This is based on work done by jeff@ and jhb@, as well as the numa.diff
patch that has been circulating when someone asks for first-touch NUMA
on -10 or -11.
* Introduce a simple set of VM policy and iterator types.
* tie the policy types into the vm_phys path for now, mirroring how
the initial first-touch allocation work was enabled.
* add syscalls to control changing thread and process defaults.
* add a global NUMA VM domain policy.
* implement a simple cascade policy order - if a thread policy exists, use it;
if a process policy exists, use it; use the default policy.
* processes inherit policies from their parent processes, threads inherit
policies from their parent threads.
* add a simple tool (numactl) to query and modify default thread/process
policities.
* add documentation for the new syscalls, for numa and for numactl.
* re-enable first touch NUMA again by default, as now policies can be
set in a variety of methods.
This is only relevant for very specific workloads.
This doesn't pretend to be a final NUMA solution.
The previous defaults in -HEAD (with MAXMEMDOM set) can be achieved by
'sysctl vm.default_policy=rr'.
This is only relevant if MAXMEMDOM is set to something other than 1.
Ie, if you're using GENERIC or a modified kernel with non-NUMA, then
this is a glorified no-op for you.
Thank you to Norse Corp for giving me access to rather large
(for FreeBSD!) NUMA machines in order to develop and verify this.
Thank you to Dell for providing me with dual socket sandybridge
and westmere v3 hardware to do NUMA development with.
Thank you to Scott Long at Netflix for providing me with access
to the two-socket, four-domain haswell v3 hardware.
Thank you to Peter Holm for running the stress testing suite
against the NUMA branch during various stages of development!
Tested:
* MIPS (regression testing; non-NUMA)
* i386 (regression testing; non-NUMA GENERIC)
* amd64 (regression testing; non-NUMA GENERIC)
* westmere, 2 socket (thankyou norse!)
* sandy bridge, 2 socket (thankyou dell!)
* ivy bridge, 2 socket (thankyou norse!)
* westmere-EX, 4 socket / 1TB RAM (thankyou norse!)
* haswell, 2 socket (thankyou norse!)
* haswell v3, 2 socket (thankyou dell)
* haswell v3, 2x18 core (thankyou scott long / netflix!)
* Peter Holm ran a stress test suite on this work and found one
issue, but has not been able to verify it (it doesn't look NUMA
related, and he only saw it once over many testing runs.)
* I've tested bhyve instances running in fixed NUMA domains and cpusets;
all seems to work correctly.
Verified:
* intel-pcm - pcm-numa.x and pcm-memory.x, whilst selecting different
NUMA policies for processes under test.
Review:
This was reviewed through phabricator (https://reviews.freebsd.org/D2559)
as well as privately and via emails to freebsd-arch@. The git history
with specific attributes is available at https://github.com/erikarn/freebsd/
in the NUMA branch (https://github.com/erikarn/freebsd/compare/local/adrian_numa_policy).
This has been reviewed by a number of people (stas, rpaulo, kib, ngie,
wblock) but not achieved a clear consensus. My hope is that with further
exposure and testing more functionality can be implemented and evaluated.
Notes:
* The VM doesn't handle unbalanced domains very well, and if you have an overly
unbalanced memory setup whilst under high memory pressure, VM page allocation
may fail leading to a kernel panic. This was a problem in the past, but it's
much more easily triggered now with these tools.
* This work only controls the path through vm_phys; it doesn't yet strongly/predictably
affect contigmalloc, KVA placement, UMA, etc. So, driver placement of memory
isn't really guaranteed in any way. That's next on my plate.
Sponsored by: Norse Corp, Inc.; Dell
a basic ACPI SLIT table parser.
For now this just exports the map via sysctl; it'll eventually be useful
to userland when there's more useful NUMA support in -HEAD.
* Add an optional mem_locality map;
* add a mapping function taking from/to domain and returning the
relative cost, or -1 if it's not available;
* Add a very basic SLIT parser to x86 ACPI.
Differential Revision: https://reviews.freebsd.org/D2460
Reviewed by: rpaulo, stas, jhb
Sponsored by: Norse Corp, Inc (hardware, coding); Dell (hardware)
managing pages from different address ranges. Generally speaking, this
feature is used to increase the likelihood that physical pages are
available that can meet special DMA requirements or can be accessed through
a limited-coverage direct mapping (e.g., MIPS). However, prior to this
change, the configuration of the free lists was static, i.e., it was
determined at compile time. Consequentally, free lists could be created
for address ranges that held no actual pages, for example, on 32-bit MIPS-
based systems with 512 MB or less of physical memory. This change makes
the creation of the free lists dynamic, i.e., it is based on the available
physical memory at boot time.
On 64-bit x86-based systems with 64 GB or more of physical memory, create
free lists for managing pages with physical addresses below 4 GB. This
change is to address reported problems with initializing devices that
require the allocation of physical pages below 4 GB on some systems with
128 GB or more of physical memory.
PR: 185727
Differential Revision: https://reviews.freebsd.org/D1274
Reviewed by: jhb, kib
MFC after: 3 weeks
Sponsored by: EMC / Isilon Storage Division
on i386 PAE. Previously, VM_PHYSSEG_SPARSE could not be used on amd64 and
i386 because vm_page_startup() would not create vm_page structures for the
kernel page table pages allocated during pmap_bootstrap() but those vm_page
structures are needed when the kernel attempts to promote the corresponding
kernel virtual addresses to superpage mappings. To address this problem, a
new public function, vm_phys_add_seg(), is introduced and vm_phys_init() is
updated to reflect the creation of vm_phys_seg structures by calls to
vm_phys_add_seg().
Discussed with: Svatopluk Kraus
MFC after: 3 weeks
Sponsored by: EMC / Isilon Storage Division
To reduce the diff struct pcu.cnt field was not renamed, so
PCPU_OP(cnt.field) is still used. pc_cnt and pcpu are also used in
kvm(3) and vmstat(8). The goal was to not affect externally used KPI.
Bump __FreeBSD_version_ in case some out-of-tree module/code relies on the
the global cnt variable.
Exp-run revealed no ports using it directly.
No objection from: arch@
Sponsored by: EMC / Isilon Storage Division
into threads each processing queue in a single domain. The structure
of the pagedaemons and queues is kept intact, most of the changes come
from the need for code to find an owning page queue for given page,
calculated from the segment containing the page.
The tie between NUMA domain and pagedaemon thread/pagequeue split is
rather arbitrary, the multithreaded daemon could be allowed for the
single-domain machines, or one domain might be split into several page
domains, to further increase concurrency.
Right now, each pagedaemon thread tries to reach the global target,
precalculated at the start of the pass. This is not optimal, since it
could cause excessive page deactivation and freeing. The code should
be changed to re-check the global page deficit state in the loop after
some number of iterations.
The pagedaemons reach the quorum before starting the OOM, since one
thread inability to meet the target is normal for split queues. Only
when all pagedaemons fail to produce enough reusable pages, OOM is
started by single selected thread.
Launder is modified to take into account the segments layout with
regard to the region for which cleaning is performed.
Based on the preliminary patch by jeff, sponsored by EMC / Isilon
Storage Division.
Reviewed by: alc
Tested by: pho
Sponsored by: The FreeBSD Foundation
freelist.
o Split the pool of free pages queues really by domain and not rely on
definition of VM_RAW_NFREELIST.
o For MAXMEMDOM > 1, wrap the RR allocation logic into a specific
function that is called when calculating the allocation domain.
The RR counter is kept, currently, per-thread.
In the future it is expected that such function evolves in a real
policy decision referee, based on specific informations retrieved by
per-thread and per-vm_object attributes.
o Add the concept of "probed domains" under the form of vm_ndomains.
It is responsibility for every architecture willing to support multiple
memory domains to correctly probe vm_ndomains along with mem_affinity
segments attributes. Those two values are supposed to remain always
consistent.
Please also note that vm_ndomains and td_dom_rr_idx are both int
because segments already store domains as int. Ideally u_int would
have much more sense. Probabilly this should be cleaned up in the
future.
o Apply RR domain selection also to vm_phys_zero_pages_idle().
Sponsored by: EMC / Isilon storage division
Partly obtained from: jeff
Reviewed by: alc
Tested by: jeff
for allocation of fictitious pages, for which PHYS_TO_VM_PAGE()
returns proper fictitious vm_page_t. The range should be de-registered
after consumer stopped using it.
De-inline the PHYS_TO_VM_PAGE() since it now carries code to iterate
over registered ranges.
A hash container might be developed instead of range registration
interface, and fake pages could be put automatically into the hash,
were PHYS_TO_VM_PAGE() could look them up later. This should be
considered before the MFC of the commit is done.
Sponsored by: The FreeBSD Foundation
Reviewed by: alc
MFC after: 1 month
yielding a new public interface, vm_page_alloc_contig(). This new function
addresses some of the limitations of the current interfaces, contigmalloc()
and kmem_alloc_contig(). For example, the physically contiguous memory that
is allocated with those interfaces can only be allocated to the kernel vm
object and must be mapped into the kernel virtual address space. It also
provides functionality that vm_phys_alloc_contig() doesn't, such as wiring
the returned pages. Moreover, unlike that function, it respects the low
water marks on the paging queues and wakes up the page daemon when
necessary. That said, at present, this new function can't be applied to all
types of vm objects. However, that restriction will be eliminated in the
coming weeks.
From a design standpoint, this change also addresses an inconsistency
between vm_phys_alloc_contig() and the other vm_phys_alloc*() functions.
Specifically, vm_phys_alloc_contig() manipulated vm_page fields that other
functions in vm/vm_phys.c didn't. Moreover, vm_phys_alloc_contig() knew
about vnodes and reservations. Now, vm_page_alloc_contig() is responsible
for these things.
Reviewed by: kib
Discussed with: jhb
eliminating duplicated code in the various pmap implementations.
Micro-optimize vm_phys_free_pages().
Introduce vm_phys_free_contig(). It is fast routine for freeing an
arbitrary number of physically contiguous pages. In particular, it
doesn't require the number of pages to be a power of two.
Use "u_long" instead of "unsigned long".
Bruce Evans (bde@) has convinced me that the "boundary" parameters
to kmem_alloc_contig(), vm_phys_alloc_contig(), and
vm_reserv_reclaim_contig() should be of type "vm_paddr_t" and not
"u_long". Make this change.
now it uses a very dumb first-touch allocation policy. This will change in
the future.
- Each architecture indicates the maximum number of supported memory domains
via a new VM_NDOMAIN parameter in <machine/vmparam.h>.
- Each cpu now has a PCPU_GET(domain) member to indicate the memory domain
a CPU belongs to. Domain values are dense and numbered from 0.
- When a platform supports multiple domains, the default freelist
(VM_FREELIST_DEFAULT) is split up into N freelists, one for each domain.
The MD code is required to populate an array of mem_affinity structures.
Each entry in the array defines a range of memory (start and end) and a
domain for the range. Multiple entries may be present for a single
domain. The list is terminated by an entry where all fields are zero.
This array of structures is used to split up phys_avail[] regions that
fall in VM_FREELIST_DEFAULT into per-domain freelists.
- Each memory domain has a separate lookup-array of freelists that is
used when fulfulling a physical memory allocation. Right now the
per-domain freelists are listed in a round-robin order for each domain.
In the future a table such as the ACPI SLIT table may be used to order
the per-domain lookup lists based on the penalty for each memory domain
relative to a specific domain. The lookup lists may be examined via a
new vm.phys.lookup_lists sysctl.
- The first-touch policy is implemented by using PCPU_GET(domain) to
pick a lookup list when allocating memory.
Reviewed by: alc
alc@.
The UMA zone based allocation is replaced by a scheme that creates
a new free page list for the KSEG0 region, and a new function
in sys/vm that allocates pages from a specific free page list.
This also fixes a race condition introduced by the UMA based page table
page allocation code. Dropping the page queue and pmap locks before
the call to uma_zfree, and re-acquiring them afterwards will introduce
a race condtion(noted by alc@).
The changes are :
- Revert the earlier changes in MIPS pmap.c that added UMA zone for
page table pages.
- Add a new freelist VM_FREELIST_HIGHMEM to MIPS vmparam.h for memory that
is not directly mapped (in 32bit kernel). Normal page allocations will first
try the HIGHMEM freelist and then the default(direct mapped) freelist.
- Add a new function 'vm_page_t vm_page_alloc_freelist(int flind, int
order, int req)' to vm/vm_page.c to allocate a page from a specified
freelist. The MIPS page table pages will be allocated using this function
from the freelist containing direct mapped pages.
- Move the page initialization code from vm_phys_alloc_contig() to a
new function vm_page_alloc_init(), and use this function to initialize
pages in vm_page_alloc_freelist() too.
- Split the function vm_phys_alloc_pages(int pool, int order) to create
vm_phys_alloc_freelist_pages(int flind, int pool, int order), and use
this function from both vm_page_alloc_freelist() and vm_phys_alloc_pages().
Reviewed by: alc
dependent memory attributes:
Rename vm_cache_mode_t to vm_memattr_t. The new name reflects the
fact that there are machine-dependent memory attributes that have
nothing to do with controlling the cache's behavior.
Introduce vm_object_set_memattr() for setting the default memory
attributes that will be given to an object's pages.
Introduce and use pmap_page_{get,set}_memattr() for getting and
setting a page's machine-dependent memory attributes. Add full
support for these functions on amd64 and i386 and stubs for them on
the other architectures. The function pmap_page_set_memattr() is also
responsible for any other machine-dependent aspects of changing a
page's memory attributes, such as flushing the cache or updating the
direct map. The uses include kmem_alloc_contig(), vm_page_alloc(),
and the device pager:
kmem_alloc_contig() can now be used to allocate kernel memory with
non-default memory attributes on amd64 and i386.
vm_page_alloc() and the device pager will set the memory attributes
for the real or fictitious page according to the object's default
memory attributes.
Update the various pmap functions on amd64 and i386 that map pages to
incorporate each page's memory attributes in the mapping.
Notes: (1) Inherent to this design are safety features that prevent
the specification of inconsistent memory attributes by different
mappings on amd64 and i386. In addition, the device pager provides a
warning when a device driver creates a fictitious page with memory
attributes that are inconsistent with the real page that the
fictitious page is an alias for. (2) Storing the machine-dependent
memory attributes for amd64 and i386 as a dedicated "int" in "struct
md_page" represents a compromise between space efficiency and the ease
of MFCing these changes to RELENG_7.
In collaboration with: jhb
Approved by: re (kib)
required by video card drivers. Specifically, this change introduces
vm_cache_mode_t with an appropriate VM_CACHE_DEFAULT definition on all
architectures. In addition, this changes adds a vm_cache_mode_t parameter
to kmem_alloc_contig() and vm_phys_alloc_contig(). These will be the
interfaces for allocating mapped kernel memory and physical memory,
respectively, with non-default cache modes.
In collaboration with: jhb
page to be in the free lists. Instead, it now returns TRUE if it
removed the page from the free lists and FALSE if the page was not
in the free lists.
This change is required to support superpage reservations. Specifically,
once reservations are introduced, a cached page can either be in the
free lists or a reservation.
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
vm_phys_free_pages(). Rename vm_phys_alloc_pages_locked() to
vm_phys_alloc_pages() and vm_phys_free_pages_locked() to
vm_phys_free_pages(). Add comments regarding the need for the free page
queues lock to be held by callers to these functions. No functional
changes.
Approved by: re (hrs)
to the build.
This allocator uses a binary buddy system with a twist. First and
foremost, this allocator is required to support the implementation of
superpages. As a side effect, it enables a more robust implementation
of contigmalloc(9). Moreover, this reimplementation of
contigmalloc(9) eliminates the acquisition of Giant by
contigmalloc(..., M_NOWAIT, ...).
The twist is that this allocator tries to reduce the number of TLB
misses incurred by accesses through a direct map to small, UMA-managed
objects and page table pages. Roughly speaking, the physical pages
that are allocated for such purposes are clustered together in the
physical address space. The performance benefits vary. In the most
extreme case, a uniprocessor kernel running on an Opteron, I measured
an 18% reduction in system time during a buildworld.
This allocator does not implement page coloring. The reason is that
superpages have much the same effect. The contiguous physical memory
allocation necessary for a superpage is inherently colored.
Finally, the one caveat is that this allocator does not effectively
support prezeroed pages. I hope this is temporary. On i386, this is
a slight pessimization. However, on amd64, the beneficial effects of
the direct-map optimization outweigh the ill effects. I speculate
that this is true in general of machines with a direct map.
Approved by: re
