Impact: reduce unnecessary memory usage on certain configurations
Embedding percpu allocator allocates unit_size *
smp_num_possible_cpus() bytes consecutively and use it for the first
chunk. However, if the static area is small, this can result in
excessive prellocated free space in the first chunk due to
PCPU_MIN_UNIT_SIZE restriction.
This patch makes embedding percpu allocator preallocate only what's
necessary as described by PERPCU_DYNAMIC_RESERVE and return the
leftover to the bootmem allocator.
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: argument semantic cleanup
In pcpu_setup_first_chunk(), zero @unit_size and @dyn_size meant
auto-sizing. It's okay for @unit_size as 0 doesn't make sense but 0
dynamic reserve size is valid. Alos, if arch @dyn_size is calculated
from other parameters, it might end up passing in 0 @dyn_size and
malfunction when the size is automatically adjusted.
This patch makes both @unit_size and @dyn_size ssize_t and use -1 for
auto sizing.
Signed-off-by: Tejun Heo <tj@kernel.org>
Recent changes in setup_percpu.c made a now meaningless DBG()
statement fail to compile and introduced a
comparison-of-different-types warning. Fix them.
Compile failure is reported by Ingo Molnar.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Ingo Molnar <mingo@elte.hu>
Impact: add better first percpu allocation for NUMA
On NUMA, embedding allocator can't be used as different units can't be
made to fall in the correct NUMA nodes. To use large page mapping,
each unit needs to be remapped. However, percpu areas are usually
much smaller than large page size and unused space hurts a lot as the
number of cpus grow. This allocator remaps large pages for each chunk
but gives back unused part to the bootmem allocator making the large
pages mapped twice.
This adds slightly to the TLB pressure but is much better than using
4k mappings while still being NUMA-friendly.
Ingo suggested that this would be the correct approach for NUMA.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Ingo Molnar <mingo@elte.hu>
Impact: add better first percpu allocation for !NUMA
On !NUMA, we can simply allocate contiguous memory and use it for the
first chunk without mapping it into vmalloc area. As the memory area
is covered by the large page physical memory mapping, it allows the
dynamic perpcu allocator to not add any TLB overhead for the static
percpu area and whatever falls into the first chunk and the
implementation is very simple too.
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: modularize percpu first chunk allocation
x86 is gonna have a few different strategies for the first chunk
allocation. Modularize it by separating out the current allocation
mechanism into pcpu_alloc_bootmem() and setup_pcpu_4k().
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: more latitude for first percpu chunk allocation
The first percpu chunk serves the kernel static percpu area and may or
may not contain extra room for further dynamic allocation.
Initialization of the first chunk needs to be done before normal
memory allocation service is up, so it has its own init path -
pcpu_setup_static().
It seems archs need more latitude while initializing the first chunk
for example to take advantage of large page mapping. This patch makes
the following changes to allow this.
* Define PERCPU_DYNAMIC_RESERVE to give arch hint about how much space
to reserve in the first chunk for further dynamic allocation.
* Rename pcpu_setup_static() to pcpu_setup_first_chunk().
* Make pcpu_setup_first_chunk() much more flexible by fetching page
pointer by callback and adding optional @unit_size, @free_size and
@base_addr arguments which allow archs to selectively part of chunk
initialization to their likings.
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: minor change to populate_extra_pte() and addition of pmd flavor
Update populate_extra_pte() to return pointer to the pte_t for the
specified address and add populate_extra_pmd() which only populates
till the pmd and returns pointer to the pmd entry for the address.
For 64bit, pud/pmd/pte fill functions are separated out from
set_pte_vaddr[_pud]() and used for set_pte_vaddr[_pud]() and
populate_extra_{pte|pmd}().
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: use new dynamic allocator, unified access to static/dynamic
percpu memory
Convert to the new dynamic percpu allocator.
* implement populate_extra_pte() for both 32 and 64
* update setup_per_cpu_areas() to use pcpu_setup_static()
* define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr()
* define config HAVE_DYNAMIC_PER_CPU_AREA
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: stack protector for x86_32
Implement stack protector for x86_32. GDT entry 28 is used for it.
It's set to point to stack_canary-20 and have the length of 24 bytes.
CONFIG_CC_STACKPROTECTOR turns off CONFIG_X86_32_LAZY_GS and sets %gs
to the stack canary segment on entry. As %gs is otherwise unused by
the kernel, the canary can be anywhere. It's defined as a percpu
variable.
x86_32 exception handlers take register frame on stack directly as
struct pt_regs. With -fstack-protector turned on, gcc copies the
whole structure after the stack canary and (of course) doesn't copy
back on return thus losing all changed. For now, -fno-stack-protector
is added to all files which contain those functions. We definitely
need something better.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: cleanup, prepare for xen boot fix.
Xen needs to call this function very early to setup the GDT and
per-cpu segments. Remove the call to smp_processor_id() and just
pass in the cpu number.
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: build fix
x86_cpu_to_apicid and x86_bios_cpu_apicid aren't defined for voyage.
Earlier patch forgot to conditionalize early percpu clearing. Fix it.
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: sync 32 and 64-bit code
Merge load_gs_base() into switch_to_new_gdt(). Load the GDT and
per-cpu state for the boot cpu when its new area is set up.
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: cleanup
Rename init_gdt() to setup_percpu_segment(), and move it to
setup_percpu.c.
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: standardize all x86 platforms on same setup code
With the preceding changes, Voyager can use the same per-cpu setup
code as all the other x86 platforms.
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: Small cleanup
Define BOOT_PERCPU_OFFSET and use it for this_cpu_offset and
__per_cpu_offset initializers.
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: Code movement
Move the variable definitions to apic.c. Ifdef the copying of
the two early per-cpu variables, since Voyager doesn't use them.
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: cleanup
The way the code is written, align is always PAGE_SIZE. Simplify
the code by removing the align variable.
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: Code movement, no functional change.
Move setup_cpu_local_masks() to kernel/cpu/common.c, where the
masks are defined.
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: Code movement, no functional change.
Move the 64-bit NUMA code from setup_percpu.c to numa_64.c
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: minor optimization
Eliminates the need for two loops over possible cpus.
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: x86_64 percpu area layout change, irq_stack now at the beginning
Now that the PDA is empty except for the stack canary, it can be removed.
The irqstack is moved to the start of the per-cpu section. If the stack
protector is enabled, the canary overlaps the bottom 48 bytes of the irqstack.
tj: * updated subject
* dropped asm relocation of irq_stack_ptr
* updated comments a bit
* rebased on top of stack canary changes
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
tj: moved cpu_number definition out of CONFIG_HAVE_SETUP_PER_CPU_AREA
for voyager.
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Move the irqstackptr variable from the PDA to per-cpu. Make the
stacks themselves per-cpu, removing some specific allocation code.
Add a seperate flag (is_boot_cpu) to simplify the per-cpu boot
adjustments.
tj: * sprinkle some underbars around.
* irq_stack_ptr is not used till traps_init(), no reason to
initialize it early. On SMP, just leaving it NULL till proper
initialization in setup_per_cpu_areas() works. Dropped
is_boot_cpu and early irq_stack_ptr initialization.
* do DECLARE/DEFINE_PER_CPU(char[IRQ_STACK_SIZE], irq_stack)
instead of (char, irq_stack[IRQ_STACK_SIZE]).
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
[ Based on original patch from Christoph Lameter and Mike Travis. ]
As pda is now allocated in percpu area, it can easily be made a proper
percpu variable. Make it so by defining per cpu symbol from linker
script and declaring it in C code for SMP and simply defining it for
UP. This change cleans up code and brings SMP and UP closer a bit.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Now that pda is allocated as part of percpu, percpu doesn't need to be
accessed through pda. Unify x86_64 SMP percpu access with x86_32 SMP
one. Other than the segment register, operand size and the base of
percpu symbols, they behave identical now.
This patch replaces now unnecessary pda->data_offset with a dummy
field which is necessary to keep stack_canary at its place. This
patch also moves per_cpu_offset initialization out of init_gdt() into
setup_per_cpu_areas(). Note that this change also necessitates
explicit per_cpu_offset initializations in voyager_smp.c.
With this change, x86_OP_percpu()'s are as efficient on x86_64 as on
x86_32 and also x86_64 can use assembly PER_CPU macros.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
[ Based on original patch from Christoph Lameter and Mike Travis. ]
Currently pdas and percpu areas are allocated separately. %gs points
to local pda and percpu area can be reached using pda->data_offset.
This patch folds pda into percpu area.
Due to strange gcc requirement, pda needs to be at the beginning of
the percpu area so that pda->stack_canary is at %gs:40. To achieve
this, a new percpu output section macro - PERCPU_VADDR_PREALLOC() - is
added and used to reserve pda sized chunk at the start of the percpu
area.
After this change, for boot cpu, %gs first points to pda in the
data.init area and later during setup_per_cpu_areas() gets updated to
point to the actual pda. This means that setup_per_cpu_areas() need
to reload %gs for CPU0 while clearing pda area for other cpus as cpu0
already has modified it when control reaches setup_per_cpu_areas().
This patch also removes now unnecessary get_local_pda() and its call
sites.
A lot of this patch is taken from Mike Travis' "x86_64: Fold pda into
per cpu area" patch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
_cpu_pda array first uses statically allocated storage in data.init
and then switches to allocated bootmem to conserve space. However,
after folding pda area into percpu area, _cpu_pda array will be
removed completely. Drop the reallocation part to simplify the code
for soon-to-follow changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
[ Based on original patch from Christoph Lameter and Mike Travis. ]
This patch makes percpu symbols zerobased on x86_64 SMP by adding
PERCPU_VADDR() to vmlinux.lds.h which helps setting explicit vaddr on
the percpu output section and using it in vmlinux_64.lds.S. A new
PHDR is added as existing ones cannot contain sections near address
zero. PERCPU_VADDR() also adds a new symbol __per_cpu_load which
always points to the vaddr of the loaded percpu data.init region.
The following adjustments have been made to accomodate the address
change.
* code to locate percpu gdt_page in head_64.S is updated to add the
load address to the gdt_page offset.
* __per_cpu_load is used in places where access to the init data area
is necessary.
* pda->data_offset is initialized soon after C code is entered as zero
value doesn't work anymore.
This patch is mostly taken from Mike Travis' "x86_64: Base percpu
variables at zero" patch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
[ Based on original patch from Christoph Lameter and Mike Travis. ]
* Ruggedize some calls in setup_percpu.c to prevent mishaps
in early calls, particularly for non-critical functions.
* Cleanup DEBUG_PER_CPU_MAPS usages and some comments.
Signed-off-by: Mike Travis <travis@sgi.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* 'x86-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip: (36 commits)
x86: fix section mismatch warnings in mcheck/mce_amd_64.c
x86: offer frame pointers in all build modes
x86: remove duplicated #include's
x86: k8 numa register active regions later
x86: update Alan Cox's email addresses
x86: rename all fields of mpc_table mpc_X to X
x86: rename all fields of mpc_oemtable oem_X to X
x86: rename all fields of mpc_bus mpc_X to X
x86: rename all fields of mpc_cpu mpc_X to X
x86: rename all fields of mpc_intsrc mpc_X to X
x86: rename all fields of mpc_lintsrc mpc_X to X
x86: rename all fields of mpc_iopic mpc_X to X
x86: irqinit_64.c init_ISA_irqs should be static
Documentation/x86/boot.txt: payload length was changed to payload_length
x86: setup_percpu.c fix style problems
x86: irqinit_64.c fix style problems
x86: irqinit_32.c fix style problems
x86: i8259.c fix style problems
x86: irq_32.c fix style problems
x86: ioport.c fix style problems
...
Impact: cleanup
Fix:
WARNING: Use #include <linux/smp.h> instead of <asm/smp.h>
WARNING: Use #include <linux/percpu.h> instead of <asm/percpu.h>
WARNING: Use #include <linux/topology.h> instead of <asm/topology.h>
WARNING: EXPORT_SYMBOL(foo); should immediately follow its function/variable
ERROR: spaces required around that '?' (ctx:VxW)
ERROR: spaces required around that ':' (ctx:VxV)
total: 2 errors, 4 warnings
Signed-off-by: Jaswinder Singh Rajput <jaswinderrajput@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: use new cpumask API to reduce memory and stack usage
Allocate the following local cpumasks based on the number of cpus that
are present. References will use new cpumask API. (Currently only
modified for x86_64, x86_32 continues to use the *_map variants.)
cpu_callin_mask
cpu_callout_mask
cpu_initialized_mask
cpu_sibling_setup_mask
Provide the following accessor functions:
struct cpumask *cpu_sibling_mask(int cpu)
struct cpumask *cpu_core_mask(int cpu)
Other changes are when setting or clearing the cpu online, possible
or present maps, use the accessor functions.
Signed-off-by: Mike Travis <travis@sgi.com>
Acked-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: cleanup
__alloc_bootmem and __alloc_bootmem_node do panic
for us in case of fail so no need for additional
checks here.
Also lets use pr_*() macros for printing.
Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: New APIs
The old node_to_cpumask/node_to_pcibus returned a cpumask_t: these
return a pointer to a struct cpumask. Part of removing cpumasks from
the stack.
Also makes __pcibus_to_node take a const pointer.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Acked-by: Ingo Molnar <mingo@elte.hu>
Impact: cleanup, better debugging
This has proven useful in debugging, *before* we try to use
for_each_possible_cpu(). It also now shows nr_cpumask_bits.
Signed-off-by: Mike Travis <travis@sgi.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Impact: change calling convention of existing cpumask APIs
Most cpumask functions started with cpus_: these have been replaced by
cpumask_ ones which take struct cpumask pointers as expected.
These four functions don't have good replacement names; fortunately
they're rarely used, so we just change them over.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Mike Travis <travis@sgi.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: paulus@samba.org
Cc: mingo@redhat.com
Cc: tony.luck@intel.com
Cc: ralf@linux-mips.org
Cc: Greg Kroah-Hartman <gregkh@suse.de>
Cc: cl@linux-foundation.org
Cc: srostedt@redhat.com
so could spare some memory with small alignment in bootmem
also tighten the alignment checking, and make print out less debug info.
Signed-off-by: Yinghai Lu <yhlu.kernel@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Clean up and optimize cpumask_of_cpu(), by sharing all the zero words.
Instead of stupidly generating all possible i=0...NR_CPUS 2^i patterns
creating a huge array of constant bitmasks, realize that the zero words
can be shared.
In other words, on a 64-bit architecture, we only ever need 64 of these
arrays - with a different bit set in one single world (with enough zero
words around it so that we can create any bitmask by just offsetting in
that big array). And then we just put enough zeroes around it that we
can point every single cpumask to be one of those things.
So when we have 4k CPU's, instead of having 4k arrays (of 4k bits each,
with one bit set in each array - 2MB memory total), we have exactly 64
arrays instead, each 8k bits in size (64kB total).
And then we just point cpumask(n) to the right position (which we can
calculate dynamically). Once we have the right arrays, getting
"cpumask(n)" ends up being:
static inline const cpumask_t *get_cpu_mask(unsigned int cpu)
{
const unsigned long *p = cpu_bit_bitmap[1 + cpu % BITS_PER_LONG];
p -= cpu / BITS_PER_LONG;
return (const cpumask_t *)p;
}
This brings other advantages and simplifications as well:
- we are not wasting memory that is just filled with a single bit in
various different places
- we don't need all those games to re-create the arrays in some dense
format, because they're already going to be dense enough.
if we compile a kernel for up to 4k CPU's, "wasting" that 64kB of memory
is a non-issue (especially since by doing this "overlapping" trick we
probably get better cache behaviour anyway).
[ mingo@elte.hu:
Converted Linus's mails into a commit. See:
http://lkml.org/lkml/2008/7/27/156http://lkml.org/lkml/2008/7/28/320
Also applied a family filter - which also has the side-effect of leaving
out the bits where Linus calls me an idio... Oh, never mind ;-)
]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
Cc: Mike Travis <travis@sgi.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* Create the cpumask_of_cpu_map statically in the init data section
using NR_CPUS but replace it during boot up with one sized by
nr_cpu_ids (num possible cpus).
Signed-off-by: Mike Travis <travis@sgi.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Jack Steiner <steiner@sgi.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
There are a couple of places where (P)Dprintk is used which is an old
compile time enabled printk wrapper. Convert it to the generic
pr_debug().
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
* Strengthen the return type for the _node_to_cpumask_ptr to be
a const pointer. This adds compiler checking to insure that
node_to_cpumask_map[] is not changed inadvertently.
Signed-off-by: Mike Travis <travis@sgi.com>
Cc: "akpm@linux-foundation.org" <akpm@linux-foundation.org>
Cc: Yinghai Lu <yhlu.kernel@gmail.com>
Acked-by: Vegard Nossum <vegard.nossum@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
call it right after we are done with MADT/mptable handling, instead of
doing that in setup_per_cpu_areas() later on...
this way for_possible_cpu() can be used early.
Signed-off-by: Yinghai Lu <yhlu.kernel@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* When CONFIG_DEBUG_PER_CPU_MAPS is set, the node passed to
node_to_cpumask and node_to_cpumask_ptr should be validated.
If invalid, then a dump_stack is performed and a zero cpumask
is returned.
v2: Slightly different version to remove a compiler warning.
v3: Redone to reflect moving setup.c -> setup_percpu.c
Signed-off-by: Mike Travis <travis@sgi.com>
Cc: Vegard Nossum <vegard.nossum@gmail.com>
Cc: "akpm@linux-foundation.org" <akpm@linux-foundation.org>
Cc: Yinghai Lu <yhlu.kernel@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
some functions need to be moved to setup_numa.c
after we merge setup32/64.c, some funcs need to be moved back to setup.c
Signed-off-by: Yinghai Lu <yhlu.kernel@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
