diff options
Diffstat (limited to 'nix/libutil')
| -rw-r--r-- | nix/libutil/seccomp.cc | 162 | ||||
| -rw-r--r-- | nix/libutil/seccomp.hh | 222 | ||||
| -rw-r--r-- | nix/libutil/spawn.cc | 36 | ||||
| -rw-r--r-- | nix/libutil/spawn.hh | 10 |
4 files changed, 430 insertions, 0 deletions
diff --git a/nix/libutil/seccomp.cc b/nix/libutil/seccomp.cc new file mode 100644 index 00000000000..585442d70b5 --- /dev/null +++ b/nix/libutil/seccomp.cc | |||
| @@ -0,0 +1,162 @@ | |||
| 1 | #if __linux__ | ||
| 2 | #include <util.hh> | ||
| 3 | #include <seccomp.hh> | ||
| 4 | #include <algorithm> | ||
| 5 | |||
| 6 | namespace nix { | ||
| 7 | |||
| 8 | struct FilterInstruction { | ||
| 9 | struct sock_filter instruction; | ||
| 10 | bool fallthroughJt = false; | ||
| 11 | bool fallthroughJf = false; | ||
| 12 | bool fallthroughK = false; | ||
| 13 | }; | ||
| 14 | |||
| 15 | /* Note: instructions in "out" should have already verified that sysno is | ||
| 16 | * >= ranges[lowIndex].low. The value to compare against should already be | ||
| 17 | * in the accumulator. */ | ||
| 18 | static void | ||
| 19 | rangeActionsToFilter(std::vector<Uint32RangeAction> & ranges, | ||
| 20 | size_t lowIndex, /* Inclusive */ | ||
| 21 | size_t end, /* Exclusive */ | ||
| 22 | std::vector<FilterInstruction> & out) | ||
| 23 | { | ||
| 24 | if(lowIndex >= end) return; | ||
| 25 | |||
| 26 | if(end == lowIndex + 1) { | ||
| 27 | FilterInstruction branch; | ||
| 28 | Uint32RangeAction range = ranges.at(lowIndex); | ||
| 29 | branch.instruction = BPF_JUMP(BPF_JMP | BPF_JGT | BPF_K, | ||
| 30 | range.high, | ||
| 31 | /* To be fixed up */ | ||
| 32 | 0, | ||
| 33 | 0); | ||
| 34 | branch.fallthroughJt = true; | ||
| 35 | out.push_back(branch); | ||
| 36 | for(auto & i : range.instructions) { | ||
| 37 | FilterInstruction f; | ||
| 38 | f.instruction = i; | ||
| 39 | out.push_back(f); | ||
| 40 | } | ||
| 41 | FilterInstruction fallthroughBranch; | ||
| 42 | fallthroughBranch.instruction = BPF_JUMP(BPF_JMP | BPF_JA | BPF_K, | ||
| 43 | /* To be fixed up */ | ||
| 44 | 0, | ||
| 45 | 0, | ||
| 46 | 0); | ||
| 47 | fallthroughBranch.fallthroughK = true; | ||
| 48 | out.push_back(fallthroughBranch); | ||
| 49 | return; | ||
| 50 | } | ||
| 51 | |||
| 52 | size_t middle = lowIndex + ((end - lowIndex) / 2); | ||
| 53 | Uint32RangeAction range = ranges.at(middle); | ||
| 54 | FilterInstruction branch; | ||
| 55 | size_t branchIndex = out.size(); | ||
| 56 | branch.instruction = BPF_JUMP(BPF_JMP | BPF_JGE | BPF_K, | ||
| 57 | range.low, | ||
| 58 | 0, | ||
| 59 | /* To be fixed up a little farther down */ | ||
| 60 | 0); | ||
| 61 | out.push_back(branch); | ||
| 62 | rangeActionsToFilter(ranges, middle, end, out); | ||
| 63 | size_t elseIndex = out.size(); | ||
| 64 | out[branchIndex].instruction.jf = (elseIndex - branchIndex - 1); | ||
| 65 | rangeActionsToFilter(ranges, lowIndex, middle, out); | ||
| 66 | } | ||
| 67 | |||
| 68 | |||
| 69 | static bool compareRanges(Uint32RangeAction a, Uint32RangeAction b) | ||
| 70 | { | ||
| 71 | return (a.low < b.low); | ||
| 72 | } | ||
| 73 | |||
| 74 | |||
| 75 | /* Produce a loop-unrolled binary search of RANGES for the u32 currently in | ||
| 76 | * the accumulator. If the binary search finds a range that contains it, it | ||
| 77 | * will execute the corresponding instructions. If these instructions fall | ||
| 78 | * through, or if no containing range is found, control resumes after the last | ||
| 79 | * instruction in the returned sequence. */ | ||
| 80 | std::vector<struct sock_filter> | ||
| 81 | rangeActionsToFilter(std::vector<Uint32RangeAction> & ranges) | ||
| 82 | { | ||
| 83 | if(ranges.size() == 0) return {}; | ||
| 84 | std::sort(ranges.begin(), ranges.end(), compareRanges); | ||
| 85 | if(ranges.size() > 1) { | ||
| 86 | for(auto & i : ranges) | ||
| 87 | if(i.low > i.high) | ||
| 88 | throw Error("Invalid range in rangeActionsToFilter"); | ||
| 89 | for(size_t j = 1; j < ranges.size(); j++) | ||
| 90 | if(ranges[j].low <= ranges[j - 1].high) | ||
| 91 | throw Error("Overlapping ranges in rangeActionsToFilter"); | ||
| 92 | } | ||
| 93 | std::vector<FilterInstruction> out; | ||
| 94 | Uint32RangeAction first = ranges.at(0); | ||
| 95 | FilterInstruction branch; | ||
| 96 | /* Verify accumulator value is >= first.low, to satisfy initial invariant */ | ||
| 97 | branch.instruction = BPF_JUMP(BPF_JMP | BPF_JGE | BPF_K, | ||
| 98 | first.low, | ||
| 99 | 0, | ||
| 100 | /* to be fixed up */ | ||
| 101 | 0); | ||
| 102 | branch.fallthroughJf = true; | ||
| 103 | out.push_back(branch); | ||
| 104 | rangeActionsToFilter(ranges, 0, ranges.size(), out); | ||
| 105 | size_t fallthrough = out.size(); | ||
| 106 | std::vector<struct sock_filter> out2; | ||
| 107 | for(size_t j = 0; j < out.size(); j++) { | ||
| 108 | if(out[j].fallthroughJt) out[j].instruction.jt = (fallthrough - j - 1); | ||
| 109 | if(out[j].fallthroughJf) out[j].instruction.jf = (fallthrough - j - 1); | ||
| 110 | if(out[j].fallthroughK) out[j].instruction.k = (fallthrough - j - 1); | ||
| 111 | out2.push_back(out[j].instruction); | ||
| 112 | } | ||
| 113 | return out2; | ||
| 114 | } | ||
| 115 | |||
| 116 | |||
| 117 | /* If the uint64 at offset OFFSET has value VALUE, run INSTRUCTIONS. | ||
| 118 | * Otherwise, or if INSTRUCTIONS falls through, continue past the last | ||
| 119 | * instruction of OUT at the time seccompMatchu64 returns. Clobbers | ||
| 120 | * accumulator! */ | ||
| 121 | std::vector<struct sock_filter> seccompMatchu64(std::vector<struct sock_filter> & out, | ||
| 122 | uint64_t value, | ||
| 123 | std::vector<struct sock_filter> instructions, | ||
| 124 | uint32_t offset) | ||
| 125 | { | ||
| 126 | /* Note: this only works where the order of bytes in uint64 is big or | ||
| 127 | * little endian, and the same order holds for uint32. */ | ||
| 128 | /* Load lower-addressed 32 bits */ | ||
| 129 | out.push_back(BPF_STMT(BPF_LD | BPF_W | BPF_ABS, offset)); | ||
| 130 | size_t jmp1Index = out.size(); | ||
| 131 | |||
| 132 | out.push_back(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, | ||
| 133 | #ifdef WORDS_BIGENDIAN | ||
| 134 | (uint32_t)((value >> 32) & 0xffffffff), | ||
| 135 | #else | ||
| 136 | (uint32_t)(value & 0xffffffff), | ||
| 137 | #endif | ||
| 138 | 0, | ||
| 139 | /* To be fixed up */ | ||
| 140 | 0)); | ||
| 141 | /* Load higher-addressed 32 bits */ | ||
| 142 | out.push_back(BPF_STMT(BPF_LD | BPF_W | BPF_ABS, offset + (uint32_t)sizeof(uint32_t))); | ||
| 143 | size_t jmp2Index = out.size(); | ||
| 144 | out.push_back(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, | ||
| 145 | #ifdef WORDS_BIGENDIAN | ||
| 146 | (uint32_t)(value & 0xffffffff), | ||
| 147 | #else | ||
| 148 | (uint32_t)((value >> 32) & 0xffffffff), | ||
| 149 | #endif | ||
| 150 | 0, | ||
| 151 | /* To be fixed up */ | ||
| 152 | 0)); | ||
| 153 | |||
| 154 | out.insert(out.end(), instructions.begin(), instructions.end()); | ||
| 155 | out[jmp1Index].jf = (out.size() - jmp1Index - 1); | ||
| 156 | out[jmp2Index].jf = (out.size() - jmp2Index - 1); | ||
| 157 | return out; | ||
| 158 | } | ||
| 159 | |||
| 160 | } | ||
| 161 | |||
| 162 | #endif | ||
diff --git a/nix/libutil/seccomp.hh b/nix/libutil/seccomp.hh new file mode 100644 index 00000000000..634dfad5f8e --- /dev/null +++ b/nix/libutil/seccomp.hh | |||
| @@ -0,0 +1,222 @@ | |||
| 1 | #pragma once | ||
| 2 | |||
| 3 | #include "util.hh" | ||
| 4 | #include <linux/audit.h> /* For AUDIT_ARCH_* */ | ||
| 5 | #include <linux/seccomp.h> | ||
| 6 | #include <linux/filter.h> | ||
| 7 | |||
| 8 | |||
| 9 | /* This file provides two preprocessor macros (among other things): | ||
| 10 | 1. AUDIT_ARCH_NATIVE, which evaluates to whichever of the AUDIT_ARCH_* | ||
| 11 | values best represents the target system. Linux's internal headers have | ||
| 12 | a SECCOMP_ARCH_NATIVE since 2020, but it's not user-visible. Detection | ||
| 13 | of this is based on src/arch.c in libseccomp. | ||
| 14 | 2. NATIVE_SYSCALL_RANGES, an array initializer for an array of two-element | ||
| 15 | objects, the first of which is an integral number representing the | ||
| 16 | start (inclusive) of a range of valid syscall numbers, and the second | ||
| 17 | of which is an integral number representing the end (inclusive) of that | ||
| 18 | range of valid syscall numbers. The ranges provided are all | ||
| 19 | non-overlapping and strictly ascending (that is, the start of a range is | ||
| 20 | strictly higher than any of the numbers in any of the ranges that | ||
| 21 | precede it). All numbers involved fit into a long. | ||
| 22 | |||
| 23 | These ranges were generated from the various syscall.tbl, | ||
| 24 | syscall_32.tbl, and syscall_64.tbl files lying around in the linux | ||
| 25 | kernel source. Some were derived from | ||
| 26 | include/uapi/asm-generic/unistd.h. The kernel source used was commit | ||
| 27 | b3ee1e460951 of https://github.com/torvalds/linux.git, read on | ||
| 28 | 2025-04-23. Not all of the gaps in the files have any comments pointing | ||
| 29 | them out, so I recommend using build-aux/extract-syscall-ranges.sh for | ||
| 30 | the *.tbl files. | ||
| 31 | |||
| 32 | The intent behind saving these ranges is to be able to use a | ||
| 33 | default-allow seccomp policy that nevertheless disallows future | ||
| 34 | syscalls. This ensures that our security analysis can work with a | ||
| 35 | static, well-defined set of system calls that won't grow in the future | ||
| 36 | unless someone explicitly revisits the system call tables to consider | ||
| 37 | the implications of the new additions. */ | ||
| 38 | |||
| 39 | /* Both ends are inclusive. Some of the .tbl files use strange entries for | ||
| 40 | * the "abi" field, check arch/$ARCH/kernel/Makefile.syscalls to see what it | ||
| 41 | * specifies for syscall_abis_32 and syscall_abis_64 in addition to 32 or 64 | ||
| 42 | * and "common" (added in Makefile.asm-headers). Also check what, if | ||
| 43 | * anything, the makefile uses as the --offset flag to syscallhdr.sh. And | ||
| 44 | * look at arch/$ARCH/include/uapi/asm/unistd.h to see what value the offset | ||
| 45 | * takes in what configurations. */ | ||
| 46 | |||
| 47 | #ifndef AUDIT_ARCH_NATIVE | ||
| 48 | |||
| 49 | #if __i386__ | ||
| 50 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_I386 | ||
| 51 | #define NATIVE_SYSCALL_RANGES { {0, 221}, {224, 250}, {252, 284}, {286, 386}, \ | ||
| 52 | {393, 414}, {416, 466} } | ||
| 53 | #elif __x86_64__ | ||
| 54 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_X86_64 | ||
| 55 | #ifdef __ILP32__ | ||
| 56 | #include <asm/unistd.h> | ||
| 57 | #define X32RANGE(low, high) { (low | __X32_SYSCALL_BIT), (high | __X32_SYSCALL_BIT) } | ||
| 58 | #define NATIVE_SYSCALL_RANGES \ | ||
| 59 | { X32RANGE(0, 12), X32RANGE(14, 14), X32RANGE(17, 18), X32RANGE(21, 44), X32RANGE(48, 53), \ | ||
| 60 | X32RANGE(56, 58), X32RANGE(60, 100), X32RANGE(102, 126), X32RANGE(130, 130), \ | ||
| 61 | X32RANGE(132, 133), X32RANGE(135, 155), X32RANGE(157, 173), X32RANGE(175, 176), \ | ||
| 62 | X32RANGE(179, 179), X32RANGE(181, 204), X32RANGE(207, 208), X32RANGE(210, 210), \ | ||
| 63 | X32RANGE(212, 213), X32RANGE(216, 221), X32RANGE(223, 235), X32RANGE(237, 243), \ | ||
| 64 | X32RANGE(245, 245), X32RANGE(248, 272), X32RANGE(275, 277), X32RANGE(280, 294), \ | ||
| 65 | X32RANGE(298, 298), X32RANGE(300, 306), X32RANGE(308, 309), X32RANGE(312, 321), \ | ||
| 66 | X32RANGE(323, 326), X32RANGE(329, 335), X32RANGE(424, 466), X32RANGE(512, 547) } | ||
| 67 | #else | ||
| 68 | #define NATIVE_SYSCALL_RANGES { {0, 335}, {424, 466} } | ||
| 69 | #endif | ||
| 70 | #elif __arm__ | ||
| 71 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_ARM | ||
| 72 | /* Note: there are at present 6 extra ARM syscall numbers not listed in | ||
| 73 | arch/arm/tools/syscall.tbl, namely __ARM_NR_breakpoint through | ||
| 74 | __ARM_NR_get_tls. */ | ||
| 75 | #ifdef __ARM_EABI__ | ||
| 76 | #include <asm/unistd.h> | ||
| 77 | #define NATIVE_SYSCALL_RANGES \ | ||
| 78 | { {0, 6}, {8, 12}, {14, 16}, {19, 21}, {23, 24}, {26, 26}, {29, 29}, \ | ||
| 79 | {33, 34}, {36, 43}, {45, 47}, {49, 52}, {54, 55}, {57, 57}, {60, 67}, \ | ||
| 80 | {70, 75}, {77, 81}, {83, 83}, {85, 88}, {91, 97}, {99, 100}, {103, 108}, \ | ||
| 81 | {111, 111}, {114, 116}, {118, 122}, {124, 126}, {128, 129}, {131, 136}, \ | ||
| 82 | {138, 165}, {168, 187}, {190, 221}, {224, 253}, {256, 401}, {403, 414}, \ | ||
| 83 | {416, 446}, {448, 466}, {(__ARM_NR_BASE + 1), (__ARM_NR_BASE + 6)} } | ||
| 84 | #else | ||
| 85 | #include <asm/unistd.h> | ||
| 86 | #define OABIRANGE(low, high) { (low | __NR_OABI_SYSCALL_BASE), (high | __NR_OABI_SYSCALL_BASE) } | ||
| 87 | #define NATIVE_SYSCALL_RANGES \ | ||
| 88 | { OABIRANGE(0, 6), OABIRANGE(8, 16), OABIRANGE(19, 27), OABIRANGE(29, 30), \ | ||
| 89 | OABIRANGE(33, 34), OABIRANGE(36, 43), OABIRANGE(45, 47), OABIRANGE(49, 52), \ | ||
| 90 | OABIRANGE(54, 55), OABIRANGE(57, 57), OABIRANGE(60, 67), OABIRANGE(70, 83), \ | ||
| 91 | OABIRANGE(85, 97), OABIRANGE(99, 100), OABIRANGE(102, 108), \ | ||
| 92 | OABIRANGE(111, 111), OABIRANGE(113, 122), OABIRANGE(124, 126), \ | ||
| 93 | OABIRANGE(128, 129), OABIRANGE(131, 136), OABIRANGE(138, 165), \ | ||
| 94 | OABIRANGE(168, 187), OABIRANGE(190, 221), OABIRANGE(224, 253), \ | ||
| 95 | OABIRANGE(256, 401), OABIRANGE(403, 414), OABIRANGE(416, 446), \ | ||
| 96 | OABIRANGE(448, 466), {(__ARM_NR_BASE + 1), (__ARM_NR_BASE + 6)} } | ||
| 97 | #endif | ||
| 98 | #elif __aarch64__ | ||
| 99 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_AARCH64 | ||
| 100 | /* extract-syscall-ranges.sh $LINUXSOURCE/arch/arm64/tools/syscall_64.tbl \ | ||
| 101 | '64|renameat|rlimit|memfd_secret' | ||
| 102 | |||
| 103 | the extra ABIs are taken from arch/arm64/kernel/Makefile.syscalls and | ||
| 104 | scripts/Makefile.asm-headers */ | ||
| 105 | #define NATIVE_SYSCALL_RANGES { {0, 243}, {260, 294}, {424, 466} } | ||
| 106 | /* To my knowledge there is no x32 equivalent for aarch64 in mainline linux */ | ||
| 107 | #elif __mips__ && _MIPS_SIM == _MIPS_SIM_ABI32 | ||
| 108 | /* o32 abi in both endianness cases */ | ||
| 109 | #include <asm/unistd.h> | ||
| 110 | #define SYSRANGE(low, high) {(low) + __NR_Linux, (high) + __NR_Linux} | ||
| 111 | #define NATIVE_SYSCALL_RANGES \ | ||
| 112 | { SYSRANGE(0, 278), SYSRANGE(280, 368), SYSRANGE(393, 414), \ | ||
| 113 | SYSRANGE(416, 446), SYSRANGE(448, 466) } | ||
| 114 | #if __MIPSEB__ | ||
| 115 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_MIPS; | ||
| 116 | #elif __MIPSEL__ | ||
| 117 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_MIPSEL | ||
| 118 | #endif | ||
| 119 | #elif __mips__ && _MIPS_SIM == _MIPS_SIM_ABI64 | ||
| 120 | /* n64 abi in both endianness cases */ | ||
| 121 | #include <asm/unistd.h> | ||
| 122 | #define SYSRANGE(low, high) {(low) + __NR_Linux, (high) + __NR_Linux} | ||
| 123 | #define NATIVE_SYSCALL_RANGES \ | ||
| 124 | { SYSRANGE(0, 237), SYSRANGE(239, 328), SYSRANGE(424, 446), SYSRANGE(448, 466) } | ||
| 125 | #if __MIPSEB__ | ||
| 126 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_MIPS64 | ||
| 127 | #elif __MIPSEL__ | ||
| 128 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_MIPSEL64 | ||
| 129 | #endif /* _MIPS_SIM_ABI64 */ | ||
| 130 | #elif __mips__ && _MIPS_SIM == _MIPS_SIM_NABI32 | ||
| 131 | /* n32 abi in both endianness cases */ | ||
| 132 | #include <asm/unistd.h> | ||
| 133 | #define SYSRANGE(low, high) {(low) + __NR_Linux, (high) + __NR_Linux} | ||
| 134 | #define NATIVE_SYSCALL_RANGES \ | ||
| 135 | { SYSRANGE(0, 241), SYSRANGE(243, 332), SYSRANGE(403, 414), \ | ||
| 136 | SYSRANGE(416, 446), SYSRANGE(448, 466) } | ||
| 137 | #if __MIPSEB__ | ||
| 138 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_MIPS64N32 | ||
| 139 | #elif __MIPSEL__ | ||
| 140 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_MIPSEL64N32 | ||
| 141 | #endif /* _MIPS_SIM_NABI32 */ | ||
| 142 | #elif __hppa64__ /* hppa64 must be checked before hppa */ | ||
| 143 | #define NATIVE_SYSCALL_RANGES \ | ||
| 144 | { {0, 101}, {103, 126}, {128, 129}, {131, 136}, {138, 166}, \ | ||
| 145 | {168, 168}, {170, 195}, {198, 202}, {206, 212}, {215, 219}, \ | ||
| 146 | {222, 262}, {264, 302}, {304, 356}, {424, 446}, {448, 466} } | ||
| 147 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_PARISC64 | ||
| 148 | #elif __hppa__ | ||
| 149 | #define NATIVE_SYSCALL_RANGES \ | ||
| 150 | { {0, 101}, {103, 126}, {128, 129}, {131, 136}, {138, 166}, \ | ||
| 151 | {168, 168}, {170, 195}, {198, 202}, {206, 212}, {215, 219}, \ | ||
| 152 | {222, 262}, {264, 302}, {304, 356}, {403, 414}, {416, 446}, {448, 466} } | ||
| 153 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_PARISC | ||
| 154 | #elif __PPC64__ | ||
| 155 | #define NATIVE_SYSCALL_RANGES \ | ||
| 156 | { {0, 191}, {198, 203}, {205, 223}, {225, 225}, {227, 253}, \ | ||
| 157 | {255, 256}, {258, 365}, {378, 388}, {392, 402}, {424, 446}, {448, 466} } | ||
| 158 | #ifdef __BIG_ENDIAN__ | ||
| 159 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_PPC64 | ||
| 160 | #else | ||
| 161 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_PPC64LE | ||
| 162 | #endif | ||
| 163 | #elif __PPC__ | ||
| 164 | #define NATIVE_SYSCALL_RANGES \ | ||
| 165 | { {0, 223}, {225, 256}, {258, 365}, {378, 388}, {393, 414}, \ | ||
| 166 | {416, 446}, {448, 466} } | ||
| 167 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_PPC | ||
| 168 | #elif __s390x__ /* s390x must be checked before s390 */ | ||
| 169 | #define NATIVE_SYSCALL_RANGES \ | ||
| 170 | { {1, 12}, {14, 15}, {19, 22}, {26, 27}, {29, 30}, {33, 34}, \ | ||
| 171 | {36, 43}, {45, 45}, {48, 48}, {51, 52}, {54, 55}, {57, 57}, \ | ||
| 172 | {60, 67}, {72, 75}, {77, 79}, {83, 83}, {85, 94}, {96, 97}, \ | ||
| 173 | {99, 100}, {102, 108}, {110, 112}, {114, 122}, {124, 137}, \ | ||
| 174 | {141, 163}, {167, 169}, {172, 181}, {183, 191}, {198, 220}, \ | ||
| 175 | {222, 222}, {224, 241}, {243, 262}, {265, 386}, {392, 402}, {424, 466} } | ||
| 176 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_S390X | ||
| 177 | #elif __s390__ | ||
| 178 | #define NATIVE_SYSCALL_RANGES \ | ||
| 179 | { {1, 16}, {19, 27}, {29, 30}, {33, 34}, {36, 43}, {45, 52}, \ | ||
| 180 | {54, 55}, {57, 57}, {60, 67}, {70, 81}, {83, 83}, {85, 97}, \ | ||
| 181 | {99, 108}, {110, 112}, {114, 122}, {124, 165}, {167, 241}, \ | ||
| 182 | {243, 262}, {264, 386}, {393, 414}, {416, 466} } | ||
| 183 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_S390 | ||
| 184 | #elif __riscv && __riscv_xlen == 64 | ||
| 185 | #define NATIVE_SYSCALL_RANGES { {0, 37}, {39, 243}, {258, 294}, {424, 466} } | ||
| 186 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_RISCV64 | ||
| 187 | #elif __riscv && __riscv_xlen == 32 | ||
| 188 | #define NATIVE_SYSCALL_RANGES \ | ||
| 189 | { {0, 3}, {5, 37}, {39, 71}, {74, 78}, {81, 85}, {89, 97}, {99, 100}, \ | ||
| 190 | {102, 107}, {109, 109}, {111, 111}, {116, 126}, {128, 136}, {138, 162}, \ | ||
| 191 | {165, 168}, {172, 181}, {184, 191}, {193, 242}, {258, 259}, {261, 265}, \ | ||
| 192 | {267, 291}, {293, 294}, {403, 414}, {416, 466} } | ||
| 193 | #define AUDIT_ARCH_NATIVE AUDIT_ARCH_RISCV32 | ||
| 194 | #else | ||
| 195 | #error cannot determine which AUDIT_ARCH_* value to use for AUDIT_ARCH_NATIVE | ||
| 196 | #endif | ||
| 197 | |||
| 198 | #else | ||
| 199 | #ifndef NATIVE_SYSCALL_RANGES | ||
| 200 | /* Fall back to default-allow if the user specified (with | ||
| 201 | -DAUDIT_ARCH_NATIVE=...) an arch but not NATIVE_SYSCALL_RANGES */ | ||
| 202 | #define NATIVE_SYSCALL_RANGES {} | ||
| 203 | #endif | ||
| 204 | #endif /* #ifndef AUDIT_ARCH_NATIVE */ | ||
| 205 | |||
| 206 | namespace nix { | ||
| 207 | |||
| 208 | struct Uint32RangeAction { | ||
| 209 | uint32_t low; /* inclusive */ | ||
| 210 | uint32_t high; /* inclusive */ | ||
| 211 | std::vector<struct sock_filter> instructions; | ||
| 212 | }; | ||
| 213 | |||
| 214 | std::vector<struct sock_filter> rangeActionsToFilter(std::vector<Uint32RangeAction> & ranges); | ||
| 215 | |||
| 216 | |||
| 217 | std::vector<struct sock_filter> | ||
| 218 | seccompMatchu64(std::vector<struct sock_filter> & out, | ||
| 219 | uint64_t value, | ||
| 220 | std::vector<struct sock_filter> instructions, | ||
| 221 | uint32_t offset); | ||
| 222 | } | ||
diff --git a/nix/libutil/spawn.cc b/nix/libutil/spawn.cc index 93bab9f59e4..414849b6f04 100644 --- a/nix/libutil/spawn.cc +++ b/nix/libutil/spawn.cc | |||
| @@ -51,6 +51,8 @@ | |||
| 51 | 51 | ||
| 52 | #ifdef __linux__ | 52 | #ifdef __linux__ |
| 53 | #include <sys/personality.h> | 53 | #include <sys/personality.h> |
| 54 | #include <linux/seccomp.h> | ||
| 55 | #include <linux/filter.h> | ||
| 54 | #endif | 56 | #endif |
| 55 | 57 | ||
| 56 | #if defined(SYS_pivot_root) | 58 | #if defined(SYS_pivot_root) |
| @@ -281,6 +283,36 @@ void setIDsAction(SpawnContext & ctx) | |||
| 281 | throw SysError("setuid failed"); | 283 | throw SysError("setuid failed"); |
| 282 | } | 284 | } |
| 283 | 285 | ||
| 286 | void setNoNewPrivsAction(SpawnContext & ctx) | ||
| 287 | { | ||
| 288 | if(ctx.setNoNewPrivs) | ||
| 289 | #if __linux__ && defined(PR_SET_NO_NEW_PRIVS) | ||
| 290 | if(prctl(PR_SET_NO_NEW_PRIVS, 0, 0, 0, 0) == -1) | ||
| 291 | throw SysError("setting PR_SET_NO_NEW_PRIVS"); | ||
| 292 | #else | ||
| 293 | throw Error("setting PR_SET_NO_NEW_PRIVS not supported on this system"); | ||
| 294 | #endif | ||
| 295 | } | ||
| 296 | |||
| 297 | void addSeccompFilterAction(SpawnContext & ctx) | ||
| 298 | { | ||
| 299 | if(ctx.addSeccompFilter) { | ||
| 300 | #if __linux__ && defined(PR_SET_SECCOMP) && defined(SECCOMP_MODE_FILTER) | ||
| 301 | /* We use no extra functionality from the seccomp system call, so | ||
| 302 | * just use prctl. */ | ||
| 303 | if(ctx.seccompFilter.size() > USHRT_MAX) | ||
| 304 | throw Error("seccomp filter too large"); | ||
| 305 | struct sock_fprog prog; | ||
| 306 | prog.len = (unsigned short) ctx.seccompFilter.size(); | ||
| 307 | prog.filter = ctx.seccompFilter.data(); | ||
| 308 | if(prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, &prog) == -1) | ||
| 309 | throw SysError("installing seccomp filter"); | ||
| 310 | #else | ||
| 311 | throw Error("setting seccomp filter not supported on this system"); | ||
| 312 | #endif | ||
| 313 | } | ||
| 314 | } | ||
| 315 | |||
| 284 | 316 | ||
| 285 | void restoreSIGPIPEAction(SpawnContext & ctx) | 317 | void restoreSIGPIPEAction(SpawnContext & ctx) |
| 286 | { | 318 | { |
| @@ -336,6 +368,8 @@ Phases getBasicSpawnPhases() | |||
| 336 | { "setPersonality", setPersonalityAction }, | 368 | { "setPersonality", setPersonalityAction }, |
| 337 | { "oomSacrifice", oomSacrificeAction }, | 369 | { "oomSacrifice", oomSacrificeAction }, |
| 338 | { "setIDs", setIDsAction }, | 370 | { "setIDs", setIDsAction }, |
| 371 | { "setNoNewPrivs", setNoNewPrivsAction }, | ||
| 372 | { "addSeccompFilter", addSeccompFilterAction }, | ||
| 339 | { "restoreSIGPIPE", restoreSIGPIPEAction }, | 373 | { "restoreSIGPIPE", restoreSIGPIPEAction }, |
| 340 | { "setupSuccess", setupSuccessAction }, | 374 | { "setupSuccess", setupSuccessAction }, |
| 341 | { "exec", execAction } }; | 375 | { "exec", execAction } }; |
| @@ -773,6 +807,8 @@ Phases getCloneSpawnPhases() | |||
| 773 | CloneSpawnContext.lockMountsMapAll = true. */ | 807 | CloneSpawnContext.lockMountsMapAll = true. */ |
| 774 | { "lockMounts", lockMountsAction }, | 808 | { "lockMounts", lockMountsAction }, |
| 775 | { "setIDs", setIDsAction }, | 809 | { "setIDs", setIDsAction }, |
| 810 | { "setNoNewPrivs", setNoNewPrivsAction }, | ||
| 811 | { "addSeccompFilter", addSeccompFilterAction }, | ||
| 776 | { "restoreSIGPIPE", restoreSIGPIPEAction }, | 812 | { "restoreSIGPIPE", restoreSIGPIPEAction }, |
| 777 | { "setupSuccess", setupSuccessAction }, | 813 | { "setupSuccess", setupSuccessAction }, |
| 778 | { "exec", execAction }}; | 814 | { "exec", execAction }}; |
diff --git a/nix/libutil/spawn.hh b/nix/libutil/spawn.hh index edc528312db..5e75bcfb097 100644 --- a/nix/libutil/spawn.hh +++ b/nix/libutil/spawn.hh | |||
| @@ -3,6 +3,9 @@ | |||
| 3 | #include <util.hh> | 3 | #include <util.hh> |
| 4 | #include <map> | 4 | #include <map> |
| 5 | #include <stddef.h> | 5 | #include <stddef.h> |
| 6 | #ifdef __linux__ | ||
| 7 | #include <linux/filter.h> | ||
| 8 | #endif | ||
| 6 | 9 | ||
| 7 | namespace nix { | 10 | namespace nix { |
| 8 | struct SpawnContext; /* Forward declaration */ | 11 | struct SpawnContext; /* Forward declaration */ |
| @@ -57,6 +60,11 @@ struct SpawnContext { | |||
| 57 | bool dropAmbientCapabilities = false; /* Whether to drop ambient | 60 | bool dropAmbientCapabilities = false; /* Whether to drop ambient |
| 58 | * capabilities if on a system that | 61 | * capabilities if on a system that |
| 59 | * supports them. */ | 62 | * supports them. */ |
| 63 | bool setNoNewPrivs = false; | ||
| 64 | bool addSeccompFilter = false; | ||
| 65 | #if __linux__ | ||
| 66 | std::vector<struct sock_filter> seccompFilter; | ||
| 67 | #endif | ||
| 60 | bool doChroot = false; | 68 | bool doChroot = false; |
| 61 | Path chrootRootDir; | 69 | Path chrootRootDir; |
| 62 | void * extraData; /* Extra user data */ | 70 | void * extraData; /* Extra user data */ |
| @@ -118,6 +126,8 @@ Action closeMostFDsAction; | |||
| 118 | Action setPersonalityAction; | 126 | Action setPersonalityAction; |
| 119 | Action oomSacrificeAction; | 127 | Action oomSacrificeAction; |
| 120 | Action setIDsAction; | 128 | Action setIDsAction; |
| 129 | Action setNoNewPrivsAction; | ||
| 130 | Action addSeccompFilterAction; | ||
| 121 | Action restoreSIGPIPEAction; | 131 | Action restoreSIGPIPEAction; |
| 122 | Action setupSuccessAction; | 132 | Action setupSuccessAction; |
| 123 | Action execAction; | 133 | Action execAction; |
