Commit af1b1a9b36 ("sparc64 mm: Fix base TSB sizing when hugetlb
pages are used") addressed the difference between hugetlb and THP
pages when computing TSB sizes. The following additional issues
were also discovered while working with the code.
In order to save memory, THP makes use of a huge zero page. This huge
zero page does not count against a task's RSS, but it does consume TSB
entries. This is similar to hugetlb pages. Therefore, count huge
zero page entries in hugetlb_pte_count.
Accounting of THP pages is done in the routine set_pmd_at().
Unfortunately, this does not catch the case where a THP page is split.
To handle this case, decrement the count in pmdp_invalidate().
pmdp_invalidate is only called when splitting a THP. However, 'sanity
checks' are added in case it is ever called for other purposes.
A more general issue exists with HPAGE_SIZE accounting.
hugetlb_pte_count tracks the number of HPAGE_SIZE (8M) pages. This
value is used to size the TSB for HPAGE_SIZE pages. However,
each HPAGE_SIZE page consists of two REAL_HPAGE_SIZE (4M) pages.
The TSB contains an entry for each REAL_HPAGE_SIZE page. Therefore,
the number of REAL_HPAGE_SIZE pages should be used to size the huge
page TSB. A new compile time constant REAL_HPAGE_PER_HPAGE is used
to multiply hugetlb_pte_count before sizing the TSB.
Changes from V1
- Fixed build issue if hugetlb or THP not configured
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
do_sparc64_fault() calculates both the base and huge page RSS sizes and
uses this information in calls to tsb_grow(). The calculation for base
page TSB size is not correct if the task uses hugetlb pages. hugetlb
pages are not accounted for in RSS, therefore the call to get_mm_rss(mm)
does not include hugetlb pages. However, the number of pages based on
huge_pte_count (which does include hugetlb pages) is subtracted from
this value. This will result in an artificially small and often negative
RSS calculation. The base TSB size is then often set to max_tsb_size
as the passed RSS is unsigned, so a negative value looks really big.
THP pages are also accounted for in huge_pte_count, and THP pages are
accounted for in RSS so the calculation in do_sparc64_fault() is correct
if a task only uses THP pages.
A single huge_pte_count is not sufficient for TSB sizing if both hugetlb
and THP pages can be used. Instead of a single counter, use two: one
for hugetlb and one for THP.
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
During hugepage map/unmap, TSB and TLB flushes are currently
issued at every PAGE_SIZE'd boundary which is unnecessary.
We now issue the flush at REAL_HPAGE_SIZE boundaries only.
Without this patch workloads which unmap a large hugepage
backed VMA region get CPU lockups due to excessive TLB
flush calls.
Orabug: 22365539, 22643230, 22995196
Signed-off-by: Nitin Gupta <nitin.m.gupta@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix following warning:
tsb.c:290:5: warning: symbol 'sysctl_tsb_ratio' was not declared. Should it be static?
Add extern declaration in asm/setup.h and remove local declaration
in kernel/sysctl.c
Signed-off-by: Sam Ravnborg <sam@ravnborg.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Access to the TSB hash tables during TLB misses requires that there be
an atomic 128-bit quad load available so that we fetch a matching TAG
and DATA field at the same time.
On cpus prior to UltraSPARC-III only virtual address based quad loads
are available. UltraSPARC-III and later provide physical address
based variants which are easier to use.
When we only have virtual address based quad loads available this
means that we have to lock the TSB into the TLB at a fixed virtual
address on each cpu when it runs that process. We can't just access
the PAGE_OFFSET based aliased mapping of these TSBs because we cannot
take a recursive TLB miss inside of the TLB miss handler without
risking running out of hardware trap levels (some trap combinations
can be deep, such as those generated by register window spill and fill
traps).
Without huge pages it's working perfectly fine, but when the huge TSB
got added another chunk of fixed virtual address space was not
allocated for this second TSB mapping.
So we were mapping both the 8K and 4MB TSBs to the same exact virtual
address, causing multiple TLB matches which gives undefined behavior.
Signed-off-by: David S. Miller <davem@davemloft.net>
This is a small patch which uses ARRAY_SIZE macro
rather than a number to make code readability better.
Signed-off-by: Doug Wilson <doug.lkml@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The impetus for this is that we would like to move to 64-bit PMDs and
PGDs, but that would result in only supporting a 42-bit address space
with the current page table layout. It'd be nice to support at least
43-bits.
The reason we'd end up with only 42-bits after making PMDs and PGDs
64-bit is that we only use half-page sized PTE tables in order to make
PMDs line up to 4MB, the hardware huge page size we use.
So what we do here is we make huge pages 8MB, and fabricate them using
4MB hw TLB entries.
Facilitate this by providing a "REAL_HPAGE_SHIFT" which is used in
places that really need to operate on hardware 4MB pages.
Use full pages (512 entries) for PTE tables, and adjust PMD_SHIFT,
PGD_SHIFT, and the build time CPP test as needed. Use a CPP test to
make sure REAL_HPAGE_SHIFT and the _PAGE_SZHUGE_* we use match up.
This makes the pgtable cache completely unused, so remove the code
managing it and the state used in mm_context_t. Now we have less
spinlocks taken in the page table allocation path.
The technique we use to fabricate the 8MB pages is to transfer bit 22
from the missing virtual address into the PTEs physical address field.
That takes care of the transparent huge pages case.
For hugetlb, we fill things in at the PTE level and that code already
puts the sub huge page physical bits into the PTEs, based upon the
offset, so there is nothing special we need to do. It all just works
out.
So, a small amount of complexity in the THP case, but this code is
about to get much simpler when we move the 64-bit PMDs as we can move
away from the fancy 32-bit huge PMD encoding and just put a real PTE
value in there.
With bug fixes and help from Bob Picco.
Signed-off-by: David S. Miller <davem@davemloft.net>
As reported by Dave Kleikamp, when we emit cross calls to do batched
TLB flush processing we have a race because we do not synchronize on
the sibling cpus completing the cross call.
So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.)
and either flushes are missed or flushes will flush the wrong
addresses.
Fix this by using generic infrastructure to synchonize on the
completion of the cross call.
This first required getting the flush_tlb_pending() call out from
switch_to() which operates with locks held and interrupts disabled.
The problem is that smp_call_function_many() cannot be invoked with
IRQs disabled and this is explicitly checked for with WARN_ON_ONCE().
We get the batch processing outside of locked IRQ disabled sections by
using some ideas from the powerpc port. Namely, we only batch inside
of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a
region, we flush TLBs synchronously.
1) Get rid of xcall_flush_tlb_pending and per-cpu type
implementations.
2) Do TLB batch cross calls instead via:
smp_call_function_many()
tlb_pending_func()
__flush_tlb_pending()
3) Batch only in lazy mmu sequences:
a) Add 'active' member to struct tlb_batch
b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE
c) Set 'active' in arch_enter_lazy_mmu_mode()
d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode()
e) Check 'active' in tlb_batch_add_one() and do a synchronous
flush if it's clear.
4) Add infrastructure for synchronous TLB page flushes.
a) Implement __flush_tlb_page and per-cpu variants, patch
as needed.
b) Likewise for xcall_flush_tlb_page.
c) Implement smp_flush_tlb_page() to invoke the cross-call.
d) Wire up global_flush_tlb_page() to the right routine based
upon CONFIG_SMP
5) It turns out that singleton batches are very common, 2 out of every
3 batch flushes have only a single entry in them.
The batch flush waiting is very expensive, both because of the poll
on sibling cpu completeion, as well as because passing the tlb batch
pointer to the sibling cpus invokes a shared memory dereference.
Therefore, in flush_tlb_pending(), if there is only one entry in
the batch perform a completely asynchronous global_flush_tlb_page()
instead.
Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
We should "|= more_flags" rather than "= more_flags".
Reported-by: David Rientjes <rientjes@google.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is relatively easy since PMD's now cover exactly 4MB of memory.
Our PMD entries are 32-bits each, so we use a special encoding. The
lowest bit, PMD_ISHUGE, determines the interpretation. This is possible
because sparc64's page tables are purely software entities so we can use
whatever encoding scheme we want. We just have to make the TLB miss
assembler page table walkers aware of the layout.
set_pmd_at() works much like set_pte_at() but it has to operate in two
page from a table of non-huge PTEs, so we have to queue up TLB flushes
based upon what mappings are valid in the PTE table. In the second regime
we are going from huge-page to non-huge-page, and in that case we need
only queue up a single TLB flush to push out the huge page mapping.
We still have 5 bits remaining in the huge PMD encoding so we can very
likely support any new pieces of THP state tracking that might get added
in the future.
With lots of help from Johannes Weiner.
Signed-off-by: David S. Miller <davem@davemloft.net>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We've split up the PTE tables so that they take up half a page instead of
a full page. This is in order to facilitate transparent huge page
support, which works much better if our PMDs cover 4MB instead of 8MB.
What we do is have a one-behind cache for PTE table allocations in the
mm struct.
This logic triggers only on allocations. For example, we don't try to
keep track of free'd up page table blocks in the style that the s390 port
does.
There were only two slightly annoying aspects to this change:
1) Changing pgtable_t to be a "pte_t *". There's all of this special
logic in the TLB free paths that needed adjustments, as did the
PMD populate interfaces.
2) init_new_context() needs to zap the pointer, since the mm struct
just gets copied from the parent on fork.
Signed-off-by: David S. Miller <davem@davemloft.net>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Narrowing the scope of the page size configurations will make the
transparent hugepage changes much simpler.
In the end what we really want to do is have the kernel support multiple
huge page sizes and use whatever is appropriate as the context dictactes.
Signed-off-by: David S. Miller <davem@davemloft.net>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Gerald Schaefer <gerald.schaefer@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
With the recent mmu_gather changes that included generic RCU freeing of
page-tables, it is now quite straightforward to implement gup_fast() on
sparc64.
This patch:
Remove the page table quicklists. They are pointless and make it harder
to use RCU page table freeing and share code with other architectures.
BTW, this is the second time this has happened, see commit 3c93646524
("[SPARC64]: Kill pgtable quicklists and use SLAB.")
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Semicolons are not necessary after switch/while/for/if braces
so remove them.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rework the sparc mmu_gather usage to conform to the new world order :-)
Sparc mmu_gather does two things:
- tracks vaddrs to unhash
- tracks pages to free
Split these two things like powerpc has done and keep the vaddrs
in per-cpu data structures and flush them on context switch.
The remaining bits can then use the generic mmu_gather.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: David Miller <davem@davemloft.net>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Richard Weinberger <richard@nod.at>
Cc: Tony Luck <tony.luck@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Add a sysctl to tweak the RSS limit used to decide when to grow
the TSB for an address space.
In order to avoid expensive divides and multiplies only simply
positive and negative powers of two are supported.
The function computed takes the number of TSB translations that will
fit at one time in the TSB of a given size, and either adds or
subtracts a percentage of entries. This final value is the
RSS limit.
See tsb_size_to_rss_limit().
Signed-off-by: David S. Miller <davem@davemloft.net>
- move all sparc64/mm/ files to arch/sparc/mm/
- commonly named files are named _64.c
- add files to sparc/mm/Makefile preserving link order
- delete now unused sparc64/mm/Makefile
- sparc64 now finds mm/ in sparc
Signed-off-by: Sam Ravnborg <sam@ravnborg.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
