Thanks Boris.
Ard, Will: An arm question came up below. Any thoughts?
On Tue, 2019-02-19 at 13:48 +0100, Borislav Petkov wrote:
> On Mon, Jan 28, 2019 at 04:34:17PM -0800, Rick Edgecombe wrote:
> > This adds a new flags VM_HAS_SPECIAL_PERMS, for enabling vfree operations
>
> s/This adds/add/ - you get the idea. :)
Yes, thanks. Fixed after your comments on other patches.
> s/flags/flag/
>
> > to immediately clear executable TLB entries to freed pages, and handle
> > freeing memory with special permissions. It also takes care of resetting
> > the direct map permissions for the pages being unmapped. So this flag is
> > useful for any kind of memory with elevated permissions, or where there can
> > be related permissions changes on the directmap. Today this is RO+X and RO
> > memory.
> >
> > Although this enables directly vfreeing RO memory now, RO memory cannot be
> > freed in an interrupt because the allocation itself is used as a node on
> > deferred free list. So when RO memory needs to be freed in an interrupt
> > the code doing the vfree needs to have its own work queue, as was the case
> > before the deferred vfree list handling was added. Today there is only one
> > case where this happens.
> >
> > For architectures with set_alias_ implementations this whole operation
> > can be done with one TLB flush when centralized like this. For others with
> > directmap permissions, currently only arm64, a backup method using
> > set_memory functions is used to reset the directmap. When arm64 adds
> > set_alias_ functions, this backup can be removed.
> >
> > When the TLB is flushed to both remove TLB entries for the vmalloc range
> > mapping and the direct map permissions, the lazy purge operation could be
> > done to try to save a TLB flush later. However today vm_unmap_aliases
> > could flush a TLB range that does not include the directmap. So a helper
> > is added with extra parameters that can allow both the vmalloc address and
> > the direct mapping to be flushed during this operation. The behavior of the
> > normal vm_unmap_aliases function is unchanged.
> >
> > Suggested-by: Dave Hansen <dave.hansen@xxxxxxxxx>
> > Suggested-by: Andy Lutomirski <luto@xxxxxxxxxx>
> > Suggested-by: Will Deacon <will.deacon@xxxxxxx>
> > Signed-off-by: Rick Edgecombe <rick.p.edgecombe@xxxxxxxxx>
> > ---
> > include/linux/vmalloc.h | 13 +++++
> > mm/vmalloc.c | 122 +++++++++++++++++++++++++++++++++-------
> > 2 files changed, 116 insertions(+), 19 deletions(-)
> >
> > diff --git a/include/linux/vmalloc.h b/include/linux/vmalloc.h
> > index 398e9c95cd61..9f643f917360 100644
> > --- a/include/linux/vmalloc.h
> > +++ b/include/linux/vmalloc.h
> > @@ -21,6 +21,11 @@ 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 */
> > +/*
> > + * Memory with VM_HAS_SPECIAL_PERMS cannot be freed in an interrupt or with
> > + * vfree_atomic.
>
> Please end function names with parentheses. You should go over the whole
> patchset - there are a bunch of places.
Ok.
> > + */
> > +#define VM_HAS_SPECIAL_PERMS 0x00000200 /* Reset directmap and
> > flush TLB on unmap */
>
> After 0x00000080 comes 0x00000100. 0x00000010 is free too. What's up?
I was just trying to follow the pattern noticed from the gap at 0x00000010. I'll
add it at 0x00000100 in case there is some reason for the other gap.
> > /* bits [20..32] reserved for arch specific ioremap internals */
> >
> > /*
> > @@ -135,6 +140,14 @@ extern struct vm_struct *__get_vm_area_caller(unsigned
> > long size,
> > extern struct vm_struct *remove_vm_area(const void *addr);
> > extern struct vm_struct *find_vm_area(const void *addr);
> >
> > +static inline void set_vm_special(void *addr)
>
> You need a different name than "special" for a vm which needs to flush
> and clear mapping perms on removal. VM_RESET_PERMS or whatever is more
> to the point than "special", for example, which could mean a lot of
> things.
I don't have a strong opinion about the name, but it has gone through some
revisions, so I'll summarize the history.
There are two intentions - not leaving a stale TLB entry to a freed page, and
then resetting the direct map for architectures that need it. Andy had pointed
out that you can do this with only one TLB flush if all the cleanup is
centralized here. The original point of this flag was the TLB flush of the
vmalloc alias, and it's something that makes sense for all architectures that
have something like NX, regardless of what they are doing with their direct map.
At one point there were two flags, one for "immediate flush" and one for "reset
direct map", but Andy had also pointed out that you would mostly use them
together, so why complicate things.
So to capture both of those intentions, maybe I'll slightly tweak your
suggestion to VM_FLUSH_RESET_PERMS?
> > +{
> > + struct vm_struct *vm = find_vm_area(addr);
> > +
> > + if (vm)
> > + vm->flags |= VM_HAS_SPECIAL_PERMS;
> > +}
> > +
> > extern int map_vm_area(struct vm_struct *area, pgprot_t prot,
> > struct page **pages);
> > #ifdef CONFIG_MMU
> > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > index 871e41c55e23..d459b5b9649b 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>
> > @@ -1055,24 +1056,11 @@ static void vb_free(const void *addr, unsigned long
> > size)
> > spin_unlock(&vb->lock);
> > }
> >
> > -/**
> > - * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
> > - *
> > - * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
> > - * to amortize TLB flushing overheads. What this means is that any page you
> > - * have now, may, in a former life, have been mapped into kernel virtual
> > - * address by the vmap layer and so there might be some CPUs with TLB
> > entries
> > - * still referencing that page (additional to the regular 1:1 kernel
> > mapping).
> > - *
> > - * vm_unmap_aliases flushes all such lazy mappings. After it returns, we
> > can
> > - * be sure that none of the pages we have control over will have any
> > aliases
> > - * from the vmap layer.
> > - */
> > -void vm_unmap_aliases(void)
> > +static void _vm_unmap_aliases(unsigned long start, unsigned long end,
> > + int must_flush)
>
> Align arguments on the opening brace. There's more places below, pls fix
> them all.
Ok.
> > {
> > - unsigned long start = ULONG_MAX, end = 0;
> > int cpu;
> > - int flush = 0;
> > + int flush = must_flush;
>
> You can't use must_flush directly because...?
>
> gcc produces the same asm here, with or without the local "flush" var.
I had thought it was easier to read. If its not the case, I'll change it as you
suggest.
> >
> > if (unlikely(!vmap_initialized))
> > return;
> > @@ -1109,6 +1097,27 @@ void vm_unmap_aliases(void)
> > flush_tlb_kernel_range(start, end);
> > mutex_unlock(&vmap_purge_lock);
> > }
> > +
> > +/**
> > + * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
> > + *
> > + * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
> > + * to amortize TLB flushing overheads. What this means is that any page you
> > + * have now, may, in a former life, have been mapped into kernel virtual
> > + * address by the vmap layer and so there might be some CPUs with TLB
> > entries
> > + * still referencing that page (additional to the regular 1:1 kernel
> > mapping).
> > + *
> > + * vm_unmap_aliases flushes all such lazy mappings. After it returns, we
> > can
> > + * be sure that none of the pages we have control over will have any
> > aliases
> > + * from the vmap layer.
> > + */
> > +void vm_unmap_aliases(void)
> > +{
> > + unsigned long start = ULONG_MAX, end = 0;
> > + int must_flush = 0;
> > +
> > + _vm_unmap_aliases(start, end, must_flush);
> > +}
> > EXPORT_SYMBOL_GPL(vm_unmap_aliases);
> >
> > /**
> > @@ -1494,6 +1503,79 @@ struct vm_struct *remove_vm_area(const void *addr)
> > return NULL;
> > }
> >
> > +static inline void set_area_alias(const struct vm_struct *area,
> > + int (*set_alias)(struct page *page))
> > +{
> > + int i;
> > +
> > + for (i = 0; i < area->nr_pages; i++) {
> > + unsigned long addr =
> > + (unsigned long)page_address(area->pages[i]);
> > +
> > + if (addr)
> > + set_alias(area->pages[i]);
>
> What's wrong with simply:
>
> for (i = 0; i < area->nr_pages; i++) {
> if (page_address(area->pages[i]))
> set_alias(area->pages[i]);
> }
>
> ?
Yes, I see that's better.
I think at one point I had played with having the set_alias()'s take an unsigned
long address, and so the address was used twice and was worth assigning to
something, so this should be fixed. Thanks.
> > + }
> > +}
> > +
> > +/* This handles removing and resetting vm mappings related to the
> > vm_struct. */
>
> s/This handles/Handle/
>
> > +static void vm_remove_mappings(struct vm_struct *area, int
> > deallocate_pages)
> > +{
> > + unsigned long addr = (unsigned long)area->addr;
> > + unsigned long start = ULONG_MAX, end = 0;
> > + int special = area->flags & VM_HAS_SPECIAL_PERMS;
> > + int i;
> > +
> > + /*
> > + * The below block can be removed when all architectures that have
> > + * direct map permissions also have set_alias_ implementations. This is
> > + * to do resetting on the directmap for any special permissions (today
> > + * only X), without leaving a RW+X window.
> > + */
> > + if (special && !IS_ENABLED(CONFIG_ARCH_HAS_SET_ALIAS)) {
> > + set_memory_nx(addr, area->nr_pages);
> > + set_memory_rw(addr, area->nr_pages);
>
> That's two not very cheap calls to the underlying worker function, for
> example change_memory_common() on ARM64, instead of calling it once with
> the respective flags. You allude to that in the commit message but you
> might wanna run it by ARM folks first.
These calls are basically getting moved to be centralized here from other
places. In the later patches they get removed from where they were, so its a net
zero (except BPF on ARM I think).
Ard had expressed interest in having the set_alias_() functions for Arm, and the
names were chosen to be arch agnostic. He didn't explicitly commit but I was
under the impression he might create an implementation for ARM and we could
remove this block.
Ard, did I misinterpret that?
> > + }
> > +
> > + remove_vm_area(area->addr);
> > +
> > + /* If this is not special memory, we can skip the below. */
> > + if (!special)
> > + return;
> > +
> > + /*
> > + * If we are not deallocating pages, we can just do the flush of the VM
> > + * area and return.
> > + */
> > + if (!deallocate_pages) {
> > + vm_unmap_aliases();
> > + return;
> > + }
> > +
> > + /*
> > + * If we are here, we need to flush the vm mapping and reset the direct
> > + * map.
> > + * First find the start and end range of the direct mappings to make
> > + * sure the vm_unmap_aliases flush includes the direct map.
> > + */
> > + for (i = 0; i < area->nr_pages; i++) {
> > + unsigned long addr =
> > + (unsigned long)page_address(area->pages[i]);
> > + if (addr) {
>
> if (page_address(area->pages[i]))
>
> as above.
Right. Thanks.
> > + start = min(addr, start);
> > + end = max(addr, end);
> > + }
> > + }
> > +
> > + /*
> > + * First we set direct map to something not valid so that it won't be
>
> Above comment says "First" too. In general, all those "we" formulations
> do not make the comments as easy to read as when you make them
> impersonal and imperative:
>
> /*
> * Set the direct map to something invalid...
>
> Just like Documentation/process/submitting-patches.rst says:
>
> "Describe your changes in imperative mood, e.g. "make xyzzy do frotz"
> instead of "[This patch] makes xyzzy do frotz" or "[I] changed xyzzy
> to do frotz", as if you are giving orders to the codebase to change
> its behaviour."
>
> you simply order your code to do stuff. :-)
I see, I will make sure to fully apply this grammar rule to commit messages and
comments for the next version.
> > + * cached if there are any accesses after the TLB flush, then we flush
> > + * the TLB, and reset the directmap permissions to the default.
> > + */
> > + set_area_alias(area, set_alias_nv_noflush);
> > + _vm_unmap_aliases(start, end, 1);
> > + set_area_alias(area, set_alias_default_noflush);
> > +}
> > +
> > static void __vunmap(const void *addr, int deallocate_pages)
> > {
> > struct vm_struct *area;
>
>
