powerpc/pseries: CMM: Drop page array
Subsystems (2)
PATCH 1/1 powerpc/pseries: CMM: Drop page array
We can simply store the pages in a list (page->lru), no need for a separate data structure (+ complicated handling). This is how most other balloon drivers store allocated pages without additional tracking data.
For the notifiers, use page_to_pfn() to check if a page is in the applicable range. plpar_page_set_loaned()/plpar_page_set_active() were called with __pa(page_address()) for now, I assume we can simply switch to page_to_phys() here. The pfn_to_kaddr() handling is now mostly gone.
---
Only compile-tested. I hope the page_to_phys() thingy is correct and I didn't mess up something else / ignoring something important why the array is needed.
I stumbled over this while looking at how the memory isolation notifier is used - and wondered why the additional array is necessary. Also, I think by switching to the generic balloon compaction mechanism, we could get rid of the memory hotplug notifier and the memory isolation notifier in this code, as the migration capability of the inflated pages is the real requirement:
commit 14b8a76b9d53346f2871bf419da2aaf219940c50
Author: Robert Jennings <[email protected]>
Date: Thu Dec 17 14:44:52 2009 +0000
powerpc: Make the CMM memory hotplug aware
The Collaborative Memory Manager (CMM) module allocates individual pages over time that are not migratable. On a long running system this can severely impact the ability to find enough pages to support a hotplug memory remove operation.
[...]
Thoughts?
--- arch/powerpc/platforms/pseries/cmm.c | 155 ++++++--------------------- 1 file changed, 31 insertions(+), 124 deletions(-)
arch/powerpc/platforms/pseries/cmm.c +31 -124
@@ -75,21 +75,13 @@ module_param_named(debug, cmm_debug, uint, 0644); MODULE_PARM_DESC(debug, "Enable module debugging logging. Set to 1 to enable. " "[Default=" __stringify(CMM_DEBUG) "]"); -#define CMM_NR_PAGES ((PAGE_SIZE - sizeof(void *) - sizeof(unsigned long)) / sizeof(unsigned long)) -#define cmm_dbg (...) if (cmm_debug) { printk(KERN_INFO "cmm: "__VA_ARGS__); } -struct cmm_page_array { -struct cmm_page_array *next; -unsigned long index; -unsigned long page[CMM_NR_PAGES ]; -}; -static unsigned long loaned_pages;static unsigned long loaned_pages_target;static unsigned long oom_freed_pages; -static struct cmm_page_array *cmm_page_list; +static LIST_HEAD (cmm_page_list);static DEFINE_SPINLOCK (cmm_lock);static DEFINE_MUTEX (hotplug_mutex);
@@ -138,8 +130,7 @@ static long plpar_page_set_active(unsigned long vpa)**/ static long cmm_alloc_pages (long nr) { -struct cmm_page_array *pa, *npa; -unsigned long addr; +struct page *page;long rc;cmm_dbg ("Begin request for %ld pages\n" , nr);
@@ -156,43 +147,20 @@ static long cmm_alloc_pages(long nr)break ; } - addr =__get_free_page (GFP_NOIO | __GFP_NOWARN | - __GFP_NORETRY | __GFP_NOMEMALLOC); -if (!addr) + page =alloc_page (GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | + __GFP_NOMEMALLOC); +if (!page)break ;spin_lock (&cmm_lock); - pa = cmm_page_list; -if (!pa || pa->index >=CMM_NR_PAGES ) { -/* Need a new page for the page list. */ -spin_unlock (&cmm_lock); - npa = (struct cmm_page_array *)__get_free_page ( -GFP_NOIO | __GFP_NOWARN | - __GFP_NORETRY | __GFP_NOMEMALLOC); -if (!npa) { -pr_info ("%s: Can not allocate new page list\n" , __func__); -free_page (addr); -break ; - } -spin_lock (&cmm_lock); - pa = cmm_page_list; - -if (!pa || pa->index >=CMM_NR_PAGES ) { - npa->next = pa; - npa->index =0 ; - pa = npa; - cmm_page_list = pa; - }else -free_page ((unsigned long ) npa); - } - -if ((rc =plpar_page_set_loaned (__pa (addr)))) { + rc =plpar_page_set_loaned (page_to_phys (page)); +if (rc) {pr_err ("%s: Can not set page to loaned. rc=%ld\n" , __func__, rc);spin_unlock (&cmm_lock); -free_page (addr); +__free_page (page);break ; } - pa->page[pa->index++] = addr; +list_add (&page->lru, &cmm_page_list); loaned_pages++;totalram_pages_dec ();spin_unlock (&cmm_lock);
@@ -212,25 +180,16 @@ static long cmm_alloc_pages(long nr)**/ static long cmm_free_pages (long nr) { -struct cmm_page_array *pa; -unsigned long addr; +struct page *page, *tmp;cmm_dbg ("Begin free of %ld pages.\n" , nr);spin_lock (&cmm_lock); - pa = cmm_page_list; -while (nr) { -if (!pa || pa->index <=0 ) +list_for_each_entry_safe (page, tmp, &cmm_page_list, lru) { +if (!nr)break ; - addr = pa->page[--pa->index]; - -if (pa->index ==0 ) { - pa = pa->next; -free_page ((unsigned long ) cmm_page_list); - cmm_page_list = pa; - } - -plpar_page_set_active (__pa (addr)); -free_page (addr); +plpar_page_set_active (page_to_phys (page)); +list_del (&page->lru); +__free_page (page); loaned_pages--; nr--;totalram_pages_inc ();
@@ -491,20 +450,13 @@ static struct notifier_block cmm_reboot_nb = {static unsigned long cmm_count_pages (void *arg) {struct memory_isolate_notify *marg = arg; -struct cmm_page_array *pa; -unsigned long start = (unsigned long )pfn_to_kaddr (marg->start_pfn); -unsigned long end = start + (marg->nr_pages <<PAGE_SHIFT ); -unsigned long idx; +struct page *page;spin_lock (&cmm_lock); - pa = cmm_page_list; -while (pa) { -if ((unsigned long )pa >= start && (unsigned long )pa < end) +list_for_each_entry (page, &cmm_page_list, lru) { +if (page_to_pfn (page) >= marg->start_pfn && +page_to_pfn (page) < marg->start_pfn + marg->nr_pages) marg->pages_found++; -for (idx =0 ; idx < pa->index; idx++) -if (pa->page[idx] >= start && pa->page[idx] < end) - marg->pages_found++; - pa = pa->next; }spin_unlock (&cmm_lock);return 0 ;
@@ -545,69 +497,24 @@ static struct notifier_block cmm_mem_isolate_nb = {static int cmm_mem_going_offline (void *arg) {struct memory_notify *marg = arg; -unsigned long start_page = (unsigned long )pfn_to_kaddr (marg->start_pfn); -unsigned long end_page = start_page + (marg->nr_pages <<PAGE_SHIFT ); -struct cmm_page_array *pa_curr, *pa_last, *npa; -unsigned long idx; +struct page *page, *tmp;unsigned long freed =0 ;cmm_dbg ("Memory going offline, searching 0x%lx (%ld pages).\n" , - start_page, marg->nr_pages); + (unsigned long )pfn_to_kaddr (marg->start_pfn), marg->nr_pages);spin_lock (&cmm_lock);/* Search the page list for pages in the range to be offlined */ - pa_last = pa_curr = cmm_page_list; -while (pa_curr) { -for (idx = (pa_curr->index -1 ); (idx +1 ) >0 ; idx--) { -if ((pa_curr->page[idx] < start_page) || - (pa_curr->page[idx] >= end_page)) -continue ; - -plpar_page_set_active (__pa (pa_curr->page[idx])); -free_page (pa_curr->page[idx]); - freed++; - loaned_pages--; -totalram_pages_inc (); - pa_curr->page[idx] = pa_last->page[--pa_last->index]; -if (pa_last->index ==0 ) { -if (pa_curr == pa_last) - pa_curr = pa_last->next; - pa_last = pa_last->next; -free_page ((unsigned long )cmm_page_list); - cmm_page_list = pa_last; - } - } - pa_curr = pa_curr->next; - } - -/* Search for page list structures in the range to be offlined */ - pa_last =NULL ; - pa_curr = cmm_page_list; -while (pa_curr) { -if (((unsigned long )pa_curr >= start_page) && - ((unsigned long )pa_curr < end_page)) { - npa = (struct cmm_page_array *)__get_free_page ( -GFP_NOIO | __GFP_NOWARN | - __GFP_NORETRY | __GFP_NOMEMALLOC); -if (!npa) { -spin_unlock (&cmm_lock); -cmm_dbg ("Failed to allocate memory for list " -"management. Memory hotplug " -"failed.\n" ); -return -ENOMEM ; - } -memcpy (npa, pa_curr,PAGE_SIZE ); -if (pa_curr == cmm_page_list) - cmm_page_list = npa; -if (pa_last) - pa_last->next = npa; -free_page ((unsigned long ) pa_curr); - freed++; - pa_curr = npa; - } - - pa_last = pa_curr; - pa_curr = pa_curr->next; +list_for_each_entry_safe (page, tmp, &cmm_page_list, lru) { +if (page_to_pfn (page) < marg->start_pfn || +page_to_pfn (page) >= marg->start_pfn + marg->nr_pages) +continue ; +plpar_page_set_active (page_to_phys (page)); +list_del (&page->lru); +__free_page (page); + freed++; + loaned_pages--; +totalram_pages_inc (); }spin_unlock (&cmm_lock);
Signature · 1 line
2.21.0
Quoted 37 lines · David Hildenbrand
We can simply store the pages in a list (
page->lru), no need for a separate data structure (+ complicated handling). This is how most other balloon drivers store allocated pages without additional tracking data.For the notifiers, use
page_to_pfn()to check if a page is in the applicable range. plpar_page_set_loaned()/plpar_page_set_active() were called with __pa(page_address()) for now, I assume we can simply switch topage_to_phys()here. Thepfn_to_kaddr()handling is now mostly gone.Cc Benjamin Herrenschmidt <[email protected]>Cc Paul Mackerras <[email protected]>Cc Michael Ellerman <[email protected]>Cc Arun KS <[email protected]>Cc Pavel Tatashin <[email protected]>Cc Thomas Gleixner <[email protected]>Cc Andrew Morton <[email protected]>Cc Vlastimil Babka <[email protected]>Signed-off-by David Hildenbrand <[email protected]>---
Only compile-tested. I hope the
page_to_phys()thingy is correct and I didn't mess up something else / ignoring something important why the array is needed.I stumbled over this while looking at how the memory isolation notifier is used - and wondered why the additional array is necessary. Also, I think by switching to the generic balloon compaction mechanism, we could get rid of the memory hotplug notifier and the memory isolation notifier in this code, as the migration capability of the inflated pages is the real requirement:
commit14b8a76b9d53346f2871bf419da2aaf219940c50
Author: Robert Jennings <[email protected]>
Date: Thu Dec 17 14:44:52 2009 +0000powerpc: Make the CMM memory hotplug aware
Quoted 7 lines
The Collaborative Memory Manager (CMM) module allocates individual pages over time that are not migratable. On a long running system this can severely impact the ability to find enough pages to support a hotplug memory remove operation.
[...]Thoughts?
Ping, is still feature still used at all?
If nobody can test, any advise on which HW I need and how to trigger it?
Signature · 3 lines
Thanks, David / dhildenb
Quoted 44 lines · David Hildenbrand
Quoted 43 lines · David Hildenbrand
We can simply store the pages in a list (
page->lru), no need for a separate data structure (+ complicated handling). This is how most other balloon drivers store allocated pages without additional tracking data.For the notifiers, use
page_to_pfn()to check if a page is in the applicable range. plpar_page_set_loaned()/plpar_page_set_active() were called with __pa(page_address()) for now, I assume we can simply switch topage_to_phys()here. Thepfn_to_kaddr()handling is now mostly gone.Cc Benjamin Herrenschmidt <[email protected]>Cc Paul Mackerras <[email protected]>Cc Michael Ellerman <[email protected]>Cc Arun KS <[email protected]>Cc Pavel Tatashin <[email protected]>Cc Thomas Gleixner <[email protected]>Cc Andrew Morton <[email protected]>Cc Vlastimil Babka <[email protected]>Signed-off-by David Hildenbrand <[email protected]>---
Only compile-tested. I hope the
page_to_phys()thingy is correct and I didn't mess up something else / ignoring something important why the array is needed.I stumbled over this while looking at how the memory isolation notifier is used - and wondered why the additional array is necessary. Also, I think by switching to the generic balloon compaction mechanism, we could get rid of the memory hotplug notifier and the memory isolation notifier in this code, as the migration capability of the inflated pages is the real requirement:
commit14b8a76b9d53346f2871bf419da2aaf219940c50
Author: Robert Jennings <[email protected]>
Date: Thu Dec 17 14:44:52 2009 +0000powerpc: Make the CMM memory hotplug aware
The Collaborative Memory Manager (CMM) module allocates individual pages over time that are not migratable. On a long running system this can severely impact the ability to find enough pages to support a hotplug memory remove operation.
[...]
Quoted 6 lines
Quoted 1 line
Thoughts?
Ping, is still feature still used at all?
If nobody can test, any advise on which HW I need and how to trigger it?
So ... if CMM is no longer alive I propose ripping it out completely. Does anybody know if this feature is still getting used? Getting rid of the memory isolation notifier sounds desirable - either by scrapping CMM or by properly wiring up balloon compaction.
Signature · 3 lines
Thanks, David / dhildenb
Quoted 64 lines · David Hildenbrand
Quoted 4 lines · David Hildenbrand
Quoted 3 lines · David Hildenbrand
We can simply store the pages in a list (
page->lru), no need for a separate data structure (+ complicated handling). This is how most other balloon drivers store allocated pages without additionaltracking
Quoted 4 lines
Quoted 4 lines
data.
For the notifiers, use
page_to_pfn()to check if a page is in the applicable range. plpar_page_set_loaned()/plpar_page_set_active()were
Quoted 1 line
Quoted 1 line
called with __pa(page_address()) for now, I assume we can simply
switch
Quoted 1 line
Quoted 1 line
to
page_to_phys()here. Thepfn_to_kaddr()handling is now mostlygone.
Quoted 12 lines
Quoted 12 lines
Cc Benjamin Herrenschmidt <[email protected]>Cc Paul Mackerras <[email protected]>Cc Michael Ellerman <[email protected]>Cc Arun KS <[email protected]>Cc Pavel Tatashin <[email protected]>Cc Thomas Gleixner <[email protected]>Cc Andrew Morton <[email protected]>Cc Vlastimil Babka <[email protected]>Signed-off-by David Hildenbrand <[email protected]>---
Quoted 1 line
Quoted 1 line
Only compile-tested. I hope the
page_to_phys()thingy is correct andI
Quoted 1 line
Quoted 1 line
didn't mess up something else / ignoring something important why the
array
Quoted 3 lines
Quoted 3 lines
is needed.
I stumbled over this while looking at how the memory isolation
notifier is
Quoted 1 line
Quoted 1 line
used - and wondered why the additional array is necessary. Also, I
think
Quoted 1 line
Quoted 1 line
by switching to the generic balloon compaction mechanism, we could
get
Quoted 1 line
Quoted 1 line
rid of the memory hotplug notifier and the memory isolation notifier
in
Quoted 1 line
Quoted 1 line
this code, as the migration capability of the inflated pages is the
real
Quoted 8 lines
Quoted 8 lines
requirement:
commit14b8a76b9d53346f2871bf419da2aaf219940c50
Author: Robert Jennings <[email protected]>
Date: Thu Dec 17 14:44:52 2009 +0000powerpc: Make the CMM memory hotplug aware
The Collaborative Memory Manager (CMM) module allocates
individual pages
Quoted 1 line
Quoted 1 line
over time that are not migratable. On a long running system
this can
Quoted 1 line
Quoted 1 line
severely impact the ability to find enough pages to support a
hotplug
Quoted 8 lines
Quoted 4 lines
memory remove operation. [...]Thoughts?
Ping, is still feature still used at all?
If nobody can test, any advise on which HW I need and how to trigger
Quoted 8 lines
it?
Quoted 1 line
So ... if CMM is no longer alive I propose ripping it out completely. Does anybody know if this feature is still getting used? Getting rid of the memory isolation notifier sounds desirable - either by scrapping CMM
or by properly wiring up balloon compaction.
It's still used AFAIK, but the people who wrote the code have left IBM, and I'm on leave.
I'll be back in a week or so and will try and track down how to test it then.
cheers
Signature · 1 line
Sent from my Android phone with K-9 Mail. Please excuse my brevity.
Quoted 76 lines · Michael Ellerman
Quoted 64 lines · David Hildenbrand
Quoted 4 lines · David Hildenbrand
Quoted 3 lines · David Hildenbrand
We can simply store the pages in a list (
page->lru), no need for a separate data structure (+ complicated handling). This is how most other balloon drivers store allocated pages without additionaltracking
Quoted 4 lines
Quoted 4 lines
data.
For the notifiers, use
page_to_pfn()to check if a page is in the applicable range. plpar_page_set_loaned()/plpar_page_set_active()were
Quoted 1 line
Quoted 1 line
called with __pa(page_address()) for now, I assume we can simply
switch
Quoted 1 line
Quoted 1 line
to
page_to_phys()here. Thepfn_to_kaddr()handling is now mostlygone.
Quoted 12 lines
Quoted 12 lines
Cc Benjamin Herrenschmidt <[email protected]>Cc Paul Mackerras <[email protected]>Cc Michael Ellerman <[email protected]>Cc Arun KS <[email protected]>Cc Pavel Tatashin <[email protected]>Cc Thomas Gleixner <[email protected]>Cc Andrew Morton <[email protected]>Cc Vlastimil Babka <[email protected]>Signed-off-by David Hildenbrand <[email protected]>---
Quoted 1 line
Quoted 1 line
Only compile-tested. I hope the
page_to_phys()thingy is correct andI
Quoted 1 line
Quoted 1 line
didn't mess up something else / ignoring something important why the
array
Quoted 3 lines
Quoted 3 lines
is needed.
I stumbled over this while looking at how the memory isolation
notifier is
Quoted 1 line
Quoted 1 line
used - and wondered why the additional array is necessary. Also, I
think
Quoted 1 line
Quoted 1 line
by switching to the generic balloon compaction mechanism, we could
get
Quoted 1 line
Quoted 1 line
rid of the memory hotplug notifier and the memory isolation notifier
in
Quoted 1 line
Quoted 1 line
this code, as the migration capability of the inflated pages is the
real
Quoted 8 lines
Quoted 8 lines
requirement:
commit14b8a76b9d53346f2871bf419da2aaf219940c50
Author: Robert Jennings <[email protected]>
Date: Thu Dec 17 14:44:52 2009 +0000powerpc: Make the CMM memory hotplug aware
The Collaborative Memory Manager (CMM) module allocates
individual pages
Quoted 1 line
Quoted 1 line
over time that are not migratable. On a long running system
this can
Quoted 1 line
Quoted 1 line
severely impact the ability to find enough pages to support a
hotplug
Quoted 8 lines
Quoted 4 lines
memory remove operation. [...]Thoughts?
Ping, is still feature still used at all?
If nobody can test, any advise on which HW I need and how to trigger
Quoted 8 lines
it?
Quoted 1 line
So ... if CMM is no longer alive I propose ripping it out completely. Does anybody know if this feature is still getting used? Getting rid of the memory isolation notifier sounds desirable - either by scrapping CMM
or by properly wiring up balloon compaction.
Quoted 4 lines
It's still used AFAIK, but the people who wrote the code have left IBM, and I'm on leave.
I'll be back in a week or so and will try and track down how to test it then.
Thanks - sure, take your time. I'll try to play with wiring up balloon compaction in the meantime.
Cheers!
Quoted 2 lines
cheers
Signature · 3 lines
Thanks, David / dhildenb