Peter Collingbourne via llvm-dev
2018-Jan-24 00:44 UTC
[llvm-dev] RFC: Using link-time optimization to eliminate retpolines
The proposed mitigation for variant 2 of CVE-2017-5715, “branch target injection”, is to send all indirect branches through an instruction sequence known as a retpoline. Because the purpose of a retpoline is to prevent attacker-controlled speculation, we also end up losing the benefits of benign speculation, which can lead to a measurable loss of performance. We can regain some of those benefits if we know that the set of possible branch targets is fixed (this is sometimes known to be the case when using whole-program devirtualization or CFI -- see https://clang.llvm.org/docs/LTOVisibility.html). In that case, we can construct a so-called “branch funnel” that selects one of the possible targets by performing a binary search on an address associated with the indirect branch (for virtual calls, this is the address of the vtable, and for indirect calls via a function pointer, this is the function pointer itself), eventually directly branching to the selected target. As a result, the processor is free to speculatively execute the virtual call, but it can only speculatively branch to addresses of valid implementations of the virtual function, as opposed to arbitrary addresses. For example, suppose that we have the following class hierarchy, which is known to be closed: struct Base { virtual void f() = 0; }; struct A : Base { virtual void f(); }; struct B : Base { virtual void f(); }; struct C : Base { virtual void f(); }; We can lay out the vtables for the derived classes in the order A, B, C, and produce an instruction sequence that directs execution to one of the targets A::f, B::f and C::f depending on the vtable address. In x86_64 assembly, a branch funnel would look like this: lea B::vtable+16(%rip), %r11 cmp %r11, %r10 jb A::f je B::f jmp C::f A caller performs a virtual call by loading the vtable address into register r10, setting up the other registers for the virtual call and directly calling the branch funnel as if it were a regular function. Because the branch funnel enforces control flow integrity by itself, we can also avoid emitting CFI checks at call sites that use branch funnels when CFI is enabled. To control the layout of vtables and function pointers, we can extend existing mechanisms for controlling layout that are used to implement CFI (see https://clang.llvm.org/docs/ControlFlowIntegrityDesign.html) so that they are also used whenever a branch funnel needs to be created. The compiler will only use branch funnels when both the retpoline mitigation (-mretpoline) and whole-program devirtualization (-fwhole-program-vtables) features are enabled (the former is on the assumption that in general a regular indirect call will be less expensive than a branch funnel, and the latter provides the necessary guarantee that the type hierarchy is closed). Even when retpolines are enabled, there is still a cost associated with executing a branch funnel that needs to be balanced against the cost of a regular CFI check and retpoline, so branch funnels are only used when there are <=10 targets (this number has not been tuned yet). Because the implementation uses some of the same mechanisms that are used to implement CFI and whole-program devirtualization, it requires LTO (it is compatible with both full LTO and ThinLTO). To measure the performance impact of branch funnels, I ran a selection of Chrome benchmark suites on Chrome binaries built with CFI, CFI + retpoline and CFI + retpoline + branch funnels, and measured the median impact over all benchmarks in each suite. The numbers are presented below. I should preface these numbers by saying that these are largely microbenchmarks, so the impact of retpoline on its own is unlikely to be characteristic of real workloads. The numbers to focus on should be the impact of retpoline + branch funnels relative to the impact of retpoline, where there is a median 5.7% regression as compared to the median 8% regression associated with retpoline. Benchmark suite CFI + retpoline impact (relative to CFI) CFI + retpoline + BF impact (relative to CFI) blink_perf.bindings 0.9% improvement 9.8% improvement blink_perf.dom 20.4% regression 17.5% regression blink_perf.layout 17.4% regression 14.3% regression blink_perf.parser 3.8% regression 5.7% regression blink_perf.svg 8.0% regression 5.4% regression Future workImplementation of branch funnels for architectures other than x86_64. Implementation of branch funnels for indirect calls via a function pointer (currently only implemented for virtual calls). This will probably require an implementation of whole-program “devirtualization” for indirect calls. Use profile data to order the comparisons in the branch funnel by frequency, to minimise the number of comparisons required for frequent virtual calls. Thanks, -- -- Peter -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180123/6fffac9c/attachment.html>
Das, Dibyendu via llvm-dev
2018-Jan-24 16:47 UTC
[llvm-dev] RFC: Using link-time optimization to eliminate retpolines
Can you run your new patch on CPU2017 INT C/C++ benchmarks and report the regression percentage ? From: llvm-dev [mailto:llvm-dev-bounces at lists.llvm.org] On Behalf Of Peter Collingbourne via llvm-dev Sent: Tuesday, January 23, 2018 6:45 PM To: llvm-dev <llvm-dev at lists.llvm.org> Cc: Vlad Tsyrklevich <vlad at tsyrklevich.net> Subject: [llvm-dev] RFC: Using link-time optimization to eliminate retpolines The proposed mitigation for variant 2 of CVE-2017-5715, “branch target injection”, is to send all indirect branches through an instruction sequence known as a retpoline. Because the purpose of a retpoline is to prevent attacker-controlled speculation, we also end up losing the benefits of benign speculation, which can lead to a measurable loss of performance. We can regain some of those benefits if we know that the set of possible branch targets is fixed (this is sometimes known to be the case when using whole-program devirtualization or CFI -- see https://clang.llvm.org/docs/LTOVisibility.html). In that case, we can construct a so-called “branch funnel” that selects one of the possible targets by performing a binary search on an address associated with the indirect branch (for virtual calls, this is the address of the vtable, and for indirect calls via a function pointer, this is the function pointer itself), eventually directly branching to the selected target. As a result, the processor is free to speculatively execute the virtual call, but it can only speculatively branch to addresses of valid implementations of the virtual function, as opposed to arbitrary addresses. For example, suppose that we have the following class hierarchy, which is known to be closed: struct Base { virtual void f() = 0; }; struct A : Base { virtual void f(); }; struct B : Base { virtual void f(); }; struct C : Base { virtual void f(); }; We can lay out the vtables for the derived classes in the order A, B, C, and produce an instruction sequence that directs execution to one of the targets A::f, B::f and C::f depending on the vtable address. In x86_64 assembly, a branch funnel would look like this: lea B::vtable+16(%rip), %r11 cmp %r11, %r10 jb A::f je B::f jmp C::f A caller performs a virtual call by loading the vtable address into register r10, setting up the other registers for the virtual call and directly calling the branch funnel as if it were a regular function. Because the branch funnel enforces control flow integrity by itself, we can also avoid emitting CFI checks at call sites that use branch funnels when CFI is enabled. To control the layout of vtables and function pointers, we can extend existing mechanisms for controlling layout that are used to implement CFI (see https://clang.llvm.org/docs/ControlFlowIntegrityDesign.html) so that they are also used whenever a branch funnel needs to be created. The compiler will only use branch funnels when both the retpoline mitigation (-mretpoline) and whole-program devirtualization (-fwhole-program-vtables) features are enabled (the former is on the assumption that in general a regular indirect call will be less expensive than a branch funnel, and the latter provides the necessary guarantee that the type hierarchy is closed). Even when retpolines are enabled, there is still a cost associated with executing a branch funnel that needs to be balanced against the cost of a regular CFI check and retpoline, so branch funnels are only used when there are <=10 targets (this number has not been tuned yet). Because the implementation uses some of the same mechanisms that are used to implement CFI and whole-program devirtualization, it requires LTO (it is compatible with both full LTO and ThinLTO). To measure the performance impact of branch funnels, I ran a selection of Chrome benchmark suites on Chrome binaries built with CFI, CFI + retpoline and CFI + retpoline + branch funnels, and measured the median impact over all benchmarks in each suite. The numbers are presented below. I should preface these numbers by saying that these are largely microbenchmarks, so the impact of retpoline on its own is unlikely to be characteristic of real workloads. The numbers to focus on should be the impact of retpoline + branch funnels relative to the impact of retpoline, where there is a median 5.7% regression as compared to the median 8% regression associated with retpoline. Benchmark suite CFI + retpoline impact (relative to CFI) CFI + retpoline + BF impact (relative to CFI) blink_perf.bindings 0.9% improvement 9.8% improvement blink_perf.dom 20.4% regression 17.5% regression blink_perf.layout 17.4% regression 14.3% regression blink_perf.parser 3.8% regression 5.7% regression blink_perf.svg 8.0% regression 5.4% regression Future work Implementation of branch funnels for architectures other than x86_64. Implementation of branch funnels for indirect calls via a function pointer (currently only implemented for virtual calls). This will probably require an implementation of whole-program “devirtualization” for indirect calls. Use profile data to order the comparisons in the branch funnel by frequency, to minimise the number of comparisons required for frequent virtual calls. Thanks, -- -- Peter -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180124/64956794/attachment.html>
Sean Silva via llvm-dev
2018-Jan-26 17:33 UTC
[llvm-dev] RFC: Using link-time optimization to eliminate retpolines
Wouldn't a branch funnel open the door to a type 1 attack? E.g. if the code looks like this, then a branch funnel basically turns into a standard type 1 pattern AFAICT: struct Base { virtual int f(long) = 0; }; struct A : Base { int f(long x) override { return 0; }; }; struct B : Base { int f(long x) override { // As in listing 1 in https://spectreattack.com/spectre.pdf return array2[array1[x] * 256]; } }; -- Sean Silva On Tue, Jan 23, 2018 at 4:44 PM, Peter Collingbourne via llvm-dev < llvm-dev at lists.llvm.org> wrote:> The proposed mitigation for variant 2 of CVE-2017-5715, “branch target > injection”, is to send all indirect branches through an instruction > sequence known as a retpoline. Because the purpose of a retpoline is to > prevent attacker-controlled speculation, we also end up losing the benefits > of benign speculation, which can lead to a measurable loss of performance. > > We can regain some of those benefits if we know that the set of possible > branch targets is fixed (this is sometimes known to be the case when using > whole-program devirtualization or CFI -- see https://clang.llvm.org/docs/ > LTOVisibility.html). In that case, we can construct a so-called “branch > funnel” that selects one of the possible targets by performing a binary > search on an address associated with the indirect branch (for virtual > calls, this is the address of the vtable, and for indirect calls via a > function pointer, this is the function pointer itself), eventually directly > branching to the selected target. As a result, the processor is free to > speculatively execute the virtual call, but it can only speculatively > branch to addresses of valid implementations of the virtual function, as > opposed to arbitrary addresses. > > For example, suppose that we have the following class hierarchy, which is > known to be closed: > > struct Base { virtual void f() = 0; }; > struct A : Base { virtual void f(); }; > struct B : Base { virtual void f(); }; > struct C : Base { virtual void f(); }; > > We can lay out the vtables for the derived classes in the order A, B, C, > and produce an instruction sequence that directs execution to one of the > targets A::f, B::f and C::f depending on the vtable address. In x86_64 > assembly, a branch funnel would look like this: > > lea B::vtable+16(%rip), %r11 > cmp %r11, %r10 > jb A::f > je B::f > jmp C::f > > A caller performs a virtual call by loading the vtable address into > register r10, setting up the other registers for the virtual call and > directly calling the branch funnel as if it were a regular function. > Because the branch funnel enforces control flow integrity by itself, we can > also avoid emitting CFI checks at call sites that use branch funnels when > CFI is enabled. > > To control the layout of vtables and function pointers, we can extend > existing mechanisms for controlling layout that are used to implement CFI > (see https://clang.llvm.org/docs/ControlFlowIntegrityDesign.html) so that > they are also used whenever a branch funnel needs to be created. > > The compiler will only use branch funnels when both the retpoline > mitigation (-mretpoline) and whole-program devirtualization > (-fwhole-program-vtables) features are enabled (the former is on the > assumption that in general a regular indirect call will be less expensive > than a branch funnel, and the latter provides the necessary guarantee that > the type hierarchy is closed). Even when retpolines are enabled, there is > still a cost associated with executing a branch funnel that needs to be > balanced against the cost of a regular CFI check and retpoline, so branch > funnels are only used when there are <=10 targets (this number has not been > tuned yet). Because the implementation uses some of the same mechanisms > that are used to implement CFI and whole-program devirtualization, it > requires LTO (it is compatible with both full LTO and ThinLTO). > > To measure the performance impact of branch funnels, I ran a selection of > Chrome benchmark suites on Chrome binaries built with CFI, CFI + retpoline > and CFI + retpoline + branch funnels, and measured the median impact over > all benchmarks in each suite. The numbers are presented below. I should > preface these numbers by saying that these are largely microbenchmarks, so > the impact of retpoline on its own is unlikely to be characteristic of real > workloads. The numbers to focus on should be the impact of retpoline + > branch funnels relative to the impact of retpoline, where there is a median > 5.7% regression as compared to the median 8% regression associated with > retpoline. > > Benchmark suite > > CFI + retpoline impact > > (relative to CFI) > > CFI + retpoline + BF impact > > (relative to CFI) > > blink_perf.bindings > > 0.9% improvement > > 9.8% improvement > > blink_perf.dom > > 20.4% regression > > 17.5% regression > > blink_perf.layout > > 17.4% regression > > 14.3% regression > > blink_perf.parser > > 3.8% regression > > 5.7% regression > > blink_perf.svg > > 8.0% regression > > 5.4% regression > Future workImplementation of branch funnels for architectures other than > x86_64. > > Implementation of branch funnels for indirect calls via a function pointer > (currently only implemented for virtual calls). This will probably require > an implementation of whole-program “devirtualization” for indirect calls. > > Use profile data to order the comparisons in the branch funnel by > frequency, to minimise the number of comparisons required for frequent > virtual calls. > > Thanks, > -- > -- > Peter > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180126/f852f3d0/attachment.html>
Paul Rouschal via llvm-dev
2018-Jan-26 18:10 UTC
[llvm-dev] RFC: Using link-time optimization to eliminate retpolines
Hi, Sean Silva via llvm-dev wrote:> Wouldn't a branch funnel open the door to a type 1 attack?Only if the code looks exactly as you wrote it. If I understand this correctly the problem with indirect branches is that the "gadget", the code leaking the data, could be *anywhere* in the binary, giving the attacker much more freedom. So restricting these calls to one of the known correct results will still be a (relative) win. Hand-Wavy Example: struct Base { virtual int f(long) = 0; }; struct A : Base { int f(long x) override { return 0; }; }; struct B : Base { int f(long x) override { return 1; }; }; static int aCompletelyUnrelatedFunction() { someOtherCode(); Gadget: int z = array2[array1[somethingInTheSameRegisterAsX] * 256]; return z; } Here the attacker could train the predictor to continue execution at "Gadget". To quote from [1] "To mistrain the BTB, the attacker finds the virtual ad- dress of the gadget in the victim’s address space, then performs indirect branches to this address. This training is done from the attacker’s address space, and it does not matter what resides at the gadget address in the attacker’s address space; all that is required is that the branch used for training branches to use the same destination virtual address." [1] Kocher et.al.: Spectre Attacks: Exploiting Speculative Execution https://spectreattack.com/spectre.pdf Best Regards, Paul> E.g. if the code looks like this, then a branch funnel basically turns > into a standard type 1 pattern AFAICT: > > struct Base { > virtual int f(long) = 0; > }; > > struct A : Base { > int f(long x) override { > return 0; > }; > }; > > struct B : Base { > int f(long x) override { > // As in listing 1 in https://spectreattack.com/spectre.pdf > return array2[array1[x] * 256]; > } > }; > > -- Sean Silva > > On Tue, Jan 23, 2018 at 4:44 PM, Peter Collingbourne via llvm-dev > <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote: > > The proposed mitigation for variant 2 of CVE-2017-5715, “branch > target injection”, is to send all indirect branches through an > instruction sequence known as a retpoline. Because the purpose of a > retpoline is to prevent attacker-controlled speculation, we also end > up losing the benefits of benign speculation, which can lead to a > measurable loss of performance. > > We can regain some of those benefits if we know that the set of > possible branch targets is fixed (this is sometimes known to be the > case when using whole-program devirtualization or CFI -- see > https://clang.llvm.org/docs/LTOVisibility.html > <https://clang.llvm.org/docs/LTOVisibility.html>). In that case, we > can construct a so-called “branch funnel” that selects one of the > possible targets by performing a binary search on an address > associated with the indirect branch (for virtual calls, this is the > address of the vtable, and for indirect calls via a function > pointer, this is the function pointer itself), eventually directly > branching to the selected target. As a result, the processor is free > to speculatively execute the virtual call, but it can only > speculatively branch to addresses of valid implementations of the > virtual function, as opposed to arbitrary addresses. > > For example, suppose that we have the following class hierarchy, > which is known to be closed: > > struct Base { virtual void f() = 0; }; > struct A : Base { virtual void f(); }; > struct B : Base { virtual void f(); }; > struct C : Base { virtual void f(); }; > > We can lay out the vtables for the derived classes in the order A, > B, C, and produce an instruction sequence that directs execution to > one of the targets A::f, B::f and C::f depending on the vtable > address. In x86_64 assembly, a branch funnel would look like this: > > lea B::vtable+16(%rip), %r11 > cmp %r11, %r10 > jb A::f > je B::f > jmp C::f > > A caller performs a virtual call by loading the vtable address into > register r10, setting up the other registers for the virtual call > and directly calling the branch funnel as if it were a regular > function. Because the branch funnel enforces control flow integrity > by itself, we can also avoid emitting CFI checks at call sites that > use branch funnels when CFI is enabled. > > To control the layout of vtables and function pointers, we can > extend existing mechanisms for controlling layout that are used to > implement CFI (see > https://clang.llvm.org/docs/ControlFlowIntegrityDesign.html > <https://clang.llvm.org/docs/ControlFlowIntegrityDesign.html>) so > that they are also used whenever a branch funnel needs to be created. > > The compiler will only use branch funnels when both the retpoline > mitigation (-mretpoline) and whole-program devirtualization > (-fwhole-program-vtables) features are enabled (the former is on the > assumption that in general a regular indirect call will be less > expensive than a branch funnel, and the latter provides the > necessary guarantee that the type hierarchy is closed). Even when > retpolines are enabled, there is still a cost associated with > executing a branch funnel that needs to be balanced against the cost > of a regular CFI check and retpoline, so branch funnels are only > used when there are <=10 targets (this number has not been tuned > yet). Because the implementation uses some of the same mechanisms > that are used to implement CFI and whole-program devirtualization, > it requires LTO (it is compatible with both full LTO and ThinLTO). > > To measure the performance impact of branch funnels, I ran a > selection of Chrome benchmark suites on Chrome binaries built with > CFI, CFI + retpoline and CFI + retpoline + branch funnels, and > measured the median impact over all benchmarks in each suite. The > numbers are presented below. I should preface these numbers by > saying that these are largely microbenchmarks, so the impact of > retpoline on its own is unlikely to be characteristic of real > workloads. The numbers to focus on should be the impact of retpoline > + branch funnels relative to the impact of retpoline, where there is > a median 5.7% regression as compared to the median 8% regression > associated with retpoline. > > Benchmark suite > > > > CFI + retpoline impact > > (relative to CFI) > > > > CFI + retpoline + BF impact > > (relative to CFI) > > blink_perf.bindings > > > > 0.9% improvement > > > > 9.8% improvement > > blink_perf.dom > > > > 20.4% regression > > > > 17.5% regression > > blink_perf.layout > > > > 17.4% regression > > > > 14.3% regression > > blink_perf.parser > > > > 3.8% regression > > > > 5.7% regression > > blink_perf.svg > > > > 8.0% regression > > > > 5.4% regression > > > Future work > > Implementation of branch funnels for architectures other than x86_64. > > Implementation of branch funnels for indirect calls via a function > pointer (currently only implemented for virtual calls). This will > probably require an implementation of whole-program > “devirtualization” for indirect calls. > > Use profile data to order the comparisons in the branch funnel by > frequency, to minimise the number of comparisons required for > frequent virtual calls. > > Thanks, > -- > -- > Peter > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev > <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev> > > > > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >
Maybe Matching Threads
- RFC: Using link-time optimization to eliminate retpolines
- RFC: Using link-time optimization to eliminate retpolines
- RFC [ThinLTO]: An embedded summary encoding to support CFI and vtable opt
- Concerns about enabling retpolines by default
- RFC: a more detailed design for ThinLTO + vcall CFI