The bitmap allocator must keep metadata consistent. The easiest way is
to scan after every allocation for each affected block and the entire
chunk. This is rather expensive.
The free path can take advantage of current contig hints to prevent
scanning within the start and end block. If a scan is needed, it can
be done by scanning backwards from the start and forwards from the end
to identify the entire free area this can be combined with. The blocks
can then be updated by some basic checks rather than complete block
scans.
A chunk scan happens when the freed area makes a page free, a block
free, or spans across blocks. This is necessary as the contig hint at
this point could span across blocks. The check uses the minimum of page
size and the block size to allow for variable sized blocks. There is a
tradeoff here with not updating after every free. It is possible a
contig hint in one block can be merged with the contig hint in the next
block. This means the contig hint can be off by up to a page. However,
if the chunk's contig hint is contained in one block, the contig hint
will be accurate.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
This patch introduces the bitmap metadata blocks and adds the skeleton
of the code that will be used to maintain these blocks. Each chunk's
bitmap is made up of full metadata blocks. These blocks maintain basic
metadata to help prevent scanning unnecssarily to update hints. Full
scanning methods are used for the skeleton and will be replaced in the
coming patches. A number of helper functions are added as well to do
conversion of pages to blocks and manage offsets. Comments will be
updated as the final version of each function is added.
There exists a relationship between PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE,
the region size, and unit_size. Every chunk's region (including offsets)
is page aligned at the beginning to preserve alignment. The end is
aligned to LCM(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE) to ensure that the end
can fit with the populated page map which is by page and every metadata
block is fully accounted for. The unit_size is already page aligned, but
must also be aligned with PCPU_BITMAP_BLOCK_SIZE to ensure full metadata
blocks.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The percpu memory allocator is experiencing scalability issues when
allocating and freeing large numbers of counters as in BPF.
Additionally, there is a corner case where iteration is triggered over
all chunks if the contig_hint is the right size, but wrong alignment.
This patch replaces the area map allocator with a basic bitmap allocator
implementation. Each subsequent patch will introduce new features and
replace full scanning functions with faster non-scanning options when
possible.
Implementation:
This patchset removes the area map allocator in favor of a bitmap
allocator backed by metadata blocks. The primary goal is to provide
consistency in performance and memory footprint with a focus on small
allocations (< 64 bytes). The bitmap removes the heavy memmove from the
freeing critical path and provides a consistent memory footprint. The
metadata blocks provide a bound on the amount of scanning required by
maintaining a set of hints.
In an effort to make freeing fast, the metadata is updated on the free
path if the new free area makes a page free, a block free, or spans
across blocks. This causes the chunk's contig hint to potentially be
smaller than what it could allocate by up to the smaller of a page or a
block. If the chunk's contig hint is contained within a block, a check
occurs and the hint is kept accurate. Metadata is always kept accurate
on allocation, so there will not be a situation where a chunk has a
later contig hint than available.
Evaluation:
I have primarily done testing against a simple workload of allocation of
1 million objects (2^20) of varying size. Deallocation was done by in
order, alternating, and in reverse. These numbers were collected after
rebasing ontop of a80099a152. I present the worst-case numbers here:
Area Map Allocator:
Object Size | Alloc Time (ms) | Free Time (ms)
----------------------------------------------
4B | 310 | 4770
16B | 557 | 1325
64B | 436 | 273
256B | 776 | 131
1024B | 3280 | 122
Bitmap Allocator:
Object Size | Alloc Time (ms) | Free Time (ms)
----------------------------------------------
4B | 490 | 70
16B | 515 | 75
64B | 610 | 80
256B | 950 | 100
1024B | 3520 | 200
This data demonstrates the inability for the area map allocator to
handle less than ideal situations. In the best case of reverse
deallocation, the area map allocator was able to perform within range
of the bitmap allocator. In the worst case situation, freeing took
nearly 5 seconds for 1 million 4-byte objects. The bitmap allocator
dramatically improves the consistency of the free path. The small
allocations performed nearly identical regardless of the freeing
pattern.
While it does add to the allocation latency, the allocation scenario
here is optimal for the area map allocator. The area map allocator runs
into trouble when it is allocating in chunks where the latter half is
full. It is difficult to replicate this, so I present a variant where
the pages are second half filled. Freeing was done sequentially. Below
are the numbers for this scenario:
Area Map Allocator:
Object Size | Alloc Time (ms) | Free Time (ms)
----------------------------------------------
4B | 4118 | 4892
16B | 1651 | 1163
64B | 598 | 285
256B | 771 | 158
1024B | 3034 | 160
Bitmap Allocator:
Object Size | Alloc Time (ms) | Free Time (ms)
----------------------------------------------
4B | 481 | 67
16B | 506 | 69
64B | 636 | 75
256B | 892 | 90
1024B | 3262 | 147
The data shows a parabolic curve of performance for the area map
allocator. This is due to the memmove operation being the dominant cost
with the lower object sizes as more objects are packed in a chunk and at
higher object sizes, the traversal of the chunk slots is the dominating
cost. The bitmap allocator suffers this problem as well. The above data
shows the inability to scale for the allocation path with the area map
allocator and that the bitmap allocator demonstrates consistent
performance in general.
The second problem of additional scanning can result in the area map
allocator completing in 52 minutes when trying to allocate 1 million
4-byte objects with 8-byte alignment. The same workload takes
approximately 16 seconds to complete for the bitmap allocator.
V2:
Fixed a bug in pcpu_alloc_first_chunk end_offset was setting the bitmap
using bytes instead of bits.
Added a comment to pcpu_cnt_pop_pages to explain bitmap_weight.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
This patch increases the minimum allocation size of percpu memory to
4-bytes. This change will help minimize the metadata overhead
associated with the bitmap allocator. The assumption is that most
allocations will be of objects or structs greater than 2 bytes with
integers or longs being used rather than shorts.
The first chunk regions are now aligned with the minimum allocation
size. The reserved region is expected to be set as a multiple of the
minimum allocation size. The static region is aligned up and the delta
is removed from the dynamic size. This works because the dynamic size is
increased to be page aligned. If the static size is not minimum
allocation size aligned, then there must be a gap that is added to the
dynamic size. The dynamic size will never be smaller than the set value.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
If a PER_CPU struct which contains a spin_lock is statically initialized
via:
DEFINE_PER_CPU(struct foo, bla) = {
.lock = __SPIN_LOCK_UNLOCKED(bla.lock)
};
then lockdep assigns a seperate key to each lock because the logic for
assigning a key to statically initialized locks is to use the address as
the key. With per CPU locks the address is obvioulsy different on each CPU.
That's wrong, because all locks should have the same key.
To solve this the following modifications are required:
1) Extend the is_kernel/module_percpu_addr() functions to hand back the
canonical address of the per CPU address, i.e. the per CPU address
minus the per CPU offset.
2) Check the lock address with these functions and if the per CPU check
matches use the returned canonical address as the lock key, so all per
CPU locks have the same key.
3) Move the static_obj(key) check into look_up_lock_class() so this check
can be avoided for statically initialized per CPU locks. That's
required because the canonical address fails the static_obj(key) check
for obvious reasons.
Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
[ Merged Dan's fixups for !MODULES and !SMP into this patch. ]
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Dan Murphy <dmurphy@ti.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20170227143736.pectaimkjkan5kow@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
printk() takes some locks and could not be used a safe way in NMI
context.
The chance of a deadlock is real especially when printing stacks from
all CPUs. This particular problem has been addressed on x86 by the
commit a9edc88093 ("x86/nmi: Perform a safe NMI stack trace on all
CPUs").
The patchset brings two big advantages. First, it makes the NMI
backtraces safe on all architectures for free. Second, it makes all NMI
messages almost safe on all architectures (the temporary buffer is
limited. We still should keep the number of messages in NMI context at
minimum).
Note that there already are several messages printed in NMI context:
WARN_ON(in_nmi()), BUG_ON(in_nmi()), anything being printed out from MCE
handlers. These are not easy to avoid.
This patch reuses most of the code and makes it generic. It is useful
for all messages and architectures that support NMI.
The alternative printk_func is set when entering and is reseted when
leaving NMI context. It queues IRQ work to copy the messages into the
main ring buffer in a safe context.
__printk_nmi_flush() copies all available messages and reset the buffer.
Then we could use a simple cmpxchg operations to get synchronized with
writers. There is also used a spinlock to get synchronized with other
flushers.
We do not longer use seq_buf because it depends on external lock. It
would be hard to make all supported operations safe for a lockless use.
It would be confusing and error prone to make only some operations safe.
The code is put into separate printk/nmi.c as suggested by Steven
Rostedt. It needs a per-CPU buffer and is compiled only on
architectures that call nmi_enter(). This is achieved by the new
HAVE_NMI Kconfig flag.
The are MN10300 and Xtensa architectures. We need to clean up NMI
handling there first. Let's do it separately.
The patch is heavily based on the draft from Peter Zijlstra, see
https://lkml.org/lkml/2015/6/10/327
[arnd@arndb.de: printk-nmi: use %zu format string for size_t]
[akpm@linux-foundation.org: min_t->min - all types are size_t here]
Signed-off-by: Petr Mladek <pmladek@suse.com>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Suggested-by: Steven Rostedt <rostedt@goodmis.org>
Cc: Jan Kara <jack@suse.cz>
Acked-by: Russell King <rmk+kernel@arm.linux.org.uk> [arm part]
Cc: Daniel Thompson <daniel.thompson@linaro.org>
Cc: Jiri Kosina <jkosina@suse.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: David Miller <davem@davemloft.net>
Cc: Daniel Thompson <daniel.thompson@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
As pure cleanup, this patch removes PERCPU_ENOUGH_ROOM which is not
used any more. That is, no code refers to the definition.
Acked-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Jungseok Lee <jungseoklee85@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
To avoid include hell, the per_cpu variable printk_func was declared
in percpu.h. But it is only defined if printk is defined.
As users of printk may also use the printk_func variable, it needs to
be defined even if CONFIG_PRINTK is not.
Also add a printk.h include in percpu.h just to be safe.
Link: http://lkml.kernel.org/r/20141121183215.01ba539c@canb.auug.org.au
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Being able to divert printk to call another function besides the normal
logging is useful for such things like NMI handling. If some functions
are to be called from NMI that does printk() it is possible to lock up
the box if the nmi handler triggers when another printk is happening.
One example of this use is to perform a stack trace on all CPUs via NMI.
But if the NMI is to do the printk() it can cause the system to lock up.
By allowing the printk to be diverted to another function that can safely
record the printk output and then print it when it in a safe context
then NMIs will be safe to call these functions like show_regs().
Link: http://lkml.kernel.org/p/20140619213952.209176403@goodmis.org
Tested-by: Jiri Kosina <jkosina@suse.cz>
Acked-by: Jiri Kosina <jkosina@suse.cz>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Petr Mladek <pmladek@suse.cz>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The percpu allocator now supports atomic allocations by only
allocating from already populated areas but the mechanism to ensure
that there's adequate amount of populated areas was missing.
This patch expands pcpu_balance_work so that in addition to freeing
excess free chunks it also populates chunks to maintain an adequate
level of populated areas. pcpu_alloc() schedules pcpu_balance_work if
the amount of free populated areas is too low or after an atomic
allocation failure.
* PERPCU_DYNAMIC_RESERVE is increased by two pages to account for
PCPU_EMPTY_POP_PAGES_LOW.
* pcpu_async_enabled is added to gate both async jobs -
chunk->map_extend_work and pcpu_balance_work - so that we don't end
up scheduling them while the needed subsystems aren't up yet.
Signed-off-by: Tejun Heo <tj@kernel.org>
Now that pcpu_alloc_area() can allocate only from populated areas,
it's easy to add atomic allocation support to [__]alloc_percpu().
Update pcpu_alloc() so that it accepts @gfp and skips all the blocking
operations and allocates only from the populated areas if @gfp doesn't
contain GFP_KERNEL. New interface functions [__]alloc_percpu_gfp()
are added.
While this means that atomic allocations are possible, this isn't
complete yet as there's no mechanism to ensure that certain amount of
populated areas is kept available and atomic allocations may keep
failing under certain conditions.
Signed-off-by: Tejun Heo <tj@kernel.org>
We're in the process of moving all percpu accessors and operations to
include/linux/percpu-defs.h so that they're available to arch headers
without having to include full include/linux/percpu.h which may cause
cyclic inclusion dependency.
This patch moves {raw|this}_cpu_*() definitions from
include/linux/percpu.h to include/linux/percpu-defs.h. The code is
moved mostly verbatim; however, raw_cpu_*() are placed above
this_cpu_*() which is more conventional as the raw operations may be
used to defined other variants.
This is pure reorganization.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Christoph Lameter <cl@linux.com>
{raw|this}_cpu_*_N() operations are expected to be provided by archs
and the generic definitions are provided as fallbacks. As such, these
firmly belong to include/asm-generic/percpu.h.
Move the generic definitions to include/asm-generic/percpu.h. The
code is moved mostly verbatim; however, raw_cpu_*_N() are placed above
this_cpu_*_N() which is more conventional as the raw operations may be
used to defined other variants.
This is pure reorganization.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Christoph Lameter <cl@linux.com>
Currently, percpu allows two separate methods for overriding
{raw|this}_cpu_*() ops - for a given operation, an arch can provide
whole replacement or sized sub operations to override specific parts
of it. e.g. arch either can provide this_cpu_add() or
this_cpu_add_4() to override only the 4 byte operation.
While quite flexible on a glance, the dual-overriding scheme
complicates the code path for no actual gain. It compilcates the
already complex operation definitions and if an arch wants to override
all sizes, it can easily provide all variants anyway. In fact, no
arch is actually making use of whole operation override.
Another oddity is that __this_cpu_*() operations are defined in the
same way as raw_cpu_*() but ignores full overrides of the raw_cpu_*()
and doesn't allow full operation override, so if an arch provides
whole overrides for raw_cpu_*() operations __this_cpu_*() ends up
using the generic implementations.
More importantly, it takes away the layering between arch-specific and
generic parts making it impossible for the generic part to implement
arch-independent features on top of arch-specific overrides.
This patch removes the support for whole operation overrides. As no
arch is using it, this doesn't cause any actual difference.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Christoph Lameter <cl@linux.com>
include/linux/percpu-defs.h is gonna host all accessors and operations
so that arch headers can make use of them too without worrying about
circular dependency through include/linux/percpu.h.
This patch moves the following accessors from include/linux/percpu.h
to include/linux/percpu-defs.h.
* get/put_cpu_var()
* get/put_cpu_ptr()
* per_cpu_ptr()
This is pure reorgniazation.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Christoph Lameter <cl@linux.com>
Pull percpu updates from Tejun Heo:
"Nothing too exciting. percpu_ref is going through some interface
changes and getting new features with more changes in the pipeline but
given its young age and few users, it's very low impact"
* 'for-3.16' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu:
percpu-refcount: implement percpu_ref_tryget()
percpu-refcount: rename percpu_ref_tryget() to percpu_ref_tryget_live()
percpu: Replace __get_cpu_var with this_cpu_ptr
The definition for raw_cpu_add_return() uses the operation prefix
"raw_add_return_", but the definitions in the various percpu.h files
expect "raw_cpu_add_return_". This commit therefore appropriately
adjusts the definition of raw_cpu_add_return().
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Christoph Lameter <cl@linux.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
__this_cpu_ptr is being phased out. Use raw_cpu_ptr instead which was
introduced in 3.15-rc1. One case of using __get_cpu_var in the
get_cpu_var macro for address calculation was remaining in
include/linux/percpu.h.
tj: Updated patch description.
Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
We define a check function in order to avoid trouble with the include
files. Then the higher level __this_cpu macros are modified to invoke
the preemption check.
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Christoph Lameter <cl@linux.com>
Acked-by: Ingo Molnar <mingo@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Tested-by: Grygorii Strashko <grygorii.strashko@ti.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The kernel has never been audited to ensure that this_cpu operations are
consistently used throughout the kernel. The code generated in many
places can be improved through the use of this_cpu operations (which
uses a segment register for relocation of per cpu offsets instead of
performing address calculations).
The patch set also addresses various consistency issues in general with
the per cpu macros.
A. The semantics of __this_cpu_ptr() differs from this_cpu_ptr only
because checks are skipped. This is typically shown through a raw_
prefix. So this patch set changes the places where __this_cpu_ptr()
is used to raw_cpu_ptr().
B. There has been the long term wish by some that __this_cpu operations
would check for preemption. However, there are cases where preemption
checks need to be skipped. This patch set adds raw_cpu operations that
do not check for preemption and then adds preemption checks to the
__this_cpu operations.
C. The use of __get_cpu_var is always a reference to a percpu variable
that can also be handled via a this_cpu operation. This patch set
replaces all uses of __get_cpu_var with this_cpu operations.
D. We can then use this_cpu RMW operations in various places replacing
sequences of instructions by a single one.
E. The use of this_cpu operations throughout will allow other arches than
x86 to implement optimized references and RMV operations to work with
per cpu local data.
F. The use of this_cpu operations opens up the possibility to
further optimize code that relies on synchronization through
per cpu data.
The patch set works in a couple of stages:
I. Patch 1 adds the additional raw_cpu operations and raw_cpu_ptr().
Also converts the existing __this_cpu_xx_# primitive in the x86
code to raw_cpu_xx_#.
II. Patch 2-4 use the raw_cpu operations in places that would give
us false positives once they are enabled.
III. Patch 5 adds preemption checks to __this_cpu operations to allow
checking if preemption is properly disabled when these functions
are used.
IV. Patches 6-20 are patches that simply replace uses of __get_cpu_var
with this_cpu_ptr. They do not depend on any changes to the percpu
code. No preemption tests are skipped if they are applied.
V. Patches 21-46 are conversion patches that use this_cpu operations
in various kernel subsystems/drivers or arch code.
VI. Patches 47/48 (not included in this series) remove no longer used
functions (__this_cpu_ptr and __get_cpu_var). These should only be
applied after all the conversion patches have made it and after we
have done additional passes through the kernel to ensure that none of
the uses of these functions remain.
This patch (of 46):
The patches following this one will add preemption checks to __this_cpu
ops so we need to have an alternative way to use this_cpu operations
without preemption checks.
raw_cpu_ops will be the basis for all other ops since these will be the
operations that do not implement any checks.
Primitive operations are renamed by this patch from __this_cpu_xxx to
raw_cpu_xxxx.
Also change the uses of the x86 percpu primitives in preempt.h.
These depend directly on asm/percpu.h (header #include nesting issue).
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Christoph Lameter <cl@linux.com>
Acked-by: Ingo Molnar <mingo@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Alex Shi <alex.shi@intel.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Bryan Wu <cooloney@gmail.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Daniel Lezcano <daniel.lezcano@linaro.org>
Cc: David Daney <david.daney@cavium.com>
Cc: David Miller <davem@davemloft.net>
Cc: David S. Miller <davem@davemloft.net>
Cc: Dimitri Sivanich <sivanich@sgi.com>
Cc: Dipankar Sarma <dipankar@in.ibm.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Fenghua Yu <fenghua.yu@intel.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: H. Peter Anvin <hpa@linux.intel.com>
Cc: Haavard Skinnemoen <hskinnemoen@gmail.com>
Cc: Hans-Christian Egtvedt <egtvedt@samfundet.no>
Cc: Hedi Berriche <hedi@sgi.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: James Hogan <james.hogan@imgtec.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Mike Frysinger <vapier@gentoo.org>
Cc: Mike Travis <travis@sgi.com>
Cc: Neil Brown <neilb@suse.de>
Cc: Nicolas Pitre <nicolas.pitre@linaro.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Robert Richter <rric@kernel.org>
Cc: Russell King <linux@arm.linux.org.uk>
Cc: Russell King <rmk+kernel@arm.linux.org.uk>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wim Van Sebroeck <wim@iguana.be>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Most of the VM_BUG_ON assertions are performed on a page. Usually, when
one of these assertions fails we'll get a BUG_ON with a call stack and
the registers.
I've recently noticed based on the requests to add a small piece of code
that dumps the page to various VM_BUG_ON sites that the page dump is
quite useful to people debugging issues in mm.
This patch adds a VM_BUG_ON_PAGE(cond, page) which beyond doing what
VM_BUG_ON() does, also dumps the page before executing the actual
BUG_ON.
[akpm@linux-foundation.org: fix up includes]
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull percpu changes from Tejun Heo:
"Two smallish changes for percpu. Two patches to remove unused
this_cpu_xor() and one to fix a bug in percpu init failure path so
that it can reach the proper BUG() instead of oopsing earlier"
* 'for-3.13' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu:
x86: remove this_cpu_xor() implementation
percpu: remove this_cpu_xor() implementation
percpu: fix bootmem error handling in pcpu_page_first_chunk()
this_cpu_sub() is implemented as negation and addition.
This patch casts the adjustment to the counter type before negation to
sign extend the adjustment. This helps in cases where the counter type
is wider than an unsigned adjustment. An alternative to this patch is
to declare such operations unsupported, but it seemed useful to avoid
surprises.
This patch specifically helps the following example:
unsigned int delta = 1
preempt_disable()
this_cpu_write(long_counter, 0)
this_cpu_sub(long_counter, delta)
preempt_enable()
Before this change long_counter on a 64 bit machine ends with value
0xffffffff, rather than 0xffffffffffffffff. This is because
this_cpu_sub(pcp, delta) boils down to this_cpu_add(pcp, -delta),
which is basically:
long_counter = 0 + 0xffffffff
Also apply the same cast to:
__this_cpu_sub()
__this_cpu_sub_return()
this_cpu_sub_return()
All percpu_test.ko passes, especially the following cases which
previously failed:
l -= ui_one;
__this_cpu_sub(long_counter, ui_one);
CHECK(l, long_counter, -1);
l -= ui_one;
this_cpu_sub(long_counter, ui_one);
CHECK(l, long_counter, -1);
CHECK(l, long_counter, 0xffffffffffffffff);
ul -= ui_one;
__this_cpu_sub(ulong_counter, ui_one);
CHECK(ul, ulong_counter, -1);
CHECK(ul, ulong_counter, 0xffffffffffffffff);
ul = this_cpu_sub_return(ulong_counter, ui_one);
CHECK(ul, ulong_counter, 2);
ul = __this_cpu_sub_return(ulong_counter, ui_one);
CHECK(ul, ulong_counter, 1);
Signed-off-by: Greg Thelen <gthelen@google.com>
Acked-by: Tejun Heo <tj@kernel.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There is not a single user in the whole kernel.
Besides that this_cpu_xor() is broken anyway since it gets
translated to this_cpu_or() (see __pcpu_size_call() line).
So instead of fixing an unused definition just remove it.
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Acked-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
Remove percpu_xxx serial functions, all of them were replaced by
this_cpu_xxx or __this_cpu_xxx serial functions
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: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
This patch adds missed "__" prefixes, otherwise these functions
works as irq/preemption safe.
Reported-by: Torsten Kaiser <just.for.lkml@googlemail.com>
Signed-off-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
It doesn't make sense to trace irq off or do irq flags
lock proving inside 'this_cpu' operations, so replace local_irq_*
with raw_local_irq_* in 'this_cpu' op.
Also the patch fixes onelockdep warning[1] by the replacement, see
below:
In commit: 933393f58fef9963eac61db8093689544e29a600(percpu:
Remove irqsafe_cpu_xxx variants), local_irq_save/restore(flags) are
added inside this_cpu_inc operation, so that trace_hardirqs_off_caller
will be called by trace_hardirqs_on_caller directly because
__debug_atomic_inc is implemented as this_cpu_inc, which may trigger
the lockdep warning[1], for example in the below ARM scenary:
kernel_thread_helper /*irq disabled*/
->trace_hardirqs_on_caller /*hardirqs_enabled was set*/
->trace_hardirqs_off_caller /*hardirqs_enabled cleared*/
__this_cpu_add(redundant_hardirqs_on)
->trace_hardirqs_off_caller /*irq disabled, so call here*/
The 'unannotated irqs-on' warning will be triggered somewhere because
irq is just enabled after the irq trace in kernel_thread_helper.
[1],
[ 0.162841] ------------[ cut here ]------------
[ 0.167694] WARNING: at kernel/lockdep.c:3493 check_flags+0xc0/0x1d0()
[ 0.174468] Modules linked in:
[ 0.177703] Backtrace:
[ 0.180328] [<c00171f0>] (dump_backtrace+0x0/0x110) from [<c0412320>] (dump_stack+0x18/0x1c)
[ 0.189086] r6:c051f778 r5:00000da5 r4:00000000 r3:60000093
[ 0.195007] [<c0412308>] (dump_stack+0x0/0x1c) from [<c00410e8>] (warn_slowpath_common+0x54/0x6c)
[ 0.204223] [<c0041094>] (warn_slowpath_common+0x0/0x6c) from [<c0041124>] (warn_slowpath_null+0x24/0x2c)
[ 0.214111] r8:00000000 r7:00000000 r6:ee069598 r5:60000013 r4:ee082000
[ 0.220825] r3:00000009
[ 0.223693] [<c0041100>] (warn_slowpath_null+0x0/0x2c) from [<c0088f38>] (check_flags+0xc0/0x1d0)
[ 0.232910] [<c0088e78>] (check_flags+0x0/0x1d0) from [<c008d348>] (lock_acquire+0x4c/0x11c)
[ 0.241668] [<c008d2fc>] (lock_acquire+0x0/0x11c) from [<c0415aa4>] (_raw_spin_lock+0x3c/0x74)
[ 0.250610] [<c0415a68>] (_raw_spin_lock+0x0/0x74) from [<c010a844>] (set_task_comm+0x20/0xc0)
[ 0.259521] r6:ee069588 r5:ee0691c0 r4:ee082000
[ 0.264404] [<c010a824>] (set_task_comm+0x0/0xc0) from [<c0060780>] (kthreadd+0x28/0x108)
[ 0.272857] r8:00000000 r7:00000013 r6:c0044a08 r5:ee0691c0 r4:ee082000
[ 0.279571] r3:ee083fe0
[ 0.282470] [<c0060758>] (kthreadd+0x0/0x108) from [<c0044a08>] (do_exit+0x0/0x6dc)
[ 0.290405] r5:c0060758 r4:00000000
[ 0.294189] ---[ end trace 1b75b31a2719ed1c ]---
[ 0.299041] possible reason: unannotated irqs-on.
[ 0.303955] irq event stamp: 5
[ 0.307159] hardirqs last enabled at (4): [<c001331c>] no_work_pending+0x8/0x2c
[ 0.314880] hardirqs last disabled at (5): [<c0089b08>] trace_hardirqs_on_caller+0x60/0x26c
[ 0.323547] softirqs last enabled at (0): [<c003f754>] copy_process+0x33c/0xef4
[ 0.331207] softirqs last disabled at (0): [< (null)>] (null)
[ 0.337585] CPU0: thread -1, cpu 0, socket 0, mpidr 80000000
Acked-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
This patch adds missed "__" into function prefix.
Otherwise on all archectures (except x86) it expands to irq/preemtion-safe
variant: _this_cpu_generic_add_return(), which do extra irq-save/irq-restore.
Optimal generic implementation is __this_cpu_generic_add_return().
Signed-off-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Acked-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
We simply say that regular this_cpu use must be safe regardless of
preemption and interrupt state. That has no material change for x86
and s390 implementations of this_cpu operations. However, arches that
do not provide their own implementation for this_cpu operations will
now get code generated that disables interrupts instead of preemption.
-tj: This is part of on-going percpu API cleanup. For detailed
discussion of the subject, please refer to the following thread.
http://thread.gmane.org/gmane.linux.kernel/1222078
Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
LKML-Reference: <alpine.DEB.2.00.1112221154380.11787@router.home>
On an architecture without CMPXCHG_LOCAL but with DEBUG_VM enabled,
the VM_BUG_ON() in __pcpu_double_call_return_bool() will cause an early
panic during boot unless we always align cpu_slab properly.
In principle we could remove the alignment-testing VM_BUG_ON() for
architectures that don't have CMPXCHG_LOCAL, but leaving it in means
that new code will tend not to break x86 even if it is introduced
on another platform, and it's low cost to require alignment.
Acked-by: David Rientjes <rientjes@google.com>
Acked-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>
Signed-off-by: Pekka Enberg <penberg@kernel.org>
The SLUB allocator use of the cmpxchg_double logic was wrong: it
actually needs the irq-safe one.
That happens automatically when we use the native unlocked 'cmpxchg8b'
instruction, but when compiling the kernel for older x86 CPUs that do
not support that instruction, we fall back to the generic emulation
code.
And if you don't specify that you want the irq-safe version, the generic
code ends up just open-coding the cmpxchg8b equivalent without any
protection against interrupts or preemption. Which definitely doesn't
work for SLUB.
This was reported by Werner Landgraf <w.landgraf@ru.ru>, who saw
instability with his distro-kernel that was compiled to support pretty
much everything under the sun. Most big Linux distributions tend to
compile for PPro and later, and would never have noticed this problem.
This also fixes the prototypes for the irqsafe cmpxchg_double functions
to use 'bool' like they should.
[ Btw, that whole "generic code defaults to no protection" design just
sounds stupid - if the code needs no protection, there is no reason to
use "cmpxchg_double" to begin with. So we should probably just remove
the unprotected version entirely as pointless. - Linus ]
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reported-and-tested-by: werner <w.landgraf@ru.ru>
Acked-and-tested-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Tejun Heo <tj@kernel.org>
Link: http://lkml.kernel.org/r/alpine.LFD.2.02.1105041539050.3005@ionos
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Introduce this_cpu_cmpxchg_double(). this_cpu_cmpxchg_double() allows
the comparison between two consecutive words and replaces them if
there is a match.
bool this_cpu_cmpxchg_double(pcp1, pcp2,
old_word1, old_word2, new_word1, new_word2)
this_cpu_cmpxchg_double does not return the old value (difficult since
there are two words) but a boolean indicating if the operation was
successful.
The first percpu variable must be double word aligned!
-tj: Updated to return bool instead of int, converted size check to
BUILD_BUG_ON() instead of VM_BUG_ON() and other cosmetic changes.
Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Generic code to provide new per cpu atomic features
this_cpu_cmpxchg
this_cpu_xchg
Fallback occurs to functions using interrupts disable/enable
to ensure correct per cpu atomicity.
Fallback to regular cmpxchg and xchg is not possible since per cpu atomic
semantics include the guarantee that the current cpus per cpu data is
accessed atomically. Use of regular cmpxchg and xchg requires the
determination of the address of the per cpu data before regular cmpxchg
or xchg which therefore cannot be atomically included in an xchg or
cmpxchg without segment override.
tj: - Relocated new ops to conform better to the general organization.
- This patch contains a trivial comment fix.
Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
- include/linux/percpu.h: this_cpu_add_return() and friends were
located next to __this_cpu_add_return(). However, the overall
organization is to first group by preemption safeness. Relocate
this_cpu_add_return() and friends to preemption-safe area.
- arch/x86/include/asm/percpu.h: Relocate percpu_add_return_op() after
other more basic operations. Relocate [__]this_cpu_add_return_8()
so that they're first grouped by preemption safeness.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Introduce generic support for this_cpu_add_return etc.
The fallback is to realize these operations with simpler __this_cpu_ops.
tj: - Reformatted __cpu_size_call_return2() to make it more consistent
with its neighbors.
- Dropped unnecessary temp variable ret__ from
__this_cpu_generic_add_return().
Reviewed-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Acked-by: H. Peter Anvin <hpa@zytor.com>
Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu:
percpu: update comments to reflect that percpu allocations are always zero-filled
percpu: Optimize __get_cpu_var()
x86, percpu: Optimize this_cpu_ptr
percpu: clear memory allocated with the km allocator
percpu: fix build breakage on s390 and cleanup build configuration tests
percpu: use percpu allocator on UP too
percpu: reduce PCPU_MIN_UNIT_SIZE to 32k
vmalloc: pcpu_get/free_vm_areas() aren't needed on UP
Fixed up trivial conflicts in include/linux/percpu.h
These are similar to {get,put}_cpu_var() except for dynamically
allocated per-cpu memory.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Tejun Heo <tj@kernel.org>
LKML-Reference: <20100917093009.252867712@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
On UP, percpu allocations were redirected to kmalloc. This has the
following problems.
* For certain amount of allocations (determined by
PERCPU_DYNAMIC_EARLY_SLOTS and PERCPU_DYNAMIC_EARLY_SIZE), percpu
allocator can be used before the usual kernel memory allocator is
brought online. On SMP, this is used to initialize the kernel
memory allocator.
* percpu allocator honors alignment upto PAGE_SIZE but kmalloc()
doesn't. For example, workqueue makes use of larger alignments for
cpu_workqueues.
Currently, users of percpu allocators need to handle UP differently,
which is somewhat fragile and ugly. Other than small amount of
memory, there isn't much to lose by enabling percpu allocator on UP.
It can simply use kernel memory based chunk allocation which was added
for SMP archs w/o MMUs.
This patch removes mm/percpu_up.c, builds mm/percpu.c on UP too and
makes UP build use percpu-km. As percpu addresses and kernel
addresses are always identity mapped and static percpu variables don't
need any special treatment, nothing is arch dependent and mm/percpu.c
implements generic setup_per_cpu_areas() for UP.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
In preparation of enabling percpu allocator for UP, reduce
PCPU_MIN_UNIT_SIZE to 32k. On UP, the first chunk doesn't have to
include static percpu variables and chunk size can be smaller which is
important as UP percpu allocator will use contiguous kernel memory to
populate chunks.
PCPU_MIN_UNIT_SIZE also determines the maximum supported allocation
size but 32k should still be enough.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Christoph Lameter <cl@linux.com>
UP accessors didn't take care of __percpu notations leading to a lot
of spurious sparse warnings on UP configurations. Fix it.
Signed-off-by: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
This patch updates percpu allocator such that it can serve limited
amount of allocation before slab comes online. This is primarily to
allow slab to depend on working percpu allocator.
Two parameters, PERCPU_DYNAMIC_EARLY_SIZE and SLOTS, determine how
much memory space and allocation map slots are reserved. If this
reserved area is exhausted, WARN_ON_ONCE() will trigger and allocation
will fail till slab comes online.
The following changes are made to implement early alloc.
* pcpu_mem_alloc() now checks slab_is_available()
* Chunks are allocated using pcpu_mem_alloc()
* Init paths make sure ai->dyn_size is at least as large as
PERCPU_DYNAMIC_EARLY_SIZE.
* Initial alloc maps are allocated in __initdata and copied to
kmalloc'd areas once slab is online.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Christoph Lameter <cl@linux-foundation.org>
In pcpu_build_alloc_info() and pcpu_embed_first_chunk(), @dyn_size was
ssize_t, -1 meant auto-size, 0 forced 0 and positive meant minimum
size. There's no use case for forcing 0 and the upcoming early alloc
support always requires non-zero dynamic size. Make @dyn_size always
mean minimum dyn_size.
While at it, make pcpu_build_alloc_info() static which doesn't have
any external caller as suggested by David Rientjes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
* 'slabh' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc:
eeepc-wmi: include slab.h
staging/otus: include slab.h from usbdrv.h
percpu: don't implicitly include slab.h from percpu.h
kmemcheck: Fix build errors due to missing slab.h
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h
iwlwifi: don't include iwl-dev.h from iwl-devtrace.h
x86: don't include slab.h from arch/x86/include/asm/pgtable_32.h
Fix up trivial conflicts in include/linux/percpu.h due to
is_kernel_percpu_address() having been introduced since the slab.h
cleanup with the percpu_up.c splitup.
percpu.h has always been including slab.h to get k[mz]alloc/free() for
UP inline implementation. percpu.h being used by very low level
headers including module.h and sched.h, this meant that a lot files
unintentionally got slab.h inclusion.
Lee Schermerhorn was trying to make topology.h use percpu.h and got
bitten by this implicit inclusion. The right thing to do is break
this ultimately unnecessary dependency. The previous patch added
explicit inclusion of either gfp.h or slab.h to the source files using
them. This patch updates percpu.h such that slab.h is no longer
included from percpu.h.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
lockdep has custom code to check whether a pointer belongs to static
percpu area which is somewhat broken. Implement proper
is_kernel/module_percpu_address() and replace the custom code.
On UP, percpu variables are regular static variables and can't be
distinguished from them. Always return %false on UP.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Ingo Molnar <mingo@redhat.com>
Commit 3b034b0d08 implemented
per_cpu_ptr_to_phys() but forgot to add UP definition. Add UP
definition which is simple wrapper around __pa().
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Reported-by: Randy Dunlap <randy.dunlap@oracle.com>
o kdump functionality reserves a per cpu area at boot time and exports the
physical address of that area to user space through sys interface. This
area stores some dump related information like cpu register states etc
at the time of crash.
o We were assuming that per cpu area always come from linearly mapped meory
region and using __pa() to determine physical address.
With percpu_alloc=page, per cpu area can come from vmalloc region also and
__pa() breaks.
o This patch implments a new function to convert per cpu address to
physical address.
Before the patch, crash_notes addresses looked as follows.
cpu0 60fffff49800
cpu1 60fffff60800
cpu2 60fffff77800
These are bogus phsyical addresses.
After the patch, address are following.
cpu0 13eb44000
cpu1 13eb43000
cpu2 13eb42000
cpu3 13eb41000
These look fine. I got 4G of memory and /proc/iomem tell me following.
100000000-13fffffff : System RAM
tj: * added missing asm/io.h include reported by Stephen Rothwell
* repositioned per_cpu_ptr_phys() in percpu.c and added comment.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Dennis Zhou (Facebook)