michael@0: // Copyright (c) 2012 The Chromium Authors. All rights reserved.
michael@0: // Use of this source code is governed by a BSD-style license that can be
michael@0: // found in the LICENSE file.
michael@0: 
michael@0: // Implementation of PreamblePatcher
michael@0: 
michael@0: #include "sandbox/win/src/sidestep/preamble_patcher.h"
michael@0: 
michael@0: #include "sandbox/win/src/sandbox_nt_util.h"
michael@0: #include "sandbox/win/src/sidestep/mini_disassembler.h"
michael@0: 
michael@0: // Definitions of assembly statements we need
michael@0: #define ASM_JMP32REL 0xE9
michael@0: #define ASM_INT3 0xCC
michael@0: 
michael@0: namespace {
michael@0: 
michael@0: // Very basic memcpy. We are copying 4 to 12 bytes most of the time, so there
michael@0: // is no attempt to optimize this code or have a general purpose function.
michael@0: // We don't want to call the crt from this code.
michael@0: inline void* RawMemcpy(void* destination, const void* source, size_t bytes) {
michael@0:   const char* from = reinterpret_cast<const char*>(source);
michael@0:   char* to = reinterpret_cast<char*>(destination);
michael@0: 
michael@0:   for (size_t i = 0; i < bytes ; i++)
michael@0:     to[i] = from[i];
michael@0: 
michael@0:   return destination;
michael@0: }
michael@0: 
michael@0: // Very basic memset. We are filling 1 to 7 bytes most of the time, so there
michael@0: // is no attempt to optimize this code or have a general purpose function.
michael@0: // We don't want to call the crt from this code.
michael@0: inline void* RawMemset(void* destination, int value, size_t bytes) {
michael@0:   char* to = reinterpret_cast<char*>(destination);
michael@0: 
michael@0:   for (size_t i = 0; i < bytes ; i++)
michael@0:     to[i] = static_cast<char>(value);
michael@0: 
michael@0:   return destination;
michael@0: }
michael@0: 
michael@0: }  // namespace
michael@0: 
michael@0: #define ASSERT(a, b) DCHECK_NT(a)
michael@0: 
michael@0: namespace sidestep {
michael@0: 
michael@0: SideStepError PreamblePatcher::RawPatchWithStub(
michael@0:     void* target_function,
michael@0:     void* replacement_function,
michael@0:     unsigned char* preamble_stub,
michael@0:     size_t stub_size,
michael@0:     size_t* bytes_needed) {
michael@0:   if ((NULL == target_function) ||
michael@0:       (NULL == replacement_function) ||
michael@0:       (NULL == preamble_stub)) {
michael@0:     ASSERT(false, (L"Invalid parameters - either pTargetFunction or "
michael@0:                    L"pReplacementFunction or pPreambleStub were NULL."));
michael@0:     return SIDESTEP_INVALID_PARAMETER;
michael@0:   }
michael@0: 
michael@0:   // TODO(V7:joi) Siggi and I just had a discussion and decided that both
michael@0:   // patching and unpatching are actually unsafe.  We also discussed a
michael@0:   // method of making it safe, which is to freeze all other threads in the
michael@0:   // process, check their thread context to see if their eip is currently
michael@0:   // inside the block of instructions we need to copy to the stub, and if so
michael@0:   // wait a bit and try again, then unfreeze all threads once we've patched.
michael@0:   // Not implementing this for now since we're only using SideStep for unit
michael@0:   // testing, but if we ever use it for production code this is what we
michael@0:   // should do.
michael@0:   //
michael@0:   // NOTE: Stoyan suggests we can write 8 or even 10 bytes atomically using
michael@0:   // FPU instructions, and on newer processors we could use cmpxchg8b or
michael@0:   // cmpxchg16b. So it might be possible to do the patching/unpatching
michael@0:   // atomically and avoid having to freeze other threads.  Note though, that
michael@0:   // doing it atomically does not help if one of the other threads happens
michael@0:   // to have its eip in the middle of the bytes you change while you change
michael@0:   // them.
michael@0:   unsigned char* target = reinterpret_cast<unsigned char*>(target_function);
michael@0: 
michael@0:   // Let's disassemble the preamble of the target function to see if we can
michael@0:   // patch, and to see how much of the preamble we need to take.  We need 5
michael@0:   // bytes for our jmp instruction, so let's find the minimum number of
michael@0:   // instructions to get 5 bytes.
michael@0:   MiniDisassembler disassembler;
michael@0:   unsigned int preamble_bytes = 0;
michael@0:   while (preamble_bytes < 5) {
michael@0:     InstructionType instruction_type =
michael@0:       disassembler.Disassemble(target + preamble_bytes, &preamble_bytes);
michael@0:     if (IT_JUMP == instruction_type) {
michael@0:       ASSERT(false, (L"Unable to patch because there is a jump instruction "
michael@0:                      L"in the first 5 bytes."));
michael@0:       return SIDESTEP_JUMP_INSTRUCTION;
michael@0:     } else if (IT_RETURN == instruction_type) {
michael@0:       ASSERT(false, (L"Unable to patch because function is too short"));
michael@0:       return SIDESTEP_FUNCTION_TOO_SMALL;
michael@0:     } else if (IT_GENERIC != instruction_type) {
michael@0:       ASSERT(false, (L"Disassembler encountered unsupported instruction "
michael@0:                      L"(either unused or unknown"));
michael@0:       return SIDESTEP_UNSUPPORTED_INSTRUCTION;
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   if (NULL != bytes_needed)
michael@0:     *bytes_needed = preamble_bytes + 5;
michael@0: 
michael@0:   // Inv: preamble_bytes is the number of bytes (at least 5) that we need to
michael@0:   // take from the preamble to have whole instructions that are 5 bytes or more
michael@0:   // in size total. The size of the stub required is cbPreamble + size of
michael@0:   // jmp (5)
michael@0:   if (preamble_bytes + 5 > stub_size) {
michael@0:     NOTREACHED_NT();
michael@0:     return SIDESTEP_INSUFFICIENT_BUFFER;
michael@0:   }
michael@0: 
michael@0:   // First, copy the preamble that we will overwrite.
michael@0:   RawMemcpy(reinterpret_cast<void*>(preamble_stub),
michael@0:             reinterpret_cast<void*>(target), preamble_bytes);
michael@0: 
michael@0:   // Now, make a jmp instruction to the rest of the target function (minus the
michael@0:   // preamble bytes we moved into the stub) and copy it into our preamble-stub.
michael@0:   // find address to jump to, relative to next address after jmp instruction
michael@0: #pragma warning(push)
michael@0: #pragma warning(disable:4244)
michael@0:   // This assignment generates a warning because it is 32 bit specific.
michael@0:   int relative_offset_to_target_rest
michael@0:     = ((reinterpret_cast<unsigned char*>(target) + preamble_bytes) -
michael@0:         (preamble_stub + preamble_bytes + 5));
michael@0: #pragma warning(pop)
michael@0:   // jmp (Jump near, relative, displacement relative to next instruction)
michael@0:   preamble_stub[preamble_bytes] = ASM_JMP32REL;
michael@0:   // copy the address
michael@0:   RawMemcpy(reinterpret_cast<void*>(preamble_stub + preamble_bytes + 1),
michael@0:             reinterpret_cast<void*>(&relative_offset_to_target_rest), 4);
michael@0: 
michael@0:   // Inv: preamble_stub points to assembly code that will execute the
michael@0:   // original function by first executing the first cbPreamble bytes of the
michael@0:   // preamble, then jumping to the rest of the function.
michael@0: 
michael@0:   // Overwrite the first 5 bytes of the target function with a jump to our
michael@0:   // replacement function.
michael@0:   // (Jump near, relative, displacement relative to next instruction)
michael@0:   target[0] = ASM_JMP32REL;
michael@0: 
michael@0:   // Find offset from instruction after jmp, to the replacement function.
michael@0: #pragma warning(push)
michael@0: #pragma warning(disable:4244)
michael@0:   int offset_to_replacement_function =
michael@0:     reinterpret_cast<unsigned char*>(replacement_function) -
michael@0:     reinterpret_cast<unsigned char*>(target) - 5;
michael@0: #pragma warning(pop)
michael@0:   // complete the jmp instruction
michael@0:   RawMemcpy(reinterpret_cast<void*>(target + 1),
michael@0:             reinterpret_cast<void*>(&offset_to_replacement_function), 4);
michael@0:   // Set any remaining bytes that were moved to the preamble-stub to INT3 so
michael@0:   // as not to cause confusion (otherwise you might see some strange
michael@0:   // instructions if you look at the disassembly, or even invalid
michael@0:   // instructions). Also, by doing this, we will break into the debugger if
michael@0:   // some code calls into this portion of the code.  If this happens, it
michael@0:   // means that this function cannot be patched using this patcher without
michael@0:   // further thought.
michael@0:   if (preamble_bytes > 5) {
michael@0:     RawMemset(reinterpret_cast<void*>(target + 5), ASM_INT3,
michael@0:               preamble_bytes - 5);
michael@0:   }
michael@0: 
michael@0:   // Inv: The memory pointed to by target_function now points to a relative
michael@0:   // jump instruction that jumps over to the preamble_stub.  The preamble
michael@0:   // stub contains the first stub_size bytes of the original target
michael@0:   // function's preamble code, followed by a relative jump back to the next
michael@0:   // instruction after the first cbPreamble bytes.
michael@0: 
michael@0:   return SIDESTEP_SUCCESS;
michael@0: }
michael@0: 
michael@0: };  // namespace sidestep
michael@0: 
michael@0: #undef ASSERT