clang 19.0.0git
MIGChecker.cpp
Go to the documentation of this file.
1//== MIGChecker.cpp - MIG calling convention checker ------------*- C++ -*--==//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file defines MIGChecker, a Mach Interface Generator calling convention
10// checker. Namely, in MIG callback implementation the following rules apply:
11// - When a server routine returns an error code that represents success, it
12// must take ownership of resources passed to it (and eventually release
13// them).
14// - Additionally, when returning success, all out-parameters must be
15// initialized.
16// - When it returns any other error code, it must not take ownership,
17// because the message and its out-of-line parameters will be destroyed
18// by the client that called the function.
19// For now we only check the last rule, as its violations lead to dangerous
20// use-after-free exploits.
21//
22//===----------------------------------------------------------------------===//
23
24#include "clang/AST/Attr.h"
33#include <optional>
34
35using namespace clang;
36using namespace ento;
37
38namespace {
39class MIGChecker : public Checker<check::PostCall, check::PreStmt<ReturnStmt>,
40 check::EndFunction> {
41 BugType BT{this, "Use-after-free (MIG calling convention violation)",
43
44 // The checker knows that an out-of-line object is deallocated if it is
45 // passed as an argument to one of these functions. If this object is
46 // additionally an argument of a MIG routine, the checker keeps track of that
47 // information and issues a warning when an error is returned from the
48 // respective routine.
49 std::vector<std::pair<CallDescription, unsigned>> Deallocators = {
50#define CALL(required_args, deallocated_arg, ...) \
51 {{{__VA_ARGS__}, required_args}, deallocated_arg}
52 // E.g., if the checker sees a C function 'vm_deallocate' that is
53 // defined on class 'IOUserClient' that has exactly 3 parameters, it knows
54 // that argument #1 (starting from 0, i.e. the second argument) is going
55 // to be consumed in the sense of the MIG consume-on-success convention.
56 CALL(3, 1, "vm_deallocate"),
57 CALL(3, 1, "mach_vm_deallocate"),
58 CALL(2, 0, "mig_deallocate"),
59 CALL(2, 1, "mach_port_deallocate"),
60 CALL(1, 0, "device_deallocate"),
61 CALL(1, 0, "iokit_remove_connect_reference"),
62 CALL(1, 0, "iokit_remove_reference"),
63 CALL(1, 0, "iokit_release_port"),
64 CALL(1, 0, "ipc_port_release"),
65 CALL(1, 0, "ipc_port_release_sonce"),
66 CALL(1, 0, "ipc_voucher_attr_control_release"),
67 CALL(1, 0, "ipc_voucher_release"),
68 CALL(1, 0, "lock_set_dereference"),
69 CALL(1, 0, "memory_object_control_deallocate"),
70 CALL(1, 0, "pset_deallocate"),
71 CALL(1, 0, "semaphore_dereference"),
72 CALL(1, 0, "space_deallocate"),
73 CALL(1, 0, "space_inspect_deallocate"),
74 CALL(1, 0, "task_deallocate"),
75 CALL(1, 0, "task_inspect_deallocate"),
76 CALL(1, 0, "task_name_deallocate"),
77 CALL(1, 0, "thread_deallocate"),
78 CALL(1, 0, "thread_inspect_deallocate"),
79 CALL(1, 0, "upl_deallocate"),
80 CALL(1, 0, "vm_map_deallocate"),
81 // E.g., if the checker sees a method 'releaseAsyncReference64()' that is
82 // defined on class 'IOUserClient' that takes exactly 1 argument, it knows
83 // that the argument is going to be consumed in the sense of the MIG
84 // consume-on-success convention.
85 CALL(1, 0, "IOUserClient", "releaseAsyncReference64"),
86 CALL(1, 0, "IOUserClient", "releaseNotificationPort"),
87#undef CALL
88 };
89
90 CallDescription OsRefRetain{{"os_ref_retain"}, 1};
91
92 void checkReturnAux(const ReturnStmt *RS, CheckerContext &C) const;
93
94public:
95 void checkPostCall(const CallEvent &Call, CheckerContext &C) const;
96
97 // HACK: We're making two attempts to find the bug: checkEndFunction
98 // should normally be enough but it fails when the return value is a literal
99 // that never gets put into the Environment and ends of function with multiple
100 // returns get agglutinated across returns, preventing us from obtaining
101 // the return value. The problem is similar to https://reviews.llvm.org/D25326
102 // but now we step into it in the top-level function.
103 void checkPreStmt(const ReturnStmt *RS, CheckerContext &C) const {
104 checkReturnAux(RS, C);
105 }
106 void checkEndFunction(const ReturnStmt *RS, CheckerContext &C) const {
107 checkReturnAux(RS, C);
108 }
109
110};
111} // end anonymous namespace
112
113// A flag that says that the programmer has called a MIG destructor
114// for at least one parameter.
115REGISTER_TRAIT_WITH_PROGRAMSTATE(ReleasedParameter, bool)
116// A set of parameters for which the check is suppressed because
117// reference counting is being performed.
118REGISTER_SET_WITH_PROGRAMSTATE(RefCountedParameters, const ParmVarDecl *)
119
121 bool IncludeBaseRegions = false) {
122 // TODO: We should most likely always include base regions here.
123 SymbolRef Sym = V.getAsSymbol(IncludeBaseRegions);
124 if (!Sym)
125 return nullptr;
126
127 // If we optimistically assume that the MIG routine never re-uses the storage
128 // that was passed to it as arguments when it invalidates it (but at most when
129 // it assigns to parameter variables directly), this procedure correctly
130 // determines if the value was loaded from the transitive closure of MIG
131 // routine arguments in the heap.
132 while (const MemRegion *MR = Sym->getOriginRegion()) {
133 const auto *VR = dyn_cast<VarRegion>(MR);
134 if (VR && VR->hasStackParametersStorage() &&
135 VR->getStackFrame()->inTopFrame())
136 return cast<ParmVarDecl>(VR->getDecl());
137
138 const SymbolicRegion *SR = MR->getSymbolicBase();
139 if (!SR)
140 return nullptr;
141
142 Sym = SR->getSymbol();
143 }
144
145 return nullptr;
146}
147
149 const LocationContext *LC = C.getLocationContext();
150 assert(LC && "Unknown location context");
151
152 const StackFrameContext *SFC;
153 // Find the top frame.
154 while (LC) {
155 SFC = LC->getStackFrame();
156 LC = SFC->getParent();
157 }
158
159 const Decl *D = SFC->getDecl();
160
161 if (std::optional<AnyCall> AC = AnyCall::forDecl(D)) {
162 // Even though there's a Sema warning when the return type of an annotated
163 // function is not a kern_return_t, this warning isn't an error, so we need
164 // an extra check here.
165 // FIXME: AnyCall doesn't support blocks yet, so they remain unchecked
166 // for now.
167 if (!AC->getReturnType(C.getASTContext())
168 .getCanonicalType()->isSignedIntegerType())
169 return false;
170 }
171
172 if (D->hasAttr<MIGServerRoutineAttr>())
173 return true;
174
175 // See if there's an annotated method in the superclass.
176 if (const auto *MD = dyn_cast<CXXMethodDecl>(D))
177 for (const auto *OMD: MD->overridden_methods())
178 if (OMD->hasAttr<MIGServerRoutineAttr>())
179 return true;
180
181 return false;
182}
183
184void MIGChecker::checkPostCall(const CallEvent &Call, CheckerContext &C) const {
185 if (OsRefRetain.matches(Call)) {
186 // If the code is doing reference counting over the parameter,
187 // it opens up an opportunity for safely calling a destructor function.
188 // TODO: We should still check for over-releases.
189 if (const ParmVarDecl *PVD =
190 getOriginParam(Call.getArgSVal(0), C, /*IncludeBaseRegions=*/true)) {
191 // We never need to clean up the program state because these are
192 // top-level parameters anyway, so they're always live.
193 C.addTransition(C.getState()->add<RefCountedParameters>(PVD));
194 }
195 return;
196 }
197
198 if (!isInMIGCall(C))
199 return;
200
201 auto I = llvm::find_if(Deallocators,
202 [&](const std::pair<CallDescription, unsigned> &Item) {
203 return Item.first.matches(Call);
204 });
205 if (I == Deallocators.end())
206 return;
207
208 ProgramStateRef State = C.getState();
209 unsigned ArgIdx = I->second;
210 SVal Arg = Call.getArgSVal(ArgIdx);
211 const ParmVarDecl *PVD = getOriginParam(Arg, C);
212 if (!PVD || State->contains<RefCountedParameters>(PVD))
213 return;
214
215 const NoteTag *T =
216 C.getNoteTag([this, PVD](PathSensitiveBugReport &BR) -> std::string {
217 if (&BR.getBugType() != &BT)
218 return "";
219 SmallString<64> Str;
220 llvm::raw_svector_ostream OS(Str);
221 OS << "Value passed through parameter '" << PVD->getName()
222 << "\' is deallocated";
223 return std::string(OS.str());
224 });
225 C.addTransition(State->set<ReleasedParameter>(true), T);
226}
227
228// Returns true if V can potentially represent a "successful" kern_return_t.
230 ProgramStateRef State = C.getState();
231
232 // Can V represent KERN_SUCCESS?
233 if (!State->isNull(V).isConstrainedFalse())
234 return true;
235
236 SValBuilder &SVB = C.getSValBuilder();
237 ASTContext &ACtx = C.getASTContext();
238
239 // Can V represent MIG_NO_REPLY?
240 static const int MigNoReply = -305;
241 V = SVB.evalEQ(C.getState(), V, SVB.makeIntVal(MigNoReply, ACtx.IntTy));
242 if (!State->isNull(V).isConstrainedTrue())
243 return true;
244
245 // If none of the above, it's definitely an error.
246 return false;
247}
248
249void MIGChecker::checkReturnAux(const ReturnStmt *RS, CheckerContext &C) const {
250 // It is very unlikely that a MIG callback will be called from anywhere
251 // within the project under analysis and the caller isn't itself a routine
252 // that follows the MIG calling convention. Therefore we're safe to believe
253 // that it's always the top frame that is of interest. There's a slight chance
254 // that the user would want to enforce the MIG calling convention upon
255 // a random routine in the middle of nowhere, but given that the convention is
256 // fairly weird and hard to follow in the first place, there's relatively
257 // little motivation to spread it this way.
258 if (!C.inTopFrame())
259 return;
260
261 if (!isInMIGCall(C))
262 return;
263
264 // We know that the function is non-void, but what if the return statement
265 // is not there in the code? It's not a compile error, we should not crash.
266 if (!RS)
267 return;
268
269 ProgramStateRef State = C.getState();
270 if (!State->get<ReleasedParameter>())
271 return;
272
273 SVal V = C.getSVal(RS);
274 if (mayBeSuccess(V, C))
275 return;
276
277 ExplodedNode *N = C.generateErrorNode();
278 if (!N)
279 return;
280
281 auto R = std::make_unique<PathSensitiveBugReport>(
282 BT,
283 "MIG callback fails with error after deallocating argument value. "
284 "This is a use-after-free vulnerability because the caller will try to "
285 "deallocate it again",
286 N);
287
288 R->addRange(RS->getSourceRange());
290 N, RS->getRetValue(), *R,
291 {bugreporter::TrackingKind::Thorough, /*EnableNullFPSuppression=*/false});
292 C.emitReport(std::move(R));
293}
294
295void ento::registerMIGChecker(CheckerManager &Mgr) {
296 Mgr.registerChecker<MIGChecker>();
297}
298
299bool ento::shouldRegisterMIGChecker(const CheckerManager &mgr) {
300 return true;
301}
#define V(N, I)
Definition: ASTContext.h:3266
#define CALL(required_args, deallocated_arg,...)
static bool mayBeSuccess(SVal V, CheckerContext &C)
Definition: MIGChecker.cpp:229
static bool isInMIGCall(CheckerContext &C)
Definition: MIGChecker.cpp:148
static const ParmVarDecl * getOriginParam(SVal V, CheckerContext &C, bool IncludeBaseRegions=false)
Definition: MIGChecker.cpp:120
#define REGISTER_SET_WITH_PROGRAMSTATE(Name, Elem)
Declares an immutable set of type NameTy, suitable for placement into the ProgramState.
#define REGISTER_TRAIT_WITH_PROGRAMSTATE(Name, Type)
Declares a program state trait for type Type called Name, and introduce a type named NameTy.
Holds long-lived AST nodes (such as types and decls) that can be referred to throughout the semantic ...
Definition: ASTContext.h:182
CanQualType IntTy
Definition: ASTContext.h:1097
static std::optional< AnyCall > forDecl(const Decl *D)
If D is a callable (Objective-C method or a function), return a constructed AnyCall object.
Definition: AnyCall.h:134
Decl - This represents one declaration (or definition), e.g.
Definition: DeclBase.h:85
bool hasAttr() const
Definition: DeclBase.h:582
It wraps the AnalysisDeclContext to represent both the call stack with the help of StackFrameContext ...
const Decl * getDecl() const
const LocationContext * getParent() const
It might return null.
const StackFrameContext * getStackFrame() const
StringRef getName() const
Get the name of identifier for this declaration as a StringRef.
Definition: Decl.h:276
Represents a parameter to a function.
Definition: Decl.h:1749
ReturnStmt - This represents a return, optionally of an expression: return; return 4;.
Definition: Stmt.h:3017
Expr * getRetValue()
Definition: Stmt.h:3048
It represents a stack frame of the call stack (based on CallEvent).
SourceRange getSourceRange() const LLVM_READONLY
SourceLocation tokens are not useful in isolation - they are low level value objects created/interpre...
Definition: Stmt.cpp:326
const BugType & getBugType() const
Definition: BugReporter.h:149
A CallDescription is a pattern that can be used to match calls based on the qualified name and the ar...
Represents an abstract call to a function or method along a particular path.
Definition: CallEvent.h:153
CHECKER * registerChecker(AT &&... Args)
Used to register checkers.
MemRegion - The root abstract class for all memory regions.
Definition: MemRegion.h:96
const SymbolicRegion * getSymbolicBase() const
If this is a symbolic region, returns the region.
Definition: MemRegion.cpp:1401
The tag upon which the TagVisitor reacts.
Definition: BugReporter.h:779
nonloc::ConcreteInt makeIntVal(const IntegerLiteral *integer)
Definition: SValBuilder.h:290
SVal evalEQ(ProgramStateRef state, SVal lhs, SVal rhs)
SVal - This represents a symbolic expression, which can be either an L-value or an R-value.
Definition: SVals.h:55
Symbolic value.
Definition: SymExpr.h:30
virtual const MemRegion * getOriginRegion() const
Find the region from which this symbol originates.
Definition: SymExpr.h:104
SymbolicRegion - A special, "non-concrete" region.
Definition: MemRegion.h:775
SymbolRef getSymbol() const
It might return null.
Definition: MemRegion.h:794
bool trackExpressionValue(const ExplodedNode *N, const Expr *E, PathSensitiveBugReport &R, TrackingOptions Opts={})
Attempts to add visitors to track expression value back to its point of origin.
The JSON file list parser is used to communicate input to InstallAPI.
#define false
Definition: stdbool.h:22