) |
| 101 | + num_distant_per_source_user[status]++; |
| 102 | + else |
| 103 | + num_local_per_source_user[status]++; |
| 104 | + } |
| 105 | + } |
| 106 | + } |
| 107 | +} |
| 108 | + |
| 109 | +void cache_show() { |
| 110 | + int i, n, sum_glob = 0, sum_stack = 0, sum_user = 0, sum_kernel = 0, sum_irq, sum_local = 0, sum_distant = 0, sum_local_user = 0, sum_distant_user = 0, sum_local_kernel = 0, sum_distant_kernel = 0; |
| 111 | + |
| 112 | + printf("\t\t"); |
| 113 | + for(i = 0; i < 12; i++) { |
| 114 | + if(i==4 || i==5 || i==6 || i==7 || i==8 || i==9) continue; |
| 115 | + printf("%d (%s)%n", i, source_name[i], &n); |
| 116 | + printf("%*.*s", 24-n, 24-n, ""); |
| 117 | + sum_glob += global_data_source[i]; |
| 118 | + sum_stack += stack_data_source[i]; |
| 119 | + sum_kernel += kernel_data_source[i]; |
| 120 | + sum_user += user_data_source[i]; |
| 121 | + sum_irq += softirq_data_source[i]; |
| 122 | + |
| 123 | + num_local_per_source[i] = num_local_per_source_user[i] + num_local_per_source_kernel[i]; |
| 124 | + sum_local += num_local_per_source[i]; |
| 125 | + sum_local_user += num_local_per_source_user[i]; |
| 126 | + sum_local_kernel += num_local_per_source_kernel[i]; |
| 127 | + |
| 128 | + num_distant_per_source[i] = num_distant_per_source_user[i] + num_distant_per_source_kernel[i]; |
| 129 | + sum_distant += num_distant_per_source[i]; |
| 130 | + sum_distant_user += num_distant_per_source_user[i]; |
| 131 | + sum_distant_kernel += num_distant_per_source_kernel[i]; |
| 132 | + } |
| 133 | + printf("\n"); |
| 134 | + if(!sum_glob) sum_glob = 1; |
| 135 | + if(!sum_stack) sum_stack = 1; |
| 136 | + if(!sum_kernel) sum_kernel = 1; |
| 137 | + if(!sum_user) sum_user = 1; |
| 138 | + if(!sum_irq) sum_irq = 1; |
| 139 | + if(!sum_distant) sum_distant = 1; |
| 140 | + if(!sum_distant_user) sum_distant_user = 1; |
| 141 | + if(!sum_distant_kernel) sum_distant_kernel = 1; |
| 142 | + |
| 143 | + printf("Total:\t\t"); |
| 144 | + for(i = 0; i < 12; i++) { |
| 145 | + if(i==4 || i==5 || i==6 || i==7 || i==8 || i==9) continue; |
| 146 | + if(global_data_source[i] > 10000000) |
| 147 | + printf("%dK\t(%5.2f%%)\t", global_data_source[i]/1000, 100.*((float)global_data_source[i])/((float)sum_glob)); |
| 148 | + else |
| 149 | + printf("%d\t(%5.2f%%)\t", global_data_source[i], 100.*((float)global_data_source[i])/((float)sum_glob)); |
| 150 | + } |
| 151 | + printf("\n"); |
| 152 | + |
| 153 | + printf("Stack:\t\t"); |
| 154 | + for(i = 0; i < 12; i++) { |
| 155 | + if(i==4 || i==5 || i==6 || i==7 || i==8 || i==9) continue; |
| 156 | + printf("%d\t(%5.2f%%)\t", stack_data_source[i], 100.*((float)stack_data_source[i])/((float)sum_stack)); |
| 157 | + } |
| 158 | + printf("\n"); |
| 159 | + |
| 160 | + printf("User:\t\t"); |
| 161 | + for(i = 0; i < 12; i++) { |
| 162 | + if(i==4 || i==5 || i==6 || i==7 || i==8 || i==9) continue; |
| 163 | + printf("%d\t(%5.2f%%)\t", user_data_source[i], 100.*((float)user_data_source[i])/((float)sum_user)); |
| 164 | + } |
| 165 | + printf("\n"); |
| 166 | + |
| 167 | + printf("Kernel:\t\t"); |
| 168 | + for(i = 0; i < 12; i++) { |
| 169 | + if(i==4 || i==5 || i==6 || i==7 || i==8 || i==9) continue; |
| 170 | + printf("%d\t(%5.2f%%)\t", kernel_data_source[i], 100.*((float)kernel_data_source[i])/((float)sum_kernel)); |
| 171 | + } |
| 172 | + printf("\n"); |
| 173 | + |
| 174 | + printf("SoftIRQ:\t"); |
| 175 | + for(i = 0; i < 12; i++) { |
| 176 | + if(i==4 || i==5 || i==6 || i==7 || i==8 || i==9) continue; |
| 177 | + printf("%d\t(%5.2f%%)\t", softirq_data_source[i], 100.*((float)softirq_data_source[i])/((float)sum_irq)); |
| 178 | + } |
| 179 | + printf("\n"); |
| 180 | + |
| 181 | + printf("Locality:\t"); |
| 182 | + for(i = 0; i < 12; i++) { |
| 183 | + if(i==4 || i==5 || i==6 || i==7 || i==8 || i==9) continue; |
| 184 | + printf("\t %5.2f%% \t", 100.*((float)num_local_per_source[i])/((float)(num_local_per_source[i]+num_distant_per_source[i]))); |
| 185 | + } |
| 186 | + printf("\n"); |
| 187 | + |
| 188 | + printf("Loc (user):\t"); |
| 189 | + for(i = 0; i < 12; i++) { |
| 190 | + if(i==4 || i==5 || i==6 || i==7 || i==8 || i==9) continue; |
| 191 | + printf("\t %5.2f%% \t", 100.*((float)num_local_per_source_user[i])/((float)(num_local_per_source_user[i]+num_distant_per_source_user[i]))); |
| 192 | + } |
| 193 | + printf("\n"); |
| 194 | + |
| 195 | + printf("Loc (kern):\t"); |
| 196 | + for(i = 0; i < 12; i++) { |
| 197 | + if(i==4 || i==5 || i==6 || i==7 || i==8 || i==9) continue; |
| 198 | + printf("\t %5.2f%% \t", 100.*((float)num_local_per_source_kernel[i])/((float)(num_local_per_source_kernel[i]+num_distant_per_source_kernel[i]))); |
| 199 | + } |
| 200 | + printf("\n"); |
| 201 | + |
| 202 | + printf("\n"); |
| 203 | + |
| 204 | + printf("%d data miss (%.2f%%)\n", data_miss, 100.*((float)data_miss)/((float)total_samples)); |
| 205 | + printf("%d load\n", load); |
| 206 | + printf("%d store (%.2f%%)\n", store, 100.*((float)store)/((float)(load+store))); |
| 207 | + printf("%d lin_valid\n", lin_valid); |
| 208 | + printf("%d phy_valid\n", phy_valid); |
| 209 | + |
| 210 | + printf("Non null kernel: %d\n", nb_non_null_phys_kern); |
| 211 | + printf("Non null user: %d\n", nb_non_null_phys_user); |
| 212 | + printf("%.2f%% of local accesses\n", 100.*((float)sum_local)/((float)sum_local+sum_distant)); |
| 213 | + printf("%.2f%% of local accesses (user)\n", 100.*((float)sum_local_user)/((float)sum_local_user+sum_distant_user)); |
| 214 | + printf("%.2f%% of local accesses (kern)\n", 100.*((float)sum_local_kernel)/((float)sum_local_kernel+sum_distant_kernel)); |
| 215 | + |
| 216 | + for(i = 4; i < 8; i++) { |
| 217 | + if(global_data_source[i]) { |
| 218 | + fprintf(stdout, "#Warning: data in global_data_source[%d] (%s)\n", i, source_name[i]); |
| 219 | + } |
| 220 | + } |
| 221 | + for(i = 8; i < 16; i++) { |
| 222 | + if(i==10 || i==11) continue; |
| 223 | + if(global_data_source[i]) { |
| 224 | + fprintf(stdout, "#Warning: data in global_data_source[%d] (%s)\n", i, source_name[i]); |
| 225 | + } |
| 226 | + } |
| 227 | +} |
| 228 | + |
0 commit comments