CVE-2022-45188: [1day to 0day] Netatalk from Pwn2own 2021 to 0x00 cent in 2022

5 minute read

TL;DR

Hi all, This is my first blog in this page. Actually this is my second blog, but first one is missed on the previous website. I will push it again in this page. Please feed free comment or create issue in github if you detect my misstakes.

Overview

Oftenly, I found a product from pwn2own to audit. I found an amazing blog which discription about CVE-2022-23121 in Pwn2own. I tried to audit the source code of Netatalk 3.1.13 to understand the vulnerability clearly. Of couse with a hope to find a new vulnerability for next Pwn2own.

Vulnerability

When finding new bug, I often focus to memory corruption, so I find some function like memcpy, strcpy, … Because the Netatalalk works with specific protocol (AFP) and file formats. Moreover, CVE-2022-23121 occurs when parsing .AppleDouble file. Therefore I found some other file format and found .appl. The function to read .appl file is afp_getappl:

    /* fake up a cname */
    cbuf = obj->newtmp;
    q = cbuf;
    *q++ = 2;   /* long path type */
    *q++ = (unsigned char)len;
    memcpy( q, p, len );

afp_getappl also has:

    buf = obj->oldtmp;
    while (( cc = read( sa.sdt_fd, buf, sizeof( appltag )
                        + sizeof( u_short ))) > 0 ) {
        p = buf + sizeof( appltag );
        memcpy( &len, p, sizeof( len ));
        len = ntohs( len );
        p += sizeof( u_short );
        if (( cc = read( sa.sdt_fd, p, len )) < len ) {
            break;
        }
        if ( sa.sdt_index == aindex ) {
            break;
        }
        sa.sdt_index++;
    }
    if ( cc <= 0 || sa.sdt_index != aindex ) {
        *rbuflen = 0;
        return( AFPERR_NOITEM );
    }
    sa.sdt_index++;

We can see len variable parse from 2 bytes (u_short types) and read len bytes to some buffer afterward. we also need understand obj (is object of AFPObj):

typedef struct AFPObj {
    const char *cmdlineconfigfile;
    int cmdlineflags;
    const void *signature;
    struct DSI *dsi;
    struct afp_options options;
    dictionary *iniconfig;
    char username[MAXUSERLEN];
    /* to prevent confusion, only use these in afp_* calls */
    char oldtmp[AFPOBJ_TMPSIZ + 1], newtmp[AFPOBJ_TMPSIZ + 1];
    void *uam_cookie; /* cookie for uams */
    struct session_info  sinfo;
    uid_t uid;  /* client login user id */
    uid_t euid; /* client effective process user id */
    int ipc_fd; /* anonymous PF_UNIX socket for IPC with afpd parent */
    gid_t *groups;
    int ngroups;
    int afp_version;
    int cnx_cnt, cnx_max;
    /* Functions */
    void (*logout)(void);
    void (*exit)(int);
    int (*reply)(void *, int);
    int (*attention)(void *, AFPUserBytes);
    int fce_version;
    char *fce_ign_names;
    char *fce_notify_script;
    struct sl_ctx *sl_ctx;
} AFPObj;

with AFPOBJ_TMPSIZ is 4096 as default, len is 0xffff as max, so we have a heap-based buffer overflow at here.

Exploit

Between parsing len and call memcpy like above, afp_getappl have one more piece:

   {
#define hextoint( c )   ( isdigit( c ) ? c - '0' : c + 10 - 'a' )
#define islxdigit(x)    (!isupper(x)&&isxdigit(x))

        char    utomname[ MAXPATHLEN + 1];
        char        *u, *m;
        int     i, h;

        u = p;
        m = utomname;
        i = len;
        while ( i ) {
            if ( *u == ':' && *(u+1) != '\0' && islxdigit( *(u+1)) &&
                    *(u+2) != '\0' && islxdigit( *(u+2))) {
                ++u, --i;
                h = hextoint( *u ) << 4;
                ++u, --i;
                h |= hextoint( *u );
                *m++ = h;
            } else {
                *m++ = *u;
            }
            ++u, --i;
        }

        len = m - utomname;
        p = utomname;

        if ( p[ len - 1 ] == '\0' ) {
            len--;
        }
    }

Basically, the above code will copy and decode byte by byte from p to a buffer in stack. If I exploit stack buffer overflow in m to override return address, overlap memory will be occured. Therefore so hard to control addresses. Therefore, I tried to find to bypass it.

I decided using

    while (( cc = read( sa.sdt_fd, buf, sizeof( appltag )
        ...
        if (( cc = read( sa.sdt_fd, p, len )) < len ) 

to overflow buffer p (point to obj->oldtmp + 6) in the first piece code which I showed above and using

    if ( cc <= 0 || sa.sdt_index != aindex ) {
        *rbuflen = 0;
        return( AFPERR_NOITEM );
    }

to break loop and exit function without reaching decoding byte to byte. Becasue aindex is parsed from AFP packet, I try to set it to some values (different 0 and 1) and it worked :))).

Finally, we need trigger shell and run command, I do not use any House of … because I am so noob :((, so I tried to override address at end of AFPObj. In first try, I tried override reply or exit function pointer in obj. I found send_reply function call:

    obj->reply(obj->dsi, err);
    obj->exit(0);

But noway to control argument. Luckily I found that send_fce_event function will check obj->fce_notify_script and run command:

    bstring cmd = bformat("%s -v %d -e %s -i %" PRIu32 "",
                          obj->fce_notify_script,
                          FCE_PACKET_VERSION,
                          fce_event_names[event],
                          event_id);
    (void)afprun_bg(1, bdata(cmd));

and afprun_bg:

int afprun_bg(int root, char *cmd){
    ...
    execl("/bin/sh","sh","-c", cmd, NULL);
    ...

I override the *fce_notify_script with pointer which point to command. It is easy to indentify because AFPObj is saved on .bss segment.

Exploit strategy

I found that .appl file will be stored in current sharing dictionary if having config (vol dbnest = yes) and it is default setting in FreeBSD. It means, this is RCE vulnerability in FreeBSD and LPE in other OS. I create a demo in TrueNAS with guest allow and SMB enable.

Checksec of afpd:

    Arch:     amd64-64-little
    RELRO:    Partial RELRO
    Stack:    Canary found
    NX:       NX enabled
    PIE:      No PIE (0x200000)
    RUNPATH:  b'/usr/local/lib'

1. Create pre .appl file with:
   • len is over 4096 and contain length of command and padding in AFPObj ( = (fce_off+8*3).to_bytes(2, ‘big’))
   • command to run
   • padding with \x00 in fce_off-len(cmd) times.
   • fce_version is 1
   • address point command.
2. using SMB to update and modify the .appl file in local in ./AppleDouble/<x>/<xyzt>.appl
3. Send AFP packet afp_getappl with aindex is a big number (like 10, 15)
4. Trigger excute command send AFP packet afp_logout

Conclusion

With hoping get a bounty because Netatalk appeared in Pwn2own 2021, I tried report this vulnerability to zdi, TrueNAS, Synology. But noone resolve this. After a half of year, I decided public this blog and no bounty :((.