On Mon, Dec 17, 2018 at 4:24 PM Edgecombe, Rick P
[off-list ref] wrote:

On Sat, 2018-12-15 at 10:52 -0800, Andy Lutomirski wrote:

quoted

On Wed, Dec 12, 2018 at 2:01 PM Edgecombe, Rick P
[off-list ref] wrote:

quoted

On Wed, 2018-12-12 at 11:57 -0800, Andy Lutomirski wrote:

quoted

On Wed, Dec 12, 2018 at 11:50 AM Edgecombe, Rick P
[off-list ref] wrote:

quoted

On Tue, 2018-12-11 at 18:20 -0800, Andy Lutomirski wrote:

quoted

On Tue, Dec 11, 2018 at 4:12 PM Rick Edgecombe
[off-list ref] wrote:

quoted

This adds two new flags VM_IMMEDIATE_UNMAP and VM_HAS_SPECIAL_PERMS,
for
enabling vfree operations to immediately clear executable TLB
entries to
freed
pages, and handle freeing memory with special permissions.

In order to support vfree being called on memory that might be RO,
the
vfree
deferred list node is moved to a kmalloc allocated struct, from
where it
is
today, reusing the allocation being freed.

arch_vunmap is a new __weak function that implements the actual
unmapping
and
resetting of the direct map permissions. It can be overridden by
more
efficient
architecture specific implementations.

For the default implementation, it uses architecture agnostic
methods
which
are
equivalent to what most usages do before calling vfree. So now it is
just
centralized here.

This implementation derives from two sketches from Dave Hansen and
Andy
Lutomirski.

Suggested-by: Dave Hansen <redacted>
Suggested-by: Andy Lutomirski <luto@kernel.org>
Suggested-by: Will Deacon <redacted>
Signed-off-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
---
 include/linux/vmalloc.h |  2 ++
 mm/vmalloc.c            | 73 +++++++++++++++++++++++++++++++++++++-
---
 2 files changed, 69 insertions(+), 6 deletions(-)

diff --git a/include/linux/vmalloc.h b/include/linux/vmalloc.h
index 398e9c95cd61..872bcde17aca 100644
--- a/include/linux/vmalloc.h
+++ b/include/linux/vmalloc.h

@@ -21,6 +21,8 @@ struct notifier_block;                /* in

notifier.h
*/
 #define VM_UNINITIALIZED       0x00000020      /* vm_struct is not
fully
initialized */
 #define VM_NO_GUARD            0x00000040      /* don't add guard
page
*/
 #define VM_KASAN               0x00000080      /* has allocated
kasan
shadow memory */
+#define VM_IMMEDIATE_UNMAP     0x00000200      /* flush before
releasing
pages */
+#define VM_HAS_SPECIAL_PERMS   0x00000400      /* may be freed with
special
perms */
 /* bits [20..32] reserved for arch specific ioremap internals */

 /*

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 97d4b25d0373..02b284d2245a 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c

@@ -18,6 +18,7 @@
 #include <linux/interrupt.h>
 #include <linux/proc_fs.h>
 #include <linux/seq_file.h>
+#include <linux/set_memory.h>
 #include <linux/debugobjects.h>
 #include <linux/kallsyms.h>
 #include <linux/list.h>

@@ -38,6 +39,11 @@

 #include "internal.h"

+struct vfree_work {
+       struct llist_node node;
+       void *addr;
+};
+
 struct vfree_deferred {
        struct llist_head list;
        struct work_struct wq;

@@ -50,9 +56,13 @@ static void free_work(struct work_struct *w)
 {
        struct vfree_deferred *p = container_of(w, struct

vfree_deferred,
wq);
        struct llist_node *t, *llnode;
+       struct vfree_work *cur;

-       llist_for_each_safe(llnode, t, llist_del_all(&p->list))
-               __vunmap((void *)llnode, 1);
+       llist_for_each_safe(llnode, t, llist_del_all(&p->list)) {
+               cur = container_of(llnode, struct vfree_work, node);
+               __vunmap(cur->addr, 1);
+               kfree(cur);
+       }
 }

 /*** Page table manipulation functions ***/

@@ -1494,6 +1504,48 @@ struct vm_struct *remove_vm_area(const void

*addr)
        return NULL;
 }

+/*
+ * This function handles unmapping and resetting the direct map as
efficiently
+ * as it can with cross arch functions. The three categories of
architectures
+ * are:
+ *   1. Architectures with no set_memory implementations and no
direct
map
+ *      permissions.
+ *   2. Architectures with set_memory implementations but no direct
map
+ *      permissions
+ *   3. Architectures with set_memory implementations and direct
map
permissions
+ */
+void __weak arch_vunmap(struct vm_struct *area, int
deallocate_pages)

My general preference is to avoid __weak functions -- they don't
optimize well.  Instead, I prefer either:

#ifndef arch_vunmap
void arch_vunmap(...);
#endif

or

#ifdef CONFIG_HAVE_ARCH_VUNMAP
...
#endif

Ok.

quoted

quoted

+{
+       unsigned long addr = (unsigned long)area->addr;
+       int immediate = area->flags & VM_IMMEDIATE_UNMAP;
+       int special = area->flags & VM_HAS_SPECIAL_PERMS;
+
+       /*
+        * In case of 2 and 3, use this general way of resetting the
permissions
+        * on the directmap. Do NX before RW, in case of X, so there
is
no
W^X
+        * violation window.
+        *
+        * For case 1 these will be noops.
+        */
+       if (immediate)
+               set_memory_nx(addr, area->nr_pages);
+       if (deallocate_pages && special)
+               set_memory_rw(addr, area->nr_pages);

Can you elaborate on the intent here?  VM_IMMEDIATE_UNMAP means "I
want that alias gone before any deallocation happens".
VM_HAS_SPECIAL_PERMS means "I mucked with the direct map -- fix it for
me, please".  deallocate means "this was vfree -- please free the
pages".  I'm not convinced that all the various combinations make
sense.  Do we really need both flags?

VM_HAS_SPECIAL_PERMS is supposed to mean, like you said, "reset the
direct
map".
Where VM_IMMEDIATE_UNMAP means, the vmalloc allocation has extra
capabilties
where we don't want to leave an enhanced capability TLB entry to the
freed
page.

I was trying to pick names that could apply more generally for potential
future
special memory capabilities. Today VM_HAS_SPECIAL_PERMS does just mean
reset
write to the directmap and VM_IMMEDIATE_UNMAP means vmalloc mapping is
executable.

A present day reason for keeping both flags is, it is more efficient in
the
arch-agnostic implementation when freeing memory that is just RO and not
executable. It saves a TLB flush.

quoted

(VM_IMMEDIATE_UNMAP is a bit of a lie, since, if in_interrupt(), it's
not immediate.)

True, maybe VM_MUST_FLUSH or something else?

quoted

If we do keep both flags, maybe some restructuring would make sense,
like this, perhaps.  Sorry about horrible whitespace damage.

if (special) {
  /* VM_HAS_SPECIAL_PERMS makes little sense without deallocate_pages.
*/
  WARN_ON_ONCE(!deallocate_pages);

  if (immediate) {
    /* It's possible that the vmap alias is X and we're about to make
the direct map RW.  To avoid a window where executable memory is
writable, first mark the vmap alias NX.  This is silly, since we're
about to *unmap* it, but this is the best we can do if all we have to
work with is the set_memory_abc() APIs.  Architectures should override
this whole function to get better behavior. */
    set_memory_nx(...);
  }

  set_memory_rw(addr, area->nr_pages);
}

Ok.

quoted

quoted

+
+       /* Always actually remove the area */
+       remove_vm_area(area->addr);
+
+       /*
+        * Need to flush the TLB before freeing pages in the case of
this
flag.
+        * As long as that's happening, unmap aliases.
+        *
+        * For 2 and 3, this will not be needed because of the
set_memory_nx
+        * above, because the stale TLBs will be NX.

I'm not sure I agree with this comment.  If the caller asked for an
immediate unmap, we should give an immediate unmap.  But I'm still not
sure I see why VM_IMMEDIATE_UNMAP needs to exist as a separate flag.

Yea. I was just trying to save a TLB flush, since for today's callers
that
have
set_memory there isn't a security downside I know of to just leaving it
NX.
Maybe its not worth the tradeoff of confusion? Or I can clarify that in
the
comment.

Don't both of the users in your series set both flags, though?  My
real objection to having them be separate is that, in the absence of
users, it's less clear exactly what they should do and the code
doesn't get exercised.

The only "just RO" user today is one of the BPF allocations. I don't have a
strong objection to combining them, just explaining the thinking. I guess if
we
could always add another flag later if it becomes more needed.

quoted

If you document that VM_IMMEDIATE_UNMAP means "I want the TLB entries
gone", then I can re-review the code in light of that.  But then I'm
unconvinced by your generic implementation, since set_memory_nx()
seems like an odd way to go about it.

Masami Hiramatsu pointed out if we don't do set_memory_nx before
set_memory_rw,
then there will be a small window of W^X violation. So that was the concern
for
the executable case, regardless of the semantics. I think the concern
applies
for any "special capability" permissions. Alternatively, if we
remove_vm_area
before we reset the direct map perms RW, maybe that would accomplish the
same
thing, if that's possible in a cross arch way. Maybe this is too much
designing
for hypothetical future... just was trying to avoid having to change the
interface, and could just update the generic implementation if new
permissions
or usages come up.

The set_memory_ stuff is really only needed for arm64 which seems to be the
only
other one with directmap permissions. So if it could eventually have its own
arch_vunmap then all of the set_memory_ parts could be dropped and the
default
would just be the simple unmap then flush logic that it was originally.

I think that's probably the best solution.  If there are only two
arches that have anything fancy here, let's just fix both of them up
for real.

quoted

Or we have up to three flushes for the generic version and meet the name
expectations and needed functionality today. I guess I'll just try that.

quoted

quoted

quoted

quoted

+        */
+       if (immediate && !IS_ENABLED(ARCH_HAS_SET_MEMORY))
+               vm_unmap_aliases();
+}
+
 static void __vunmap(const void *addr, int deallocate_pages)
 {
        struct vm_struct *area;

@@ -1515,7 +1567,8 @@ static void __vunmap(const void *addr, int

deallocate_pages)
        debug_check_no_locks_freed(area->addr,
get_vm_area_size(area));
        debug_check_no_obj_freed(area->addr,
get_vm_area_size(area));

-       remove_vm_area(addr);
+       arch_vunmap(area, deallocate_pages);
+
        if (deallocate_pages) {
                int i;

@@ -1542,8 +1595,15 @@ static inline void __vfree_deferred(const

void
*addr)
         * nother cpu's list.  schedule_work() should be fine with
this
too.
         */
        struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
+       struct vfree_work *w = kmalloc(sizeof(struct vfree_work),
GFP_ATOMIC);
+
+       /* If no memory for the deferred list node, give up */
+       if (!w)
+               return;

That's nasty.  I see what you're trying to do here, but I think you're
solving a problem that doesn't need solving quite so urgently.  How
about dropping this part and replacing it with a comment like "NB:
this writes a word to a potentially executable address.  It would be
nice if we could avoid doing this."  And maybe a future patch could
more robustly avoid it without risking memory leaks.

Yea, sorry I should have called this out, because I wasn't sure on how
likely
that was to happen. I did find some other places in the kernel with the
same
ignoring logic.

I'll have to think though, I am not sure what the alternative is. Since
the
memory can be RO in the module_memfree case, the old method of re-using
the
allocation will no longer work. The list node could be stuffed on the
vm_struct,
but then the all of the spin_lock(&vmap_area_lock)'s need to be changed
to
work
with interrupts so that the struct could be looked up. Not sure of the
implications of that. Or maybe have some slow backup that resets the
permissions
and re-uses the allocation if kmalloc fails?

I guess it could also go back to the old v1 implementation that doesn't
handle
RO and the directmap, and leave the W^X violation window during
teardown.
Then
solve that problem when modules are loaded via something like Nadav's
stuff.

Hmm.  Switching to spin_lock_irqsave() doesn't seem so bad to me.

Ok.

Actually, I think I have a better solution.  Just declare the
problematic case to be illegal: say that you may not free memory with
the new flags set while IRQs are off.  Enforce this with a VM_WARN_ON
in the code that reads the vfree_deferred list.

Thanks. Yea just making a rule for the one case seems better that disabling
interrupts all over. It turned out to lock in quite a few places, including the
longish lazy purge operation. Reading a little history on the deferred free list
- 6 years ago vfree used to not support interrupts at all, and different clients
had their own work queues. So this will just be having the original situation
for the new vm flag.

I think we only need to move the module init section free from the RCU callback
to a work queue, to get to the point where, functionally wise, everything should
work with the existing deferred free list implementation (since then we will
just not use it for the new special memory case).

We could also just add a
WARN_ON(in_interrupt()) in module_memfree to give context to some callers to
what will soon be a "BUG: unable to handle kernel paging request". Any objection
to leaving it there?

Seems reasonable.

Over the above, moving the vfree_deferred list to the struct and the associated
cost of the lookup in every interrupt/atomic vfree would enable a WARN and the
handling of a (declared) illegal case that could deadlock anyway, right? Is it
worth it?

I suspect it's not worth it.

I'm not sure we why we wouldn't have deadlocks in normal interrupt vfrees if we
don't use irq spinlocks everywhere...I may be missing your insight.

I may be misunderstanding your question, but: I suspect that we can
easily deadlock if vfree() is called with IRQs off but
!in_interrupt().  Perhaps no one does that?  At the very least, I
assume that lockdep would scream loudly if this happened.

--Andy