: Containers(V2)- Generic Linux kernel changes
Subsystems (10)
This patch contains changes to generic part of kernel code. These changes tracks events like new task creating, task's exit, new page allocation (both file and anonymous) etc.
fs/inode.c | 3 +++ include/linux/fs.h | 5 +++++ include/linux/mm_inline.h | 4 ++++ include/linux/mm_types.h | 4 ++++ include/linux/sched.h | 6 ++++++ init/Kconfig | 8 ++++++++ kernel/Makefile | 1 + kernel/exit.c | 2 ++ kernel/fork.c | 9 +++++++++ mm/Makefile | 2 ++ mm/filemap.c | 4 ++++ mm/page_alloc.c | 3 +++ mm/rmap.c | 8 +++++++- mm/swap.c | 1 + mm/vmscan.c | 1 + 15 files changed, 60 insertions(+), 1 deletion(-)
include/linux/mm_inline.h +4 -0
@@ -1,9 +1,11 @@ +#include <linux/container.h> static inline void add_page_to_active_list (struct zone *zone,struct page *page) {list_add (&page->lru, &zone->active_list); zone->nr_active++; +container_inc_activepage_count (page); } static inline void
@@ -23,6 +25,7 @@ add_page_to_inactive_list_tail(struct zostatic inline void del_page_from_active_list (struct zone *zone,struct page *page) { +container_dec_activepage_count (page);list_del (&page->lru); zone->nr_active--; }
@@ -41,6 +44,7 @@ del_page_from_lru(struct zone *zone, strif (PageActive (page)) {__ClearPageActive (page); zone->nr_active--; +container_dec_activepage_count (page); }else { zone->nr_inactive--; }
include/linux/fs.h +5 -0
@@ -449,6 +449,7 @@ struct address_space_operations { };struct backing_dev_info ; +struct container_struct ;struct address_space {struct inode *host;/* owner: inode, block_device */ struct radix_tree_root page_tree;/* radix tree of all pages */
@@ -465,6 +466,10 @@ struct address_space {struct backing_dev_info *backing_dev_info;/* device readahead, etc */ spinlock_t private_lock;/* for use by the address_space */ struct list_head private_list;/* ditto */ +#ifdef CONFIG_CONTAINERS +struct container_struct *ctn;/* Pointer to container */ +struct list_head ctn_mapping_list;/* List of files belonging to same container*/ +#endif struct address_space *assoc_mapping;/* ditto */ } __attribute__((aligned(sizeof(long))));/*
include/linux/sched.h +6 -0
@@ -298,6 +298,8 @@ typedef unsigned long mm_counter_t; (mm)->hiwater_vm = (mm)->total_vm; \ }while (0 ) +struct container_struct; +struct mm_struct {struct vm_area_struct * mmap;/* list of VMAs */ struct rb_root mm_rb;
@@ -1046,6 +1048,10 @@ struct task_struct {#ifdef CONFIG_TASK_DELAY_ACCT struct task_delay_info *delays;#endif +#ifdef CONFIG_CONTAINERS +struct container_struct *ctn; +struct list_head ctn_task_list;/*List of processes belonging to container */ +#endif };static inline pid_t process_group (struct task_struct *tsk)
include/linux/mm_types.h +4 -0
@@ -6,6 +6,7 @@#include <linux/list.h> #include <linux/spinlock.h> +struct container_struct ;struct address_space ;/*
@@ -48,6 +49,9 @@ struct page {struct list_head lru;/* Pageout list, eg. active_list * protected by zone->lru_lock ! */ +#ifdef CONFIG_CONTAINERS +struct container_struct *ctn;/* Pointer to container, may be NULL */ +#endif /* * On machines where all RAM is mapped into kernel address space, * we can simply calculate the virtual address. On machines with
mm/filemap.c +4 -0
@@ -30,6 +30,7 @@#include <linux/security.h> #include <linux/syscalls.h> #include <linux/cpuset.h> +#include <linux/container.h> #include "filemap.h" #include "internal.h"
@@ -126,6 +127,7 @@ void __remove_from_page_cache(struct pag page->mapping =NULL ; mapping->nrpages--;__dec_zone_page_state (page,NR_FILE_PAGES ); +container_dec_filepage_count (page); }EXPORT_SYMBOL (__remove_from_page_cache);
@@ -461,6 +463,8 @@ int add_to_page_cache(struct page *page, }write_unlock_irq (&mapping->tree_lock);radix_tree_preload_end (); +if (!error) +container_inc_filepage_count (mapping, page); }return error; }
init/Kconfig +8 -0
@@ -560,6 +560,14 @@ config STOP_MACHINE depends on (SMP && MODULE_UNLOAD) || HOTPLUG_CPU help Need stop_machine() primitive. + +config CONTAINERS + bool "Containers" + def_bool y + depends on CONFIGFS_FS + help + This option allows grouping of resources like memory and tasks. It + depends on CONFIGFS_FS support in psuedo filesystem. endmenu menu "Block layer"
mm/Makefile +2 -0
@@ -29,3 +29,5 @@ obj-$(CONFIG_MEMORY_HOTPLUG) += memory_h obj-$(CONFIG_FS_XIP) += filemap_xip.o obj-$(CONFIG_MIGRATION) += migrate.o obj-$(CONFIG_SMP) += allocpercpu.o +obj-$(CONFIG_CONTAINERS) += container.o container_mm.o +
mm/page_alloc.c +3 -0
@@ -39,6 +39,7 @@#include <linux/stop_machine.h> #include <linux/sort.h> #include <linux/pfn.h> +#include <linux/container.h> #include <asm/tlbflush.h> #include <asm/div64.h>
@@ -502,6 +503,7 @@ static void free_one_page(struct zone *z zone->pages_scanned =0 ;__free_one_page (page, zone ,order);spin_unlock (&zone->lock); +container_init_page_ptr (page,NULL ); }static void __free_pages_ok (struct page *page,unsigned int order)
@@ -798,6 +800,7 @@ static void fastcall free_hot_cold_page(arch_free_page (page,0 ); +container_init_page_ptr (page,NULL );if (PageAnon (page)) page->mapping =NULL ;if (free_pages_check (page))
mm/rmap.c +7 -1
@@ -53,6 +53,7 @@#include <linux/rmap.h> #include <linux/rcupdate.h> #include <linux/module.h> +#include <linux/container.h> #include <asm/tlbflush.h>
@@ -521,6 +522,7 @@ static void __page_set_anon_rmap(struct* interrupts because it is not modified via interrupt. */ __inc_zone_page_state (page,NR_ANON_PAGES ); +container_inc_page_count (page); }/**
@@ -563,8 +565,10 @@ void page_add_new_anon_rmap(struct page*/ void page_add_file_rmap (struct page *page) { -if (atomic_inc_and_test (&page->_mapcount)) +if (atomic_inc_and_test (&page->_mapcount)) {__inc_zone_page_state (page,NR_FILE_MAPPED ); +container_inc_page_count (page); + } }/**
@@ -598,6 +602,8 @@ void page_remove_rmap(struct page *page)set_page_dirty (page);__dec_zone_page_state (page,PageAnon (page) ?NR_ANON_PAGES :NR_FILE_MAPPED ); +container_dec_page_count (page); + } }
mm/swap.c +1 -0
@@ -196,6 +196,7 @@ void fastcall lru_cache_add_active(strucstruct pagevec *pvec = &get_cpu_var (lru_add_active_pvecs);page_cache_get (page); +container_init_page_ptr (page, current);if (!pagevec_add (pvec, page))__pagevec_lru_add_active (pvec);put_cpu_var (lru_add_active_pvecs);
mm/vmscan.c +1 -0
@@ -818,6 +818,7 @@ force_reclaim_mapped:SetPageLRU (page);VM_BUG_ON (!PageActive (page));ClearPageActive (page); +container_dec_activepage_count (page);list_move (&page->lru, &zone->inactive_list); pgmoved++;
kernel/exit.c +2 -0
@@ -41,6 +41,7 @@#include <linux/audit.h> /* for audit_free() */ #include <linux/resource.h> #include <linux/blkdev.h> +#include <linux/container.h> #include <asm/uaccess.h> #include <asm/unistd.h>
@@ -171,6 +172,7 @@ repeat:sched_exit (p);write_unlock_irq (&tasklist_lock); +container_remove_task (p,NULL );proc_flush_task (p);release_thread (p);call_rcu (&p->rcu, delayed_put_task_struct);
kernel/fork.c +9 -0
@@ -47,6 +47,7 @@#include <linux/cn_proc.h> #include <linux/delayacct.h> #include <linux/taskstats_kern.h> +#include <linux/container.h> #include <linux/random.h> #include <asm/pgtable.h>
@@ -175,6 +176,13 @@ static struct task_struct *dup_task_stru } *tsk = *orig; + +container_init_task_ptr (tsk); +if (container_add_task (tsk, orig,NULL ) == -ENOSPC ) { +free_task_struct (tsk); +free_thread_info (ti); +return NULL ; + } tsk->thread_info = ti;setup_thread_stack (tsk, orig);
@@ -1295,6 +1303,7 @@ bad_fork_cleanup_count:atomic_dec (&p->user->processes);free_uid (p->user); bad_fork_free: +container_remove_task (p,NULL );free_task (p); fork_out:return ERR_PTR (retval);
kernel/Makefile +1 -0
@@ -52,6 +52,7 @@ obj-$(CONFIG_RELAY) += relay.o obj-$(CONFIG_UTS_NS) += utsname.o obj-$(CONFIG_TASK_DELAY_ACCT) += delayacct.o obj-$(CONFIG_TASKSTATS) += taskstats.o tsacct.o +obj-$(CONFIG_CONTAINERS) += container_configfs.o ifneq ($(CONFIG_SCHED_NO_NO_OMIT_FRAME_POINTER),y) # According to Alan Modra <[email protected]>, the -fno-omit-frame-pointer is
fs/inode.c +3 -0
@@ -22,6 +22,7 @@#include <linux/bootmem.h> #include <linux/inotify.h> #include <linux/mount.h> +#include <linux/container.h> /* * This is needed for the following functions:
@@ -164,6 +165,7 @@ static struct inode *alloc_inode(struct } inode->i_private =0 ; inode->i_mapping = mapping; +container_add_file (mapping,NULL ); }return inode; }
@@ -172,6 +174,7 @@ void destroy_inode(struct inode *inode) {BUG_ON (inode_has_buffers (inode));security_inode_free (inode); +container_remove_file (inode->i_mapping);if (inode->i_sb->s_op->destroy_inode) inode->i_sb->s_op->destroy_inode (inode);else
include/linux/mm_types.h struct page { · line 54
Quoted 4 lines · Rohit Seth
include/linux/mm_types.h +3
-0*/ +#ifdef CONFIG_CONTAINERS +struct container_struct *ctn;/* Pointer to container, may be NULL */ +#endif
I still don't think it's a good idea to add a pointer to struct page for this. This means any kernel that enables the config would need to carry this significant overhead, no matter if containers are used to not.
Better would be to store them in some other data structure that is only allocated on demand or figure out a way to store them in the sometimes not all used fields in struct page.
BTW your patchkit seems to be also in wrong order in that when 02 is applied it won't compile.
-Andi
Quoted 7 lines · Andi Kleen
include/linux/mm_types.h struct page { · line 54
Quoted 4 lines · Rohit Seth
include/linux/mm_types.h +3
-0*/ +#ifdef CONFIG_CONTAINERS +struct container_struct *ctn;/* Pointer to container, may be NULL */ +#endif I still don't think it's a good idea to add a pointer to struct page for this.
I thought last time you supported adding a pointer to struct page (when you mentioned next gen slab will also consume page->mapping). May be I missed your point.
Quoted 3 lines
This means any kernel that enables the config would need to carry this significant overhead, no matter if containers are used to not.
Sure this is non-zero overhead but I think this is the logical place to track the memory.
Quoted 4 lines
Better would be to store them in some other data structure that is only allocated on demand or figure out a way to store them in the sometimes not all used fields in struct page.
which one...I think the fields in page structure are already getting doubly used.
Quoted 2 lines
BTW your patchkit seems to be also in wrong order in that when 02 is applied it won't compile.
Not sure if I understood that. I've myself tested these patches on 2.6.18-rc6-mm2 kernel and they apply just fine. Are you just trying to apply 02....if so then that wouldn't suffice.
-rohit
Quoted 11 lines · Rohit Seth
Quoted 7 lines · Andi Kleen
include/linux/mm_types.h struct page { · line 54
Quoted 4 lines · Rohit Seth
include/linux/mm_types.h +3
-0*/ +#ifdef CONFIG_CONTAINERS +struct container_struct *ctn;/* Pointer to container, may be NULL */ +#endif I still don't think it's a good idea to add a pointer to struct page for this.
I thought last time you supported adding a pointer to struct page (when you mentioned next gen slab will also consume
page->mapping).
I didn't. Alternative was a separate data structure.
Quoted 2 lines
which one...I think the fields in page structure are already getting doubly used.
There are lots of different cases. At least for anonymous memory ->mapping should be free. Perhaps that could be used for anonymous memory and a separate data structure for the important others.
slab should have at least one field free too, although it might be a different one (iirc Christoph's rewrite uses more than the current slab, but it would surprise me if he needed all)
Quoted 6 lines
Quoted 2 lines
BTW your patchkit seems to be also in wrong order in that when 02 is applied it won't compile.
Not sure if I understood that. I've myself tested these patches on 2.6.18-rc6-mm2 kernel and they apply just fine. Are you just trying to apply 02....if so then that wouldn't suffice.
I meant assuming the patchkit was applied you would break binary search inbetween because not each piece compiles on its own.
-Andi
Quoted 13 lines · Andi Kleen
Quoted 11 lines · Rohit Seth
Quoted 7 lines · Andi Kleen
include/linux/mm_types.h struct page { · line 54
Quoted 4 lines · Rohit Seth
include/linux/mm_types.h +3
-0*/ +#ifdef CONFIG_CONTAINERS +struct container_struct *ctn;/* Pointer to container, may be NULL */ +#endif I still don't think it's a good idea to add a pointer to struct page for this.
I thought last time you supported adding a pointer to struct page (when you mentioned next gen slab will also consume
page->mapping).
Quoted 9 lines
I didn't. Alternative was a separate data structure.
Quoted 2 lines
which one...I think the fields in page structure are already getting doubly used.
There are lots of different cases. At least for anonymous memory ->mapping should be free. Perhaps that could be used for anonymous memory and a separate data structure for the important others.
It is not free for anonymous memory as it is overloaded with pointer to anon_vma. I think one single pointer consistent across all page usages is just so much cleaner and simple...
Quoted 7 lines
slab should have at least one field free too, although it might be a different one (iirc Christoph's rewrite uses more than the current slab, but it would surprise me if he needed all)
Quoted 3 lines
Quoted 2 lines
BTW your patchkit seems to be also in wrong order in that when 02 is applied it won't compile.
Quoted 6 lines
Quoted 3 lines
Not sure if I understood that. I've myself tested these patches on 2.6.18-rc6-mm2 kernel and they apply just fine. Are you just trying to apply 02....if so then that wouldn't suffice.
I meant assuming the patchkit was applied you would break binary search inbetween because not each piece compiles on its own.
I've currently separated the patches based on where the changes are made and something that makes each a logical block. I will reorder the patches next time so that each subsequent patch compiles.
-rohit
Quoted 3 lines
It is not free for anonymous memory as it is overloaded with pointer to
anon_vma.
Sorry. But ->index should be free there.
Quoted 2 lines
I think one single pointer consistent across all page usages is just so much cleaner and simple...
It just costs you 0.2% of all your memory in all cases. Too much imho.
-Andi
Quoted 3 lines · Andi Kleen
There are lots of different cases. At least for anonymous memory ->mapping should be free. Perhaps that could be used for anonymous memory and a separate data structure for the important others.
mapping is used for swap and to point to the anon vma.
Quoted 3 lines
slab should have at least one field free too, although it might be a different one (iirc Christoph's rewrite uses more than the current slab, but it would surprise me if he needed all)
slab currently has lots of fields free but my rewrite uses all of them. And AFAICT this patchset does not track slab pages.
Hmm.... Build a radix tree with pointers to the pages?
Quoted 6 lines · Christoph Lameter
Quoted 3 lines · Andi Kleen
There are lots of different cases. At least for anonymous memory ->mapping should be free. Perhaps that could be used for anonymous memory and a separate data structure for the important others.
Quoted 9 lines
mapping is used for swap and to point to the anon vma.
Quoted 3 lines
slab should have at least one field free too, although it might be a different one (iirc Christoph's rewrite uses more than the current slab, but it would surprise me if he needed all)
slab currently has lots of fields free but my rewrite uses all of them. And AFAICT this patchset does not track slab pages.
Currently it doesn't track kernel memory. That is why I don't want to over load any of the existing fields.
Quoted 2 lines
Hmm.... Build a radix tree with pointers to the pages?
...my preference is to leave it in page struct...that has non-zero cost.
-rohit
Quoted 1 line · Rohit Seth
...my preference is to leave it in page struct...that has non-zero cost.
Oh. Making struct page bigger than a cacheline or not fitting exactly in a cacheline has some costs.
Quoted 7 lines · Christoph Lameter
Quoted 1 line · Rohit Seth
...my preference is to leave it in page struct...that has non-zero cost.
Oh. Making struct page bigger than a cacheline or not fitting exactly in a cacheline has some costs.
Isn't most of the architectures not defining WANT_PAGE_VIRTUAL and thus the page structure on these is not fitting cache line well. In most of these cases the additional pointer will actually make the structure 64 bytes on 64-bit machines.
-rohit
Quoted 8 lines · Christoph Lameter
Quoted 3 lines · Andi Kleen
There are lots of different cases. At least for anonymous memory ->mapping should be free. Perhaps that could be used for anonymous memory and a separate data structure for the important others.
mapping is used for swap and to point to the anon vma.
Quoted 8 lines
Quoted 3 lines
slab should have at least one field free too, although it might be a different one (iirc Christoph's rewrite uses more than the current slab, but it would surprise me if he needed all)
slab currently has lots of fields free but my rewrite uses all of them. And AFAICT this patchset does not track slab pages.
Hmm.... Build a radix tree with pointers to the pages?
Yes, that would be a way to isolate the overhead.
Signature · 4 lines
----------------------------------------------------------------------
Chandra Seetharaman | Be careful what you choose....
- [email protected] | .......you may get it.
----------------------------------------------------------------------Quoted 13 lines · Chandra Seetharaman
Quoted 12 lines · Christoph Lameter
Quoted 3 lines · Andi Kleen
There are lots of different cases. At least for anonymous memory ->mapping should be free. Perhaps that could be used for anonymous memory and a separate data structure for the important others.
mapping is used for swap and to point to the anon vma.
Quoted 3 lines
slab should have at least one field free too, although it might be a different one (iirc Christoph's rewrite uses more than the current slab, but it would surprise me if he needed all)
Quoted 7 lines
Quoted 4 lines
slab currently has lots of fields free but my rewrite uses all of them. And AFAICT this patchset does not track slab pages.
Hmm.... Build a radix tree with pointers to the pages?
Yes, that would be a way to isolate the overhead.
As said earlier, the additional cost is not really that much. And if we do get additional benefit of proper cache alignment for page struct because of this additional pointer then it could be a almost zero cost.
And in mm kernel, I also see PAGE_OWNER defined which puts the whole page struct out of sync with caches.
Christoph, Please let me know if there is still a placeholder for container pointer ONLY for kernel pages. I think it is easy to do away with per page pointer for user land as it can be stored in anon_vma or mappings.
-rohit