| 1 | /* Initialize CPU feature data. AArch64 version. |
| 2 | This file is part of the GNU C Library. |
| 3 | Copyright (C) 2017-2024 Free Software Foundation, Inc. |
| 4 | |
| 5 | The GNU C Library is free software; you can redistribute it and/or |
| 6 | modify it under the terms of the GNU Lesser General Public |
| 7 | License as published by the Free Software Foundation; either |
| 8 | version 2.1 of the License, or (at your option) any later version. |
| 9 | |
| 10 | The GNU C Library is distributed in the hope that it will be useful, |
| 11 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 13 | Lesser General Public License for more details. |
| 14 | |
| 15 | You should have received a copy of the GNU Lesser General Public |
| 16 | License along with the GNU C Library; if not, see |
| 17 | <https://www.gnu.org/licenses/>. */ |
| 18 | |
| 19 | #include <array_length.h> |
| 20 | #include <cpu-features.h> |
| 21 | #include <sys/auxv.h> |
| 22 | #include <elf/dl-hwcaps.h> |
| 23 | #include <sys/prctl.h> |
| 24 | #include <sys/utsname.h> |
| 25 | #include <dl-tunables-parse.h> |
| 26 | |
| 27 | #define DCZID_DZP_MASK (1 << 4) |
| 28 | #define DCZID_BS_MASK (0xf) |
| 29 | |
| 30 | /* The maximal set of permitted tags that the MTE random tag generation |
| 31 | instruction may use. We exclude tag 0 because a) we want to reserve |
| 32 | that for the libc heap structures and b) because it makes it easier |
| 33 | to see when pointer have been correctly tagged. */ |
| 34 | #define MTE_ALLOWED_TAGS (0xfffe << PR_MTE_TAG_SHIFT) |
| 35 | |
| 36 | struct cpu_list |
| 37 | { |
| 38 | const char *name; |
| 39 | size_t len; |
| 40 | uint64_t midr; |
| 41 | }; |
| 42 | |
| 43 | static const struct cpu_list cpu_list[] = |
| 44 | { |
| 45 | #define CPU_LIST_ENTRY(__str, __num) { __str, sizeof (__str) - 1, __num } |
| 46 | CPU_LIST_ENTRY ("thunderxt88" , 0x430F0A10), |
| 47 | CPU_LIST_ENTRY ("thunderx2t99" , 0x431F0AF0), |
| 48 | CPU_LIST_ENTRY ("thunderx2t99p1" , 0x420F5160), |
| 49 | CPU_LIST_ENTRY ("ares" , 0x411FD0C0), |
| 50 | CPU_LIST_ENTRY ("emag" , 0x503F0001), |
| 51 | CPU_LIST_ENTRY ("kunpeng920" , 0x481FD010), |
| 52 | CPU_LIST_ENTRY ("a64fx" , 0x460F0010), |
| 53 | CPU_LIST_ENTRY ("generic" , 0x0), |
| 54 | }; |
| 55 | |
| 56 | static uint64_t |
| 57 | get_midr_from_mcpu (const struct tunable_str_t *mcpu) |
| 58 | { |
| 59 | for (int i = 0; i < array_length (cpu_list); i++) |
| 60 | if (tunable_strcmp (strval: mcpu, str: cpu_list[i].name, len: cpu_list[i].len)) |
| 61 | return cpu_list[i].midr; |
| 62 | |
| 63 | return UINT64_MAX; |
| 64 | } |
| 65 | |
| 66 | #if __LINUX_KERNEL_VERSION < 0x060200 |
| 67 | |
| 68 | /* Return true if we prefer using SVE in string ifuncs. Old kernels disable |
| 69 | SVE after every system call which results in unnecessary traps if memcpy |
| 70 | uses SVE. This is true for kernels between 4.15.0 and before 6.2.0, except |
| 71 | for 5.14.0 which was patched. For these versions return false to avoid using |
| 72 | SVE ifuncs. |
| 73 | Parse the kernel version into a 24-bit kernel.major.minor value without |
| 74 | calling any library functions. If uname() is not supported or if the version |
| 75 | format is not recognized, assume the kernel is modern and return true. */ |
| 76 | |
| 77 | static inline bool |
| 78 | prefer_sve_ifuncs (void) |
| 79 | { |
| 80 | struct utsname buf; |
| 81 | const char *p = &buf.release[0]; |
| 82 | int kernel = 0; |
| 83 | int val; |
| 84 | |
| 85 | if (__uname (&buf) < 0) |
| 86 | return true; |
| 87 | |
| 88 | for (int shift = 16; shift >= 0; shift -= 8) |
| 89 | { |
| 90 | for (val = 0; *p >= '0' && *p <= '9'; p++) |
| 91 | val = val * 10 + *p - '0'; |
| 92 | kernel |= (val & 255) << shift; |
| 93 | if (*p++ != '.') |
| 94 | break; |
| 95 | } |
| 96 | |
| 97 | if (kernel >= 0x060200 || kernel == 0x050e00) |
| 98 | return true; |
| 99 | if (kernel >= 0x040f00) |
| 100 | return false; |
| 101 | return true; |
| 102 | } |
| 103 | |
| 104 | #endif |
| 105 | |
| 106 | static inline void |
| 107 | init_cpu_features (struct cpu_features *cpu_features) |
| 108 | { |
| 109 | register uint64_t midr = UINT64_MAX; |
| 110 | |
| 111 | /* Get the tunable override. */ |
| 112 | const struct tunable_str_t *mcpu = TUNABLE_GET (glibc, cpu, name, |
| 113 | struct tunable_str_t *, |
| 114 | NULL); |
| 115 | if (mcpu != NULL) |
| 116 | midr = get_midr_from_mcpu (mcpu); |
| 117 | |
| 118 | /* If there was no useful tunable override, query the MIDR if the kernel |
| 119 | allows it. */ |
| 120 | if (midr == UINT64_MAX) |
| 121 | { |
| 122 | if (GLRO (dl_hwcap) & HWCAP_CPUID) |
| 123 | asm volatile ("mrs %0, midr_el1" : "=r" (midr)); |
| 124 | else |
| 125 | midr = 0; |
| 126 | } |
| 127 | |
| 128 | cpu_features->midr_el1 = midr; |
| 129 | |
| 130 | /* Check if ZVA is enabled. */ |
| 131 | unsigned dczid; |
| 132 | asm volatile ("mrs %0, dczid_el0" : "=r" (dczid)); |
| 133 | |
| 134 | if ((dczid & DCZID_DZP_MASK) == 0) |
| 135 | cpu_features->zva_size = 4 << (dczid & DCZID_BS_MASK); |
| 136 | |
| 137 | /* Check if BTI is supported. */ |
| 138 | cpu_features->bti = GLRO (dl_hwcap2) & HWCAP2_BTI; |
| 139 | |
| 140 | /* Setup memory tagging support if the HW and kernel support it, and if |
| 141 | the user has requested it. */ |
| 142 | cpu_features->mte_state = 0; |
| 143 | |
| 144 | #ifdef USE_MTAG |
| 145 | int mte_state = TUNABLE_GET (glibc, mem, tagging, unsigned, 0); |
| 146 | cpu_features->mte_state = (GLRO (dl_hwcap2) & HWCAP2_MTE) ? mte_state : 0; |
| 147 | /* If we lack the MTE feature, disable the tunable, since it will |
| 148 | otherwise cause instructions that won't run on this CPU to be used. */ |
| 149 | TUNABLE_SET (glibc, mem, tagging, cpu_features->mte_state); |
| 150 | |
| 151 | if (cpu_features->mte_state & 4) |
| 152 | /* Enable choosing system-preferred faulting mode. */ |
| 153 | __prctl (PR_SET_TAGGED_ADDR_CTRL, |
| 154 | (PR_TAGGED_ADDR_ENABLE | PR_MTE_TCF_SYNC | PR_MTE_TCF_ASYNC |
| 155 | | MTE_ALLOWED_TAGS), |
| 156 | 0, 0, 0); |
| 157 | else if (cpu_features->mte_state & 2) |
| 158 | __prctl (PR_SET_TAGGED_ADDR_CTRL, |
| 159 | (PR_TAGGED_ADDR_ENABLE | PR_MTE_TCF_SYNC | MTE_ALLOWED_TAGS), |
| 160 | 0, 0, 0); |
| 161 | else if (cpu_features->mte_state) |
| 162 | __prctl (PR_SET_TAGGED_ADDR_CTRL, |
| 163 | (PR_TAGGED_ADDR_ENABLE | PR_MTE_TCF_ASYNC | MTE_ALLOWED_TAGS), |
| 164 | 0, 0, 0); |
| 165 | #endif |
| 166 | |
| 167 | /* Check if SVE is supported. */ |
| 168 | cpu_features->sve = GLRO (dl_hwcap) & HWCAP_SVE; |
| 169 | |
| 170 | cpu_features->prefer_sve_ifuncs = cpu_features->sve; |
| 171 | |
| 172 | #if __LINUX_KERNEL_VERSION < 0x060200 |
| 173 | if (cpu_features->sve) |
| 174 | cpu_features->prefer_sve_ifuncs = prefer_sve_ifuncs (); |
| 175 | #endif |
| 176 | |
| 177 | /* Check if MOPS is supported. */ |
| 178 | cpu_features->mops = GLRO (dl_hwcap2) & HWCAP2_MOPS; |
| 179 | } |
| 180 | |