|
| 1 | +# Optimizing binaries with pac-ret hardening |
| 2 | + |
| 3 | +This is a design document about processing the `DW_CFA_AARCH64_negate_ra_state` |
| 4 | +DWARF instruction in BOLT. As it describes internal design decisions, the |
| 5 | +intended audience is BOLT developers. The document is an updated version of the |
| 6 | +[RFC posted on the LLVM Discourse](https://discourse.llvm.org/t/rfc-bolt-aarch64-handle-opnegaterastate-to-enable-optimizing-binaries-with-pac-ret-hardening/86594). |
| 7 | + |
| 8 | + |
| 9 | +`DW_CFA_AARCH64_negate_ra_state` is also referred to as `.cfi_negate_ra_state` |
| 10 | +in assembly, or `OpNegateRAState` in BOLT sources. In this document, I will use |
| 11 | +**negate-ra-state** as a shorthand. |
| 12 | + |
| 13 | +## Introduction |
| 14 | + |
| 15 | +### Pointer Authentication |
| 16 | + |
| 17 | +For more information, see the [pac-ret section of the BOLT-binary-analysis document](BinaryAnalysis.md#pac-ret-analysis). |
| 18 | + |
| 19 | +### DW_CFA_AARCH64_negate_ra_state |
| 20 | + |
| 21 | +The negate-ra-state CFI is a vendor-specific Call Frame Instruction defined in |
| 22 | +the [Arm ABI](https://github.com/ARM-software/abi-aa/blob/main/aadwarf64/aadwarf64.rst#id1). |
| 23 | + |
| 24 | +``` |
| 25 | +The DW_CFA_AARCH64_negate_ra_state operation negates bit[0] of the RA_SIGN_STATE pseudo-register. |
| 26 | +``` |
| 27 | + |
| 28 | +This bit indicates to the unwinder whether the current return address is signed |
| 29 | +or not (hence the name). The unwinder uses this information to authenticate the |
| 30 | +pointer, and remove the Pointer Authentication Code (PAC) bits. |
| 31 | +Incorrect placement of negate-ra-state CFIs causes the unwinder to either attempt |
| 32 | +to authenticate an unsigned pointer (resulting in a segmentation fault), or skip |
| 33 | +authentication on a signed pointer, which can also cause a fault. |
| 34 | + |
| 35 | +Note: some unwinders use the `xpac` instruction to strip the PAC bits without |
| 36 | +authenticating the pointer. This is an incorrect (incomplete) implementation, |
| 37 | +as it allows control-flow modification in the case of unwinding. |
| 38 | + |
| 39 | +There are no DWARF instructions to directly set or clear the RA State. However, |
| 40 | +two other CFIs can also affect the RA state: |
| 41 | +- `DW_CFA_remember_state`: this CFI stores register rules onto an implicit stack. |
| 42 | +- `DW_CFA_restore_state`: this CFI pops rules from this stack. |
| 43 | + |
| 44 | +Example: |
| 45 | + |
| 46 | +| CFI | Effect on RA state | |
| 47 | +| ------------------------------ | ------------------------------ | |
| 48 | +| (default) | 0 | |
| 49 | +| DW_CFA_AARCH64_negate_ra_state | 0 -> 1 | |
| 50 | +| DW_CFA_remember_state | 1 pushed to the stack | |
| 51 | +| DW_CFA_AARCH64_negate_ra_state | 1 -> 0 | |
| 52 | +| DW_CFA_restore_state | 0 -> 1 (popped from the stack) | |
| 53 | + |
| 54 | +The Arm ABI also defines the DW_CFA_AARCH64_negate_ra_state_with_pc CFI, but it |
| 55 | +is not widely used, and is [likely to become deprecated](https://github.com/ARM-software/abi-aa/issues/327). |
| 56 | + |
| 57 | +### Where are these CFIs needed? |
| 58 | + |
| 59 | +Whenever two consecutive instructions have different RA states, the unwinder must |
| 60 | +be informed of the change. This typically occurs during pointer signing or |
| 61 | +authentication. If adjacent instructions differ in RA state but neither signs |
| 62 | +nor authenticates the return address, they must belong to different control flow |
| 63 | +paths. One is part of an execution path with signed RA, the other is part of a |
| 64 | +path with an unsigned RA. |
| 65 | + |
| 66 | +In the example below, the first BasicBlock ends in a conditional branch, and |
| 67 | +jumps to two different BasicBlocks, each with their own authentication, and |
| 68 | +return. The instructions on the border of the second and third BasicBlock have |
| 69 | +different RA states. The `ret` at the end of the second BasicBlock is in unsigned |
| 70 | +state. The start of the third BasicBlock is after the `paciasp` in the control |
| 71 | +flow, but before the authentication. In this case, a negate-ra-state is needed |
| 72 | +at the end of the second BasicBlock. |
| 73 | + |
| 74 | +``` |
| 75 | + +----------------+ |
| 76 | + | paciasp | |
| 77 | + | | |
| 78 | + | b.cc | |
| 79 | + +--------+-------+ |
| 80 | + | |
| 81 | ++----------------+ |
| 82 | +| | |
| 83 | +| +--------v-------+ |
| 84 | +| | | |
| 85 | +| | autiasp | |
| 86 | +| | ret | // RA: unsigned |
| 87 | +| +----------------+ |
| 88 | ++----------------+ |
| 89 | + | |
| 90 | + +--------v-------+ // RA: signed |
| 91 | + | | |
| 92 | + | autiasp | |
| 93 | + | ret | |
| 94 | + +----------------+ |
| 95 | +``` |
| 96 | + |
| 97 | +> [!important] |
| 98 | +> The unwinder does not follow the control flow graph. It reads unwind |
| 99 | +> information in the layout order. |
| 100 | +
|
| 101 | +Because these locations are dependent on how the function layout looks, |
| 102 | +negate-ra-state CFIs will become invalid during BasicBlock reordering. |
| 103 | + |
| 104 | +## Solution design |
| 105 | + |
| 106 | +The implementation introduces two new passes: |
| 107 | +1. `MarkRAStatesPass`: assigns the RA state to each instruction based on the CFIs |
| 108 | + in the input binary |
| 109 | +2. `InsertNegateRAStatePass`: reads those assigned instruction RA states after |
| 110 | + optimizations, and emits `DW_CFA_AARCH64_negate_ra_state` CFIs at the correct |
| 111 | + places: wherever there is a state change between two consecutive instructions |
| 112 | + in the layout order. |
| 113 | + |
| 114 | +To track metadata on individual instructions, the `MCAnnotation` class was |
| 115 | +extended. These also have helper functions in `MCPlusBuilder`. |
| 116 | + |
| 117 | +### Saving annotations at CFI reading |
| 118 | + |
| 119 | +CFIs are read and added to BinaryFunctions in `CFIReaderWriter::FillCFIInfoFor`. |
| 120 | +At this point, we add MCAnnotations about negate-ra-state, remember-state and |
| 121 | +restore-state CFIs to the instructions they refer to. This is to not interfere |
| 122 | +with the CFI processing that already happens in BOLT (e.g. remember-state and |
| 123 | +restore-state CFIs are removed in `normalizeCFIState` for reasons unrelated to PAC). |
| 124 | + |
| 125 | +As we add the MCAnnotations *to instructions*, we have to account for the case |
| 126 | +where the function starts with a CFI altering the RA state. As CFIs modify the RA |
| 127 | +state of the instructions before them, we cannot add the annotation to the first |
| 128 | +instruction. |
| 129 | +This special case is handled by adding an `initialRAState` bool to each BinaryFunction. |
| 130 | +If the `Offset` the CFI refers to is zero, we don't store an annotation, but set |
| 131 | +the `initialRAState` in `FillCFIInfoFor`. This information is then used in |
| 132 | +`MarkRAStates`. |
| 133 | + |
| 134 | +### Binaries without DWARF info |
| 135 | + |
| 136 | +In some cases, the DWARF tables are stripped from the binary. These programs |
| 137 | +usually have some other unwind-mechanism. |
| 138 | +These passes only run on functions that include at least one negate-ra-state CFI. |
| 139 | +This avoids processing functions that do not use Pointer Authentication, or on |
| 140 | +functions that use Pointer Authentication, but do not have DWARF info. |
| 141 | + |
| 142 | +In summary: |
| 143 | +- pointer auth is not used: no change, the new passes do not run. |
| 144 | +- pointer auth is used, but DWARF info is stripped: no change, the new passes |
| 145 | + do not run. |
| 146 | +- pointer auth is used, and we have DWARF CFIs: passes run, and rewrite the |
| 147 | + negate-ra-state CFI. |
| 148 | + |
| 149 | +### MarkRAStates pass |
| 150 | + |
| 151 | +This pass runs before optimizations reorder anything. |
| 152 | + |
| 153 | +It processes MCAnnotations generated during the CFI reading stage to check if |
| 154 | +instructions have either of the three CFIs that can modify RA state: |
| 155 | +- negate-ra-state, |
| 156 | +- remember-state, |
| 157 | +- restore-state. |
| 158 | + |
| 159 | +Then it adds new MCAnnotations to each instruction, indicating their RA state. |
| 160 | +Those annotations are: |
| 161 | +- Signed, |
| 162 | +- Unsigned. |
| 163 | + |
| 164 | +Below is a simple example, that shows the two different type of annotations: |
| 165 | +what we have before the pass, and after it. |
| 166 | + |
| 167 | +| Instruction | Before | After | |
| 168 | +| ----------------------------- | --------------- | -------- | |
| 169 | +| paciasp | negate-ra-state | unsigned | |
| 170 | +| stp x29, x30, [sp, #-0x10]! | | signed | |
| 171 | +| mov x29, sp | | signed | |
| 172 | +| ldp x29, x30, [sp], #0x10 | | signed | |
| 173 | +| autiasp | negate-ra-state | signed | |
| 174 | +| ret | | unsigned | |
| 175 | + |
| 176 | +##### Error handling in MarkRAState Pass: |
| 177 | + |
| 178 | +Whenever the MarkRAStates pass finds inconsistencies in the current |
| 179 | +BinaryFunction, it marks the function as ignored using `BF.setIgnored()`. BOLT |
| 180 | +will not optimize this function but will emit it unchanged in the original section |
| 181 | +(`.bolt.org.text`). |
| 182 | + |
| 183 | +The inconsistencies are as follows: |
| 184 | +- finding a `pac*` instruction when already in signed state |
| 185 | +- finding an `aut*` instruction when already in unsigned state |
| 186 | +- finding `pac*` and `aut*` instructions without `.cfi_negate_ra_state`. |
| 187 | + |
| 188 | +Users will be informed about the number of ignored functions in the pass, the |
| 189 | +exact functions ignored, and the found inconsistency. |
| 190 | + |
| 191 | +### InsertNegateRAStatePass |
| 192 | + |
| 193 | +This pass runs after optimizations. It performns the _inverse_ of MarkRAState pa s: |
| 194 | +1. it reads the RA state annotations attached to the instructions, and |
| 195 | +2. whenever the state changes, it adds a PseudoInstruction that holds an |
| 196 | + OpNegateRAState CFI. |
| 197 | + |
| 198 | +##### Covering newly generated instructions: |
| 199 | + |
| 200 | +Some BOLT passes can add new Instructions. In InsertNegateRAStatePass, we have |
| 201 | +to know what RA state these have. |
| 202 | + |
| 203 | +The current solution has the `inferUnknownStates` function to cover these, using |
| 204 | +a fairly simple strategy: unknown states inherit the last known state. |
| 205 | + |
| 206 | +This will be updated to a more robust solution. |
| 207 | + |
| 208 | +> [!important] |
| 209 | +> As issue #160989 describes, unwind info is incorrect in stubs with multiple callers. |
| 210 | +> For this same reason, we cannot generate correct pac-specific unwind info: the signess |
| 211 | +> of the _incorrect_ return address is meaningless. |
| 212 | +
|
| 213 | +### Optimizations requiring special attention |
| 214 | + |
| 215 | +Marking states before optimizations ensure that instructions can be moved around |
| 216 | +freely. The only special case is function splitting. When a function is split, |
| 217 | +the split part becomes a new function in the emitted binary. For unwinding to |
| 218 | +work, it needs to "replay" all CFIs that lead up to the split point. BOLT does |
| 219 | +this for other CFIs. As negate-ra-state is not read (only stored as an Annotation), |
| 220 | +we have to do this manually in InsertNegateRAStatePass. Here, if the split part |
| 221 | +starts with an instruction that has Signed RA state, we add a negate-ra-state CFI |
| 222 | +to indicate this. |
| 223 | + |
| 224 | +## Option to disallow the feature |
| 225 | + |
| 226 | +The feature can be guarded with the `--update-branch-prediction` flag, which is |
| 227 | +on by default. If the flag is set to false, and a function |
| 228 | +`containedNegateRAState()` after `FillCFIInfoFor()`, BOLT exits with an error. |
0 commit comments