| Message ID | 202005041205.C7AF4AF@keescook (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | sysctl: Make sure proc handlers can't expose heap memory | expand |
On Mon, May 04, 2020 at 12:08:55PM -0700, Kees Cook wrote: > Just as a precaution, make sure that proc handlers don't accidentally > grow "count" beyond the allocated kbuf size. > > Signed-off-by: Kees Cook <keescook@chromium.org> > --- > This applies to hch's sysctl cleanup tree... > --- > fs/proc/proc_sysctl.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/fs/proc/proc_sysctl.c b/fs/proc/proc_sysctl.c > index 15030784566c..535ab26473af 100644 > --- a/fs/proc/proc_sysctl.c > +++ b/fs/proc/proc_sysctl.c > @@ -546,6 +546,7 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > struct inode *inode = file_inode(filp); > struct ctl_table_header *head = grab_header(inode); > struct ctl_table *table = PROC_I(inode)->sysctl_entry; > + size_t count_max = count; > void *kbuf; > ssize_t error; > > @@ -590,6 +591,8 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > > if (!write) { > error = -EFAULT; > + if (WARN_ON(count > count_max)) > + count = count_max; That crash a system with panic-on-warn. I don't think we want that? Luis
On Mon, May 04, 2020 at 07:59:37PM +0000, Luis Chamberlain wrote: > On Mon, May 04, 2020 at 12:08:55PM -0700, Kees Cook wrote: > > Just as a precaution, make sure that proc handlers don't accidentally > > grow "count" beyond the allocated kbuf size. > > > > Signed-off-by: Kees Cook <keescook@chromium.org> > > --- > > This applies to hch's sysctl cleanup tree... > > --- > > fs/proc/proc_sysctl.c | 3 +++ > > 1 file changed, 3 insertions(+) > > > > diff --git a/fs/proc/proc_sysctl.c b/fs/proc/proc_sysctl.c > > index 15030784566c..535ab26473af 100644 > > --- a/fs/proc/proc_sysctl.c > > +++ b/fs/proc/proc_sysctl.c > > @@ -546,6 +546,7 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > > struct inode *inode = file_inode(filp); > > struct ctl_table_header *head = grab_header(inode); > > struct ctl_table *table = PROC_I(inode)->sysctl_entry; > > + size_t count_max = count; > > void *kbuf; > > ssize_t error; > > > > @@ -590,6 +591,8 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > > > > if (!write) { > > error = -EFAULT; > > + if (WARN_ON(count > count_max)) > > + count = count_max; > > That crash a system with panic-on-warn. I don't think we want that? Eh? None of the handlers should be making this mistake currently and it's not a mistake that can be controlled from userspace. WARN() is absolutely what's wanted here: report an impossible situation (and handle it gracefully for the bulk of users that don't have panic_on_warn set).
On Mon, May 04, 2020 at 01:32:07PM -0700, Kees Cook wrote: > On Mon, May 04, 2020 at 07:59:37PM +0000, Luis Chamberlain wrote: > > On Mon, May 04, 2020 at 12:08:55PM -0700, Kees Cook wrote: > > > Just as a precaution, make sure that proc handlers don't accidentally > > > grow "count" beyond the allocated kbuf size. > > > > > > Signed-off-by: Kees Cook <keescook@chromium.org> > > > --- > > > This applies to hch's sysctl cleanup tree... > > > --- > > > fs/proc/proc_sysctl.c | 3 +++ > > > 1 file changed, 3 insertions(+) > > > > > > diff --git a/fs/proc/proc_sysctl.c b/fs/proc/proc_sysctl.c > > > index 15030784566c..535ab26473af 100644 > > > --- a/fs/proc/proc_sysctl.c > > > +++ b/fs/proc/proc_sysctl.c > > > @@ -546,6 +546,7 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > > > struct inode *inode = file_inode(filp); > > > struct ctl_table_header *head = grab_header(inode); > > > struct ctl_table *table = PROC_I(inode)->sysctl_entry; > > > + size_t count_max = count; > > > void *kbuf; > > > ssize_t error; > > > > > > @@ -590,6 +591,8 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > > > > > > if (!write) { > > > error = -EFAULT; > > > + if (WARN_ON(count > count_max)) > > > + count = count_max; > > > > That would crash a system with panic-on-warn. I don't think we want that? > > Eh? None of the handlers should be making this mistake currently and > it's not a mistake that can be controlled from userspace. WARN() is > absolutely what's wanted here: report an impossible situation (and > handle it gracefully for the bulk of users that don't have > panic_on_warn set). Alrighty, Greg are you OK with this type of WARN_ON()? You recently expressed concerns over its use due to panic-on-warn on another patch. LUis
On Mon, May 04, 2020 at 12:08:55PM -0700, Kees Cook wrote: > Just as a precaution, make sure that proc handlers don't accidentally > grow "count" beyond the allocated kbuf size. > > Signed-off-by: Kees Cook <keescook@chromium.org> > --- > This applies to hch's sysctl cleanup tree... This looks ok o me. You should probably add Al to the Cc list as he has picked up my series into a branch of vfs.git. Acked-by: Christoph Hellwig <hch@lst.de>
On Mon, May 04, 2020 at 09:59:03PM +0000, Luis Chamberlain wrote: > On Mon, May 04, 2020 at 01:32:07PM -0700, Kees Cook wrote: > > On Mon, May 04, 2020 at 07:59:37PM +0000, Luis Chamberlain wrote: > > > On Mon, May 04, 2020 at 12:08:55PM -0700, Kees Cook wrote: > > > > Just as a precaution, make sure that proc handlers don't accidentally > > > > grow "count" beyond the allocated kbuf size. > > > > > > > > Signed-off-by: Kees Cook <keescook@chromium.org> > > > > --- > > > > This applies to hch's sysctl cleanup tree... > > > > --- > > > > fs/proc/proc_sysctl.c | 3 +++ > > > > 1 file changed, 3 insertions(+) > > > > > > > > diff --git a/fs/proc/proc_sysctl.c b/fs/proc/proc_sysctl.c > > > > index 15030784566c..535ab26473af 100644 > > > > --- a/fs/proc/proc_sysctl.c > > > > +++ b/fs/proc/proc_sysctl.c > > > > @@ -546,6 +546,7 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > > > > struct inode *inode = file_inode(filp); > > > > struct ctl_table_header *head = grab_header(inode); > > > > struct ctl_table *table = PROC_I(inode)->sysctl_entry; > > > > + size_t count_max = count; > > > > void *kbuf; > > > > ssize_t error; > > > > > > > > @@ -590,6 +591,8 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > > > > > > > > if (!write) { > > > > error = -EFAULT; > > > > + if (WARN_ON(count > count_max)) > > > > + count = count_max; > > > > > > That would crash a system with panic-on-warn. I don't think we want that? > > > > Eh? None of the handlers should be making this mistake currently and > > it's not a mistake that can be controlled from userspace. WARN() is > > absolutely what's wanted here: report an impossible situation (and > > handle it gracefully for the bulk of users that don't have > > panic_on_warn set). > > Alrighty, Greg are you OK with this type of WARN_ON()? You recently > expressed concerns over its use due to panic-on-warn on another patch. We should never call WARN() on any path that a user can trigger. If it is just a "the developer called this api in a foolish way" then we could use a WARN_ON() to have them realize their mistake, but in my personal experience, foolish developers don't even notice that kind of mistake :( thanks, greg k-h
On Tue, May 05, 2020 at 08:34:41AM +0200, Greg KH wrote: > On Mon, May 04, 2020 at 09:59:03PM +0000, Luis Chamberlain wrote: > > On Mon, May 04, 2020 at 01:32:07PM -0700, Kees Cook wrote: > > > On Mon, May 04, 2020 at 07:59:37PM +0000, Luis Chamberlain wrote: > > > > On Mon, May 04, 2020 at 12:08:55PM -0700, Kees Cook wrote: > > > > > Just as a precaution, make sure that proc handlers don't accidentally > > > > > grow "count" beyond the allocated kbuf size. > > > > > > > > > > Signed-off-by: Kees Cook <keescook@chromium.org> > > > > > --- > > > > > This applies to hch's sysctl cleanup tree... > > > > > --- > > > > > fs/proc/proc_sysctl.c | 3 +++ > > > > > 1 file changed, 3 insertions(+) > > > > > > > > > > diff --git a/fs/proc/proc_sysctl.c b/fs/proc/proc_sysctl.c > > > > > index 15030784566c..535ab26473af 100644 > > > > > --- a/fs/proc/proc_sysctl.c > > > > > +++ b/fs/proc/proc_sysctl.c > > > > > @@ -546,6 +546,7 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > > > > > struct inode *inode = file_inode(filp); > > > > > struct ctl_table_header *head = grab_header(inode); > > > > > struct ctl_table *table = PROC_I(inode)->sysctl_entry; > > > > > + size_t count_max = count; > > > > > void *kbuf; > > > > > ssize_t error; > > > > > > > > > > @@ -590,6 +591,8 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > > > > > > > > > > if (!write) { > > > > > error = -EFAULT; > > > > > + if (WARN_ON(count > count_max)) > > > > > + count = count_max; > > > > > > > > That would crash a system with panic-on-warn. I don't think we want that? > > > > > > Eh? None of the handlers should be making this mistake currently and > > > it's not a mistake that can be controlled from userspace. WARN() is > > > absolutely what's wanted here: report an impossible situation (and > > > handle it gracefully for the bulk of users that don't have > > > panic_on_warn set). > > > > Alrighty, Greg are you OK with this type of WARN_ON()? You recently > > expressed concerns over its use due to panic-on-warn on another patch. > > We should never call WARN() on any path that a user can trigger. > > If it is just a "the developer called this api in a foolish way" then we > could use a WARN_ON() to have them realize their mistake, but in my > personal experience, foolish developers don't even notice that kind of > mistake :( Right -- while it'd be nice if the developer noticed it, it is _usually_ an unsuspecting end user (or fuzzer), in which case we absolutely want a WARN (and not a BUG![1]) and have the situations handled gracefully, so it can be reported and fixed. -Kees [1] https://www.kernel.org/doc/html/latest/process/deprecated.html#bug-and-bug-on
On Tue, May 05, 2020 at 01:41:44PM -0700, Kees Cook wrote: > On Tue, May 05, 2020 at 08:34:41AM +0200, Greg KH wrote: > > On Mon, May 04, 2020 at 09:59:03PM +0000, Luis Chamberlain wrote: > > > On Mon, May 04, 2020 at 01:32:07PM -0700, Kees Cook wrote: > > > > On Mon, May 04, 2020 at 07:59:37PM +0000, Luis Chamberlain wrote: > > > > > On Mon, May 04, 2020 at 12:08:55PM -0700, Kees Cook wrote: > > > > > > Just as a precaution, make sure that proc handlers don't accidentally > > > > > > grow "count" beyond the allocated kbuf size. > > > > > > > > > > > > Signed-off-by: Kees Cook <keescook@chromium.org> > > > > > > --- > > > > > > This applies to hch's sysctl cleanup tree... > > > > > > --- > > > > > > fs/proc/proc_sysctl.c | 3 +++ > > > > > > 1 file changed, 3 insertions(+) > > > > > > > > > > > > diff --git a/fs/proc/proc_sysctl.c b/fs/proc/proc_sysctl.c > > > > > > index 15030784566c..535ab26473af 100644 > > > > > > --- a/fs/proc/proc_sysctl.c > > > > > > +++ b/fs/proc/proc_sysctl.c > > > > > > @@ -546,6 +546,7 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > > > > > > struct inode *inode = file_inode(filp); > > > > > > struct ctl_table_header *head = grab_header(inode); > > > > > > struct ctl_table *table = PROC_I(inode)->sysctl_entry; > > > > > > + size_t count_max = count; > > > > > > void *kbuf; > > > > > > ssize_t error; > > > > > > > > > > > > @@ -590,6 +591,8 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, > > > > > > > > > > > > if (!write) { > > > > > > error = -EFAULT; > > > > > > + if (WARN_ON(count > count_max)) > > > > > > + count = count_max; > > > > > > > > > > That would crash a system with panic-on-warn. I don't think we want that? > > > > > > > > Eh? None of the handlers should be making this mistake currently and > > > > it's not a mistake that can be controlled from userspace. WARN() is > > > > absolutely what's wanted here: report an impossible situation (and > > > > handle it gracefully for the bulk of users that don't have > > > > panic_on_warn set). > > > > > > Alrighty, Greg are you OK with this type of WARN_ON()? You recently > > > expressed concerns over its use due to panic-on-warn on another patch. > > > > We should never call WARN() on any path that a user can trigger. > > > > If it is just a "the developer called this api in a foolish way" then we > > could use a WARN_ON() to have them realize their mistake, but in my > > personal experience, foolish developers don't even notice that kind of > > mistake :( > > Right -- while it'd be nice if the developer noticed it, it is _usually_ > an unsuspecting end user (or fuzzer), in which case we absolutely want a > WARN (and not a BUG![1]) and have the situations handled gracefully, so > it can be reported and fixed. I've been using WARN*() for this exact purpose before, so I am as surprised as you are bout these concerns. However if we have folks shipping with panic-on-warn this would be rather detrimental to our goals. Greg, are you aware of folks shipping with panic-on-warn on some products? Luis
On Tue, May 05, 2020 at 10:03:27PM +0000, Luis Chamberlain wrote: > On Tue, May 05, 2020 at 01:41:44PM -0700, Kees Cook wrote: > > Right -- while it'd be nice if the developer noticed it, it is _usually_ > > an unsuspecting end user (or fuzzer), in which case we absolutely want a > > WARN (and not a BUG![1]) and have the situations handled gracefully, so > > it can be reported and fixed. > > I've been using WARN*() for this exact purpose before, so I am as > surprised as you are bout these concerns. However if we have folks I don't see any mismatch here: it's not user-reachable, which is what Greg said. WARN is for non-user-reachable "impossible situations". We want to know if those can be hit (via bad API usage, races, etc). If it's reachable from userspace, then it can't be a WARN() any more and needs to be pr_warn(). > shipping with panic-on-warn this would be rather detrimental to our > goals. > > Greg, are you aware of folks shipping with panic-on-warn on some products? People shipping with panic_on_warn are expecting to panic for WARNs like this. :P
diff --git a/fs/proc/proc_sysctl.c b/fs/proc/proc_sysctl.c index 15030784566c..535ab26473af 100644 --- a/fs/proc/proc_sysctl.c +++ b/fs/proc/proc_sysctl.c @@ -546,6 +546,7 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, struct inode *inode = file_inode(filp); struct ctl_table_header *head = grab_header(inode); struct ctl_table *table = PROC_I(inode)->sysctl_entry; + size_t count_max = count; void *kbuf; ssize_t error; @@ -590,6 +591,8 @@ static ssize_t proc_sys_call_handler(struct file *filp, void __user *ubuf, if (!write) { error = -EFAULT; + if (WARN_ON(count > count_max)) + count = count_max; if (copy_to_user(ubuf, kbuf, count)) goto out_free_buf; }
Just as a precaution, make sure that proc handlers don't accidentally grow "count" beyond the allocated kbuf size. Signed-off-by: Kees Cook <keescook@chromium.org> --- This applies to hch's sysctl cleanup tree... --- fs/proc/proc_sysctl.c | 3 +++ 1 file changed, 3 insertions(+)