Impact: fix sporadic slowdowns and warning messages
This patch fixes a performance issue reported by Linus on his
Nehalem system. While Linus reverted the PAT patch (commit
58dab916df) which exposed the issue,
existing cpa() code can potentially still cause wrong(page attribute
corruption) behavior.
This patch also fixes the "WARNING: at arch/x86/mm/pageattr.c:560" that
various people reported.
In 64bit kernel, kernel identity mapping might have holes depending
on the available memory and how e820 reports the address range
covering the RAM, ACPI, PCI reserved regions. If there is a 2MB/1GB hole
in the address range that is not listed by e820 entries, kernel identity
mapping will have a corresponding hole in its 1-1 identity mapping.
If cpa() happens on the kernel identity mapping which falls into these holes,
existing code fails like this:
__change_page_attr_set_clr()
__change_page_attr()
returns 0 because of if (!kpte). But doesn't
set cpa->numpages and cpa->pfn.
cpa_process_alias()
uses uninitialized cpa->pfn (random value)
which can potentially lead to changing the page
attribute of kernel text/data, kernel identity
mapping of RAM pages etc. oops!
This bug was easily exposed by another PAT patch which was doing
cpa() more often on kernel identity mapping holes (physical range between
max_low_pfn_mapped and 4GB), where in here it was setting the
cache disable attribute(PCD) for kernel identity mappings aswell.
Fix cpa() to handle the kernel identity mapping holes. Retain
the WARN() for cpa() calls to other not present address ranges
(kernel-text/data, ioremap() addresses)
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: right-align /proc/meminfo consistent with other fields
When the split-LRU patches added Inactive(anon) and Inactive(file) lines
to /proc/meminfo, all counts were moved two columns rightwards to fit in.
Now move x86's DirectMap lines two columns rightwards to line up.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Do a global flush tlb after splitting the large page and before we do the
actual change page attribute in the PTE.
With out this, we violate the TLB application note, which says
"The TLBs may contain both ordinary and large-page translations for
a 4-KByte range of linear addresses. This may occur if software
modifies the paging structures so that the page size used for the
address range changes. If the two translations differ with respect
to page frame or attributes (e.g., permissions), processor behavior
is undefined and may be implementation-specific."
And also serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity
mappings) using cpa_lock. So that we don't allow any other cpu, with stale
large tlb entries change the page attribute in parallel to some other cpu
splitting a large page entry along with changing the attribute.
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: arjan@linux.intel.com
Cc: venkatesh.pallipadi@intel.com
Cc: jeremy@goop.org
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Interrupt context no longer splits large page in cpa(). So we can do away
with cpa memory pool code.
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: arjan@linux.intel.com
Cc: venkatesh.pallipadi@intel.com
Cc: jeremy@goop.org
Signed-off-by: Ingo Molnar <mingo@elte.hu>
No alias checking needed for setting present/not-present mapping. Otherwise,
we may need to break large pages for 64-bit kernel text mappings (this adds to
complexity if we want to do this from atomic context especially, for ex:
with CONFIG_DEBUG_PAGEALLOC). Let's keep it simple!
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: arjan@linux.intel.com
Cc: venkatesh.pallipadi@intel.com
Cc: jeremy@goop.org
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Export set_memory_ro() and set_memory_rw() calls for use by drivers that need
to have more debug information about who might be writing to memory space.
this was initially developed for use while debugging a memory corruption
problem with e1000e.
Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Fix the start addr for free_memtype calls in the error path.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Acked-by: Rene Herman <rene.herman@keyaccess.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Actually, might as well simply reconstruct the memtype list at free time
I guess. How is this for a coalescing version of the array functions?
Compiles, boots and provides me with:
root@7ixe4:~# wc -l /debug/x86/pat_memtype_list
53 /debug/x86/pat_memtype_list
otherwise (down from 16384+).
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The new set_memory_array_{uc,wb}() pass virtual addresses to
{reserve,free}_memtype() it seems.
Signed-off-by: Rene Herman <rene.herman@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add array interface APIs of pageattr. page based cache flush is quite
slow for a lot of pages. If pages are more than 1024 (4M), the patch
will use a wbinvd(). We have a simple test here (run a 3d game - open
arena), nearly all agp memory allocation are small (< 1M), so suppose
this will not impact runtime performance.
Signed-off-by: Dave Airlie <airlied@gmail.com>
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Big thinko in pat memtype tracking code. reserve_memtype should be called
with physical address and not virtual address.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Highmem code can leave ptes and tlb entries around for a given page even after
kunmap, and after it has been freed.
>From what I can gather, the PAT code may change the cache attributes of
arbitrary physical addresses (ie. including highmem pages), which would result
in aliases in the case that it operates on one of these lazy tlb highmem
pages.
Flushing kmaps should solve the problem.
I've also just added code for conditional flushing if we haven't got
any dangling highmem aliases -- this should help performance if we
change page attributes frequently or systems that aren't using much
highmem pages (eg. if < 4G RAM). Should be turned into 2 patches, but
just for RFC...
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Introduce two APIs for page attribute. flushing tlb/cache in every page
attribute is expensive. AGP gart usually will do a lot of operations to
change a page to uc, new APIs can reduce flush.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Cc: airlied@linux.ie
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Arjan van de Ven <arjan@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Do we actually want these DirectMap lines in the x86 /proc/meminfo?
I can see they're interesting to CPA developers and TLB optimizers,
but they don't fit its usual "where has all my memory gone?" usage.
If they are to stay, here are some fixes.
1. On x86_32 without PAE, they're not 2M but 4M pages: no need to
mess with the internal enum, but show the right name to users.
2. Many machines can never show anything but 0 for DirectMap1G,
so suppress that line unless direct_gbpages are really enabled.
3. The unit in /proc/meminfo is kB not number of pages: HugePages
messed that up, but they're an example to regret not to follow.
4. Once we use kB, it's easy to see that 1GB has gone missing (which
explains why CONFIG_CPA_DEBUG=y soon wraps DirectMap2M negative):
because head_64.S's level2_ident_pgt entries were not counted.
My fix is not ideal, but works for more and for less than 1G,
and avoids interfering with early bootup pagetable contortions.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Use WARN() instead of a printk+WARN_ON() pair; this way the message becomes
part of the warning section for better reporting/collection.
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: akpm@linux-foundation.org
Cc: arjan@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
tighten the boundary checks around max_low_pfn_mapped - dont overmap
nor undermap into holes.
also print out tseg for AMD cpus, for diagnostic purposes.
(this is an SMM area, and we split up any big mappings around that area)
Signed-off-by: Yinghai Lu <yhlu.kernel@gmail.com>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
when more than 4g memory is installed, don't map the big hole below 4g.
Signed-off-by: Yinghai Lu <yhlu.kernel@gmail.com>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
1) Remove __meminit from update_pages_count. It is used inside
split_pages()
2) Make the code depend on PROC_FS. Doing statistics for nothing is
useless and not adding useless code is nice to the Linux tiny folks.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add information about the mapping state of the direct mapping to
/proc/meminfo. I chose /proc/meminfo because that is where all the other
memory statistics are too and it is a generally useful metric even
outside debugging situations. A lot of split kernel pages means the
kernel will run slower.
This way we can see how many large pages are really used for it and how
many are split.
Useful for general insight into the kernel.
v2: Add hotplug locking to 64bit to plug a very obscure theoretical race.
32bit doesn't need it because it doesn't support hotadd for lowmem.
Fix some typos
v3: Rename dpages_cnt
Add CONFIG ifdef for count update as requested by tglx
Expand description
v4: Fix stupid bugs added in v3
Move update_page_count to pageattr.c
Signed-off-by: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
It's not even passed on to smp_call_function() anymore, since that
was removed. So kill it.
Acked-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
BTW, what does pat_wc_enabled stand for? Does it mean
"write-combining"?
Currently it is used to globally switch on or off PAT support.
Thus I renamed it to pat_enabled.
I think this increases readability (and hope that I didn't miss
something).
Signed-off-by: Andreas Herrmann <andreas.herrmann3@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Use lookup_address() from pageattr.c instead of doing the same
manually. Also had to EXPORT_SYMBOL_GPL(lookup_address) to make this
work for modules. This also fixes "undefined symbol 'init_mm'"
compile error for x86_32.
Signed-off-by: Pekka Paalanen <pq@iki.fi>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Use UC_MINUS for ioremap(), ioremap_nocache() instead of strong UC.
Once all the X drivers move to ioremap_wc(), we can go back to strong
UC semantics for ioremap() and ioremap_nocache().
To avoid attribute aliasing issues, pci_mmap_page_range() will also
use UC_MINUS for default non write-combining mapping request.
Next steps:
a) change all the video drivers using ioremap() or ioremap_nocache()
and adding WC MTTR using mttr_add() to ioremap_wc()
b) for strict usage, we can go back to strong uc semantics
for ioremap() and ioremap_nocache() after some grace period for
completing step-a.
c) user level X server needs to use the appropriate method for setting
up WC mapping (like using resourceX_wc sysfs file instead of
adding MTRR for WC and using /dev/mem or resourceX under /sys)
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/x86/linux-2.6-xen-next: (52 commits)
xen: add balloon driver
xen: allow compilation with non-flat memory
xen: fold xen_sysexit into xen_iret
xen: allow set_pte_at on init_mm to be lockless
xen: disable preemption during tlb flush
xen pvfb: Para-virtual framebuffer, keyboard and pointer driver
xen: Add compatibility aliases for frontend drivers
xen: Module autoprobing support for frontend drivers
xen blkfront: Delay wait for block devices until after the disk is added
xen/blkfront: use bdget_disk
xen: Make xen-blkfront write its protocol ABI to xenstore
xen: import arch generic part of xencomm
xen: make grant table arch portable
xen: replace callers of alloc_vm_area()/free_vm_area() with xen_ prefixed one
xen: make include/xen/page.h portable moving those definitions under asm dir
xen: add resend_irq_on_evtchn() definition into events.c
Xen: make events.c portable for ia64/xen support
xen: move events.c to drivers/xen for IA64/Xen support
xen: move features.c from arch/x86/xen/features.c to drivers/xen
xen: add missing definitions in include/xen/interface/vcpu.h which ia64/xen needs
...
Rename (alloc|release)_(pt|pd) to pte/pmd to explicitly match the name
of the appropriate pagetable level structure.
[ x86.git merge work by Mark McLoughlin <markmc@redhat.com> ]
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
do not return a -EINVAL when mmap()-ing PCI holes.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Acked-by: Suresh Siddha <suresh.b.siddha@intel.com>
Acked-by: H. Peter Anvin <hpa@zytor.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Arjan van de Ven <arjan@linux.intel.com>
Add a new function to force split large pages into 4k pages.
This is needed for some followup optimizations.
I had to add a new field to cpa_data to pass down the information
that try_preserve_large_page should not run.
Right now no set_page_4k() because I didn't need it and all the
specialized users I have in mind would be more comfortable with
pure addresses. I also didn't export it because it's unlikely
external code needs it.
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: andreas.herrmann3@amd.com
Cc: mingo@elte.hu
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add a set_memory_wc interface(), similar to set_memory_uc interface.
Callers has to call set_memory_uc, set_memory_wb and
set_memory_wc, set_memory_wb as pairs.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Use reserve_memtype and free_memtype interfaces in set_memory_uc/set_memory_wb
interfaces to avoid aliasing.
Usage model of set_memory_uc and set_memory_wb is for RAM memory and users
will first call set_memory_uc and call set_memory_wb after use to reset the
attribute.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Sets up pat_init() infrastructure.
PAT MSR has following setting.
PAT
|PCD
||PWT
|||
000 WB _PAGE_CACHE_WB
001 WC _PAGE_CACHE_WC
010 UC- _PAGE_CACHE_UC_MINUS
011 UC _PAGE_CACHE_UC
We are effectively changing WT from boot time setting to WC.
UC_MINUS is used to provide backward compatibility to existing /dev/mem
users(X).
reserve_memtype and free_memtype are new interfaces for maintaining alias-free
mapping. It is currently implemented in a simple way with a linked list and
not optimized. reserve and free tracks the effective memory type, as a result
of PAT and MTRR setting rather than what is actually requested in PAT.
pat_init piggy backs on mtrr_init as the rules for setting both pat and mtrr
are same.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add debug information for DEBUG_PAGEALLOC to get some statistics about
the pool usage and split status.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
fix the 3D performance drop reported at:
http://bugzilla.kernel.org/show_bug.cgi?id=10328
fb drivers are using ioremap()/ioremap_nocache(), followed by mtrr_add with
WC attribute. Recent changes in page attribute code made both
ioremap()/ioremap_nocache() mappings as UC (instead of previous UC-). This
breaks the graphics performance, as the effective memory type is UC instead
of expected WC.
The correct way to fix this is to add ioremap_wc() (which uses UC- in the
absence of PAT kernel support and WC with PAT) and change all the
fb drivers to use this new ioremap_wc() API.
We can take this correct and longer route for post 2.6.25. For now,
revert back to the UC- behavior for ioremap/ioremap_nocache.
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Revert:
commit 8be8f54bae
Author: Thomas Gleixner <tglx@linutronix.de>
Date: Sat Feb 23 20:43:21 2008 +0100
x86: CPA: avoid split of alias mappings
because it clearly mishandles the case when __change_page_attr(), called
from __change_page_attr_set_clr(), changes cpa->processed to 1 and
cpa_process_alias(cpa) is executed right after that.
This crashes my x86-64 test box early in the boot process
(ref. http://bugzilla.kernel.org/show_bug.cgi?id=10140#c4).
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
avoid over-eager large page splitup.
When the target area needs to be split or is split already (ioremap)
then the current code enforces the split of large mappings in the alias
regions even if we could avoid it.
Use a separate variable processed in the cpa_data structure to carry
the number of pages which have been processed instead of reusing the
numpages variable. This keeps numpages intact and gives the alias code
a chance to keep large mappings intact.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Use PF_MEMALLOC to prevent recursive calls in the DBEUG_PAGEALLOC
case. This makes the code simpler and more robust against allocation
failures.
This fixes the following fallback to non-mmconfig:
http://lkml.org/lkml/2008/2/20/551http://bugzilla.kernel.org/show_bug.cgi?id=10083
Also, for DEBUG_PAGEALLOC=n reduce the pool size to one page.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Make hibernation work with CONFIG_DEBUG_PAGEALLOC set on x86, by
checking if the pages to be copied are marked as present in the
kernel mapping and temporarily marking them as present if that's not
the case. No functional modifications are introduced if
CONFIG_DEBUG_PAGEALLOC is unset.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
New implementation does not use lru for anything so there is no need
to reject pages that are in the LRU. Similar for compound pages (which
were checked because they also use page->lru)
[ tglx@linutronix.de: removed unused variable ]
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
When the CPA code is called with an virtual address in the range of
the direct mapping or the high alias then we do not need to run
through the alias check for this range.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
NX settings are not required to be consistent across alias mappings.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
c_p_a() did not discover all aliases correctly. (such as when called
on vmalloc()-ed areas or ioremap()-ed areas)
Push the alias checks to the lower, physical level and consistently
discover all aliases that might exist: the low direct mappings and
the high linear kernel-text mappings (on 64-bit).
Thanks to Andi Kleen for pointing out that this was buggy.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
the cpa API is page aligned - warn about any weird alignments.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
[ mingo@elte.hu: while gbpages cannot be enabled on mainline currently,
keep the code uptodate and this fix is easy enough. ]
Use correct page sizes and masks for GB pages in try_preserve_large_page()
This prevents a boot hang on a GB capable system with CONFIG_DIRECT_GBPAGES
enabled.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Right now, we check only the first 4k page for static required protections.
This does not take overlapping regions into account. So we might end up
setting the wrong permissions/protections for other parts of this large page.
This can be optimized further, but correctness is the important part.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Switch the split page code to use the page pool. We do this
unconditionally to avoid different behaviour with and without
DEBUG_PAGEALLOC enabled.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
DEBUG_PAGEALLOC was not possible on 64-bit due to its early-bootup
hardcoded reliance on PSE pages, and the unrobustness of the runtime
splitup of large pages. The splitup ended in recursive calls to
alloc_pages() when a page for a pte split was requested.
Avoid the recursion with a preallocated page pool, which is used to
split up large mappings and gets refilled in the return path of
kernel_map_pages after the split has been done. The size of the page
pool is adjusted to the available memory.
This part just implements the page pool and the initialization w/o
using it yet.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Adjust the definition of lookup_address to take an unsigned long
level argument. Adjust callers in xen/mmu.c that pass in a
dummy variable.
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The .rodata section really should just be read only; the config option
is there to make breaking up the 2Mb page an option (so people whos machines
give more performance for the 2Mb case can opt to do so).
But when the page gets split anyway, this is no longer an issue, so
clean up the code and remove the ifdefs
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
When change_page_attr splits a large page on x86_32 (without PAE), it is
currently corrupting every process's page directory: fix that by removing
the thinko which passes down a physical instead of a virtual address.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Move the readout of the large entry into the spinlock section to
prevent an unlikely but possible race.
Mark the pmd/pud entry present after the split. We preserved the
non present bit in the new split mapping.
Remove the stale gfp_flags double initialization.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
lookup_address() returns a wrong level and a wrong pointer to a non
existing pte, when pmd or pud entries are marked !present. This
happens for example due to boot time mapping of GART into the low
memory space.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
An Athlon 64 X2 test system showed hard hangs shortly after marking
the kernel text read-only, if we tried to preserve largepages and
changed the PSE entry from RW to RO. The pagetable code itself is
correct, it's the CPU that locked up hard (and not even the NMI
watchdog could punch through that hard hang).
So be conservative and always do splitups - like we did in the past.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
When CPA is called on a range which fits into a large page mapping,
avoid to split the page when:
1) There is no change of attributes
2) The range to change is a complete large mapping
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The number of arguments which need to be transported is increasing
and we want to add flush optimizations and large page preserving.
Create struct cpa data and pass a pointer instead of increasing the
number of arguments further.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We only need to flush the caches in cpa() if the the caching attributes
have changed. Otherwise only flush the TLBs.
This checks the PAT bits too although they are currently not used by
the kernel.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Mask out the not supported bits (e.g. NX). If the clr/set masks
are empty after the mask return without changing anything.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
When splitting large pages, we ge the pfn from the existing entry
instead of calculating it ourself.
This removes the last remaining range restriction of the cpa code.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
When changing the attributes of a pte, we should use the PFN from the
existing PTE rather than going through hoops calculating what we think
it might have been; this is both fragile and totally unneeded. It also
makes it more hairy to call any of these functions on non-direct maps
for no good reason whatsover.
With this change, __change_page_attr() no longer takes a pfn as argument,
which simplifies all the callers.
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@tglx.de>
Right now, enforcing that the high mapping of the kernel text doesn't
get the NX bit is done deep in the guts of CPA, rather than in the
static_protection() function that enforces all other per-arch sanity
checks.
This patch moves this sanity check into the central static_protection()
function instead, and makes it apply ONLY to the kernel text, not to all
other areas in the high mapping.
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
In split_large_page we clear the NX bit for the new split ptes, but we
need to preserve the original setting of it for the split ptes.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch replaces __change_page_attr_set_clr() with
change_page_attr_set_clr() in change_page_attr_clear() to flush the
TLB/cache properly.
Signed-off-by: Huang Ying <ying.huang@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Use a standard list threaded through page->lru for maintaining the pgd
list on PAE. This is the same as 64-bit, and seems saner than using a
non-standard list via page->index.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cleanup the address calculations, which are necessary to identify the
high/low alias mappings of the kernel on 64 bit machines. Instead of
calling __pa/__va back and forth, calculate the physical address once
and base the other calculations on it. Add understandable constants so
we can use the already available within() helper. Also add comments,
which help mere mortals to understand what this code does.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
clflush is sufficient to be issued on one CPU. The invalidation is
broadcast throughout the coherence domain.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
clflush is an unordered operation with respect to other memory
traffic, including other CLFLUSH instructions. This needs proper
fencing with mfence.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The function name global_flush_tlb() suggests something different from
what the function really does. Rename it to cpa_flush_all(), which is an
understandable counterpart to cpa_flush_range().
no global visibility of the old API anymore.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Use clflush on CPUs which support this.
clflush is only used when the page attribute operation has been
successful. On CPUs which do not support clflush and in the case of
error the old fashioned global_flush_tlb() is called.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Convert cpa_set and cpa_clear to call the new set_clr function.
Seperate out the debug helpers.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Create a set_and_clr function to avoid the duplicate loops. Allows
also to do combined operations for optimization.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
To avoid the modification of the flush code for the clflush
implementation, move the flush into the set and clear functions and
provide helper functions for the debugging code.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Suresh Siddha