| Age | Commit message (Collapse) | Author |
|---|
|
The slab allocator has grown to use multiple ways to allocate slabs
as well as track them, which got a little messy. One consequence is
the breaking of the KMEM_CF_VERIFY option. In order to make the code
less confusing, this change expresses all options as explicit cache
flags and clearly defines their relationships.
The special kmem_slab and vm_map_entry caches are initialized
accordingly.
* kern/slab.c (KMEM_CF_DIRECTMAP): Rename to ...
(KMEM_CF_PHYSMEM): ... this new macro.
(KMEM_CF_DIRECT): Restore macro.
(KMEM_CF_USE_TREE, KMEM_CF_USE_PAGE): New macros.
(KMEM_CF_VERIFY): Update value.
(kmem_pagealloc_directmap): Rename to...
(kmem_pagealloc_physmem): ... this new function.
(kmem_pagefree_directmap): Rename to ...
(kmem_pagefree_physmem): ... this new function.
(kmem_pagealloc, kmem_pagefree): Update macro names.
(kmem_slab_use_tree): Remove function.
(kmem_slab_create, kmem_slab_destroy): Update according to the new
cache flags.
(kmem_cache_compute_sizes): Rename to ...
(kmem_cache_compute_properties): ... this new function, and update
to properly set cache flags.
(kmem_cache_init): Update call to kmem_cache_compute_properties.
(kmem_cache_alloc_from_slab): Check KMEM_CF_USE_TREE instead of
calling the defunct kmem_slab_use_tree function.
(kmem_cache_free_to_slab): Update according to the new cache flags.
kmem_cache_free_verify): Add assertion.
(slab_init): Update initialization of kmem_slab_cache.
* kern/slab.h (KMEM_CACHE_DIRECTMAP): Rename to ...
(KMEM_CACHE_PHYSMEM): ... this new macro.
* vm/vm_map.c (vm_map_init): Update initialization of vm_map_entry_cache.
|
|
Since the slab allocator has been changed to sit directly on top of the
physical allocator, failures caused by fragmentation have been observed,
as one could expect. This change makes the slab allocator revert to
kernel virtual memory when allocating larger-than-page slabs. This
solution is motivated in part to avoid the complexity of other solutions
such as page mobility, and also because a microkernel cannot be extended
to new arbitrary uncontrolled usage patterns such as a monolithic kernel
with loadable modules. As such, large objects are rare, and their use
infrequent, which is compatible with the use of kernel virtual memory.
* kern/slab.c: Update module description.
(KMEM_CF_SLAB_EXTERNAL, KMEM_CF_VERIFY): Update values.
(KMEM_CF_DIRECT): Remove macro.
(KMEM_CF_DIRECTMAP): New macro.
(kmem_pagealloc_directmap, kmem_pagefree_directmap,
kmem_pagealloc_virtual, kmem_pagefree_virtual): New functions.
(kmem_pagealloc, kmem_pagefree, kmem_slab_create, kmem_slab_destroy,
kalloc, kfree): Update to use the new pagealloc functions.
(kmem_cache_compute_sizes): Update the algorithm used to determine slab
size and other cache properties.
(kmem_slab_use_tree, kmem_cache_free_to_slab, host_slab_info): Update to
correctly use the cache flags.
(slab_init): Add KMEM_CACHE_DIRECTMAP to the kmem_slab_cache init flags.
* kern/slab.h (KMEM_CACHE_VERIFY): Change value.
(KMEM_CACHE_DIRECTMAP): New macro.
* vm/vm_map.c (vm_map_init): Add KMEM_CACHE_DIRECTMAP to the
vm_map_entry_cache init flags.
|
|
Now that the slab allocator doesn't use kernel virtual memory any more,
this map has become irrelevant.
* kern/slab.c (KMEM_MAP_SIZE): Remove macro.
(kmem_map_store, kmem_map): Remove variables.
(slab_init): Remove call kmem_submap.
* kern/slab.h (kmem_map): Remove extern declaration.
|
|
A few errors were introduced in the latest changes.
o Add VM_PAGE_WAIT calls around physical allocation attempts in case of
memory exhaustion.
o Fix stack release.
o Fix memory exhaustion report.
o Fix free page accounting.
* kern/slab.c (kmem_pagealloc, kmem_pagefree): New functions
(kmem_slab_create, kmem_slab_destroy, kalloc, kfree): Use kmem_pagealloc
and kmem_pagefree instead of the raw page allocation functions.
(kmem_cache_compute_sizes): Don't store slab order.
* kern/slab.h (struct kmem_cache): Remove `slab_order' member.
* kern/thread.c (stack_alloc): Call VM_PAGE_WAIT in case of memory
exhaustion.
(stack_collect): Call vm_page_free_contig instead of kmem_free to
release pages.
* vm/vm_page.c (vm_page_seg_alloc): Fix memory exhaustion report.
(vm_page_setup): Don't update vm_page_free_count.
(vm_page_free_pa): Check page parameter.
(vm_page_mem_free): New function.
* vm/vm_page.h (vm_page_free_count): Remove extern declaration.
(vm_page_mem_free): New prototype.
* vm/vm_pageout.c: Update comments not to refer to vm_page_free_count.
(vm_pageout_scan, vm_pageout_continue, vm_pageout): Use vm_page_mem_free
instead of vm_page_free_count, update types accordingly.
* vm/vm_resident.c (vm_page_free_count, vm_page_free_count_minimum):
Remove variables.
(vm_page_free_avail): New variable.
(vm_page_bootstrap, vm_page_grab, vm_page_release, vm_page_grab_contig,
vm_page_free_contig, vm_page_wait): Use vm_page_mem_free instead of vm_page_free_count,
update types accordingly, don't set vm_page_free_count_minimum.
* vm/vm_user.c (vm_statistics): Likewise.
|
|
In order to increase the amount of memory available for kernel objects,
without reducing the amount of memory available for user processes,
a new allocation strategy is introduced in this change.
Instead of allocating kernel objects out of kernel virtual memory,
the slab allocator directly uses the direct mapping of physical
memory as its backend. This largely increases the kernel heap, and
removes the need for address translation updates.
In order to allow this strategy, an assumption made by the interrupt
code had to be removed. In addition, kernel stacks are now also
allocated directly from the physical allocator.
* i386/i386/db_trace.c: Include i386at/model_dep.h
(db_i386_reg_value): Update stack check.
* i386/i386/locore.S (trap_from_kernel, all_intrs,
int_from_intstack): Update interrupt handling.
* i386/i386at/model_dep.c: Include kern/macros.h.
(int_stack, int_stack_base): New variables.
(int_stack_high): Remove variable.
(i386at_init): Update interrupt stack initialization.
* i386/i386at/model_dep.h: Include i386/vm_param.h.
(int_stack_top, int_stack_base): New extern declarations.
(ON_INT_STACK): New macro.
* kern/slab.c: Include vm/vm_page.h
(KMEM_CF_NO_CPU_POOL, KMEM_CF_NO_RECLAIM): Remove macros.
(kmem_pagealloc, kmem_pagefree, kalloc_pagealloc, kalloc_pagefree): Remove
functions.
(kmem_slab_create): Allocate slab pages directly from the physical allocator.
(kmem_slab_destroy): Release slab pages directly to the physical allocator.
(kmem_cache_compute_sizes): Update the slab size computation algorithm to
return a power-of-two suitable for the physical allocator.
(kmem_cache_init): Remove custom allocation function pointers.
(kmem_cache_reap): Remove check on KMEM_CF_NO_RECLAIM.
(slab_init, kalloc_init): Update calls to kmem_cache_init.
(kalloc, kfree): Directly fall back on the physical allocator for big
allocation sizes.
(host_slab_info): Remove checks on defunct flags.
* kern/slab.h (kmem_slab_alloc_fn_t, kmem_slab_free_fn_t): Remove types.
(struct kmem_cache): Add `slab_order' member, remove `slab_alloc_fn' and
`slab_free_fn' members.
(KMEM_CACHE_NOCPUPOOL, KMEM_CACHE_NORECLAIM): Remove macros.
(kmem_cache_init): Update prototype, remove custom allocation functions.
* kern/thread.c (stack_alloc): Allocate stacks from the physical allocator.
* vm/vm_map.c (vm_map_kentry_cache, kentry_data, kentry_data_size): Remove
variables.
(kentry_pagealloc): Remove function.
(vm_map_init): Update calls to kmem_cache_init, remove initialization of
vm_map_kentry_cache.
(vm_map_create, _vm_map_entry_dispose, vm_map_copyout): Unconditionnally
use vm_map_entry_cache.
* vm/vm_map.h (kentry_data, kentry_data_size, kentry_count): Remove extern
declarations.
* vm/vm_page.h (VM_PT_STACK): New page type.
* device/dev_lookup.c (dev_lookup_init): Update calls to kmem_cache_init.
* device/dev_pager.c (dev_pager_hash_init, device_pager_init): Likewise.
* device/ds_routines.c (mach_device_init, mach_device_trap_init): Likewise.
* device/net_io.c (net_io_init): Likewise.
* i386/i386/fpu.c (fpu_module_init): Likewise.
* i386/i386/machine_task.c (machine_task_module_init): Likewise.
* i386/i386/pcb.c (pcb_module_init): Likewise.
* i386/intel/pmap.c (pmap_init): Likewise.
* ipc/ipc_init.c (ipc_bootstrap): Likewise.
* ipc/ipc_marequest.c (ipc_marequest_init): Likewise.
* kern/act.c (global_act_init): Likewise.
* kern/processor.c (pset_sys_init): Likewise.
* kern/rdxtree.c (rdxtree_cache_init): Likewise.
* kern/task.c (task_init): Likewise.
* vm/memory_object_proxy.c (memory_object_proxy_init): Likewise.
* vm/vm_external.c (vm_external_module_initialize): Likewise.
* vm/vm_fault.c (vm_fault_init): Likewise.
* vm/vm_object.c (vm_object_bootstrap): Likewise.
* vm/vm_resident.c (vm_page_module_init): Likewise.
(vm_page_bootstrap): Remove initialization of kentry_data.
|
|
This change replaces the historical page allocator with a buddy allocator
implemented in vm/vm_page.c. This allocator allows easy contiguous allocations
and also manages memory inside segments. In a future change, these segments
will be used to service requests with special constraints, such as "usable
for 16-bits DMA" or "must be part of the direct physical mapping".
* Makefrag.am (libkernel_a_SOURCES): Add vm/vm_page.c.
* i386/Makefrag.am (libkernel_a_SOURCES): Add i386/i386at/biosmem.{c,h}.
* i386/i386/vm_param.h: Include kern/macros.h.
(VM_PAGE_DMA_LIMIT, VM_PAGE_MAX_SEGS, VM_PAGE_DMA32_LIMIT,
VM_PAGE_DIRECTMAP_LIMIT, VM_PAGE_HIGHMEM_LIMIT, VM_PAGE_SEG_DMA,
VM_PAGE_SEG_DMA32, VM_PAGE_SEG_DIRECTMAP, VM_PAGE_SEG_HIGHMEM): New macros.
* i386/i386at/model_dep.c: Include i386at/biosmem.h.
(avail_next, avail_remaining): Remove variables.
(mem_size_init): Remove function.
(i386at_init): Initialize and use the biosmem module for early physical
memory management.
(pmap_free_pages): Return phys_last_addr instead of avail_remaining.
(init_alloc_aligned): Turn into a wrapper for biosmem_bootalloc.
(pmap_grab_page): Directly call init_alloc_aligned instead of pmap_next_page.
* i386/include/mach/i386/vm_types.h (phys_addr_t): New type.
* kern/bootstrap.c (free_bootstrap_pages): New function.
(bootstrap_create): Call free_bootstrap_pages instead of vm_page_create.
* kern/cpu_number.h (CPU_L1_SIZE): New macro.
* kern/slab.h: Include kern/cpu_number.h.
(CPU_L1_SIZE): Remove macro, moved to kern/cpu_number.h.
* kern/startup.c (setup_main): Change the value of machine_info.memory_size.
* linux/dev/glue/glue.h (alloc_contig_mem, free_contig_mem): Update prototypes.
* linux/dev/glue/kmem.c (linux_kmem_init): Don't use defunct page queue.
* linux/dev/init/main.c (linux_init): Don't free unused memory.
(alloc_contig_mem, free_contig_mem): Turn into wrappers for the vm_page
allocator.
* linux/pcmcia-cs/glue/ds.c (PAGE_SHIFT): Don't undefine.
* vm/pmap.h (pmap_startup, pmap_next_page): Remove prototypes.
* vm/vm_fault.c (vm_fault_page): Update calls to vm_page_convert.
* vm/vm_init.c (vm_mem_init): Call vm_page_info_all.
* vm/vm_object.c (vm_object_page_map): Update call to vm_page_init.
* vm/vm_page.h (vm_page_queue_free): Remove variable declaration.
(vm_page_create, vm_page_release_fictitious, vm_page_release): Remove
declarations.
(vm_page_convert, vm_page_init): Update prototypes.
(vm_page_grab_contig, vm_page_free_contig): New prototypes.
* vm/vm_resident.c (vm_page_template, vm_page_queue_free,
vm_page_big_pagenum): Remove variables.
(vm_page_bootstrap): Update and call vm_page_setup.
(pmap_steal_memory): Update and call vm_page_bootalloc.
(pmap_startup, vm_page_create, vm_page_grab_contiguous_pages): Remove functions.
(vm_page_init_template, vm_page_grab_contig,
vm_page_free_contig): New functions.
(vm_page_init): Update and call vm_page_init_template.
(vm_page_release_fictitious): Make static.
(vm_page_more_fictitious): Update call to vm_page_init.
(vm_page_convert): Rewrite to comply with vm_page.
(vm_page_grab): Update and call vm_page_alloc_pa.
(vm_page_release): Update and call vm_page_free_pa.
|
|
* ddb/db_command.c (db_show_cmds): Add `slabinfo'.
* kern/slab.c (slab_info): Generalize so that it can be used with
different printf-like functions, and turn it into a static function.
(slab_info): New wrapper retaining the old behaviour.
(db_show_slab_info): New wrapper that uses `db_printf' instead.
* kern/slab.h (db_show_slab_info): New declaration.
|
|
* kern/slab.h (struct kmem_cache): Align kmem_cache objects using
__cacheline_aligned.
|
|
* kern/slab.h (struct kmem_cache): Reorder the fields so that all hot
fields are within the first cache line.
|
|
* kern/slab.h (KMEM_CACHE_NAME_SIZE): Explain the significance of the
chosen length.
|
|
Previously, the size of struct kmem_cache was 136 bytes, just eight
bytes larger than 128 bytes, which is typically two cache lines on
today's CPUs.
By reducing the size of the name field which holds a human-readable
description by eight bytes to 24 bytes, the struct kmem_cache can be
made fit into two cache lines. This change should not affect the
usefulness of this field. For reference, the length of the largest
hard-coded name is 17.
* kern/slab.h (KMEM_CACHE_NAME_SIZE): Define to 24.
|
|
Don't enforce strong ordering of partial slabs. Separating partial slabs
from free slabs is already effective against fragmentation, and sorting
would sometimes cause pathological scalability issues. In addition, store
new slabs (whether free or partial) in LIFO order for better cache usage.
* kern/slab.c (kmem_cache_grow): Insert new slab at the head of the slabs list.
(kmem_cache_alloc_from_slab): Likewise. In addition, don't sort partial slabs.
(kmem_cache_free_to_slab): Likewise.
* kern/slab.h: Remove comment about partial slabs sorting.
|
|
The purpose of this function is to allow kernel code to display the state
of the slab caches in situations where the host_slab_info RPC wouldn't be
available, e.g. before a panic.
* kern/slab.c (slab_info): New function.
* kern/slab.h: Add declaration for slab_info.
|
|
Maksym and I have assigned copyright to the Free Software Foundation.
In addition, restore the original upstream copyrights for correctness.
* kern/list.h: Fix copyright assignment.
* kern/rbtree.c: Likewise.
* kern/rbtree.h: Likewise.
* kern/rbtree_i.h: Likewise.
* kern/slab.c: Likewise.
* kern/slab.h: Likewise.
|
|
As it is intended to completely replace the zone allocator, remove it on
the way. So long to the venerable code !
* Makefrag.am (libkernel_a_SOURCES): Add kern/slab.{c,h}, remove kern/kalloc.c
and kern/zalloc.{c,h}.
* configfrag.ac (SLAB_VERIFY, SLAB_USE_CPU_POOLS): Add defines.
* i386/Makefrag.am (libkernel_a_SOURCES): Remove i386/i386/zalloc.h.
* i386/configfrag.ac (CPU_L1_SHIFT): Remove define.
* include/mach_debug/slab_info.h: New file.
* kern/slab.c: Likewise.
* kern/slab.h: Likewise.
* i386/i386/zalloc.h: Remove file.
* include/mach_debug/zone_info.h: Likewise.
* kern/kalloc.c: Likewise.
* kern/zalloc.c: Likewise.
* kern/zalloc.h: Likewise.
|
