devicefilter.go

Documentation: github.com/containerd/cgroups/v2

     1  /*
     2     Copyright The containerd Authors.
     3  
     4     Licensed under the Apache License, Version 2.0 (the "License");
     5     you may not use this file except in compliance with the License.
     6     You may obtain a copy of the License at
     7  
     8         http://www.apache.org/licenses/LICENSE-2.0
     9  
    10     Unless required by applicable law or agreed to in writing, software
    11     distributed under the License is distributed on an "AS IS" BASIS,
    12     WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    13     See the License for the specific language governing permissions and
    14     limitations under the License.
    15  */
    16  
    17  // Devicefilter containes eBPF device filter program
    18  //
    19  // The implementation is based on https://github.com/containers/crun/blob/0.10.2/src/libcrun/ebpf.c
    20  //
    21  // Although ebpf.c is originally licensed under LGPL-3.0-or-later, the author (Giuseppe Scrivano)
    22  // agreed to relicense the file in Apache License 2.0: https://github.com/opencontainers/runc/issues/2144#issuecomment-543116397
    23  //
    24  // This particular Go implementation based on runc version
    25  // https://github.com/opencontainers/runc/blob/master/libcontainer/cgroups/ebpf/devicefilter/devicefilter.go
    26  
    27  package v2
    28  
    29  import (
    30  	"errors"
    31  	"fmt"
    32  	"math"
    33  
    34  	"github.com/cilium/ebpf/asm"
    35  	"github.com/opencontainers/runtime-spec/specs-go"
    36  	"golang.org/x/sys/unix"
    37  )
    38  
    39  const (
    40  	// license string format is same as kernel MODULE_LICENSE macro
    41  	license = "Apache"
    42  )
    43  
    44  // DeviceFilter returns eBPF device filter program and its license string
    45  func DeviceFilter(devices []specs.LinuxDeviceCgroup) (asm.Instructions, string, error) {
    46  	p := &program{}
    47  	p.init()
    48  	for i := len(devices) - 1; i >= 0; i-- {
    49  		if err := p.appendDevice(devices[i]); err != nil {
    50  			return nil, "", err
    51  		}
    52  	}
    53  	insts, err := p.finalize()
    54  	return insts, license, err
    55  }
    56  
    57  type program struct {
    58  	insts       asm.Instructions
    59  	hasWildCard bool
    60  	blockID     int
    61  }
    62  
    63  func (p *program) init() {
    64  	// struct bpf_cgroup_dev_ctx: https://elixir.bootlin.com/linux/v5.3.6/source/include/uapi/linux/bpf.h#L3423
    65  	/*
    66  		u32 access_type
    67  		u32 major
    68  		u32 minor
    69  	*/
    70  	// R2 <- type (lower 16 bit of u32 access_type at R1[0])
    71  	p.insts = append(p.insts,
    72  		asm.LoadMem(asm.R2, asm.R1, 0, asm.Half))
    73  
    74  	// R3 <- access (upper 16 bit of u32 access_type at R1[0])
    75  	p.insts = append(p.insts,
    76  		asm.LoadMem(asm.R3, asm.R1, 0, asm.Word),
    77  		// RSh: bitwise shift right
    78  		asm.RSh.Imm32(asm.R3, 16))
    79  
    80  	// R4 <- major (u32 major at R1[4])
    81  	p.insts = append(p.insts,
    82  		asm.LoadMem(asm.R4, asm.R1, 4, asm.Word))
    83  
    84  	// R5 <- minor (u32 minor at R1[8])
    85  	p.insts = append(p.insts,
    86  		asm.LoadMem(asm.R5, asm.R1, 8, asm.Word))
    87  }
    88  
    89  // appendDevice needs to be called from the last element of OCI linux.resources.devices to the head element.
    90  func (p *program) appendDevice(dev specs.LinuxDeviceCgroup) error {
    91  	if p.blockID < 0 {
    92  		return errors.New("the program is finalized")
    93  	}
    94  	if p.hasWildCard {
    95  		// All entries after wildcard entry are ignored
    96  		return nil
    97  	}
    98  
    99  	bpfType := int32(-1)
   100  	hasType := true
   101  	switch dev.Type {
   102  	case string('c'):
   103  		bpfType = int32(unix.BPF_DEVCG_DEV_CHAR)
   104  	case string('b'):
   105  		bpfType = int32(unix.BPF_DEVCG_DEV_BLOCK)
   106  	case string('a'):
   107  		hasType = false
   108  	default:
   109  		// if not specified in OCI json, typ is set to DeviceTypeAll
   110  		return fmt.Errorf("invalid DeviceType %q", dev.Type)
   111  	}
   112  	if *dev.Major > math.MaxUint32 {
   113  		return fmt.Errorf("invalid major %d", *dev.Major)
   114  	}
   115  	if *dev.Minor > math.MaxUint32 {
   116  		return fmt.Errorf("invalid minor %d", *dev.Major)
   117  	}
   118  	hasMajor := *dev.Major >= 0 // if not specified in OCI json, major is set to -1
   119  	hasMinor := *dev.Minor >= 0
   120  	bpfAccess := int32(0)
   121  	for _, r := range dev.Access {
   122  		switch r {
   123  		case 'r':
   124  			bpfAccess |= unix.BPF_DEVCG_ACC_READ
   125  		case 'w':
   126  			bpfAccess |= unix.BPF_DEVCG_ACC_WRITE
   127  		case 'm':
   128  			bpfAccess |= unix.BPF_DEVCG_ACC_MKNOD
   129  		default:
   130  			return fmt.Errorf("unknown device access %v", r)
   131  		}
   132  	}
   133  	// If the access is rwm, skip the check.
   134  	hasAccess := bpfAccess != (unix.BPF_DEVCG_ACC_READ | unix.BPF_DEVCG_ACC_WRITE | unix.BPF_DEVCG_ACC_MKNOD)
   135  
   136  	blockSym := fmt.Sprintf("block-%d", p.blockID)
   137  	nextBlockSym := fmt.Sprintf("block-%d", p.blockID+1)
   138  	prevBlockLastIdx := len(p.insts) - 1
   139  	if hasType {
   140  		p.insts = append(p.insts,
   141  			// if (R2 != bpfType) goto next
   142  			asm.JNE.Imm(asm.R2, bpfType, nextBlockSym),
   143  		)
   144  	}
   145  	if hasAccess {
   146  		p.insts = append(p.insts,
   147  			// if (R3 & bpfAccess == 0 /* use R1 as a temp var */) goto next
   148  			asm.Mov.Reg32(asm.R1, asm.R3),
   149  			asm.And.Imm32(asm.R1, bpfAccess),
   150  			asm.JEq.Imm(asm.R1, 0, nextBlockSym),
   151  		)
   152  	}
   153  	if hasMajor {
   154  		p.insts = append(p.insts,
   155  			// if (R4 != major) goto next
   156  			asm.JNE.Imm(asm.R4, int32(*dev.Major), nextBlockSym),
   157  		)
   158  	}
   159  	if hasMinor {
   160  		p.insts = append(p.insts,
   161  			// if (R5 != minor) goto next
   162  			asm.JNE.Imm(asm.R5, int32(*dev.Minor), nextBlockSym),
   163  		)
   164  	}
   165  	if !hasType && !hasAccess && !hasMajor && !hasMinor {
   166  		p.hasWildCard = true
   167  	}
   168  	p.insts = append(p.insts, acceptBlock(dev.Allow)...)
   169  	// set blockSym to the first instruction we added in this iteration
   170  	p.insts[prevBlockLastIdx+1] = p.insts[prevBlockLastIdx+1].Sym(blockSym)
   171  	p.blockID++
   172  	return nil
   173  }
   174  
   175  func (p *program) finalize() (asm.Instructions, error) {
   176  	if p.hasWildCard {
   177  		// acceptBlock with asm.Return() is already inserted
   178  		return p.insts, nil
   179  	}
   180  	blockSym := fmt.Sprintf("block-%d", p.blockID)
   181  	p.insts = append(p.insts,
   182  		// R0 <- 0
   183  		asm.Mov.Imm32(asm.R0, 0).Sym(blockSym),
   184  		asm.Return(),
   185  	)
   186  	p.blockID = -1
   187  	return p.insts, nil
   188  }
   189  
   190  func acceptBlock(accept bool) asm.Instructions {
   191  	v := int32(0)
   192  	if accept {
   193  		v = 1
   194  	}
   195  	return []asm.Instruction{
   196  		// R0 <- v
   197  		asm.Mov.Imm32(asm.R0, v),
   198  		asm.Return(),
   199  	}
   200  }
   201
View as plain text