[Dovecot] Bug in dovecot 2.2.5: segfault due to bad alignment

[Andreas F. Borchert]

Take a look at the sources, hmac.h declares struct hmac_context:

   struct hmac_context {
           char ctx[HMAC_MAX_CONTEXT_SIZE];
           char ctxo[HMAC_MAX_CONTEXT_SIZE];
           const struct hash_method *hash;
   };

If compiled for a 32 bit virtual address space, this has an alignment
requirement of 4 due to the hash pointer.

In line 171 of auth-token.c, we have following declaration of ctx
as a local variable in auth_token_get():

   struct hmac_context ctx;

This is put on an address with an alignment requirement of 4. In
lines 174 and 175 hmac_init is invoked with hash_method_sha1:

   hmac_init(&ctx, (const unsigned char*)username, strlen(username),
             &hash_method_sha1);

In hmac.c, lines 43 and following, ctx->ctx with an alignment of
4 is passed to meth->init and meth->loop where meth refers to
hash_method_sha1:

   meth->init(ctx->ctx);
   meth->loop(ctx->ctx, k_ipad, 64);

These functions refer now to sha1_init and sha1_loop where the
first parameter is expected to be a pointer to struct sha1_ctxt,
a data structure which is declared in sha1.h:

   struct sha1_ctxt {
           union {
                   uint8_t b8[20];
                   uint32_t b32[5];
           } h;
           union {
                   uint8_t b8[8];
                   uint64_t b64[1];
           } c;
           union {
                   uint8_t b8[64];
                   uint32_t b32[16];
           } m;
           uint8_t count;
   };

Here we have with b64 one uint64_t which has on a SPARC platform
an alignment requirement of 8. In consequence, struct sha1_ctxt
has an alignment requirement of 8. With the invocations of
meth->init and meth->loop above we pass a pointer to a data
structure of alignment 4 to a function expecting a pointer to
a data structure of alignment 8. Chances are that the alignment
requirement is not met, causing a segmentation violation.

This must be solved by declaring struct hmac_context such that
is not just big enough but respects also the highest alignment
required for one of the hashing data structures.

There are several options to do this:

 * Beginning with C11, you are free to use an alignment specifier,
   i.e. add _Alignas ( uint64_t )
   (see section 6.7.5 in ISO 9899-2011)

 * GCC supports alignment attributes,
   i.e. add __attribute__ ((aligned (8)))
   or whatever is required instead of 8,
   see http://gcc.gnu.org/onlinedocs/gcc/Variable-Attributes.html

 * Do not use a local variable for it, allocate the data structure
   using malloc instead.

If you want to see a live crash, here is the relevant output of gdb
that debugs ${prefix}/libexec/dovecot/auth.

Program received signal SIGSEGV, Segmentation fault.
sha1_loop (ctxt=0xffbff63c, input=0xffbff548, len=64) at sha1.c:224
224     sha1.c: No such file or directory.
(gdb) where
#0  sha1_loop (ctxt=0xffbff63c, input=0xffbff548, len=64) at sha1.c:224
#1  0xff2e218c in hmac_init (ctx=ctx at entry=0xffbff63c,
    key=key at entry=0x6a698 "borchert", key_len=8,
    meth=0x555d0 <hash_method_sha1>) at hmac.c:44
#2  0x00023310 in auth_token_get (service=service at entry=0x6a648 "imap",
    session_pid=0x56071 "26272", username=0x6a698 "borchert",
    session_id=0x6a650 "AErv6nbk6gB/AAAB") at auth-token.c:174
#3  0x00021708 in userdb_callback (result=USERDB_RESULT_OK, request=0x6a530)
    at auth-request-handler.c:668
#4  0x0001f144 in auth_request_userdb_callback (result=<optimized out>,
    result at entry=USERDB_RESULT_OK, request=request at entry=0x6a530)
    at auth-request.c:1039
#5  0x000312c8 in prefetch_lookup (auth_request=0x6a530,
    callback=0x1f058 <auth_request_userdb_callback>) at userdb-prefetch.c:40
#6  0x0001f37c in auth_request_lookup_user (request=0x6a530,
    callback=callback at entry=0x2150c <userdb_callback>) at auth-request.c:1072
#7  0x00022034 in auth_request_handler_master_request (
    handler=<optimized out>, master=master at entry=0x6b120, id=1292369921,
    client_id=1, params=0x55bfc) at auth-request-handler.c:758
#8  0x0001be98 in master_input_request (args=<optimized out>, conn=0x6b120)
    at auth-master-connection.c:127
#9  auth_master_input_line (line=<optimized out>, conn=0x6b120)
    at auth-master-connection.c:598
#10 master_input (conn=0x6b120) at auth-master-connection.c:653
#11 0xff2ecca4 in io_loop_call_io (io=io at entry=0x6b398) at ioloop.c:387
#12 0xff2ed604 in io_loop_handler_run (ioloop=ioloop at entry=0x5e5e8)
    at ioloop-poll.c:211
#13 0xff2ec7a8 in io_loop_run (ioloop=0x5e5e8) at ioloop.c:406
#14 0xff29ad7c in master_service_run (service=0x5e128,
    callback=0x27b40 <client_connected>) at master-service.c:566
#15 0x0001852c in main (argc=1, argv=0xffbffd54) at main.c:393
(gdb) print ctxt
$1 = (struct sha1_ctxt *) 0xffbff63c

As you can see, ctxt is on a 4-byte boundary, not on an 8-byte boundary.
The crash happens at sha1.c:224 where the 8-byte-alignment is indeed
mandatory on a SPARC architecture:

   ctxt->c.b64[0] += copysiz * 8;

The environment is Solaris 10 on SPARCv9. The sources have been compiled
using gcc 4.8.0 for 32 bit.

Andreas.