Message-ID: <CALCETrWScgJpdnzNswJSKioQ93Oyw+Y_dJLoRxPX2Z=REVV1Ug@mail.gmail.com>
Date:   Fri, 21 Dec 2018 09:12:23 -0800
From:   Andy Lutomirski <luto@...nel.org>
To:     Ard Biesheuvel <ard.biesheuvel@...aro.org>
Cc:     Andy Lutomirski <luto@...nel.org>,
        Rick Edgecombe <rick.p.edgecombe@...el.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Will Deacon <will.deacon@....com>,
        Linux-MM <linux-mm@...ck.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Kernel Hardening <kernel-hardening@...ts.openwall.com>,
        "Naveen N . Rao" <naveen.n.rao@...ux.vnet.ibm.com>,
        Anil S Keshavamurthy <anil.s.keshavamurthy@...el.com>,
        "David S. Miller" <davem@...emloft.net>,
        Masami Hiramatsu <mhiramat@...nel.org>,
        Steven Rostedt <rostedt@...dmis.org>,
        Ingo Molnar <mingo@...hat.com>,
        Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Jessica Yu <jeyu@...nel.org>, Nadav Amit <namit@...are.com>,
        Network Development <netdev@...r.kernel.org>,
        Jann Horn <jannh@...gle.com>,
        Kristen Carlson Accardi <kristen@...ux.intel.com>,
        Dave Hansen <dave.hansen@...el.com>,
        "Dock, Deneen T" <deneen.t.dock@...el.com>
Subject: Re: [PATCH v2 1/4] vmalloc: New flags for safe vfree on special perms

> On Dec 21, 2018, at 9:39 AM, Ard Biesheuvel <ard.biesheuvel@...aro.org> wrote:
>
>> On Wed, 12 Dec 2018 at 03:20, Andy Lutomirski <luto@...nel.org> wrote:
>>
>> On Tue, Dec 11, 2018 at 4:12 PM Rick Edgecombe
>> <rick.p.edgecombe@...el.com> wrote:
>>>
>>> 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 <dave.hansen@...el.com>
>>> Suggested-by: Andy Lutomirski <luto@...nel.org>
>>> Suggested-by: Will Deacon <will.deacon@....com>
>>> Signed-off-by: Rick Edgecombe <rick.p.edgecombe@...el.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
>>
>>
>>> +{
>>> +       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_IMMEDIATE_UNMAP is a bit of a lie, since, if in_interrupt(), it's
>> not immediate.)
>>
>> 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. */
>
> So can't we fix this first? Assuming that architectures that bother to
> implement them will not have executable mappings in the linear region,
> all we'd need is set_linear_range_ro/rw() routines that default to
> doing nothing, and encapsulate the existing code for x86 and arm64.
> That way, we can handle do things in the proper order, i.e., release
> the vmalloc mapping (without caring about the permissions), restore
> the linear alias attributes, and finally release the pages.

Seems reasonable, except that I think it should be
set_linear_range_not_present() and set_linear_range_rw(), for three
reasons:

1. It’s not at all clear to me that we need to keep the linear mapping
around for modules.

2. At least on x86, the obvious algorithm to do the free operation
with a single flush requires it.  Someone should probably confirm that
arm’s TLB works the same way, i.e. that no flush is needed when
changing from not-present (or whatever ARM calls it) to RW.

3. Anyone playing with XPFO wants this facility anyway.  In fact, with
this change, Rick’s series will more or less implement XPFO for
vmalloc memory :)

Does that seem reasonable to you?

Powered by blists - more mailing lists