Guest page hinting version 6.

Greetings,
I've dedusted the guest page hinting patches and ported them to todays upstream git tree. There is one reject if applied to 2.6.24-rc5-mm1 but that is easy to fix. The code stills works as expected on my test system.

Our z/VM performance team recently published a report on guest page hinting vs. the ballooner approach on SLES10 for a farm of web servers. The code on SLES10 differs a bit from the upstream variant but the performance results should be still valid. You will find the report here:

http://www.vm.ibm.com/perf/reports/zvm/html/530cmm.html

(the VMRM-CMM the web page speaks about is the balloon approach,
CMMA is the guest page hinting).

Both approaches to the memory overcommit problem show comparable benefits for this workload, with an advantage for guest page hinting for large number of guests. For other workloads your mileage may vary.

The main benefit for guest page hinting vs. the ballooner is that there is no need for a monitor that keeps track of the memory usage of all the guests, a complex algorithm that calculates the working set sizes and for the calls into the guest kernel to control the size of the balloons. The host just does normal LRU based paging. If the host picks one of the pages the guest can recreate, the host can throw it away instead of writing it to the paging device. Simple and elegant.
The main disadvantage is the added complexity that is introduced to the guests memory management code to do the page state changes and to deal with discard faults.

The last versions of the patches do not differ much, I consider the code to be stable. My question now is how to proceed with the code. I sure would love to see the code going upstream some day but that depends on the mm developers as the code adds complexity that needs to be supported. If the general feeling is that the advantages of this approach do not warrent for the added complexity this will likely be the last time you will hear about guest page hinting.

--
blue skies,
Martin.

"Reality continues to ruin my life." - Calvin.

From: Martin Schwidefsky <[email protected]>
From: Hubertus Franke <[email protected]>
From: Himanshu Raj

The guest page hinting patchset introduces code that passes guest page usage information to the host system that virtualizes the memory of its guests. There are three different page states:
* Unused: The page content is of no interest to the guest. The host can forget the page content and replace it with a page containing zeroes.
* Stable: The page content is needed by the guest and has to be preserved by the host.
* Volatile: The page content is useful to the guest but not essential. The host can discard the page but has to deliver a special kind of fault to the guest if the guest accesses a page discarded by the host.

The unused state is used for free pages, it allows the host to avoid the paging of empty pages. The default state for non-free pages is stable. The host can write stable pages to a swap device but has to restore the page if the guest accesses it. The volatile page state is used for clean uptodate page cache pages. The host can choose to discard volatile pages as part of its vmscan operation instead of writing them to the hosts paging device. The guest system doesn't notice that a volatile page is gone until it tries to access the page or if it tries to make the page stable again. For a guest access to a discarded page the host generates a discard fault to notify the guest. The guest has to remove the page from the cache and reload the page from its backing device.

The volatile state is used for all page cache pages, even for pages which are referenced by writable ptes. The host needs to be able to check the dirty state of the pages. Since the host doesn't know where the page table entries of the guest are located, the volatile state as introduced by this patch is only usable on architectures with per-page dirty bits (s390 only). For per-pte dirty bit architectures some additional code is needed, see patch #4.

The main question is where to put the state transitions between the volatile and the stable state. The simple solution is to make a page stable whenever a lookup is done or a page reference is derived from a page table entry. Attempts to make pages volatile are added at strategic points. The conditions that prevent a page from being made volatile:
1) The page is reserved. Some sort of special page.
2) The page is marked dirty in the struct page. The page content is more recent than the data on the backing device. The host cannot access the linux internal dirty bit so the page needs to be stable.
3) The page is in writeback. The page content is needed for i/o.
4) The page is locked. Someone has exclusive access to the page.
5) The page is anonymous. Swap cache support needs additional code. See patch #2.
6) The page has no mapping. Without a backing the page cannot be recreated.
7) The page is not uptodate.
8) The page has private information. try_to_release_page can fail, e.g. in case the private information is journaling data. The discard fault need to be able to remove the page.
9) The page is already discarded.
10) The page is not on the LRU list. The page has been isolated, some processing is done.
11) The page map count is not equal to the page reference count - 1. The discard fault handler can remove the page cache reference and all mappers of a page. It cannot remove the page reference for any other user of the page.

The transitions to stable are done by find_get_pages() and its variants, in follow_page if the FOLL_GET flag is set, by copy-on-write in do_wp_page, and by the early copy-on-write in do_no_page. For page cache page this is always done with a call to page_make_stable().
To make enough pages discardable by the host an attempt to do the transition to volatile state is done at several places:
1) When a page gets unlocked (unlock_page).
2) When writeback has finished (test_clear_page_writeback).
3) When the page reference counter is decreased (__free_pages, page_cache_release alias put_page_check and __pagevec_release_nonlru right before the put_page_testzero call).
4) When the map counter in increased (page_add_file_rmap).
5) When a page is moved from the active list to the inactive list.
6) In filemap_nopage after the wait for readpage has finished. This try is necessary because filemap_nopage held an additional reference to the page so that the page_make_volatile call in unlock_page could not do the state transition.
The function for the state transitions to volatile is page_make_volatile().

The major obstacles that need to get addressed:
* Concurrent page state changes:
To guard against concurrent page state updates some kind of lock is needed. If page_make_volatile() has already done the 11 checks it will issue the state change primitive. If in the meantime one of the conditions has changed the user that requires that page in stable state will have to wait in the page_make_stable() function until the make volatile operation has finished. It is up to the architecture to define how this is done with the three primitives page_test_set_state_change, page_clear_state_change and page_state_change.
There are some alternatives how this can be done, e.g. a global lock, or lock per segment in the kernel page table, or the per page bit PG_arch_1 if it is still free.

* Page references acquired from page tables:
All page references acquired with find_get_page and friends can be used to access the page frame content. A page reference grabbed from a page table cannot be used to access the page content, the page has to be made stable first. If the make stable operation fails because the page has been discarded it has to be removed from page cache. That removes the page table entry as well.

* Page discard vs. __remove_from_page_cache race
A new page flag PG_discarded is added. This bit is set for discarded pages. It prevents multiple removes of a page from the page cache due to concurrent discard faults and/or normal page removals. It also prevents the re-add of isolated pages to the lru list in vmscan if the page has been discarded while it was not on the lru list.

* Page discard vs. pte establish
The discard fault handler does three things: 1) set the PG_discarded bit for the page, 2) remove the page from all page tables and 3) remove the page from the page cache. All page references of the discarded page that are still around after step 2 may not be used to establish new mappings because step 3 clears the page->mapping field that is required to find the mappers. Code that establishes new ptes to pages that might be discarded has to check the PG_discarded bit. Step 2 has to check all possible location for a pte of a particular page and check if the pte exists or another processor might be in the process of establishing one. To do that the page table lock for the pte is used. See page_unmap_all and the modified quick check in page_check_address for the details.

* copy_one_pte vs. discarded pages
The code that copies the page tables may not copy ptes for discarded pages because this races with the discard fault handler. copy_one_pte cannot back out either since there is no automatic repeat of the fault that caused the pte modification. Ptes to discarded pages only show up in copy_one_pte if a fork races with a discard fault. In this case copy_one_pte has to create a pte in the new page table that looks like the one that the discard fault handler would have created in the original page table if copy_one_pte would not have grabed the page table lock first.

* get_user_pages with FOLL_GET
If get_user_pages is called with a non-NULL pages argument the caller has to be able to access the page content using the references returned in the pages array. This is done with a check in follow_page for the FOLL_GET bit and a call to page_make_stable.
If get_user_pages is called with NULL as the pages argument the pages are not made stable. The caller cannot expect that the pages are available after the call because vmscan might have removed them.

* buffer heads / page_private
A page that is modified with sys_write will get a buffer-head to keep track of the dirty state. The existence of a buffer-head makes PagePrivate(page) return true. Pages with private information cannot be made volatile. Until the buffer-head is removed the page will stay stable. The standard logic is to call try_to_release_page which frees the buffer-head only if more than 10% of GFP_USER memory are used for buffer heads. Without high memory every page can have a buffer-head without running over the limit. The result is that every page written to with sys_write will stay stable until it is removed. To get these pages volatile again max_buffer_heads is set to zero (!) to force a call to try_to_release_page whenever a page is moved from the active to the inactive list.

* page_free_discarded hook
The architecture might want/need to do special things for discarded pages before they are freed. E.g. s390 has to delay the freeing of discarded pages. To allow this a hook in added to free_hot_cold_page.

Another noticable change is that the first few lines of code in try_to_unmap_one that calculates the address from the page and the vma is moved out of try_to_unmap_one to the callers. This is done to make try_to_unmap_one usable for the removal of discarded pages in page_unmap_all.

---

 fs/buffer.c                 |   12 ++
 include/linux/mm.h          |    1 
 include/linux/page-flags.h  |   13 ++
 include/linux/page-states.h |  123 +++++++++++++++++++++++++++
 include/linux/pagemap.h     |    6 +
 mm/Makefile                 |    1 
 mm/filemap.c                |   74 +++++++++++++++-
 mm/memory.c                 |   58 ++++++++++++
 mm/page-states.c            |  197 ++++++++++++++++++++++++++++++++++++++++++++
 mm/page-writeback.c         |    2 
 mm/page_alloc.c             |   14 ++-
 mm/rmap.c                   |   93 ++++++++++++++++++--
 mm/swap.c                   |   14 +++
 mm/vmscan.c                 |   68 ++++++++++-----
 14 files changed, 640 insertions(+), 36 deletions(-)

Index: linux-2.6/fs/buffer.c
===================================================================
@@ -3275,11 +3275,23 @@ void __init buffer_init(void)
 				SLAB_MEM_SPREAD),
 				init_buffer_head);
 
+#ifdef CONFIG_PAGE_STATES
+	/*
+	 * If volatile page cache is enabled we want to get as many
+	 * pages into volatile state as possible. Pages with private
+	 * information cannot be made stable. Set max_buffer_heads
+	 * to zero to make shrink_active_list to release the private
+	 * information when moving page from the active to the inactive
+	 * list.
+	 */
+	max_buffer_heads = 0;
+#else
 	/*
 	 * Limit the bh occupancy to 10% of ZONE_NORMAL
 	 */
 	nrpages = (nr_free_buffer_pages() * 10) / 100;
 	max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
+#endif
 	hotcpu_notifier(buffer_cpu_notify, 0);
 }
 

Index: linux-2.6/include/linux/mm.h
===================================================================
@@ -280,6 +280,7 @@ static inline void init_page_count(struc
 }
 
 void put_page(struct page *page);
+void put_page_check(struct page *page);
 void put_pages_list(struct list_head *pages);
 
 void split_page(struct page *page, unsigned int order);

Index: linux-2.6/include/linux/page-flags.h
===================================================================
@@ -108,6 +108,8 @@
 #define PG_uncached		31	/* Page has been mapped as uncached */
 #endif
 
+#define PG_discarded		20	/* Page discarded by the hypervisor. */
+
 /*
  * Manipulation of page state flags
  */

@@ -296,6 +298,17 @@ static inline void __ClearPageTail(struc
 #define SetPageUncached(page)	set_bit(PG_uncached, &(page)->flags)
 #define ClearPageUncached(page)	clear_bit(PG_uncached, &(page)->flags)
 
+#ifdef CONFIG_PAGE_STATES
+#define PageDiscarded(page)	test_bit(PG_discarded, &(page)->flags)
+#define ClearPageDiscarded(page) clear_bit(PG_discarded, &(page)->flags)
+#define TestSetPageDiscarded(page) \
+		test_and_set_bit(PG_discarded, &(page)->flags)
+#else
+#define PageDiscarded(page)		0
+#define ClearPageDiscarded(page)	do { } while (0)
+#define TestSetPageDiscarded(page)	0
+#endif
+
 struct page;	/* forward declaration */
 
 extern void cancel_dirty_page(struct page *page, unsigned int account_size);

Index: linux-2.6/include/linux/pagemap.h
===================================================================
@@ -12,6 +12,7 @@
 #include <asm/uaccess.h>
 #include <linux/gfp.h>
 #include <linux/bitops.h>
+#include <linux/page-states.h>
 
 /*
  * Bits in mapping->flags.  The lower __GFP_BITS_SHIFT bits are the page

@@ -59,7 +60,11 @@ static inline void mapping_set_gfp_mask(
 #define PAGE_CACHE_ALIGN(addr)	(((addr)+PAGE_CACHE_SIZE-1)&PAGE_CACHE_MASK)
 
 #define page_cache_get(page)		get_page(page)
+#ifdef CONFIG_PAGE_STATES
+#define page_cache_release(page)	put_page_check(page)
+#else
 #define page_cache_release(page)	put_page(page)
+#endif
 void release_pages(struct page **pages, int nr, int cold);
 
 #ifdef CONFIG_NUMA

@@ -139,6 +144,7 @@ int add_to_page_cache_lru(struct page *p
 				pgoff_t index, gfp_t gfp_mask);
 extern void remove_from_page_cache(struct page *page);
 extern void __remove_from_page_cache(struct page *page);
+extern void __remove_from_page_cache_nocheck(struct page *page);
 
 /*
  * Return byte-offset into filesystem object for page.

Index: linux-2.6/mm/Makefile
===================================================================
@@ -33,4 +33,5 @@ obj-$(CONFIG_MIGRATION) += migrate.o
 obj-$(CONFIG_SMP) += allocpercpu.o
 obj-$(CONFIG_QUICKLIST) += quicklist.o
 obj-$(CONFIG_CGROUP_MEM_RES_CTLR) += memcontrol.o
+obj-$(CONFIG_PAGE_STATES) += page-states.o
 

Index: linux-2.6/mm/page_alloc.c
===================================================================
@@ -45,6 +45,7 @@
 #include <linux/fault-inject.h>
 #include <linux/page-isolation.h>
 #include <linux/memcontrol.h>
+#include <linux/page-states.h>
 
 #include <asm/tlbflush.h>
 #include <asm/div64.h>

@@ -245,7 +246,8 @@ static void bad_page(struct page *page)
 			1 << PG_slab    |
 			1 << PG_swapcache |
 			1 << PG_writeback |
-			1 << PG_buddy );
+			1 << PG_buddy |
+			1 << PG_discarded );
 	set_page_count(page, 0);
 	reset_page_mapcount(page);
 	page->mapping = NULL;

@@ -531,6 +533,7 @@ static void __free_pages_ok(struct page 
 		reserved += free_pages_check(page + i);
 	if (reserved)
 		return;
+	page_set_unused(page, order);
 
 	if (!PageHighMem(page))
 		debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);

@@ -990,10 +993,16 @@ static void free_hot_cold_page(struct pa
 	struct per_cpu_pages *pcp;
 	unsigned long flags;
 
+	if (unlikely(PageDiscarded(page))) {
+		if (page_free_discarded(page))
+			return;
+	}
+
 	if (PageAnon(page))
 		page->mapping = NULL;
 	if (free_pages_check(page))
 		return;
+	page_set_unused(page, 0);
 
 	if (!PageHighMem(page))
 		debug_check_no_locks_freed(page_address(page), PAGE_SIZE);

@@ -1107,6 +1116,7 @@ again:
 	put_cpu();
 
 	VM_BUG_ON(bad_range(zone, page));
+	page_set_stable(page, order);
 	if (prep_new_page(page, order, gfp_flags))
 		goto again;
 	return page;

@@ -1690,6 +1700,8 @@ void __pagevec_free(struct pagevec *pvec
 
 void __free_pages(struct page *page, unsigned int order)
 {
+	if (page_count(page) > 1)
+		page_make_volatile(page, 2);
 	if (put_page_testzero(page)) {
 		if (order == 0)
 			free_hot_page(page);

Index: linux-2.6/mm/page-writeback.c
===================================================================
@@ -34,6 +34,7 @@
 #include <linux/syscalls.h>
 #include <linux/buffer_head.h>
 #include <linux/pagevec.h>
+#include <linux/page-states.h>
 
 /*
  * The maximum number of pages to writeout in a single bdflush/kupdate

@@ -1192,6 +1193,7 @@ int test_clear_page_writeback(struct pag
 			radix_tree_tag_clear(&mapping->page_tree,
 						page_index(page),
 						PAGECACHE_TAG_WRITEBACK);
+			page_make_volatile(page, 1);
 			if (bdi_cap_writeback_dirty(bdi)) {
 				__dec_bdi_stat(bdi, BDI_WRITEBACK);
 				__bdi_writeout_inc(bdi);

Index: linux-2.6/mm/swap.c
===================================================================
@@ -30,6 +30,7 @@
 #include <linux/notifier.h>
 #include <linux/backing-dev.h>
 #include <linux/memcontrol.h>
+#include <linux/page-states.h>
 
 /* How many pages do we try to swap or page in/out together? */
 int page_cluster;

@@ -77,6 +78,16 @@ void put_page(struct page *page)
 }
 EXPORT_SYMBOL(put_page);
 
+#ifdef CONFIG_PAGE_STATES
+void put_page_check(struct page *page)
+{
+	if (page_count(page) > 1)
+		page_make_volatile(page, 2);
+	put_page(page);
+}
+EXPORT_SYMBOL(put_page_check);
+#endif
+
 /**
  * put_pages_list(): release a list of pages
  *

@@ -382,6 +393,8 @@ void __pagevec_release_nonlru(struct pag
 		struct page *page = pvec->pages[i];
 
 		VM_BUG_ON(PageLRU(page));
+		if (page_count(page) > 1)
+			page_make_volatile(page, 2);
 		if (put_page_testzero(page))
 			pagevec_add(&pages_to_free, page);
 	}

@@ -411,6 +424,7 @@ void __pagevec_lru_add(struct pagevec *p
 		VM_BUG_ON(PageLRU(page));
 		SetPageLRU(page);
 		add_page_to_inactive_list(zone, page);
+		page_make_volatile(page, 2);
 	}
 	if (zone)
 		spin_unlock_irq(&zone->lru_lock);

blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.

In this patch yes, but a later patch adds a condition:

- if (flags & FOLL_GET) {
+ if ((flags & FOLL_GET) || (vma->vm_flags & VM_LOCKED)) {

Fixed.

That is a major nit. This should be an #ifdef. In previous versions the complete "if (!pte_present(*pte)) { }" is ifdefed, the later versions use the !spin_is_locked condition. Only I forgot to invert the #ifndef. Fixed.

I take the small review anytime. Already found one major nit.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

From: Martin Schwidefsky <[email protected]>
From: Hubertus Franke <[email protected]>
From: Himanshu Raj

The volatile page state can be used for anonymous pages as well, if they have been added to the swap cache and the swap write is finished. The tricky bit is in free_swap_and_cache. The call to find_get_page dead-locks with the discard handler. If the page has been discarded find_get_page will try to remove it. To do that it needs the page table lock of all mappers but one is held by the caller of free_swap_and_cache. A special variant of find_get_page is needed that does not check the page state and returns a page reference even if the page is discarded. The second pitfall is that the page needs to be made stable before the swap slot gets freed. If the page cannot be made stable because it has been discarded the swap slot may not be freed because it is still needed to reload the discarded page from the swap device.

---

 include/linux/pagemap.h |    3 ++
 include/linux/swap.h    |    5 ++++
 mm/filemap.c            |   19 +++++++++++++++++
 mm/memory.c             |   13 +++++++++++-
 mm/page-states.c        |   34 +++++++++++++++++++++++---------
 mm/rmap.c               |   51 ++++++++++++++++++++++++++++++++++++++++++++----
 mm/swap_state.c         |   25 ++++++++++++++++++++++-
 mm/swapfile.c           |   30 ++++++++++++++++++++++------
 mm/vmscan.c             |    3 ++
 9 files changed, 162 insertions(+), 21 deletions(-)

Index: linux-2.6/include/linux/pagemap.h
===================================================================
@@ -61,8 +61,11 @@ static inline void mapping_set_gfp_mask(
 
 #define page_cache_get(page)		get_page(page)
 #ifdef CONFIG_PAGE_STATES
+extern struct page * find_get_page_nodiscard(struct address_space *mapping,
+					     unsigned long index);
 #define page_cache_release(page)	put_page_check(page)
 #else
+#define find_get_page_nodiscard(mapping, index) find_get_page(mapping, index)
 #define page_cache_release(page)	put_page(page)
 #endif
 void release_pages(struct page **pages, int nr, int cold);

Index: linux-2.6/include/linux/swap.h
===================================================================
@@ -226,6 +226,7 @@ extern void show_swap_cache_info(void);
 extern int add_to_swap(struct page *, gfp_t);
 extern int add_to_swap_cache(struct page *, swp_entry_t, gfp_t);
 extern void __delete_from_swap_cache(struct page *);
+extern void __delete_from_swap_cache_nocheck(struct page *);
 extern void delete_from_swap_cache(struct page *);
 extern void free_page_and_swap_cache(struct page *);
 extern void free_pages_and_swap_cache(struct page **, int);

@@ -331,6 +332,10 @@ static inline void __delete_from_swap_ca
 {
 }
 
+static inline void __delete_from_swap_cache_nocheck(struct page *page)
+{
+}
+
 static inline void delete_from_swap_cache(struct page *page)
 {
 }

Index: linux-2.6/mm/filemap.c
===================================================================
@@ -537,6 +537,25 @@ static int __sleep_on_page_lock(void *wo
 	return 0;
 }
 
+#ifdef CONFIG_PAGE_STATES
+
+struct page * find_get_page_nodiscard(struct address_space *mapping,
+				      unsigned long offset)
+{
+	struct page *page;
+
+	read_lock_irq(&mapping->tree_lock);
+	page = radix_tree_lookup(&mapping->page_tree, offset);
+	if (page)
+		page_cache_get(page);
+	read_unlock_irq(&mapping->tree_lock);
+	return page;
+}
+
+EXPORT_SYMBOL(find_get_page_nodiscard);
+
+#endif
+
 /*
  * In order to wait for pages to become available there must be
  * waitqueues associated with pages. By using a hash table of

Index: linux-2.6/mm/memory.c
===================================================================
@@ -507,7 +507,18 @@ out_discard_pte:
 	 * in page cache anymore. Do what try_to_unmap_one would do
 	 * if the copy_one_pte had taken place before page_discard.
 	 */
-	if (page->index != linear_page_index(vma, addr))
+	if (PageAnon(page)) {
+		swp_entry_t entry = { .val = page_private(page) };
+		swap_duplicate(entry);
+		if (list_empty(&dst_mm->mmlist)) {
+			spin_lock(&mmlist_lock);
+			if (list_empty(&dst_mm->mmlist))
+				list_add(&dst_mm->mmlist, &init_mm.mmlist);
+			spin_unlock(&mmlist_lock);
+		}
+		pte = swp_entry_to_pte(entry);
+		set_pte_at(dst_mm, addr, dst_pte, pte);
+	} else if (page->index != linear_page_index(vma, addr))
 		/* If nonlinear, store the file page offset in the pte. */
 		set_pte_at(dst_mm, addr, dst_pte, pgoff_to_pte(page->index));
 	else

Index: linux-2.6/mm/page-states.c
===================================================================
@@ -19,6 +19,7 @@
 #include <linux/buffer_head.h>
 #include <linux/pagevec.h>
 #include <linux/page-states.h>
+#include <linux/swap.h>
 
 #include "internal.h"
 

@@ -34,7 +35,16 @@ static inline int check_bits(struct page
 	 */
 	if (PageDirty(page) || PageReserved(page) || PageWriteback(page) ||
 	    PageLocked(page) || PagePrivate(page) || PageDiscarded(page) ||
-	    !PageUptodate(page) || !PageLRU(page) || PageAnon(page))
+	    !PageUptodate(page) || !PageLRU(page) ||
+	    (PageAnon(page) && !PageSwapCache(page)))
+		return 0;
+
+	/*
+	 * Special case shared memory: page is PageSwapCache but not
+	 * PageAnon. page_unmap_all failes for swapped shared memory
+	 * pages.
+	 */
+	if (PageSwapCache(page) && !PageAnon(page))
 		return 0;
 
 	/*

@@ -168,15 +178,21 @@ static void __page_discard(struct page *
 	}
 	spin_unlock_irq(&zone->lru_lock);
 
-	/* We can't handle swap cache pages (yet). */
-	VM_BUG_ON(PageSwapCache(page));
-
-	/* Remove page from page cache. */
+	/* Remove page from page cache/swap cache. */
  	mapping = page->mapping;
-	write_lock_irq(&mapping->tree_lock);
-	__remove_from_page_cache_nocheck(page);
-	write_unlock_irq(&mapping->tree_lock);
-	__put_page(page);
+	if (PageSwapCache(page)) {
+		swp_entry_t entry = { .val = page_private(page) };
+		write_lock_irq(&swapper_space.tree_lock);
+		__delete_from_swap_cache_nocheck(page);
+		write_unlock_irq(&swapper_space.tree_lock);
+		swap_free(entry);
+		page_cache_release(page);
+	} else {
+		write_lock_irq(&mapping->tree_lock);
+		__remove_from_page_cache_nocheck(page);
+		write_unlock_irq(&mapping->tree_lock);
+ 		__put_page(page);
+	}
 }
 
 /**

Index: linux-2.6/mm/swapfile.c
===================================================================
@@ -28,6 +28,7 @@
 #include <linux/capability.h>
 #include <linux/syscalls.h>
 #include <linux/memcontrol.h>
+#include <linux/page-states.h>
 
 #include <asm/pgtable.h>
 #include <asm/tlbflush.h>

@@ -370,9 +371,11 @@ int remove_exclusive_swap_page(struct pa
 		/* Recheck the page count with the swapcache lock held.. */
 		write_lock_irq(&swapper_space.tree_lock);
 		if ((page_count(page) == 2) && !PageWriteback(page)) {
-			__delete_from_swap_cache(page);
-			SetPageDirty(page);
-			retval = 1;
+			if (likely(page_make_stable(page))) {
+				__delete_from_swap_cache(page);
+				SetPageDirty(page);
+				retval = 1;
+			}
 		}
 		write_unlock_irq(&swapper_space.tree_lock);
 	}

@@ -401,7 +404,13 @@ void free_swap_and_cache(swp_entry_t ent
 	p = swap_info_get(entry);
 	if (p) {
 		if (swap_entry_free(p, swp_offset(entry)) == 1) {
-			page = find_get_page(&swapper_space, entry.val);
+			/*
+			 * Use find_get_page_nodiscard to avoid the deadlock
+			 * on the swap_lock and the page table lock if the
+			 * page has been discarded.
+			 */
+			page = find_get_page_nodiscard(&swapper_space,
+						       entry.val);
 			if (page && unlikely(TestSetPageLocked(page))) {
 				page_cache_release(page);
 				page = NULL;

@@ -418,8 +427,17 @@ void free_swap_and_cache(swp_entry_t ent
 		/* Also recheck PageSwapCache after page is locked (above) */
 		if (PageSwapCache(page) && !PageWriteback(page) &&
 					(one_user || vm_swap_full())) {
-			delete_from_swap_cache(page);
-			SetPageDirty(page);
+			/*
+			 * To be able to reload the page from swap the
+			 * swap slot may not be freed. The caller of
+			 * free_swap_and_cache holds a page table lock
+			 * for this page. The discarded page can not be
+			 * removed here.
+			 */
+			if (likely(page_make_stable(page))) {
+				delete_from_swap_cache(page);
+				SetPageDirty(page);
+			}
 		}
 		unlock_page(page);
 		page_cache_release(page);

Index: linux-2.6/mm/swap_state.c
===================================================================
@@ -17,6 +17,7 @@
 #include <linux/backing-dev.h>
 #include <linux/pagevec.h>
 #include <linux/migrate.h>
+#include <linux/page-states.h>
 
 #include <asm/pgtable.h>
 

@@ -97,7 +98,7 @@ int add_to_swap_cache(struct page *page,
  * This must be called only on pages that have
  * been verified to be in the swap cache.
  */
-void __delete_from_swap_cache(struct page *page)
+void inline __delete_from_swap_cache_nocheck(struct page *page)
 {
 	BUG_ON(!PageLocked(page));
 	BUG_ON(!PageSwapCache(page));

@@ -112,6 +113,28 @@ void __delete_from_swap_cache(struct pag
 	INC_CACHE_INFO(del_total);
 }
 
+void __delete_from_swap_cache(struct page *page)
+{
+	/*
+	 * Check if the discard fault handler already removed
+	 * the page from the page cache. If not set the discard
+	 * bit in the page flags to prevent double page free if
+	 * a discard fault is racing with normal page free.
+	 */
+	if (TestSetPageDiscarded(page))
+		return;
+
+	__delete_from_swap_cache_nocheck(page);
+
+	/*
+	 * Check the hardware page state and clear the discard
+	 * bit in the page flags only if the page is not
+	 * discarded.
+	 */
+	if (!page_discarded(page))
+		ClearPageDiscarded(page);
+}
+
 /**
  * add_to_swap - allocate swap space for a page
  * @page: page we want to move to swap

Index: linux-2.6/mm/vmscan.c
===================================================================
@@ -492,6 +492,9 @@ static unsigned long shrink_page_list(st
 
 		sc->nr_scanned++;
 
+		if (unlikely(PageDiscarded(page)))
+			goto free_it;
+
 		if (!sc->may_swap && page_mapped(page))
 			goto keep_locked;
 

blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.

From: Martin Schwidefsky <[email protected]>
From: Hubertus Franke <[email protected]>
From: Himanshu Raj

Add code to get mlock() working with guest page hinting. The problem with mlock is that locked pages may not be removed from page cache. That means they need to be stable. page_make_volatile needs a way to check if a page has been locked. To avoid traversing vma lists - which would hurt performance a lot - a field is added in the struct address_space. This field is set in mlock_fixup if a vma gets mlocked. The bit never gets removed - once a file had an mlocked vma all future pages added to it will stay stable.

The pages of an mlocked area are made present in the linux page table by a call to make_pages_present which calls get_user_pages and follow_page. The follow_page function is called for each page in the mlocked vma, if the VM_LOCKED bit in the vma flags is set the page is made stable.

---

 include/linux/fs.h |   10 ++++++++++
 mm/memory.c        |    5 +++--
 mm/mlock.c         |    3 +++
 mm/page-states.c   |    5 ++++-
 mm/rmap.c          |   13 +++++++++++--
 5 files changed, 31 insertions(+), 5 deletions(-)

Index: linux-2.6/include/linux/fs.h
===================================================================
@@ -513,6 +513,9 @@ struct address_space {
 	spinlock_t		private_lock;	/* for use by the address_space */
 	struct list_head	private_list;	/* ditto */
 	struct address_space	*assoc_mapping;	/* ditto */
+#ifdef CONFIG_PAGE_STATES
+	unsigned int		mlocked;	/* set if VM_LOCKED vmas present */
+#endif
 } __attribute__((aligned(sizeof(long))));
 	/*
 	 * On most architectures that alignment is already the case; but

@@ -520,6 +523,13 @@ struct address_space {
 	 * of struct page's "mapping" pointer be used for PAGE_MAPPING_ANON.
 	 */
 
+static inline void mapping_set_mlocked(struct address_space *mapping)
+{
+#ifdef CONFIG_PAGE_STATES
+	mapping->mlocked = 1;
+#endif
+}
+
 struct block_device {
 	dev_t			bd_dev;  /* not a kdev_t - it's a search key */
 	struct inode *		bd_inode;	/* will die */

Index: linux-2.6/mm/memory.c
===================================================================
@@ -987,9 +987,10 @@ struct page *follow_page(struct vm_area_
 	if (flags & FOLL_GET)
 		get_page(page);
 
-	if (flags & FOLL_GET) {
+	if ((flags & FOLL_GET) || (vma->vm_flags & VM_LOCKED)) {
 		/*
-		 * The page is made stable if a reference is acquired.
+		 * The page is made stable if a reference is acquired or
+		 * the vm area is locked.
 		 * If the caller does not get a reference it implies that
 		 * the caller can deal with page faults in case the page
 		 * is swapped out. In this case the caller can deal with

Index: linux-2.6/mm/mlock.c
===================================================================
@@ -12,6 +12,7 @@
 #include <linux/syscalls.h>
 #include <linux/sched.h>
 #include <linux/module.h>
+#include <linux/fs.h>
 
 int can_do_mlock(void)
 {

@@ -71,6 +72,8 @@ success:
 	 */
 	pages = (end - start) >> PAGE_SHIFT;
 	if (newflags & VM_LOCKED) {
+		if (vma->vm_file && vma->vm_file->f_mapping)
+			mapping_set_mlocked(vma->vm_file->f_mapping);
 		pages = -pages;
 		if (!(newflags & VM_IO))
 			ret = make_pages_present(start, end);

Index: linux-2.6/mm/page-states.c
===================================================================
@@ -29,6 +29,8 @@
  */
 static inline int check_bits(struct page *page)
 {
+	struct address_space *mapping;
+
 	/*
 	 * There are several conditions that prevent a page from becoming
 	 * volatile. The first check is for the page bits.

@@ -52,7 +54,8 @@ static inline int check_bits(struct page
 	 * it volatile. It will be freed soon. And if the mapping ever
 	 * had locked pages all pages of the mapping will stay stable.
 	 */
-	return page_mapping(page) != NULL;
+	mapping = page_mapping(page);
+	return mapping && !mapping->mlocked;
 }
 
 /*

Index: linux-2.6/mm/rmap.c
===================================================================
@@ -722,8 +722,17 @@ static int try_to_unmap_one(struct page 
 	 */
 	if (!migration && ((vma->vm_flags & VM_LOCKED) ||
 			(ptep_clear_flush_young(vma, address, pte)))) {
-		ret = SWAP_FAIL;
-		goto out_unmap;
+		/*
+		 * Check for discarded pages. This can happen if there have
+		 * been discarded pages before a vma gets mlocked. The code
+		 * in make_pages_present will force all discarded pages out
+		 * and reload them. That happens after the VM_LOCKED bit
+		 * has been set.
+		 */
+		if (likely(!PageDiscarded(page))) {
+			ret = SWAP_FAIL;
+			goto out_unmap;
+		}
 	}
 
 	/* Nuke the page table entry. */

blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.

Yes, makes sense. I moved the #ifdef.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

From: Martin Schwidefsky <[email protected]>
From: Hubertus Franke <[email protected]>
From: Himanshu Raj

The volatile state for page cache and swap cache pages requires that the host system needs to be able to determine if a volatile page is dirty before removing it. This excludes almost all platforms from using the scheme. What is needed is a way to distinguish between pages that are purely read-only and pages that might get written to. This allows platforms with per-pte dirty bits to use the scheme and platforms with per-page dirty bits a small optimization.

Whenever a writable pte is created a check is added that allows to move the page into the correct state. This needs to be done before the writable pte is established. To avoid unnecessary state transitions and the need for a counter, a new page flag PG_writable is added. Only the creation of the first writable pte will do a page state change. Even if all the writable ptes pointing to a page are removed again, the page stays in the safe state until all read-only users of the page have unmapped it as well. Only then is the PG_writable bit reset.

The state a page needs to have if a writable pte is present depends on the platform. A platform with per-pte dirty bits wants to move the page into stable state, a platform with per-page dirty bits like s390 can decide to move the page into a special state that requires the host system to check the dirty bit before discarding a page.

---

 fs/exec.c                   |    1 
 include/linux/page-flags.h  |    6 ++++
 include/linux/page-states.h |   27 +++++++++++++++++++-
 mm/memory.c                 |    5 +++
 mm/mprotect.c               |    2 +
 mm/page-states.c            |   58 ++++++++++++++++++++++++++++++++++++++++++--
 mm/page_alloc.c             |    3 +-
 mm/rmap.c                   |    1 
 8 files changed, 99 insertions(+), 4 deletions(-)

Index: linux-2.6/fs/exec.c
===================================================================
@@ -51,6 +51,7 @@
 #include <linux/tsacct_kern.h>
 #include <linux/cn_proc.h>
 #include <linux/audit.h>
+#include <linux/page-states.h>
 
 #include <asm/uaccess.h>
 #include <asm/mmu_context.h>

Index: linux-2.6/include/linux/page-flags.h
===================================================================
@@ -109,6 +109,7 @@
 #endif
 
 #define PG_discarded		20	/* Page discarded by the hypervisor. */
+#define PG_writable		21	/* Page is mapped writable. */
 
 /*
  * Manipulation of page state flags

@@ -309,6 +310,11 @@ static inline void __ClearPageTail(struc
 #define TestSetPageDiscarded(page)	0
 #endif
 
+#define PageWritable(page) test_bit(PG_writable, &(page)->flags)
+#define TestSetPageWritable(page) \
+		test_and_set_bit(PG_writable, &(page)->flags)
+#define ClearPageWritable(page) clear_bit(PG_writable, &(page)->flags)
+
 struct page;	/* forward declaration */
 
 extern void cancel_dirty_page(struct page *page, unsigned int account_size);

Index: linux-2.6/include/linux/page-states.h
===================================================================
@@ -57,6 +57,9 @@ extern void page_discard(struct page *pa
 extern int  __page_make_stable(struct page *page);
 extern void __page_make_volatile(struct page *page, int offset);
 extern void __pagevec_make_volatile(struct pagevec *pvec);
+extern void __page_check_writable(struct page *page, pte_t pte,
+				  unsigned int offset);
+extern void __page_reset_writable(struct page *page);
 
 /*
  * Extended guest page hinting functions defined by using the

@@ -78,6 +81,12 @@ extern void __pagevec_make_volatile(stru
  *     from the LRU list and the radix tree of its mapping.
  *     page_discard uses page_unmap_all to remove all page table
  *     entries for a page.
+ * - page_check_writable:
+ *     Checks if the page states needs to be adapted because a new
+ *     writable page table entry refering to the page is established.
+ * - page_reset_writable:
+ *     Resets the page state after the last writable page table entry
+ *     refering to the page has been removed.
  */
 
 static inline int page_make_stable(struct page *page)

@@ -97,12 +106,26 @@ static inline void pagevec_make_volatile
 		__pagevec_make_volatile(pvec);
 }
 
+static inline void page_check_writable(struct page *page, pte_t pte,
+				       unsigned int offset)
+{
+	if (page_host_discards() && pte_write(pte) &&
+	    !test_bit(PG_writable, &page->flags))
+		__page_check_writable(page, pte, offset);
+}
+
+static inline void page_reset_writable(struct page *page)
+{
+	if (page_host_discards() && test_bit(PG_writable, &page->flags))
+		__page_reset_writable(page);
+}
+
 #else
 
 #define page_host_discards()			(0)
 #define page_set_unused(_page,_order)		do { } while (0)
 #define page_set_stable(_page,_order)		do { } while (0)
-#define page_set_volatile(_page)		do { } while (0)
+#define page_set_volatile(_page,_writable)	do { } while (0)
 #define page_set_stable_if_present(_page)	(1)
 #define page_discarded(_page)			(0)
 #define page_volatile(_page)			(0)

@@ -117,6 +140,8 @@ static inline void pagevec_make_volatile
 #define page_make_volatile(_page, offset)	do { } while (0)
 #define pagevec_make_volatile(_pagevec)	do { } while (0)
 #define page_discard(_page)			do { } while (0)
+#define page_check_writable(_page,_pte,_off)	do { } while (0)
+#define page_reset_writable(_page)		do { } while (0)
 
 #endif
 

Index: linux-2.6/mm/memory.c
===================================================================
@@ -1681,6 +1681,7 @@ static int do_wp_page(struct mm_struct *
 		flush_cache_page(vma, address, pte_pfn(orig_pte));
 		entry = pte_mkyoung(orig_pte);
 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
+		page_check_writable(old_page, entry, 1);
 		if (ptep_set_access_flags(vma, address, page_table, entry,1))
 			update_mmu_cache(vma, address, entry);
 		ret |= VM_FAULT_WRITE;

@@ -1728,6 +1729,7 @@ gotten:
 		flush_cache_page(vma, address, pte_pfn(orig_pte));
 		entry = mk_pte(new_page, vma->vm_page_prot);
 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
+		page_check_writable(new_page, entry, 2);
 		/*
 		 * Clear the pte entry and flush it first, before updating the
 		 * pte with the new entry. This will avoid a race condition

@@ -2147,6 +2149,7 @@ static int do_swap_page(struct mm_struct
 	}
 
 	flush_icache_page(vma, page);
+	page_check_writable(page, pte, 2);
 	set_pte_at(mm, address, page_table, pte);
 	page_add_anon_rmap(page, vma, address);
 

@@ -2204,6 +2207,7 @@ static int do_anonymous_page(struct mm_s
 
 	entry = mk_pte(page, vma->vm_page_prot);
 	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
+	page_check_writable(page, entry, 2);
 
 	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
 	if (!pte_none(*page_table))

@@ -2365,6 +2369,7 @@ retry:
 		entry = mk_pte(page, vma->vm_page_prot);
 		if (flags & FAULT_FLAG_WRITE)
 			entry = maybe_mkwrite(pte_mkdirty(entry), vma);
+		page_check_writable(page, entry, 2);
 		set_pte_at(mm, address, page_table, entry);
 		if (anon) {
                         inc_mm_counter(mm, anon_rss);

Index: linux-2.6/mm/mprotect.c
===================================================================
@@ -21,6 +21,7 @@
 #include <linux/syscalls.h>
 #include <linux/swap.h>
 #include <linux/swapops.h>
+#include <linux/page-states.h>
 #include <asm/uaccess.h>
 #include <asm/pgtable.h>
 #include <asm/cacheflush.h>

@@ -52,6 +53,7 @@ static void change_pte_range(struct mm_s
 			 */
 			if (dirty_accountable && pte_dirty(ptent))
 				ptent = pte_mkwrite(ptent);
+			page_check_writable(pte_page(ptent), ptent, 1);
 			set_pte_at(mm, addr, pte, ptent);
 #ifdef CONFIG_MIGRATION
 		} else if (!pte_file(oldpte)) {

Index: linux-2.6/mm/page_alloc.c
===================================================================
@@ -637,7 +637,8 @@ static int prep_new_page(struct page *pa
 
 	page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_readahead |
 			1 << PG_referenced | 1 << PG_arch_1 |
-			1 << PG_owner_priv_1 | 1 << PG_mappedtodisk);
+			1 << PG_owner_priv_1 | 1 << PG_mappedtodisk |
+			1 << PG_writable);
 	set_page_private(page, 0);
 	set_page_refcounted(page);
 

Index: linux-2.6/mm/rmap.c
===================================================================
@@ -695,6 +695,7 @@ void page_remove_rmap(struct page *page,
 
 		__dec_zone_page_state(page,
 				PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
+		page_reset_writable(page);
 	}
 }
 

blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.

True. I removed the include.

It is an error on the safe side. If the page writable bit is set then the page state has to indicate to the host that the page dirty bit needs to be checked.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

From: Martin Schwidefsky <[email protected]>
From: Hubertus Franke <[email protected]>
From: Himanshu Raj

On of the challenges of the guest page hinting scheme is the cost for the state transitions. If the cost gets too high the whole concept of page state information is in question. Therefore it is important to avoid the state transitions when possible. One place where the state transitions can be avoided are minor faults. Why change the page state to stable in find_get_page and back in page_add_anon_rmap/ page_add_file_rmap if the discarded pages can be handled by the discard fault handler? If the page is in page/swap cache just map it even if it is already discarded. The first access to the page will cause a discard fault which needs to be able to deal with this kind of situation anyway because of other races in the memory management.

The special find_get_page_nodiscard variant introduced for volatile swap cache is used which does not change the page state. The calls to find_get_page in filemap_nopage and lookup_swap_cache are replaced with find_get_page_nodiscard. By the use of this function a new race is created. If a minor fault races with the discard of a page the page may not get mapped to the page table because the discard handler removed the page from the cache which removes the page->mapping that is needed to find the page table entry. A check for the discarded bit is added to do_swap_page and do_no_page. The page table lock for the pte takes care of the synchronization.

That removes the state transitions on the minor fault path. A page that has been mapped will eventually be unmapped again. On the unmap path each page that has been removed from the page table is freed with a call to page_cache_release. In general that causes an unnecessary page state transition from volatile to volatile. To get rid of these state transitions as well a special variants of page_cache_release is added that does not attempt to make the page volatile. page_cache_release_nocheck is then used in free_page_and_swap_cache and release_pages. This makes the unmap of ptes state transitions free.

---

 include/linux/pagemap.h |    4 ++++
 include/linux/swap.h    |    2 +-
 mm/filemap.c            |   35 ++++++++++++++++++++++++++++++++---
 mm/fremap.c             |    1 +
 mm/memory.c             |    4 ++--
 mm/rmap.c               |    4 +---
 mm/shmem.c              |    7 +++++++
 mm/swap_state.c         |    4 ++--
 8 files changed, 50 insertions(+), 11 deletions(-)

Index: linux-2.6/include/linux/pagemap.h
===================================================================
@@ -63,11 +63,15 @@ static inline void mapping_set_gfp_mask(
 #ifdef CONFIG_PAGE_STATES
 extern struct page * find_get_page_nodiscard(struct address_space *mapping,
 					     unsigned long index);
+extern struct page * find_lock_page_nodiscard(struct address_space *mapping,
+					      unsigned long index);
 #define page_cache_release(page)	put_page_check(page)
 #else
 #define find_get_page_nodiscard(mapping, index) find_get_page(mapping, index)
+#define find_lock_page_nodiscard(mapping, index) find_lock_page(mapping, index)
 #define page_cache_release(page)	put_page(page)
 #endif
+#define page_cache_release_nocheck(page)	put_page(page)
 void release_pages(struct page **pages, int nr, int cold);
 
 #ifdef CONFIG_NUMA

Index: linux-2.6/include/linux/swap.h
===================================================================
@@ -288,7 +288,7 @@ static inline void disable_swap_token(vo
 /* only sparc can not include linux/pagemap.h in this file
  * so leave page_cache_release and release_pages undeclared... */
 #define free_page_and_swap_cache(page) \
-	page_cache_release(page)
+	page_cache_release_nocheck(page)
 #define free_pages_and_swap_cache(pages, nr) \
 	release_pages((pages), (nr), 0);
 

Index: linux-2.6/mm/filemap.c
===================================================================
@@ -554,6 +554,35 @@ struct page * find_get_page_nodiscard(st
 
 EXPORT_SYMBOL(find_get_page_nodiscard);
 
+struct page *find_lock_page_nodiscard(struct address_space *mapping,
+				      unsigned long offset)
+{
+	struct page *page;
+
+	read_lock_irq(&mapping->tree_lock);
+repeat:
+	page = radix_tree_lookup(&mapping->page_tree, offset);
+	if (page) {
+		page_cache_get(page);
+		if (TestSetPageLocked(page)) {
+			read_unlock_irq(&mapping->tree_lock);
+			__lock_page(page);
+			read_lock_irq(&mapping->tree_lock);
+
+			/* Has the page been truncated while we slept? */
+			if (unlikely(page->mapping != mapping ||
+				     page->index != offset)) {
+				unlock_page(page);
+				page_cache_release(page);
+				goto repeat;
+			}
+		}
+	}
+	read_unlock_irq(&mapping->tree_lock);
+	return page;
+}
+EXPORT_SYMBOL(find_lock_page_nodiscard);
+
 #endif
 
 /*

@@ -1424,7 +1453,7 @@ int filemap_fault(struct vm_area_struct 
 	 * Do we have something in the page cache already?
 	 */
 retry_find:
-	page = find_lock_page(mapping, vmf->pgoff);
+	page = find_lock_page_nodiscard(mapping, vmf->pgoff);
 	/*
 	 * For sequential accesses, we use the generic readahead logic.
 	 */

@@ -1432,7 +1461,7 @@ retry_find:
 		if (!page) {
 			page_cache_sync_readahead(mapping, ra, file,
 							   vmf->pgoff, 1);
-			page = find_lock_page(mapping, vmf->pgoff);
+			page = find_lock_page_nodiscard(mapping, vmf->pgoff);
 			if (!page)
 				goto no_cached_page;
 		}

@@ -1471,7 +1500,7 @@ retry_find:
 				start = vmf->pgoff - ra_pages / 2;
 			do_page_cache_readahead(mapping, file, start, ra_pages);
 		}
-		page = find_lock_page(mapping, vmf->pgoff);
+		page = find_lock_page_nodiscard(mapping, vmf->pgoff);
 		if (!page)
 			goto no_cached_page;
 	}

Index: linux-2.6/mm/fremap.c
===================================================================
@@ -15,6 +15,7 @@
 #include <linux/rmap.h>
 #include <linux/module.h>
 #include <linux/syscalls.h>
+#include <linux/page-states.h>
 
 #include <asm/mmu_context.h>
 #include <asm/cacheflush.h>

Index: linux-2.6/mm/memory.c
===================================================================
@@ -2131,7 +2131,7 @@ static int do_swap_page(struct mm_struct
 	 * Back out if somebody else already faulted in this pte.
 	 */
 	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
-	if (unlikely(!pte_same(*page_table, orig_pte)))
+	if (unlikely(!pte_same(*page_table, orig_pte) || PageDiscarded(page)))
 		goto out_nomap;
 
 	if (unlikely(!PageUptodate(page))) {

@@ -2364,7 +2364,7 @@ retry:
 	 * handle that later.
 	 */
 	/* Only go through if we didn't race with anybody else... */
-	if (likely(pte_same(*page_table, orig_pte))) {
+	if (likely(pte_same(*page_table, orig_pte) && !PageDiscarded(page))) {
 		flush_icache_page(vma, page);
 		entry = mk_pte(page, vma->vm_page_prot);
 		if (flags & FAULT_FLAG_WRITE)

Index: linux-2.6/mm/rmap.c
===================================================================
@@ -593,7 +593,6 @@ void page_add_anon_rmap(struct page *pag
 		 */
 		mem_cgroup_uncharge_page(page);
 	}
-	page_make_volatile(page, 1);
 }
 
 /*

@@ -630,7 +629,6 @@ void page_add_file_rmap(struct page *pag
 		 * This takes care of balancing the reference counts
 		 */
 		mem_cgroup_uncharge_page(page);
-	page_make_volatile(page, 1);
 }
 
 #ifdef CONFIG_DEBUG_VM

@@ -795,7 +793,7 @@ static int try_to_unmap_one(struct page 
 	}
 
 	page_remove_rmap(page, vma);
-	page_cache_release(page);
+	page_cache_release_nocheck(page);
 
 out_unmap:
 	pte_unmap_unlock(pte, ptl);

Index: linux-2.6/mm/shmem.c
===================================================================
@@ -50,6 +50,7 @@
 #include <linux/migrate.h>
 #include <linux/highmem.h>
 #include <linux/seq_file.h>
+#include <linux/page-states.h>
 
 #include <asm/uaccess.h>
 #include <asm/div64.h>

@@ -1290,6 +1291,12 @@ repeat:
 	if (swap.val) {
 		/* Look it up and read it in.. */
 		swappage = lookup_swap_cache(swap);
+		if (swappage && unlikely(!page_make_stable(swappage))) {
+			shmem_swp_unmap(entry);
+			spin_unlock(&info->lock);
+			page_discard(swappage);
+			goto repeat;
+		}
 		if (!swappage) {
 			shmem_swp_unmap(entry);
 			/* here we actually do the io */

Index: linux-2.6/mm/swap_state.c
===================================================================
@@ -228,7 +228,7 @@ static inline void free_swap_cache(struc
 void free_page_and_swap_cache(struct page *page)
 {
 	free_swap_cache(page);
-	page_cache_release(page);
+	page_cache_release_nocheck(page);
 }
 
 /*

@@ -262,7 +262,7 @@ struct page * lookup_swap_cache(swp_entr
 {
 	struct page *page;
 
-	page = find_get_page(&swapper_space, entry.val);
+	page = find_get_page_nodiscard(&swapper_space, entry.val);
 
 	if (page)
 		INC_CACHE_INFO(find_success);

blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.

From: Martin Schwidefsky <[email protected]>
From: Hubertus Franke <[email protected]>
From: Himanshu Raj

s390 uses the milli-coded ESSA instruction to set the page state. The page state is formed by four guest page states called block usage states and three host page states called block content states.

The guest states are:
- stable (S): there is essential content in the page
- unused (U): there is no useful content and any access to the page will cause an addressing exception
- volatile (V): there is useful content in the page. The host system is allowed to discard the content anytime, but has to deliver a discard fault with the absolute address of the page if the guest tries to access it.
- potential volatile (P): the page has useful content. The host system is allowed to discard the content after it has checked the dirty bit of the page. It has to deliver a discard fault with the absolute address of the page if the guest tries to access it.

The host states are:
- resident: the page is present in real memory.
- preserved: the page is not present in real memory but the content is preserved elsewhere by the machine, e.g. on the paging device.
- zero: the page is not present in real memory. The content of the page is logically-zero.

There are 12 combinations of guest and host state, currently only 8 are valid page states:
Sr: a stable, resident page.
Sp: a stable, preserved page.
Sz: a stable, logically zero page. A page filled with zeroes will be
allocated on first access.
Ur: an unused but resident page. The host could make it Uz anytime but
it doesn't have to.
Uz: an unused, logically zero page.
Vr: a volatile, resident page. The guest can access it normally.
Vz: a volatile, logically zero page. This is a discarded page. The host
will deliver a discard fault for any access to the page.
Pr: a potential volatile, resident page. The guest can access it normally.

The remaining 4 combinations can't occur:
Up: an unused, preserved page. If the host tries to get rid of a Ur page
it will remove it without writing the page content to disk and set the page to Uz.
Vp: a volatile, preserved page. If the host picks a Vr page for eviction
it will discard it and set the page state to Vz.
Pp: a potential volatile, preserved page. There are two cases for page out:
1) if the page is dirty then the host will preserved the page and set it to Sp or 2) if the page is clean then the host will discard it and set the page state to Vz.
Pz: a potential volatile, logically zero page. The host system will always
use Vz instead of Pz.

The state transitions (a diagram would be nicer but that is too hard to do in ascii art...):
{Ur,Sr,Vr,Pr}: a resident page will change its block usage state if the
guest requests it with page_set_{unused,stable,volatile}.
{Uz,Sz,Vz}: a logically zero page will change its block usage state if the
guest requests it with page_set_{unused,stable,volatile}. The guest can't create the Pz state, the state will be Vz instead.
Ur -> Uz: the host system can remove an unused, resident page from memory Sz -> Sr: on first access a stable, logically zero page will become resident Sr -> Sp: the host system can swap a stable page to disk
Sp -> Sr: a guest access to a Sp page forces the host to retrieve it Vr -> Vz: the host can discard a volatile page
Sp -> Uz: a page preserved by the host will be removed if the guest sets
the block usage state to unused.
Sp -> Vz: a page preserved by the host will be discarded if the guest sets
the block usage state to volatile.
Pr -> Sp: the host can move a page from Pr to Sp if it discovers that the
page is dirty while trying to discard the page. The page content is written to the paging device.
Pr -> Vz: the host can discard a Pr page. The Pz state is replaced by the
Vz state.

The are some hazards the code has to deal with:
1) For potential volatile pages the transfer of the hardware dirty bit to the software dirty bit needs to make sure that the page gets into the stable state before the hardware dirty bit is cleared. Between the page_test_dirty and the page_clear_dirty call a page_make_stable is required.

2) Since the access of unused pages causes addressing exceptions we need to take care with /dev/mem. The copy_{from_to}_user functions need to be able to cope with addressing exceptions for the kernel address space.

3) The discard fault on a s390 machine delivers the absolute address of the page that caused the fault instead of the virtual address. With the virtual address we could have used the page table entry of the current process to safely get a reference to the discarded page. We can get to the struct page from the absolute page address but it is rather hard to get to a proper page reference. The page that caused the fault could already have been freed and reused for a different purpose. None of the fields in the struct page would be reliable to use. The freeing of discarded pages therefore has to be postponed until all pending discard faults for this page have been dealt with. The discard fault handler is called disabled for interrupts and tries to get a page reference with get_page_unless_zero. A discarded page is only freed after all cpus have been enabled for interrupts at least once since the detection of the discarded page. This is done using the timer interrupts and the cpu-idle notifier.

---

 arch/s390/Kconfig              |    3 
 arch/s390/kernel/time.c        |   11 ++
 arch/s390/kernel/traps.c       |    4 
 arch/s390/lib/uaccess_mvcos.c  |   10 +
 arch/s390/lib/uaccess_std.c    |    7 -
 arch/s390/mm/fault.c           |  210 +++++++++++++++++++++++++++++++++++++++++
 include/asm-s390/page-states.h |  117 ++++++++++++++++++++++
 mm/rmap.c                      |    9 +
 8 files changed, 364 insertions(+), 7 deletions(-)

Index: linux-2.6/arch/s390/Kconfig
===================================================================
@@ -411,6 +411,9 @@ config CMM_IUCV
 	  Select this option to enable the special message interface to
 	  the cooperative memory management.
 
+config PAGE_STATES
+	bool "Enable support for guest page hinting."
+
 config VIRT_TIMER
 	bool "Virtual CPU timer support"
 	help

Index: linux-2.6/arch/s390/kernel/time.c
===================================================================
@@ -30,6 +30,7 @@
 #include <linux/timex.h>
 #include <linux/notifier.h>
 #include <linux/clocksource.h>
+#include <linux/page-states.h>
 
 #include <asm/uaccess.h>
 #include <asm/delay.h>

@@ -222,6 +223,9 @@ static int nohz_idle_notify(struct notif
 	switch (action) {
 	case S390_CPU_IDLE:
 		stop_hz_timer();
+#ifdef CONFIG_PAGE_STATES
+		page_shrink_discard_list();
+#endif
 		break;
 	case S390_CPU_NOT_IDLE:
 		start_hz_timer();

@@ -270,6 +274,9 @@ void init_cpu_timer(void)
 
 static void clock_comparator_interrupt(__u16 code)
 {
+#ifdef CONFIG_PAGE_STATES
+	page_shrink_discard_list();
+#endif
 	/* set clock comparator for next tick */
 	set_clock_comparator(S390_lowcore.jiffy_timer + CPU_DEVIATION);
 }

@@ -349,6 +356,10 @@ void __init time_init(void)
 #ifdef CONFIG_VIRT_TIMER
 	vtime_init();
 #endif
+
+#ifdef CONFIG_PAGE_STATES
+	page_discard_init();
+#endif
 }
 
 /*

Index: linux-2.6/arch/s390/kernel/traps.c
===================================================================
@@ -61,6 +61,7 @@ extern pgm_check_handler_t do_protection
 extern pgm_check_handler_t do_dat_exception;
 extern pgm_check_handler_t do_monitor_call;
 extern pgm_check_handler_t do_asce_exception;
+extern pgm_check_handler_t do_discard_fault;
 
 #define stack_pointer ({ void **sp; asm("la %0,0(15)" : "=&d" (sp)); sp; })
 

@@ -740,5 +741,8 @@ void __init trap_init(void)
         pgm_check_table[0x1C] = &space_switch_exception;
         pgm_check_table[0x1D] = &hfp_sqrt_exception;
 	pgm_check_table[0x40] = &do_monitor_call;
+#ifdef CONFIG_PAGE_STATES
+	pgm_check_table[0x1a] = &do_discard_fault;
+#endif
 	pfault_irq_init();
 }

Index: linux-2.6/arch/s390/lib/uaccess_mvcos.c
===================================================================
@@ -36,7 +36,7 @@ static size_t copy_from_user_mvcos(size_
 	tmp1 = -4096UL;
 	asm volatile(
 		"0: .insn ss,0xc80000000000,0(%0,%2),0(%1),0\n"
-		"   jz    7f\n"
+		"10:jz    7f\n"
 		"1:"ALR"  %0,%3\n"
 		"  "SLR"  %1,%3\n"
 		"  "SLR"  %2,%3\n"

@@ -47,7 +47,7 @@ static size_t copy_from_user_mvcos(size_
 		"  "CLR"  %0,%4\n"	/* copy crosses next page boundary? */
 		"   jnh   4f\n"
 		"3: .insn ss,0xc80000000000,0(%4,%2),0(%1),0\n"
-		"  "SLR"  %0,%4\n"
+		"11:"SLR"  %0,%4\n"
 		"  "ALR"  %2,%4\n"
 		"4:"LHI"  %4,-1\n"
 		"  "ALR"  %4,%0\n"	/* copy remaining size, subtract 1 */

@@ -62,6 +62,7 @@ static size_t copy_from_user_mvcos(size_
 		"7:"SLR"  %0,%0\n"
 		"8: \n"
 		EX_TABLE(0b,2b) EX_TABLE(3b,4b)
+		EX_TABLE(10b,8b) EX_TABLE(11b,8b)
 		: "+a" (size), "+a" (ptr), "+a" (x), "+a" (tmp1), "=a" (tmp2)
 		: "d" (reg0) : "cc", "memory");
 	return size;

@@ -82,7 +83,7 @@ static size_t copy_to_user_mvcos(size_t 
 	tmp1 = -4096UL;
 	asm volatile(
 		"0: .insn ss,0xc80000000000,0(%0,%1),0(%2),0\n"
-		"   jz    4f\n"
+		"6: jz    4f\n"
 		"1:"ALR"  %0,%3\n"
 		"  "SLR"  %1,%3\n"
 		"  "SLR"  %2,%3\n"

@@ -93,11 +94,12 @@ static size_t copy_to_user_mvcos(size_t 
 		"  "CLR"  %0,%4\n"	/* copy crosses next page boundary? */
 		"   jnh   5f\n"
 		"3: .insn ss,0xc80000000000,0(%4,%1),0(%2),0\n"
-		"  "SLR"  %0,%4\n"
+		"7:"SLR"  %0,%4\n"
 		"   j     5f\n"
 		"4:"SLR"  %0,%0\n"
 		"5: \n"
 		EX_TABLE(0b,2b) EX_TABLE(3b,5b)
+		EX_TABLE(6b,5b) EX_TABLE(7b,5b)
 		: "+a" (size), "+a" (ptr), "+a" (x), "+a" (tmp1), "=a" (tmp2)
 		: "d" (reg0) : "cc", "memory");
 	return size;

Index: linux-2.6/arch/s390/lib/uaccess_std.c
===================================================================
@@ -36,12 +36,12 @@ size_t copy_from_user_std(size_t size, c
 	tmp1 = -256UL;
 	asm volatile(
 		"0: mvcp  0(%0,%2),0(%1),%3\n"
-		"   jz    8f\n"
+		"10:jz    8f\n"
 		"1:"ALR"  %0,%3\n"
 		"   la    %1,256(%1)\n"
 		"   la    %2,256(%2)\n"
 		"2: mvcp  0(%0,%2),0(%1),%3\n"
-		"   jnz   1b\n"
+		"11:jnz   1b\n"
 		"   j     8f\n"
 		"3: la    %4,255(%1)\n"	/* %4 = ptr + 255 */
 		"  "LHI"  %3,-4096\n"

@@ -50,7 +50,7 @@ size_t copy_from_user_std(size_t size, c
 		"  "CLR"  %0,%4\n"	/* copy crosses next page boundary? */
 		"   jnh   5f\n"
 		"4: mvcp  0(%4,%2),0(%1),%3\n"
-		"  "SLR"  %0,%4\n"
+		"12:"SLR"  %0,%4\n"
 		"  "ALR"  %2,%4\n"
 		"5:"LHI"  %4,-1\n"
 		"  "ALR"  %4,%0\n"	/* copy remaining size, subtract 1 */

@@ -65,6 +65,7 @@ size_t copy_from_user_std(size_t size, c
 		"8:"SLR"  %0,%0\n"
 		"9: \n"
 		EX_TABLE(0b,3b) EX_TABLE(2b,3b) EX_TABLE(4b,5b)
+		EX_TABLE(10b,9b) EX_TABLE(11b,9b) EX_TABLE(12b,9b)
 		: "+a" (size), "+a" (ptr), "+a" (x), "+a" (tmp1), "=a" (tmp2)
 		: : "cc", "memory");
 	return size;

Index: linux-2.6/mm/rmap.c
===================================================================
@@ -686,6 +686,15 @@ void page_remove_rmap(struct page *page,
 		 * faster for those pages still in swapcache.
 		 */
 		if (page_test_dirty(page)) {
+			int stable = page_make_stable(page);
+			VM_BUG_ON(!stable);
+			/*
+			 * We decremented the mapcount so we now have an
+			 * extra reference for the page. That prevents
+			 * page_make_volatile from making the page
+			 * volatile again while the dirty bit is in
+			 * transit.
+			 */
 			page_clear_dirty(page);
 			set_page_dirty(page);
 		}

blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.

I created the attached .dot graph based purely on this description. It looks reasonable, but I didn't see how a page enters a Pr state.

J

digraph gph {
	Ur -> Sr [ label="page_set_stable" ];
	Ur -> Vr [ label="page_set_volatile" ];
	Ur -> Ur [ label="page_set_unused" ];

	Sr -> Sr [ label="page_set_stable" ];
	Sr -> Vr [ label="page_set_volatile" ];
	Sr -> Ur [ label="page_set_unused" ];

	Vr -> Sr [ label="page_set_stable" ];
	Vr -> Vr [ label="page_set_volatile" ];
	Vr -> Ur [ label="page_set_unused" ];

	Uz -> Sz [ label="page_set_stable" ];
	Uz -> Vz [ label="page_set_volatile" ];
	Uz -> Uz [ label="page_set_unused" ];

	Sz -> Sz [ label="page_set_stable" ];
	Sz -> Vz [ label="page_set_volatile" ];
	Sz -> Uz [ label="page_set_unused" ];

	Vz -> Sz [ label="page_set_stable" ];
	Vz -> Vz [ label="page_set_volatile" ];
	Vz -> Uz [ label="page_set_unused" ];

	Ur -> Uz [ label="host evict" ];

	Sz -> Sr [ label="guest write" ];
	Sr -> Sp [ label="host swap" ];
	Sp -> Sr [ label="guest access" ];

	Sp -> Uz [ label="guest discard" ];
	Sp -> Vz [ label="page_set_volatile" ];

	Pr -> Sp [ label="host discard dirty" ];
	Pr -> Vz [ label="host discard clean" ];
}

That is the first block of state transitions: {Ur,Sr,Vr,Pr} You can go from any of the four states to any of the remaining three.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

You only mention page_set_{unused,stable,volatile}. Is page_set_stable_if_present() the fourth. And shouldn't that be "stable_if_clean":

- potential volatile (P): the page has useful content. The host system is allowed to discard the content after it has checked the dirty bit of the page. It has to deliver a discard fault with the absolute address of the page if the guest tries to access it.

The use of "stable" in the function call and "volatile" in this description is a bit confusing. My understanding is that a page in this state is either stable or volatile depending on whether its dirty, which makes sense, but it would be good to consistently refer to it in the same way.

Updated .dot attached.

J

digraph gph {
	Ur -> Sr [ label="set stable" ];
	Ur -> Vr [ label="set volatile" ];
	Ur -> Ur [ label="set unused" ];
	Ur -> Pr [ label="set stable_if_present" ];

	Sr -> Sr [ label="set stable" ];
	Sr -> Vr [ label="set volatile" ];
	Sr -> Ur [ label="set unused" ];
	Sr -> Pr [ label="set stable_if_present" ];

	Vr -> Sr [ label="set stable" ];
	Vr -> Vr [ label="set volatile" ];
	Vr -> Ur [ label="set unused" ];
	Vr -> Pr [ label="set stable_if_present" ];

	Pr -> Sr [ label="set stable" ];
	Pr -> Vr [ label="set volatile" ];
	Pr -> Ur [ label="set unused" ];
	Pr -> Pr [ label="set stable_if_present" ];

	Uz -> Sz [ label="set stable" ];
	Uz -> Vz [ label="set volatile" ];
	Uz -> Uz [ label="set unused" ];

	Sz -> Sz [ label="set stable" ];
	Sz -> Vz [ label="set volatile" ];
	Sz -> Uz [ label="set unused" ];

	Vz -> Sz [ label="set stable" ];
	Vz -> Vz [ label="set volatile" ];
	Vz -> Uz [ label="set unused" ];

	Ur -> Uz [ label="host evict" ];

	Sz -> Sr [ label="guest write" ];
	Sr -> Sp [ label="host swap" ];
	Sp -> Sr [ label="guest access" ];

	Sp -> Uz [ label="guest discard" ];
	Sp -> Vz [ label="set volatile" ];

	Pr -> Sp [ label="host discard dirty" ];
	Pr -> Vz [ label="host discard clean" ];
}

page_set_volatile has a "writable" argument. For writable==0 you get a Vx page, for writable==1 you get a Px page.

With stable_if_clean you are refering to stable_if_present? If yes the answer is that this operation is used to get a page from Vx/Px back to Sx but only if the page has not been discarded. The operation will fail if the page state is Vz/Pz. The dirty bit only matters for the hosts decision to discard the page, these are the state transitions from Vr/Pr to Vz.

Your understanding is good, but how can I make this less confusing? A Px page that is dirty may not be discarded which makes it basically stable. The guest state still is potential volatile though as it does not have a state of Sx.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

Hm. But a Vx page is writable isn't it? It's just that its contents can go away at any time. Or does the kernel treat Vx pages as strictly RO cached copies of other things?

It also seems to me that given you talking about "potentially volatile" as a distinct state, it would would be best to have a distinct state-setting function associated with it, so there's a 1:1 correspondence between the code and the description.

No. I misunderstood and thought that stable_if_present sets the Px state. I'd overlooked the writable flag on page_set_volatile().

So you mean it will change Vr/Pr to Sr but everything else will fail? Are there there any other non-discarded states for Vx/Px?

Do mean just Vz here? You say that Pz is never used.

Mainly, use identical terminology in code and description so they can be easily compared. I found the diagram was quite helpful in understanding what's going on; feel free to include it in your documentation.

Updated .dot attached; I've updated it to include the page_set_volatile writable argument and the stable_if_present transitions; commented it, removed the self-edges which were cluttering things up.

Also, does a page go from Vz->Vr on guest memory write? If so, does a clean page which goes from Pr->Vz->Vr lose its Px state in the process?

J

digraph gph {
	/* Guest state changes on resident pages */
	Ur -> Sr [ label="set stable" ];
	Ur -> Vr [ label="set volatile\n(w=0)" ];
	Ur -> Pr [ label="set volatile\n(w=1)" ];

	Sr -> Ur [ label="set unused" ];
	Sr -> Vr [ label="set volatile\n(w=0)" ];
	Sr -> Pr [ label="set volatile\n(w=1)" ];

	Vr -> Sr [ label="set stable(_if_present)" ];
	Vr -> Ur [ label="set unused" ];
	Vr -> Pr [ label="set volatile\n(w=1)" ];

	Pr -> Sr [ label="set stable(_if_present)" ];
	Pr -> Vr [ label="set volatile\n(w=0)" ];
	Pr -> Ur [ label="set unused" ];

	/* Guest state changes on zero pages */
	Uz -> Sz [ label="set stable" ];
	Uz -> Vz [ label="set volatile" ];

	Sz -> Vz [ label="set volatile" ];
	Sz -> Uz [ label="set unused" ];

	Vz -> Sz [ label="set stable" ];
	Vz -> Uz [ label="set unused" ];

	/* Guest state changes on host-swapped pages */
	Sp -> Uz [ label="set unused" ];
	Sp -> Vz [ label="set volatile" ];

	/* Guest touches pages */
	Sz -> Sr [ label="guest write" ];
	Sp -> Sr [ label="guest access" ];
	Vz -> Vr [ label="guest write" ];

	/* Host actions */
	Sr -> Sp [ label="host swap" ];
	Ur -> Uz [ label="host discard" ];
	Vr -> Vz [ label="host discard" ];
	Pr -> Sp [ label="host discard\n(dirty)" ];
	Pr -> Vz [ label="host discard\n(clean)" ];
}

Well presumably Vp/Pr => Sp? Is is true that from the guest's perspective, all of the 'p' states are identical to the 'r' states?

Do the host states even really need visibility to the guest at all? It may be useful for the guest to be able to distinguish between Ur and Uz but it doesn't seem necessary.

BTW Jeremy, the .dot was very useful!

Regards,

Anthony Liguori

Vp should never happen, since you'd never preserve a V page. And surely it would be Pr -> Sr, since the hypervisor wouldn't push the page to backing store when you change the client state.

Well, you implicitly see the hypervisor state. If you touch a [UV]z page then you get a fault telling you that the page has been taken away from you (I think). And it would definitely help with debugging (seems likely there's lots of scope for race conditions if you prematurely tell the hypervisor you don't need the page any more...).

Yes, there's no way I'd be able to get my head around this otherwise. BTW, here's an updated one with the host-driven events as dashed lines, and a couple of extra transitions I think should be in there (but waiting for Martin's confirmation).

J

digraph gph {
	/* Guest state changes on resident pages */
	Ur -> Sr [ label="set stable" ];
	Ur -> Vr [ label="set volatile\n(w=0)" ];
	Ur -> Pr [ label="set volatile\n(w=1)" ];

	Sr -> Ur [ label="set unused" ];
	Sr -> Vr [ label="set volatile\n(w=0)" ];
	Sr -> Pr [ label="set volatile\n(w=1)" ];

	Vr -> Sr [ label="set stable(_if_present)" ];
	Vr -> Ur [ label="set unused" ];
	Vr -> Pr [ label="set volatile\n(w=1)" ];

	Pr -> Sr [ label="set stable(_if_present)" ];
	Pr -> Vr [ label="set volatile\n(w=0)" ];
	Pr -> Ur [ label="set unused" ];

	/* Guest state changes on zero pages */
	Uz -> Sz [ label="set stable" ];
	Uz -> Vz [ label="set volatile" ];

	Sz -> Vz [ label="set volatile" ];
	Sz -> Uz [ label="set unused" ];

	Vz -> Sz [ label="set stable" ];
	Vz -> Uz [ label="set unused" ];

	/* Guest state changes on host-swapped pages */
	Sp -> Uz [ label="set unused" ];
	Sp -> Vz [ label="set volatile" ];

	/* Guest touches pages */
	Sz -> Sr [ label="guest write" ];
	Sp -> Sr [ label="guest access" ];
	Vz -> Vr [ label="guest write" ];

	/* Host actions */
	Sr -> Sp [ label="host swap", style=dashed ];
	Ur -> Uz [ label="host discard", style=dashed ];
	Vr -> Vz [ label="host discard", style=dashed ];
	Pr -> Sp [ label="host discard\n(dirty)", style=dashed ];
	Pr -> Vz [ label="host discard\n(clean)", style=dashed ];
}

You're right, I meant Vp/Pp but they are invalid states. I think one of the things that keeps tripping me up is that the host can change both the host and guest page states. My initial impression was that the host handled the host state and the guest handled the guest state.

I was thinking that it may be useful to know a Ur verses a Uz when allocating memory. In this case, you'd rather allocate Ur pages verses Uz to avoid the fault. I don't read s390 arch code well, is the host state explicit to the guest?

Excellent!

Regards,

Anthony LIguori

Yes. And it seems to me that you get unfortunate outcomes if you have a Pr->Vz->Vr transition.

Yes, reusing Ur pages might well be better, but who knows - they've probably got an instruction which makes Uz cheap...

Stuff like this suggets that both parts of the state are packed together, and are guest-visible:

+ return (state & ESSA_USTATE_MASK) == ESSA_USTATE_VOLATILE &&
+ (state & ESSA_CSTATE_MASK) == ESSA_CSTATE_ZERO;

J

Vz->Vr cannot happen. This would be a bug in the host.

Yes, faulting in a Uz page is cheap on s390. Isn't it a lovely architecture :-)

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

Does that mean that Vz is effectively identical to Uz?

J

Hm, on further thought:

If guests writes to Vz pages are disallowed, then the only way out of Vz is if the guest sets it to something else (Uz,Sz). If so, what's the point of using that state? Why not make:

   Vr -> Uz      host discard
   Pr -> Uz      host discard clean
   Sp -> Uz      set volatile
   Uz -> Uz      set volatile

But given how you've described V-state pages, I really would expect writes to a Vz to work, or alternatively, all writes to V-state pages to be disallowed. Are there any real uses for a writable Vr page?

On the other hand, removing Vz->Vr does clean up the dot graph a lot...

J

digraph gph {
	/* Guest state changes on resident pages */
	Ur -> Sr [ label="set stable" ];
	Ur -> Vr [ label="set volatile\n(w=0)" ];
	Ur -> Pr [ label="set volatile\n(w=1)" ];

	Sr -> Ur [ label="set unused" ];
	Sr -> Vr [ label="set volatile\n(w=0)" ];
	Sr -> Pr [ label="set volatile\n(w=1)" ];

	Vr -> Sr [ label="set stable(_if_present)" ];
	Vr -> Ur [ label="set unused" ];
	Vr -> Pr [ label="set volatile\n(w=1)" ];

	Pr -> Sr [ label="set stable(_if_present)" ];
	Pr -> Vr [ label="set volatile\n(w=0)" ];
	Pr -> Ur [ label="set unused" ];

	/* Guest state changes on zero pages */
	Uz -> Sz [ label="set stable" ];
	Uz -> Vz [ label="set volatile" ];

	Sz -> Vz [ label="set volatile" ];
	Sz -> Uz [ label="set unused" ];

	Vz -> Sz [ label="set stable" ];
	Vz -> Uz [ label="set unused" ];

	/* Guest state changes on host-swapped pages */
	Sp -> Uz [ label="set unused" ];
	Sp -> Vz [ label="set volatile" ];

	/* Guest touches pages */
	Sz -> Sr [ label="guest write" ];
	Sp -> Sr [ label="guest access" ];

	/* Host actions */
	Sr -> Sp [ label="host swap", style=dashed ];
	Ur -> Uz [ label="host discard", style=dashed ];
	Vr -> Vz [ label="host discard", style=dashed ];
	Pr -> Sp [ label="host discard\n(dirty)", style=dashed ];
	Pr -> Vz [ label="host discard\n(clean)", style=dashed ];
}

Vz is the page discarded state. The difference to Uz is slim, both states will cause a program check on access. Vz generates a discard fault, Uz generates an addressing exception which is nice for debugging. But I don't see a reason why an implementation that uses Uz instead of Vz shouldn't work.

You mean in the section that speaks about the guests states S/U/V/P ? Always keep in mind that you can access a V/P page only until it gets discarded. Then the useful content of the page frame is lost and any read of write to the not Vz page will be answered with a discard fault.

A Vr page is read-only. If a page gets mapped for writing it needs to get into the Pr state. This is the hint for the host to look at the dirty bit before it discards a page.
So yes, there is no use for a writable Vr page.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

How do you handle these different cases in Linux? Do you use Vr pages in the pagecache, and then shoot down the pagecache entry if the host steals the page?

The Uz access exception presumably just generates a normal oops.

(I should probably make time to read the rest of the series.)

Presumably reads from a Vz page also generate a discard fault?

OK, thanks, that clears things up. I was assuming that Vr was technically writable but that writes could be discarded at any time (ie, allowing guests to merrily shoot themselves in the foot ;). Making it forced RO is much more sensible.

J

The environment where we currently run all this is z/VM as the host and Linux as the guest. We have two page tables on s390, a host page table and a guest page table. If the host discards a page it simple removes the entry for the page in the host page table. If the guest comes along and accesses the page the host gets the fault and generates the appropriate fault.

Yes, the handler for an addressing exception will call die() for a kernel check without a fixup.

Yes.

Well, technically you could write to a Vr page via the kernel address space. The thing is that the host can just discard the page although it is dirty. The Vr state is used for page cache pages which do not have any writable mapping.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

In principle only the guest changes the guest state and only the host changes the host state. The simplified state diagram shows exceptions for Pr->Sp and Pr->Vz.

This is the second optimization you might want to think about. The other is to avoid the page clearing for Uz.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

Vp does not happen in the current implementation. But it actually may be useful. z/VM has multiple layers of paging, the first goes to expanded storage which is very fast. If you make the page Vz and the guests needs it you have to do a standard Linux I/O to get retrieve the page. This can be slower than a read and a write to expanded storage.

You get an addressing exception if you touch a Uz page. This indicates a BUG in the Linux code because this is a use after free. If the guests touches a Vz page you get a discard fault.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

In the extended version Vp/Pp to Sr as well but the current z/VM code will discard a page if the host picks a Vr/Pr page to swap it.

Basically yes. The guest doesn't care about the host state.

It is very useful for debugging to have the host state in the guest as well. There is one possible optimization: if the guests finds a Uz page in the free list, it can make it Sz and doesn't have to clear it because the host will provide an already empty page (not yet implemented though).

I've search on my disk and found the state diagrams we've used for the OLS paper. You may find these useful as well.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

Well, I want this feature, but I agree about complexity.

AFAICT, the trivial subset of this is the hinting of Unused pages. It seems that would buy us something, and perhaps be a stepping stone to full page hinting?

Cheers,
Rusty.

I've been there but the unused page thing is so small that it doesn't make sense to separate it from the patches. If I don't see any progress then I will come up with a patch that adds the Unused state transitions to the arch files of s390.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

Oh, that would be such a shame. Your guest page hinting patches remind me of that childhood thrill, when once a year the circus comes to town ;)

But seriously, I'm ashamed to see my name in the Cc list: it would be very unfair if your patches never made it in, just because I've failed to find the time to wrap my own puny brain around them.

It's very encouraging to see Jeremy and Rusty weighing in. I hope Zach will too, and I've added Andrea: their support would count a lot. You have Nick on the list, good, I've added Christoph and Peter (if you do resend, linux-mm might prove more useful than linux-kernel).

With support from rival virtualizers,
I do think you've a good chance of getting in.

Hugh

It is an effort to get you head around it the first time. It gets easiert the more you talk about it :-)

Grr, did I really forgot to copy linux-mm?!? (..insert you favourite four letter word here..). I absolutely intended to copy linux-mm but somehow replaced it with linux-s390.

Yes, it would be great if we can find another user for it.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

I like the circus too.

Bah! So modest.

I agree the page hinting technique is generally useful, even cross-architecture.

What doesn't appear to be useful however, is support for this under VMware. It can be done, even without the writable pte support (yes, really). But due to us exploiting optimizations at lower layers, it doesn't appear that it will gain us any performance - and we must already have the complex working set algorithms to support non-paravirtualized guests.

I would say we support it, but I don't expect us to make use of the infrastructure anytime soon. For us it would make more sense to use the swap-fault optimization, but this requires some significant design changes in our monitor.

Either way, these are both great ideas and I would not want to be held responsible for blocking their upstream progress. Someday, with the evolving x86 architecture (if we ever get per-page dirty bits), they might make sense for us to do as well.

Zach

With non-paravirt all you can do is to swap the guest physical memory (mmu notifiers allows linux to do that) or share memory (mmu notifiers
+ ksm allows linux to do that too). We also have complex working set algorithms that we use to finds which parts of the guest physical address space are best to swap first: the core linux VM.

What paravirt allows us to do (and that's the whole point of the paper I guess), is to go one step further than just guest swapping and to ask the guest if the page really need to be swapped or if it can be freed right away. So this would be an extension of the mmu notifiers (this also shows how EMM API is too restrictive, while MMU notifiers will allow that extension in the future) to avoid I/O sometime if guest tells us it's not necessary to swap through paravirt ops.

When talking with friends about ballooning I already once suggested to auto inflate the balloon with pages in the freelist.

Now this paper goes well beyond the pages in the freelist (called U/unused in the paper), this also covers cache and mapped-clean cache in the guest. That would have been the next step.

Anyway plain ballooning remains useful as rss limiting or numa compartments in the linux hypervisor, to provide unfariness to certain guests.

I didn't read the patch yet, but I think paravirt knowledge about U/unused pages is needed to avoid guest swapping. The cache and mapped cache in the guest is a gray area, because linux as hypervisor will be extremely efficient at swapping out and swapping in the guest cache (host swapping guest cache, may be faster than re-issuing a read-I/O to refill the cache by itself, clearly with guest using paravirt). Let's say I'm mostly interested about page-hinting for the U pages initially.

I'm currently busy with other two features and trying to get mmu notifier #v9 into mainline which is orders of magnitude more important than avoiding a few swapouts sometime (without mmu notifiers everything else is irrelevant, including guest page hinting and including ballooning too cause madvise(don't need) won't clear sptes and invalidate guest tlbs).

We can tap into those algorithms just as effectively using ballooning, and we've optimized the sharing and working set models from outside of the guest. So while CMM gives slightly better information for a random forced page eviction, the complexity doesn't appear to justify the savings for a VMware implementation.

Things would change quite a bit if we had hardware supported per-page dirty bits.

Ballooning still works if you use a kernel based balloon driver. Using madvise wouldn't be a reliable way to balloon anyway. Are you talking about an API to manage working sets and such from userspace?

Cheers,

Zach

I like the idea and it seems basically sound, but unfortunately Xen won't be able to make use of it in the near term, because it doesn't support any kind of backing for guest domain memory. There has been some thought about adding this kind of functionality to Xen. Keir, Ian: do you think this kind of support in the kernel be useful to us?

One concern I have is that 4k is really a very fine grain. We're thinking about moving Xen's memory management to operate in 2M chunk units, which would allow guests to directly use large pages with the corresponding reduction in TLB pressure. One side-effect of this is that we'd need to change ballooning to be in 2M rather than 4k units in order to prevent physical memory fragmentation.

Page hinting at 4k resolution poses the same problem. Would this technique still be useful operating on 2M chunks? Certainly it seems less likely that you could easily get a whole 2M area with the same fine-grained properties that these patches track. Would some kind of page/sub-page tracking be useful?

My other concern is just correctness over time on the Linux side. We already have enough trouble keeping things like the pte and page structure state in sync, with resulting rare data-loss bugs. Adding another layer which only applies in specific environments raises the possibility for new bugs to be un-noticed for a long time. How can we structure the VM changes to make sure that its robust in the face of maintenance?

J

Yes, that's the main concern, as whenever lots of subtlety is added. I wonder if there's any chance of a CONFIG_DEBUG mode, which could be run on anybody's x86 machine, without involving any virtualization, but in which the PAGE_STATEs become essential to the correct working of the mm.

Hugh

How about a fake hypervisor, which is really just a random page evictor, following the rules of CMM?

Probably simpler to just have variants of the page_set_* functions which simulate the worst-possible host action immediately (ie, stealing pages, logically swapping them, etc). That wouldn't give you full coverage, but it would go some way. An async variant which schedules a change in a few milliseconds would help too.

I guess that's equivalent to having a special-purpose hypervisor built into the kernel (hm, sounds familiar...).

J

It needn't be that hard on s390, I believe you don't need to worry about PTEs becoming asynchronous when stealing a page, since if I understand the hypervisor architecture, there is a per-page mapping level available, allowing you to generate discard faults on access. It might be possible to use this mapping layer without implementing a full blown hypervisor. Martin?

For x86, at discard time, you would have to manually walk and invalidate any PTEs potentially mapping the discarded page, but there is already this great thing called Xen paravirt-ops which actually does that for completely different reasons (PT page protection).

I think a random exponential distribution for discard would be needed to catch all the racey failure modes.

Zach

Yes, I don't expect its a problem for s390, but the point is making something workalike enough to make sure there's an evenly distributed number of explosions-in-face when things go wrong.

Not sure I follow. Xen pvops pays attention to whether a particular page is being used as part of a pagetable, and changes its permissions accordingly. But because pagetable pages are strictly kernel-only, we can get away with updating a single kernel-mapping pte which is shared across all processes. In the guest page hinting case, we need to deal with general pages which can be mapped anywhere, so that really does require a full traversal of the pagetables. Presumably rmap would be helpful here.

J

Yes, on s390 the PTEs cannot be asynchronous because there is no need to synchronize them in the first place. A mapping layer with all primitives without using the SIE instruction will be difficult. For one we cannot use the ESSA instruction which isolates the state changes and host page table is tied to the SIE. The page state is stored in the page table extension and the discard state is basically a specially marked invalid pte in the host table. A mapping layer with some restrictions is certainly possible.

If you have to walk the guest page tables you call into the guest, no ? I would characterize this more as a ballooner since you need guest activity to do the page stealing. The trick with guest page hinting is that you do NOT call into the guest to do the discard. You'll a nested page table for that I'm afraid.

We had quite a few of these racy failures. Nasty. Hard to find.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

Traffic on the guest page hinting patches died down again. Until another user shows up I guess that's it for the full version. In the meantime I push the patch below which is the poor mans version that can be done without common code change. It uses the arch_alloc_page/arch_free_page hooks to do the stable/unused state transitions. Better than nothing ..

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.

---
Subject: [PATCH] guest page hinting light

From: Martin Schwidefsky <[email protected]>

Use the existing arch_alloc_page/arch_free_page callbacks to do
the guest page state transitions between stable and unused.

Signed-off-by: Martin Schwidefsky <[email protected]>
---

 arch/s390/Kconfig          |    7 +++
 arch/s390/mm/Makefile      |    1 
 arch/s390/mm/init.c        |    3 +
 arch/s390/mm/page-states.c |   79 +++++++++++++++++++++++++++++++++++++++++++++
 include/asm-s390/page.h    |   11 ++++++
 include/asm-s390/system.h  |    6 +++
 6 files changed, 107 insertions(+)

diff -urpN linux-2.6/arch/s390/Kconfig linux-2.6-patched/arch/s390/Kconfig
--- linux-2.6/arch/s390/Kconfig	2008-05-06 17:38:14.000000000 +0200
+++ linux-2.6-patched/arch/s390/Kconfig	2008-05-06 17:38:28.000000000 +0200
@@ -430,6 +430,13 @@ config CMM_IUCV
 	  Select this option to enable the special message interface to
 	  the cooperative memory management.
 
+config PAGE_STATES
+	bool "Unused page notification"
+	help
+	  This enables the notification of unused pages to the
+	  hypervisor. The ESSA instruction is used to do the states
+	  changes between a page that has content and the unused state.
+
 config VIRT_TIMER
 	bool "Virtual CPU timer support"
 	help
diff -urpN linux-2.6/arch/s390/mm/init.c linux-2.6-patched/arch/s390/mm/init.c
--- linux-2.6/arch/s390/mm/init.c	2008-05-06 17:38:14.000000000 +0200
+++ linux-2.6-patched/arch/s390/mm/init.c	2008-05-06 17:38:28.000000000 +0200
@@ -126,6 +126,9 @@ void __init mem_init(void)
         /* clear the zero-page */
         memset(empty_zero_page, 0, PAGE_SIZE);
 
+	/* Setup guest page hinting */
+	cmma_init();
+
 	/* this will put all low memory onto the freelists */
 	totalram_pages += free_all_bootmem();
 
diff -urpN linux-2.6/arch/s390/mm/Makefile linux-2.6-patched/arch/s390/mm/Makefile
--- linux-2.6/arch/s390/mm/Makefile	2008-05-06 17:38:14.000000000 +0200
+++ linux-2.6-patched/arch/s390/mm/Makefile	2008-05-06 17:38:28.000000000 +0200
@@ -5,3 +5,4 @@
 obj-y	 := init.o fault.o extmem.o mmap.o vmem.o pgtable.o
 obj-$(CONFIG_CMM) += cmm.o
 obj-$(CONFIG_HUGETLB_PAGE) += hugetlbpage.o
+obj-$(CONFIG_PAGE_STATES) += page-states.o
diff -urpN linux-2.6/arch/s390/mm/page-states.c linux-2.6-patched/arch/s390/mm/page-states.c
--- linux-2.6/arch/s390/mm/page-states.c	1970-01-01 01:00:00.000000000 +0100
+++ linux-2.6-patched/arch/s390/mm/page-states.c	2008-05-06 17:38:28.000000000 +0200
@@ -0,0 +1,79 @@
+/*
+ * arch/s390/mm/page-states.c
+ *
+ * (C) Copyright IBM Corp. 2008
+ *
+ * Guest page hinting for unused pages.
+ *
+ * Author(s): Martin Schwidefsky <[email protected]>
+ */
+
+#include <linux/kernel.h>
+#include <linux/errno.h>
+#include <linux/types.h>
+#include <linux/mm.h>
+#include <linux/init.h>
+
+#define ESSA_SET_STABLE		1
+#define ESSA_SET_UNUSED		2
+
+int cmma_flag;
+
+static int __init cmma(char *str)
+{
+	char *parm;
+	parm = strstrip(str);
+	if (strcmp(parm, "yes") == 0 || strcmp(parm, "on") == 0) {
+		cmma_flag = 1;
+		return 1;
+	}
+	cmma_flag = 0;
+	if (strcmp(parm, "no") == 0 || strcmp(parm, "off") == 0)
+		return 1;
+	return 0;
+}
+
+__setup("cmma=", cmma);
+
+void __init cmma_init(void)
+{
+	register unsigned long tmp asm("0") = 0;
+	register int rc asm("1") = -ENOSYS;
+
+	if (!cmma_flag)
+		return;
+	asm volatile(
+		"       .insn rrf,0xb9ab0000,%1,%1,0,0\n"
+		"0:     la      %0,0\n"
+		"1:\n"
+		EX_TABLE(0b,1b)
+		: "+&d" (rc), "+&d" (tmp));
+	if (rc)
+		cmma_flag = 0;
+}
+
+void arch_free_page(struct page *page, int order)
+{
+	int i, rc;
+
+	if (!cmma_flag)
+		return;
+	for (i = 0; i < (1 << order); i++)
+		asm volatile(".insn rrf,0xb9ab0000,%0,%1,%2,0"
+			     : "=&d" (rc)
+			     : "a" ((page_to_pfn(page) + i) << PAGE_SHIFT),
+			       "i" (ESSA_SET_UNUSED));
+}
+
+void arch_alloc_page(struct page *page, int order)
+{
+	int i, rc;
+
+	if (!cmma_flag)
+		return;
+	for (i = 0; i < (1 << order); i++)
+		asm volatile(".insn rrf,0xb9ab0000,%0,%1,%2,0"
+			     : "=&d" (rc)
+			     : "a" ((page_to_pfn(page) + i) << PAGE_SHIFT),
+			       "i" (ESSA_SET_STABLE));
+}
diff -urpN linux-2.6/include/asm-s390/page.h linux-2.6-patched/include/asm-s390/page.h
--- linux-2.6/include/asm-s390/page.h	2008-05-06 17:38:17.000000000 +0200
+++ linux-2.6-patched/include/asm-s390/page.h	2008-05-06 17:38:28.000000000 +0200
@@ -125,6 +125,17 @@ page_get_storage_key(unsigned long addr)
 	return skey;
 }
 
+#ifdef CONFIG_PAGE_STATES
+
+struct page;
+void arch_free_page(struct page *page, int order);
+void arch_alloc_page(struct page *page, int order);
+
+#define HAVE_ARCH_FREE_PAGE
+#define HAVE_ARCH_ALLOC_PAGE
+
+#endif
+
 #endif /* !__ASSEMBLY__ */
 
 /* to align the pointer to the (next) page boundary */
diff -urpN linux-2.6/include/asm-s390/system.h linux-2.6-patched/include/asm-s390/system.h
--- linux-2.6/include/asm-s390/system.h	2008-05-06 17:38:17.000000000 +0200
+++ linux-2.6-patched/include/asm-s390/system.h	2008-05-06 17:38:28.000000000 +0200
@@ -116,6 +116,12 @@ extern void pfault_fini(void);
 #define pfault_fini()		do { } while (0)
 #endif /* CONFIG_PFAULT */
 
+#ifdef CONFIG_PAGE_STATES
+extern void cmma_init(void);
+#else
+#define cmma_init()		do { } while (0)
+#endif
+
 #define finish_arch_switch(prev) do {					     \
 	set_fs(current->thread.mm_segment);				     \
 	account_vtime(prev);						     \

I suspect one of the problems is that there are too many state transitions to have it implemented with a low overhead on anything but S390, and even there you need milicoded instructions to handle things.

If the number of transitions can be reduced, page hinting could be useful for KVM, too.

All Rights Reversed

Spot on Rik, if every transition becomes a hypercall (and a synchronous one at that), it isn't workable for us. If, on the other hand, you share the state bits between the guest and hypervisor, you need a giant (standalone) bit array for per-page state, which is neither convenient for Linux nor the hypervisor. I believe s390 has an 'instruction' to migrate the state bits into the hypervisor per-physical-page data without requiring a hypercall.

Zach

That is why we invented the millicoded ESSA instruction on s390. We had an emulation of the instruction to test things. It worked but was awfully slow.

blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.