a bunch of lustre bugs...
Al Viro
maintainer 1) ECHO_IOC_GET_STRIPE starts withcopy_to_user (ulsm, lsm, sizeof(*ulsm)), where ulsm is a user-supplied pointer to struct lov_stripe_md. Which starts with
struct lov_stripe_md {
atomic_t lsm_refc;
spinlock_t lsm_lock;and since sizeof(spinlock_t) depends on a slew of CONFIG_..., so do the offsets of everything after it. May I politely inquire how the hell does it manage to be of any use for userland code?
2) echo_copyout_lsm() proceeds to do the following:
for (i = 0; i < lsm->lsm_stripe_count; i++) {
if (copy_to_user (ulsm->lsm_oinfo[i], lsm->lsm_oinfo[i],
sizeof(lsm->lsm_oinfo[0])))
return -EFAULT;
}What do you think will happen if &(ulsm->lsm_oinfo) happens to be on a page boundary, with the next page unmapped? Or, for that matter, what happens if that gets used on an architecture with separate address spaces for kernel and userland. Sparc, for example... That one is trivial to fix - it's just missing get_user(up, ulsm->lsm_oinfo + i), with copy_to_user(up, ....) following it.
3) echo_copyin_lsm() has the same issues (both of them).
4) fld_proc_hash_seq_write() does this:
if (!strncmp(fld_hash[i].fh_name, buffer, count)) {
'buffer' is a userland pointer - argument of write(2), actually.
5) ll_fiemap() does
memcpy(fieinfo->fi_extents_start, &fiemap->fm_extents[0],
fiemap->fm_mapped_extents *
sizeof(struct ll_fiemap_extent));It's _not_ a nice thing to do, seeing that fi_extents_start is a userland pointer. Granted, it has passed access_ok() in ioctl_fiemap(), so it's not an instant roothole on x86. On anything with separate ASI for kernel and userland it might very well be, depending on whether any kernel addresses pass access_ok() there. parisc, for example, has access_ok() always 1 and there it *is* a roothole. And it's certainly oopsable on x86.
Incidentally, WTF are ll_fiemap_extent and ll_user_fiemap? AFAICS these are identical copies on include/uapi/linux/fiemap.h stuff, which had been there for 6 years already...
Anyway, fixes for missing get_user() and for strncmp() on userland pointers follow. The rest is a bit trickier.
Al, really annoyed by swimming through the lustre sewerful of ioctls...
Quoted 1 line
From
fee276ea51f61386438e8e65f8e39babad8c6a25Mon Sep 17 00:00:00 2001
From: Al Viro <[email protected]>
Date: Sun, 30 Nov 2014 00:12:37 -0500
Subject: [PATCH 1/2] lustre: strncmp() on user-supplied address is a Bad
Thing(tm)
--- drivers/staging/lustre/lustre/fld/lproc_fld.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-)
drivers/staging/lustre/lustre/fld/lproc_fld.c +7 -1
@@ -92,16 +92,21 @@ fld_proc_hash_seq_write(struct file *file, const char *buffer, {struct lu_client_fld *fld;struct lu_fld_hash *hash =NULL ; +char *s;int i; fld = ((struct seq_file *)file->private_data)->private;LASSERT (fld !=NULL ); + s =memdup_user (buffer, count); +if (IS_ERR (s)) +return PTR_ERR (s); +for (i =0 ; fld_hash[i].fh_name !=NULL ; i++) {if (count !=strlen (fld_hash[i].fh_name))continue ; -if (!strncmp (fld_hash[i].fh_name, buffer, count)) { +if (!strncmp (fld_hash[i].fh_name, s, count)) { hash = &fld_hash[i];break ; }
@@ -115,6 +120,7 @@ fld_proc_hash_seq_write(struct file *file, const char *buffer,CDEBUG (D_INFO ,"%s: Changed hash to \"%s\"\n" , fld->lcf_name, hash->fh_name); } +kfree (s);return count; }
Signature · 1 line
1.7.10.4
Quoted 1 line
From
fc00a7396d279f77ef192fb442dc05daecb6136dMon Sep 17 00:00:00 2001
From: Al Viro <[email protected]>
Date: Sun, 30 Nov 2014 16:02:33 -0500
Subject: [PATCH 2/2] lustre: echo_copy.._lsm() dereferences userland pointers
directly
missing get_user()
--- .../staging/lustre/lustre/obdecho/echo_client.c | 20 +++++++++++--------- 1 file changed, 11 insertions(+), 9 deletions(-)
drivers/staging/lustre/lustre/obdecho/echo_client.c +11 -9
@@ -1251,6 +1251,7 @@ static int echo_copyout_lsm (struct lov_stripe_md *lsm, void *_ulsm, int ulsm_nob) { struct lov_stripe_md *ulsm = _ulsm; +struct lov_oinfo **p;int nob, i; nob = offsetof (struct lov_stripe_md, lsm_oinfo[lsm->lsm_stripe_count]);
@@ -1260,9 +1261,10 @@ echo_copyout_lsm (struct lov_stripe_md *lsm, void *_ulsm, int ulsm_nob)if (copy_to_user (ulsm, lsm,sizeof (*ulsm)))return -EFAULT ; -for (i =0 ; i < lsm->lsm_stripe_count; i++) { -if (copy_to_user (ulsm->lsm_oinfo[i], lsm->lsm_oinfo[i], -sizeof (lsm->lsm_oinfo[0 ]))) +for (i =0 , p = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++, p++) { +struct lov_oinfo __user *up; +if (get_user (up, ulsm->lsm_oinfo + i) || +copy_to_user (up, *p,sizeof (struct lov_oinfo )))return -EFAULT ; }return 0 ;
@@ -1270,9 +1272,10 @@ echo_copyout_lsm (struct lov_stripe_md *lsm, void *_ulsm, int ulsm_nob)static int echo_copyin_lsm (struct echo_device *ed,struct lov_stripe_md *lsm, -void *ulsm,int ulsm_nob) +struct lov_stripe_md __user *ulsm,int ulsm_nob) {struct echo_client_obd *ec = ed->ed_ec; +struct lov_oinfo **p;int i;if (ulsm_nob <sizeof (*lsm))
@@ -1288,11 +1291,10 @@ echo_copyin_lsm (struct echo_device *ed, struct lov_stripe_md *lsm,return -EINVAL ; -for (i =0 ; i < lsm->lsm_stripe_count; i++) { -if (copy_from_user (lsm->lsm_oinfo[i], - ((struct lov_stripe_md *)ulsm)-> \ - lsm_oinfo[i], -sizeof (lsm->lsm_oinfo[0 ]))) +for (i =0 , p = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++, p++) { +struct lov_oinfo __user *up; +if (get_user (up, ulsm->lsm_oinfo + i) || +copy_from_user (*p, up,sizeof (struct lov_oinfo )))return -EFAULT ; }return 0 ;
Signature · 1 line
1.7.10.4
Quoted 11 lines · Al Viro
1)
ECHO_IOC_GET_STRIPEstarts withcopy_to_user(ulsm, lsm,sizeof(*ulsm)), where ulsm is a user-supplied pointer to structlov_stripe_md. Which starts withstruct lov_stripe_md { atomic_t lsm_refc; spinlock_t lsm_lock;and since
sizeof(spinlock_t)depends on a slew ofCONFIG_..., so do the offsets of everything after it. May I politely inquire how the hell does it manage to be of any use for userland code?
This is old testing code that was recently deleted from our development tree. John, could you push your patch http://review.whamcloud.com/10139
upstream to Greg KH? It cleans up a bunch of ioctl cruft.
Quoted 7 lines
2)
echo_copyout_lsm()proceeds to do the following:for (i = 0; i < lsm->lsm_stripe_count; i++) { if (copy_to_user (ulsm->lsm_oinfo[i], lsm->lsm_oinfo[i], sizeof(lsm->lsm_oinfo[0]))) return -EFAULT; }
Quoted 6 lines
What do you think will happen if &(ulsm->lsm_oinfo) happens to be on a page boundary, with the next page unmapped? Or, for that matter, what happens if that gets used on an architecture with separate address spaces for kernel and userland. Sparc, for example... That one is trivial to fix - it's just missing get_user(up,
ulsm->lsm_oinfo
+ i), withcopy_to_user(up, ....)following it.
This code was deleted in the same patch.
Quoted 1 line
3)
echo_copyin_lsm()has the same issues (both of them).
This needs the fix, thanks.
Quoted 3 lines
4) fld_proc_hash_seq_write() does this: if (!strncmp(fld_hash[i].fh_name, buffer, count)) {'buffer' is a userland pointer - argument of write(2), actually.
We just started adding __user annotation to our code, and see that "buffer" is annotated already, but it doesn't look like we have a patch to fix this issue yet. Thanks for pointing it out.
Quoted 5 lines
5) ll_fiemap() does
memcpy(fieinfo->fi_extents_start, &fiemap->fm_extents[0],
fiemap->fm_mapped_extents *
sizeof(struct ll_fiemap_extent));Quoted 8 lines
It's _not_ a nice thing to do, seeing that
fi_extents_startis a userland pointer. Granted, it has passedaccess_ok()inioctl_fiemap(), so it's not an instant roothole on x86. On anythingwith separate ASI for kernel and userland it might very well be, depending on whether any kernel addresses pass
access_ok()there. parisc, for example, hasaccess_ok()always 1 and there it *is* a roothole. And it's certainly oopsable on x86.
Looks like this needs copy_from_user() here and copy_to_user() on the other side. Fortunately, Lustre isn't being run on embedded or other strange systems w/o proper MMUs.
Quoted 2 lines
Incidentally, WTF are
ll_fiemap_extentandll_user_fiemap? AFAICS these are identical copies oninclude/uapi/linux/fiemap.hstuff, which had been there for 6 years already...
You're right - ll_*fiemap are for older kernels (FIEMAP was first added to the upstream kernel in 2.6.27, long after it was in Lustre) but they aren't needed in the upstream kernel.
Quoted 128 lines
Anyway, fixes for missing
get_user()and forstrncmp()on userland pointers follow. The rest is a bit trickier.Al, really annoyed by swimming through the lustre sewerful of ioctls...
From
fee276ea51f61386438e8e65f8e39babad8c6a25Mon Sep 17 00:00:00 2001
From: Al Viro <[email protected]>
Date: Sun, 30 Nov 2014 00:12:37 -0500
Subject: [PATCH 1/2] lustre:strncmp()on user-supplied address is a Bad Thing(tm)Signed-off-by Al Viro <[email protected]>---
drivers/staging/lustre/lustre/fld/lproc_fld.c| 8 +++++++-1 file changed, 7 insertions(+), 1 deletion(-)drivers/staging/lustre/lustre/fld/lproc_fld.c +7
-1@@ -92,16 +92,21 @@ fld_proc_hash_seq_write(struct file *file, const char *buffer, {struct lu_client_fld *fld;struct lu_fld_hash *hash =NULL ; +char *s;int i; fld = ((struct seq_file *)file->private_data)->private;LASSERT (fld !=NULL ); + s =memdup_user (buffer, count); +if (IS_ERR (s)) +return PTR_ERR (s); +for (i =0 ; fld_hash[i].fh_name !=NULL ; i++) {if (count !=strlen (fld_hash[i].fh_name))continue ; -if (!strncmp (fld_hash[i].fh_name, buffer, count)) { +if (!strncmp (fld_hash[i].fh_name, s, count)) { hash = &fld_hash[i];break ; }@@ -115,6 +120,7 @@ fld_proc_hash_seq_write(struct file *file, const char *buffer,CDEBUG (D_INFO ,"%s: Changed hash to \"%s\"\n" , fld->lcf_name, hash->fh_name); } +kfree (s);return count; }Signature · 15 lines
1.7.10.4 From fc00a7396d279f77ef192fb442dc05daecb6136d Mon Sep 17 00:00:00 2001 From: Al Viro <[email protected]> Date: Sun, 30 Nov 2014 16:02:33 -0500 Subject: [PATCH 2/2] lustre: echo_copy.._lsm() dereferences userland pointers directly missing get_user() Signed-off-by: Al Viro <[email protected]> --- .../staging/lustre/lustre/obdecho/echo_client.c | 20 +++++++++++--------- 1 file changed, 11 insertions(+), 9 deletions(-)drivers/staging/lustre/lustre/obdecho/echo_client.c +5
-3@@ -1251,6 +1251,7 @@ static int echo_copyout_lsm (struct lov_stripe_md *lsm, void *_ulsm, int ulsm_nob) { struct lov_stripe_md *ulsm = _ulsm; +struct lov_oinfo **p;int nob, i; nob = offsetof (struct lov_stripe_md, lsm_oinfo[lsm->lsm_stripe_count]);@@ -1260,9 +1261,10 @@ echo_copyout_lsm (struct lov_stripe_md *lsm, void *_ulsm, int ulsm_nob)if (copy_to_user (ulsm, lsm,sizeof (*ulsm)))return -EFAULT ; -for (i =0 ; i < lsm->lsm_stripe_count; i++) { -if (copy_to_user (ulsm->lsm_oinfo[i], lsm->lsm_oinfo[i], -sizeof (lsm->lsm_oinfo[0 ]))) +for (i =0 , p = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++, p++) { +struct lov_oinfo __user *up; +if (get_user (up, ulsm->lsm_oinfo + i) || +copy_to_user (up, *p,sizeof (struct lov_oinfo )))return -EFAULT ; }return 0 ;@@ -1270,9 +1272,10 @@ echo_copyout_lsm (struct lov_stripe_md *lsm, void *_ulsm, int ulsm_nob)static int
echo_copyin_lsm(structecho_device*ed, structlov_stripe_md*lsm,drivers/staging/lustre/lustre/obdecho/echo_client.c +1
-1- void *ulsm, int ulsm_nob) + struct lov_stripe_md __user *ulsm, int ulsm_nob){ struct echo_client_obd *ec = ed->ed_ec;drivers/staging/lustre/lustre/obdecho/echo_client.c +1
-0+struct lov_oinfo **p;int i; if (ulsm_nob < sizeof (*lsm))drivers/staging/lustre/lustre/obdecho/echo_client.c +4
-5@@ -1288,11 +1291,10 @@ echo_copyin_lsm (struct echo_device *ed, struct lov_stripe_md *lsm,return -EINVAL ; -for (i =0 ; i < lsm->lsm_stripe_count; i++) { -if (copy_from_user (lsm->lsm_oinfo[i], - ((struct lov_stripe_md *)ulsm)-> \ - lsm_oinfo[i], -sizeof (lsm->lsm_oinfo[0 ]))) +for (i =0 , p = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++, p++) { +struct lov_oinfo __user *up; +if (get_user (up, ulsm->lsm_oinfo + i) || +copy_from_user (*p, up,sizeof (struct lov_oinfo )))return -EFAULT ; }return 0 ;Signature · 6 lines
1.7.10.4 -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
Cheers, Andreas