The code is correct, but only for a rather subtle reason. This
confused me for quite a while when I read switch_mm, so clarify the
code to avoid confusing other people, too.
TBH, I wouldn't be surprised if this code was only correct by
accident.
Signed-off-by: Andy Lutomirski <luto@amacapital.net>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Kees Cook <keescook@chromium.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Vince Weaver <vince@deater.net>
Cc: "hillf.zj" <hillf.zj@alibaba-inc.com>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/0db86397f968996fb772c443c251415b0b430ddd.1414190806.git.luto@amacapital.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The 3.19 merge window saw some TLB modifications merged which caused a
performance regression. They were fixed in commit 045bbb9fa.
Once that fix was applied, I also noticed that there was a small
but intermittent regression still present. It was not present
consistently enough to bisect reliably, but I'm fairly confident
that it came from (my own) MPX patches. The source was reading
a relatively unused field in the mm_struct via arch_unmap.
I also noted that this code was in the main instruction flow of
do_munmap() and probably had more icache impact than we want.
This patch does two things:
1. Adds a static (via Kconfig) and dynamic (via cpuid) check
for MPX with cpu_feature_enabled(). This keeps us from
reading that cacheline in the mm and trades it for a check
of the global CPUID variables at least on CPUs without MPX.
2. Adds an unlikely() to ensure that the MPX call ends up out
of the main instruction flow in do_munmap(). I've added
a detailed comment about why this was done and why we want
it even on systems where MPX is present.
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Cc: luto@amacapital.net
Cc: Dave Hansen <dave@sr71.net>
Link: http://lkml.kernel.org/r/20150108223021.AEEAB987@viggo.jf.intel.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
asm-generic/mm_hooks.h provides some generic fillers for the 90%
of architectures that do not need to hook some mmap-manipulation
functions. A comment inside says:
> Define generic no-op hooks for arch_dup_mmap and
> arch_exit_mmap, to be included in asm-FOO/mmu_context.h
> for any arch FOO which doesn't need to hook these.
So, does x86 need to hook these? It depends on CONFIG_PARAVIRT.
We *conditionally* include this generic header if we have
CONFIG_PARAVIRT=n. That's madness.
With this patch, x86 stops using asm-generic/mmu_hooks.h entirely.
We use our own copies of the functions. The paravirt code
provides some stubs if it is disabled, and we always call those
stubs in our x86-private versions of arch_exit_mmap() and
arch_dup_mmap().
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Dave Hansen <dave@sr71.net>
Cc: x86@kernel.org
Link: http://lkml.kernel.org/r/20141118182349.14567FA5@viggo.jf.intel.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The previous patch allocates bounds tables on-demand. As noted in
an earlier description, these can add up to *HUGE* amounts of
memory. This has caused OOMs in practice when running tests.
This patch adds support for freeing bounds tables when they are no
longer in use.
There are two types of mappings in play when unmapping tables:
1. The mapping with the actual data, which userspace is
munmap()ing or brk()ing away, etc...
2. The mapping for the bounds table *backing* the data
(is tagged with VM_MPX, see the patch "add MPX specific
mmap interface").
If userspace use the prctl() indroduced earlier in this patchset
to enable the management of bounds tables in kernel, when it
unmaps the first type of mapping with the actual data, the kernel
needs to free the mapping for the bounds table backing the data.
This patch hooks in at the very end of do_unmap() to do so.
We look at the addresses being unmapped and find the bounds
directory entries and tables which cover those addresses. If
an entire table is unused, we clear associated directory entry
and free the table.
Once we unmap the bounds table, we would have a bounds directory
entry pointing at empty address space. That address space might
now be allocated for some other (random) use, and the MPX
hardware might now try to walk it as if it were a bounds table.
That would be bad. So any unmapping of an enture bounds table
has to be accompanied by a corresponding write to the bounds
directory entry to invalidate it. That write to the bounds
directory can fault, which causes the following problem:
Since we are doing the freeing from munmap() (and other paths
like it), we hold mmap_sem for write. If we fault, the page
fault handler will attempt to acquire mmap_sem for read and
we will deadlock. To avoid the deadlock, we pagefault_disable()
when touching the bounds directory entry and use a
get_user_pages() to resolve the fault.
The unmapping of bounds tables happends under vm_munmap(). We
also (indirectly) call vm_munmap() to _do_ the unmapping of the
bounds tables. We avoid unbounded recursion by disallowing
freeing of bounds tables *for* bounds tables. This would not
occur normally, so should not have any practical impact. Being
strict about it here helps ensure that we do not have an
exploitable stack overflow.
Based-on-patch-by: Qiaowei Ren <qiaowei.ren@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Cc: linux-mm@kvack.org
Cc: linux-mips@linux-mips.org
Cc: Dave Hansen <dave@sr71.net>
Link: http://lkml.kernel.org/r/20141114151831.E4531C4A@viggo.jf.intel.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This is really the meat of the MPX patch set. If there is one patch to
review in the entire series, this is the one. There is a new ABI here
and this kernel code also interacts with userspace memory in a
relatively unusual manner. (small FAQ below).
Long Description:
This patch adds two prctl() commands to provide enable or disable the
management of bounds tables in kernel, including on-demand kernel
allocation (See the patch "on-demand kernel allocation of bounds tables")
and cleanup (See the patch "cleanup unused bound tables"). Applications
do not strictly need the kernel to manage bounds tables and we expect
some applications to use MPX without taking advantage of this kernel
support. This means the kernel can not simply infer whether an application
needs bounds table management from the MPX registers. The prctl() is an
explicit signal from userspace.
PR_MPX_ENABLE_MANAGEMENT is meant to be a signal from userspace to
require kernel's help in managing bounds tables.
PR_MPX_DISABLE_MANAGEMENT is the opposite, meaning that userspace don't
want kernel's help any more. With PR_MPX_DISABLE_MANAGEMENT, the kernel
won't allocate and free bounds tables even if the CPU supports MPX.
PR_MPX_ENABLE_MANAGEMENT will fetch the base address of the bounds
directory out of a userspace register (bndcfgu) and then cache it into
a new field (->bd_addr) in the 'mm_struct'. PR_MPX_DISABLE_MANAGEMENT
will set "bd_addr" to an invalid address. Using this scheme, we can
use "bd_addr" to determine whether the management of bounds tables in
kernel is enabled.
Also, the only way to access that bndcfgu register is via an xsaves,
which can be expensive. Caching "bd_addr" like this also helps reduce
the cost of those xsaves when doing table cleanup at munmap() time.
Unfortunately, we can not apply this optimization to #BR fault time
because we need an xsave to get the value of BNDSTATUS.
==== Why does the hardware even have these Bounds Tables? ====
MPX only has 4 hardware registers for storing bounds information.
If MPX-enabled code needs more than these 4 registers, it needs to
spill them somewhere. It has two special instructions for this
which allow the bounds to be moved between the bounds registers
and some new "bounds tables".
They are similar conceptually to a page fault and will be raised by
the MPX hardware during both bounds violations or when the tables
are not present. This patch handles those #BR exceptions for
not-present tables by carving the space out of the normal processes
address space (essentially calling the new mmap() interface indroduced
earlier in this patch set.) and then pointing the bounds-directory
over to it.
The tables *need* to be accessed and controlled by userspace because
the instructions for moving bounds in and out of them are extremely
frequent. They potentially happen every time a register pointing to
memory is dereferenced. Any direct kernel involvement (like a syscall)
to access the tables would obviously destroy performance.
==== Why not do this in userspace? ====
This patch is obviously doing this allocation in the kernel.
However, MPX does not strictly *require* anything in the kernel.
It can theoretically be done completely from userspace. Here are
a few ways this *could* be done. I don't think any of them are
practical in the real-world, but here they are.
Q: Can virtual space simply be reserved for the bounds tables so
that we never have to allocate them?
A: As noted earlier, these tables are *HUGE*. An X-GB virtual
area needs 4*X GB of virtual space, plus 2GB for the bounds
directory. If we were to preallocate them for the 128TB of
user virtual address space, we would need to reserve 512TB+2GB,
which is larger than the entire virtual address space today.
This means they can not be reserved ahead of time. Also, a
single process's pre-popualated bounds directory consumes 2GB
of virtual *AND* physical memory. IOW, it's completely
infeasible to prepopulate bounds directories.
Q: Can we preallocate bounds table space at the same time memory
is allocated which might contain pointers that might eventually
need bounds tables?
A: This would work if we could hook the site of each and every
memory allocation syscall. This can be done for small,
constrained applications. But, it isn't practical at a larger
scale since a given app has no way of controlling how all the
parts of the app might allocate memory (think libraries). The
kernel is really the only place to intercept these calls.
Q: Could a bounds fault be handed to userspace and the tables
allocated there in a signal handler instead of in the kernel?
A: (thanks to tglx) mmap() is not on the list of safe async
handler functions and even if mmap() would work it still
requires locking or nasty tricks to keep track of the
allocation state there.
Having ruled out all of the userspace-only approaches for managing
bounds tables that we could think of, we create them on demand in
the kernel.
Based-on-patch-by: Qiaowei Ren <qiaowei.ren@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Cc: linux-mm@kvack.org
Cc: linux-mips@linux-mips.org
Cc: Dave Hansen <dave@sr71.net>
Link: http://lkml.kernel.org/r/20141114151829.AD4310DE@viggo.jf.intel.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The code is correct, but only for a rather subtle reason. This
confused me for quite a while when I read switch_mm, so clarify
the code to avoid confusing other people, too.
TBH, I wouldn't be surprised if this code was only correct by
accident.
[ I wouldn't normally send a comment-only patch, but it took me a long
time to first figure out wtf was going on here, and then to figure
out why this wasn't exploitable by malicious code, and then to
figure out why this oddity had no user-visible effect at all. Let's
spare future readers the same confusion. ]
Signed-off-by: Andy Lutomirski <luto@amacapital.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/36275c99801a87d8dcf0502a41cf4e2ad81aae46.1412623954.git.luto@amacapital.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We don't have any good way to figure out what kinds of flushes
are being attempted. Right now, we can try to use the vm
counters, but those only tell us what we actually did with the
hardware (one-by-one vs full) and don't tell us what was actually
_requested_.
This allows us to select out "interesting" TLB flushes that we
might want to optimize (like the ranged ones) and ignore the ones
that we have very little control over (the ones at context
switch).
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Link: http://lkml.kernel.org/r/20140731154059.4C96CBA5@viggo.jf.intel.com
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Dick Fowles, Don Zickus and Joe Mario have been working on
improvements to perf, and noticed heavy cache line contention
on the mm_cpumask, running linpack on a 60 core / 120 thread
system.
The cause turned out to be unnecessary atomic accesses to the
mm_cpumask. When in lazy TLB mode, the CPU is only removed from
the mm_cpumask if there is a TLB flush event.
Most of the time, no such TLB flush happens, and the kernel
skips the TLB reload. It can also skip the atomic memory
set & test.
Here is a summary of Joe's test results:
* The __schedule function dropped from 24% of all program cycles down
to 5.5%.
* The cacheline contention/hotness for accesses to that bitmask went
from being the 1st/2nd hottest - down to the 84th hottest (0.3% of
all shared misses which is now quite cold)
* The average load latency for the bit-test-n-set instruction in
__schedule dropped from 10k-15k cycles down to an average of 600 cycles.
* The linpack program results improved from 133 GFlops to 144 GFlops.
Peak GFlops rose from 133 to 153.
Reported-by: Don Zickus <dzickus@redhat.com>
Reported-by: Joe Mario <jmario@redhat.com>
Tested-by: Joe Mario <jmario@redhat.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Paul Turner <pjt@google.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/20130731221421.616d3d20@annuminas.surriel.com
[ Made the comments consistent around the modified code. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Since percpu_xxx() serial functions are duplicated with this_cpu_xxx().
Removing percpu_xxx() definition and replacing them by this_cpu_xxx()
in code. There is no function change in this patch, just preparation for
later percpu_xxx serial function removing.
On x86 machine the this_cpu_xxx() serial functions are same as
__this_cpu_xxx() without no unnecessary premmpt enable/disable.
Thanks for Stephen Rothwell, he found and fixed a i386 build error in
the patch.
Also thanks for Andrew Morton, he kept updating the patchset in Linus'
tree.
Signed-off-by: Alex Shi <alex.shi@intel.com>
Acked-by: Christoph Lameter <cl@gentwo.org>
Acked-by: Tejun Heo <tj@kernel.org>
Acked-by: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
This allows us to move duplicated code in <asm/atomic.h>
(atomic_inc_not_zero() for now) to <linux/atomic.h>
Signed-off-by: Arun Sharma <asharma@fb.com>
Reviewed-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: Mike Frysinger <vapier@gentoo.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Clearing the cpu in prev's mm_cpumask early will avoid the flush tlb
IPI's while the cr3 is still pointing to the prev mm. And this window
can lead to the possibility of bogus TLB fills resulting in strange
failures. One such problematic scenario is mentioned below.
T1. CPU-1 is context switching from mm1 to mm2 context and got a NMI
etc between the point of clearing the cpu from the mm_cpumask(mm1)
and before reloading the cr3 with the new mm2.
T2. CPU-2 is tearing down a specific vma for mm1 and will proceed with
flushing the TLB for mm1. It doesn't send the flush TLB to CPU-1
as it doesn't see that cpu listed in the mm_cpumask(mm1).
T3. After the TLB flush is complete, CPU-2 goes ahead and frees the
page-table pages associated with the removed vma mapping.
T4. CPU-2 now allocates those freed page-table pages for something
else.
T5. As the CR3 and TLB caches for mm1 is still active on CPU-1, CPU-1
can potentially speculate and walk through the page-table caches
and can insert new TLB entries. As the page-table pages are
already freed and being used on CPU-2, this page walk can
potentially insert a bogus global TLB entry depending on the
(random) contents of the page that is being used on CPU-2.
T6. This bogus TLB entry being global will be active across future CR3
changes and can result in weird memory corruption etc.
To avoid this issue, for the prev mm that is handing over the cpu to
another mm, clear the cpu from the mm_cpumask(prev) after the cr3 is
changed.
Marking it for -stable, though we haven't seen any reported failure that
can be attributed to this.
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: stable@kernel.org [v2.6.32+]
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Makes code futureproof against the impending change to mm->cpu_vm_mask (to be a pointer).
It's also a chance to use the new cpumask_ ops which take a pointer
(the older ones are deprecated, but there's no hurry for arch code).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Impact: pt_regs changed, lazy gs handling made optional, add slight
overhead to SAVE_ALL, simplifies error_code path a bit
On x86_32, %gs hasn't been used by kernel and handled lazily. pt_regs
doesn't have place for it and gs is saved/loaded only when necessary.
In preparation for stack protector support, this patch makes lazy %gs
handling optional by doing the followings.
* Add CONFIG_X86_32_LAZY_GS and place for gs in pt_regs.
* Save and restore %gs along with other registers in entry_32.S unless
LAZY_GS. Note that this unfortunately adds "pushl $0" on SAVE_ALL
even when LAZY_GS. However, it adds no overhead to common exit path
and simplifies entry path with error code.
* Define different user_gs accessors depending on LAZY_GS and add
lazy_save_gs() and lazy_load_gs() which are noop if !LAZY_GS. The
lazy_*_gs() ops are used to save, load and clear %gs lazily.
* Define ELF_CORE_COPY_KERNEL_REGS() which always read %gs directly.
xen and lguest changes need to be verified.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jeremy Fitzhardinge <jeremy@xensource.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: cleanup
On x86_32, %gs is handled lazily. It's not saved and restored on
kernel entry/exit but only when necessary which usually is during task
switch but there are few other places. Currently, it's done by
calling savesegment() and loadsegment() explicitly. Define
get_user_gs(), set_user_gs() and task_user_gs() and use them instead.
While at it, clean up register access macros in signal.c.
This cleans up code a bit and will help future changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: cleanup
tj: * changed cpu to unsigned as was done on mmu_context_64.h as cpu
id is officially unsigned int
* added missing ';' to 32bit version of deactivate_mm()
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Change header guards named "ASM_X86__*" to "_ASM_X86_*" since:
a. the double underscore is ugly and pointless.
b. no leading underscore violates namespace constraints.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Andy Lutomirski