cifs.upcall: fix regression with krb5 + creduid
Subsystems (2)
PATCH 1/3 cifs.upcall: fix regression with krb5 + creduid
With 'mount.cifs -o sec=krb5,cruid=XXX' (where XXX > 0), get_uidgid() maps XXX to 0, as mount.cifs is running as root.
This makes kerberos lookup for credentials for UID 0, which, if not existent, will fail with -ENOKEY.
To fix this, simply ignore UID/GID mapping (get_uidgid() call) when arg->uid == 0, as cifs.upcall is already running as root anyway, and then kerberos will use 'uid' (which might be 0 or cruid) for lookup.
--- cifs.upcall.c | 30 +++++++++++++++++++++--------- 1 file changed, 21 insertions(+), 9 deletions(-)
cifs.upcall.c +21 -9
@@ -1708,15 +1708,27 @@ int main(const int argc, char *const argv[])if (uid ==0 ) env_probe = false; -/* -* FIXME: this only works if we haven't switched PID namespaces. -* If we did, /proc/arg->pid/ might not exist, or worse, point to something else. -*/ - rc =get_uidgid (arg->pid, &uid, &gid); -if (rc) { -syslog (LOG_ERR ,"get_uidgid (NS): %s" ,strerror (errno)); - rc =1 ; - goto out; + gid =INVALID_UIDGID ; +if (arg->uid ==0 ) { +/* +* By now, 'uid' is either 0 or 'arg->creduid'. +* In either case, we just need to set GID=0 to make setgid() happy, as kerberos +* credential cache lookup relies only on UID anyway (in case arg->creduid != 0). +*/ + gid =0 ; + }else { +/* +* For any other case, we need to get UID/GID from procfs. +* +* FIXME: this only works if we haven't switched PID namespaces. +* If we did, /proc/arg->pid/ might not exist, or worse, point to something else. +*/ + rc =get_uidgid (arg->pid, &uid, &gid); +if (rc) { +syslog (LOG_ERR ,"get_uidgid (NS): %s" ,strerror (errno)); + rc =1 ; + goto out; + } } rc =setgid (gid);
Signature · 1 line
2.54.0
PATCH 2/3 cifs.upcall: namespace-aware key authority
spnego key is allocated before switching namespaces (i.e. as root/host credentials), but instantiation is done only later.
If we have successfully switched user namespaces, key instantiation returns EPERM because credentials have changed.
Call keyctl_assume_authority() after switching NS/UID/GID so key is instantiated with correct creds.
--- cifs.upcall.c | 23 +++++++++++++++++++++++ 1 file changed, 23 insertions(+)
cifs.upcall.c +23 -0
@@ -1539,6 +1539,7 @@ int main(const int argc, char *const argv[])krb5_ccache ccache =NULL ;unsigned expire_time =DNS_RESOLVER_DEFAULT_TIMEOUT ;const char *key_descr =NULL ; +bool same_ns = true; hostbuf[0 ] ='\0' ;
@@ -1688,6 +1689,7 @@ int main(const int argc, char *const argv[])if (arg->upcall_target ==UPTARGET_APP || arg->upcall_target ==UPTARGET_UNSPECIFIED ) {syslog (LOG_INFO ,"upcall_target=app, switching namespaces to application thread" ); arg->upcall_target =UPTARGET_APP ; + same_ns =in_same_user_ns (getpid (), arg->pid); rc =switch_to_process_ns (arg->pid);if (rc ==-1 ) {syslog (LOG_ERR ,"unable to switch to process namespace: %s" ,strerror (errno));
@@ -1756,6 +1758,27 @@ int main(const int argc, char *const argv[])if (rc) goto out; +/* +* Keyring keys' credentials are per user NS. If we switched user namespaces, we need to +* re-acquire authority (with correct UID/GID) over it so we're able to instantiate it +* later (in setup_key()). +*/ +if (!same_ns) { + errno =0 ; + rc =keyctl_assume_authority (key); + +/* +* EINPROGRESS and ENOENT are non-fatal in most cases. +* keyctl_instantiate() (in setup_key()) will fail if something was really +* unexpected. +*/ +if (rc <0 && errno !=EINPROGRESS && errno !=ENOENT ) { +syslog (LOG_ERR ,"keyctl_assume_authority: %s" ,strerror (errno)); + rc =1 ; + goto out; + } + } + rc =krb5_init_context (&context);if (rc) {syslog (LOG_ERR ,"unable to init krb5 context: %ld" , rc);
Signature · 1 line
2.54.0
PATCH 3/3 cifs.upcall: add support for different PID namespaces
In reply to Enzo MatsumiyaCurrently, applications are not able to access multiuser mounts from different PID namespace (e.g. docker/podman containers).
That's because arg->pid is set in the kernel, who uses host PID NS (init_pid_ns) value, so cifs.upcall fails when opening/parsing /proc/<arg->pid>/* entries for proper setup (because that path doesn't exist on the target PID NS, or points to a different task).
Add get_ns_pid() helper to parse the 'NSpid' field from /proc/<arg->pid>/status (before changing NS) in order to get the corresponding NS PID.
* Note: nested/shared namespaces are unsupported for now
Changes:
- rename in_same_user_ns() to in_same_ns(); also, it now takes an @nsname string to check arbitrary namespace names
- add 'struct pid_status_field' and parse_proc_pid_status() as a common helper for get_ns_pid() and get_uidgid()
--- cifs.upcall.c | 296 +++++++++++++++++++++++++++++++++++++------------- 1 file changed, 223 insertions(+), 73 deletions(-)
cifs.upcall.c +223 -73
@@ -302,26 +302,25 @@ static struct namespace_file {#define NS_PATH_MAXLEN (6 + 10 + 4 + 6 + 1)/** -* in_same_user_ns - return true if two processes are in the same user -* namespace. +* in_same_ns - Return true if two processes are in the same namespace. +* @nsname: namespace name * @pid_a: the pid of the first process * @pid_b: the pid of the second process * -* Works by comparing the inode numbers for /proc/<pid>/user. +* Works by comparing the inode numbers for /proc/<pid>/ns/@nsname. */ -static int -in_same_user_ns (pid_t pid_a,pid_t pid_b) +static int in_same_ns (const char *nsname,pid_t pid_a,pid_t pid_b) {char path[NS_PATH_MAXLEN ];ino_t a_ino, b_ino;struct stat st; -snprintf (path,sizeof (path),NS_PATH_FMT , pid_a,"user" ); +snprintf (path,sizeof (path),NS_PATH_FMT , pid_a, nsname);if (stat (path, &st) !=0 )return 0 ; a_ino = st.st_ino; -snprintf (path,sizeof (path),NS_PATH_FMT , pid_b,"user" ); +snprintf (path,sizeof (path),NS_PATH_FMT , pid_b, nsname);if (stat (path, &st) !=0 )return 0 ; b_ino = st.st_ino;
@@ -353,7 +352,7 @@ switch_to_process_ns(pid_t pid)#ifdef CLONE_NEWUSER if (namespace_files[n].nstype ==CLONE_NEWUSER - &&in_same_user_ns (getpid (), pid)) { + &&in_same_ns ("user" ,getpid (), pid)) {/* Switching to the same user namespace is forbidden, because switching to a user namespace grants all capabilities in that namespace regardless of uid. */
@@ -1380,6 +1379,72 @@ static int ip_to_fqdn(const char *addrstr, char *host, size_t hostlen)/* max valid UID/GID is (UINT_MAX - 1) */ #define INVALID_UIDGID UINT_MAX +struct pid_status_field { +const char *name; +char *value; +}; + +/* Search for @fields[].name in /proc/pid/status, save matching lines in @fields[].value. */ +static int parse_proc_pid_status (pid_t pid,struct pid_status_field *fields,int n) +{ +char path[PROC_PID_PATH_MAXLEN ] = {}, buf[256 ]; +FILE *fp =NULL ; +int i, ret; + + errno =0 ; +if (pid <0 || !fields || n <1 ) { + errno =EINVAL ; +return -1 ; + } + +for (i =0 ; i < n; i++) { +if (!fields[i].name) { + errno =EINVAL ; +return -1 ; + } + } + + ret =snprintf (path,PROC_PID_PATH_MAXLEN ,"/proc/%d/status" , pid); +if (ret <0 || ret >=PROC_PID_PATH_MAXLEN ) { +if (!errno) + errno =ENAMETOOLONG ; +return -1 ; + } + + fp =fopen (path,"r" ); +if (!fp) +return -1 ; + + errno =0 ; + ret =-1 ; +while (fgets (buf,256 , fp)) { +for (i =0 ; i < n; i++) { +size_t len =strlen (fields[i].name); + +if (fields[i].value ||strncmp (buf, fields[i].name, len)) +continue ; + + fields[i].value =strdup (buf); +if (!fields[i].value) + goto out; + } + } +out: +fclose (fp); + +if (errno) +return -1 ; + +for (i =0 ; i < n; i++) { +if (!fields[i].value) { + errno =ENODATA ; +return -1 ; + } + } + +return 0 ; +} +/* * get_uidgid - Get @pid's (real) UID and/or GID. * @pid: process to get UID/GID from
@@ -1397,94 +1462,161 @@ static int ip_to_fqdn(const char *addrstr, char *host, size_t hostlen)*/ static int get_uidgid (pid_t pid,uid_t *uidp,gid_t *gidp) { -char path[PROC_PID_PATH_MAXLEN ] = {}, buf[256 ]; -FILE *fp =NULL ; +struct pid_status_field buf[2 ] = {}; +unsigned long long val;int ret; errno =0 ; -if (pid <0 || (!uidp && !gidp)) { +if (pid <0 || !uidp || !gidp) { errno =EINVAL ;return -1 ; } -if (uidp) - *uidp =INVALID_UIDGID ; + *uidp =INVALID_UIDGID ; + *gidp =INVALID_UIDGID ; + buf[0 ].name ="Uid:" ; + buf[1 ].name ="Gid:" ; -if (gidp) - *gidp =INVALID_UIDGID ; + ret =parse_proc_pid_status (pid, buf,2 ); +if (ret) + goto out; - ret =snprintf (path,PROC_PID_PATH_MAXLEN ,"/proc/%d/status" , pid); -if (ret <0 || ret >=PROC_PID_PATH_MAXLEN ) { +/* +* Example line format (same for both Uid/Gid): +* "Uid:\t%u\t%u\t%u\t%u" +* +* Where the numbers represents: +* <real> <effective> <saved> <fsuid> +* +* We're only interested in the <real> value. +* +* (field names "Uid:"/"Gid:" parsed above, skip it) +*/ + ret =sscanf (buf[0 ].value +4 ,"%llu" , &val); +if (ret !=1 ) { + ret =-1 ;if (!errno) - errno =ENAMETOOLONG ; -return -1 ; + errno =ENODATA ; + goto out; } - fp =fopen (path,"r" ); -if (!fp) { +if (val >=UINT_MAX ) { + ret =-1 ; + errno =EINVAL ; + goto out; + } + + *uidp = (uid_t )val; + + ret =sscanf (buf[1 ].value +4 ,"%llu" , &val); +if (ret !=1 ) { + ret =-1 ; +if (!errno) + errno =EINVAL ; + goto out; + } + +if (val >=UINT_MAX ) { ret =-1 ; + errno =EINVAL ; goto out; } -/* Parse /proc/pid/status fields */ + *gidp = (gid_t )val; + ret =0 ; +out: +free (buf[0 ].value); +free (buf[1 ].value); + +if (ret) { +syslog (LOG_DEBUG ,"%s(pid=%d): %s" , __func__, pid,strerror (errno)); + *uidp =INVALID_UIDGID ; + *gidp =INVALID_UIDGID ; + } + +return ret; +} + +/* cf. kernel's include/threads.h */ +#define PID_MAX (4 * 1024 * 1024) +#define INVALID_PID ((pid_t)-1) + +/* +* get_ns_pid - Map host @pid to target-NS PID. +* @pid: host (kernel) PID +* @nspidp: pointer to store mapped NS PID +* +* Note: because /proc/pid/status information is based on the namespace the reader is coming from, +* this function must be called _before_ setns() in order to get the PID from the correct PID NS. +* +* Return: 0 on success, -1 otherwise (errno set). +* +* On errors, *@nspidp is set to INVALID_PID. +*/ +static int get_ns_pid (pid_t pid,pid_t *nspidp) +{ +struct pid_status_field buf = {}; +pid_t host_pid, ns_pid; +int ret; + errno =0 ; - ret =-1 ; -while (fgets (buf,256 , fp)) { -unsigned long long val; +if (pid <0 || !nspidp) { + errno =EINVAL ; +return -1 ; + } - errno =ENODATA ; -if ((!uidp ||strncmp (buf,"Uid:" ,4 )) && (!gidp ||strncmp (buf,"Gid:" ,4 ))) -continue ; +if (in_same_ns ("pid" ,getpid (), pid)) { + *nspidp = pid; +return 0 ; + } - errno =0 ; + *nspidp =INVALID_PID ; + buf.name ="NSpid:" ; -/* -* Example line format (same for both Uid/Gid): -* "Uid:\t%u\t%u\%u\%u" -* -* Where the numbers represents: -* <real> <effective> <saved> <fsuid> -* -* We're only interested in the <real> value. -* -* (field names "Uid:"/"Gid:" parsed above, skip it) -*/ - ret =sscanf (&buf[0 ] +4 ,"%llu" , &val); -if (ret !=1 ) { - ret =-1 ; -if (errno) -break ; -continue ; - } + ret =parse_proc_pid_status (pid, &buf,1 ); +if (ret) + goto out; +/* +* NSpid field format: +* "NSpid:\t%d\t%d" +* +* Where the numbers represents: +* (for NS process seen from host NS): +* <host PID> <NS pid> +* +* (for NS process seen from same NS): +* <NS/host pid> +* +* We're only interested in the second value. +* But we still scan the first one so we can check host_pid == @pid. +* +* Note: sscanf() returning 1 is a bug because it means this is being called +* _after_ setns(). Beware. +*/ + ret =sscanf (buf.value +6 ,"%d %d" , &host_pid, &ns_pid); +free (buf.value); +if (ret !=2 ) { ret =-1 ; -if (val >=UINT_MAX ) { +if (!errno) errno =EINVAL ; -break ; - } - -if (uidp && !strncmp (buf,"Uid:" ,4 )) - *uidp = (uid_t )val; -else - *gidp = (gid_t )val; + goto out; + } -if ((!uidp || *uidp !=INVALID_UIDGID ) && (!gidp || *gidp !=INVALID_UIDGID )) { - errno =0 ; - ret =0 ; -break ; - } +if (host_pid <0 || host_pid >=PID_MAX || host_pid != pid || ns_pid <0 || + ns_pid >=PID_MAX ) { + errno =EINVAL ; + ret =-1 ; + goto out; } -out: -if (fp) -fclose (fp); + *nspidp = ns_pid; + errno =0 ; + ret =0 ; +out:if (ret) {syslog (LOG_DEBUG ,"%s(pid=%d): %s" , __func__, pid,strerror (errno)); -if (uidp) - *uidp =INVALID_UIDGID ; - -if (gidp) - *gidp =INVALID_UIDGID ; + *nspidp =INVALID_PID ; }return ret;
@@ -1687,10 +1819,28 @@ int main(const int argc, char *const argv[])* filesystem and have the correct network configuration. */ if (arg->upcall_target ==UPTARGET_APP || arg->upcall_target ==UPTARGET_UNSPECIFIED ) { -syslog (LOG_INFO ,"upcall_target=app, switching namespaces to application thread" ); +pid_t pid = arg->pid; + arg->upcall_target =UPTARGET_APP ; - same_ns =in_same_user_ns (getpid (), arg->pid); - rc =switch_to_process_ns (arg->pid); + same_ns =in_same_ns ("user" ,getpid (), pid); +if (!same_ns) +syslog (LOG_INFO ,"upcall_target=app, switching namespaces to application thread" ); +#ifdef CLONE_NEWPID +/* +* arg->pid holds host PID (gotten from kernel). +* If we're going to switch PID namespaces, we need to map the PID from the correct +* target PID NS before switching, so UID/GID are also fetched from the correct NS. +*/ + rc =get_ns_pid (arg->pid, &pid); +if (rc) { +syslog (LOG_ERR ,"get_ns_pid: %s" ,strerror (errno)); + rc =1 ; + goto out; + } + + arg->pid = pid; +#endif /* CLONE_NEWPID */ + rc =switch_to_process_ns (pid);if (rc ==-1 ) {syslog (LOG_ERR ,"unable to switch to process namespace: %s" ,strerror (errno)); rc =1 ;
@@ -1722,8 +1872,8 @@ int main(const int argc, char *const argv[])/* * For any other case, we need to get UID/GID from procfs. * -* FIXME: this only works if we haven't switched PID namespaces. -* If we did, /proc/arg->pid/ might not exist, or worse, point to something else. +* If we switched PID namespaces, arg->pid already points to the correct +* NS PID value. */ rc =get_uidgid (arg->pid, &uid, &gid);if (rc) {
Signature · 1 line
2.54.0
Quoted 10 lines · Enzo Matsumiya
With 'mount.cifs -o sec=krb5,cruid=XXX' (where XXX > 0),
get_uidgid()maps XXX to 0, as mount.cifs is running as root.This makes kerberos lookup for credentials for UID 0, which, if not existent, will fail with
-ENOKEY.To fix this, simply ignore UID/GID mapping (
get_uidgid()call) whenarg->uid== 0, as cifs.upcall is already running as root anyway, and then kerberos will use 'uid' (which might be 0 or cruid) for lookup.
Quoted 6 lines
Fixes 972c5b5ff95e ("cifs.upcall: remove getpwuid() dependency")Reported-by Paulo Alcantara <[email protected]>Signed-off-by Enzo Matsumiya <[email protected]>--- cifs.upcall.c | 30 +++++++++++++++++++++--------- 1 file changed, 21 insertions(+), 9 deletions(-)