diff mbox series

[v3,03/34] s390: Use pt_frag_refcount for pagetables

Message ID 20230531213032.25338-4-vishal.moola@gmail.com (mailing list archive)
State Superseded
Headers show
Series Split ptdesc from struct page | expand

Commit Message

Vishal Moola May 31, 2023, 9:30 p.m. UTC
s390 currently uses _refcount to identify fragmented page tables.
The page table struct already has a member pt_frag_refcount used by
powerpc, so have s390 use that instead of the _refcount field as well.
This improves the safety for _refcount and the page table tracking.

This also allows us to simplify the tracking since we can once again use
the lower byte of pt_frag_refcount instead of the upper byte of _refcount.

Signed-off-by: Vishal Moola (Oracle) <vishal.moola@gmail.com>
---
 arch/s390/mm/pgalloc.c | 38 +++++++++++++++-----------------------
 1 file changed, 15 insertions(+), 23 deletions(-)

Comments

Gerald Schaefer June 1, 2023, 1:19 p.m. UTC | #1
On Wed, 31 May 2023 14:30:01 -0700
"Vishal Moola (Oracle)" <vishal.moola@gmail.com> wrote:

> s390 currently uses _refcount to identify fragmented page tables.
> The page table struct already has a member pt_frag_refcount used by
> powerpc, so have s390 use that instead of the _refcount field as well.
> This improves the safety for _refcount and the page table tracking.
> 
> This also allows us to simplify the tracking since we can once again use
> the lower byte of pt_frag_refcount instead of the upper byte of _refcount.

This would conflict with s390 impact of pte_free_defer() work from Hugh Dickins
https://lore.kernel.org/lkml/35e983f5-7ed3-b310-d949-9ae8b130cdab@google.com/
https://lore.kernel.org/lkml/6dd63b39-e71f-2e8b-7e0-83e02f3bcb39@google.com/

There he uses pt_frag_refcount, or rather pt_mm in the same union, to save
the mm_struct for deferred pte_free().

I still need to look closer into both of your patch series, but so far it
seems that you have no hard functional requirement to switch from _refcount
to pt_frag_refcount here, for s390.

If this is correct, and you do not e.g. need this to make some other use
of _refcount, I would suggest to drop this patch.
Vishal Moola June 1, 2023, 8:43 p.m. UTC | #2
On Thu, Jun 1, 2023 at 6:19 AM Gerald Schaefer
<gerald.schaefer@linux.ibm.com> wrote:
>
>  On Wed, 31 May 2023 14:30:01 -0700
> "Vishal Moola (Oracle)" <vishal.moola@gmail.com> wrote:
>
> > s390 currently uses _refcount to identify fragmented page tables.
> > The page table struct already has a member pt_frag_refcount used by
> > powerpc, so have s390 use that instead of the _refcount field as well.
> > This improves the safety for _refcount and the page table tracking.
> >
> > This also allows us to simplify the tracking since we can once again use
> > the lower byte of pt_frag_refcount instead of the upper byte of _refcount.
>
> This would conflict with s390 impact of pte_free_defer() work from Hugh Dickins
> https://lore.kernel.org/lkml/35e983f5-7ed3-b310-d949-9ae8b130cdab@google.com/
> https://lore.kernel.org/lkml/6dd63b39-e71f-2e8b-7e0-83e02f3bcb39@google.com/
>
> There he uses pt_frag_refcount, or rather pt_mm in the same union, to save
> the mm_struct for deferred pte_free().
>
> I still need to look closer into both of your patch series, but so far it
> seems that you have no hard functional requirement to switch from _refcount
> to pt_frag_refcount here, for s390.
>
> If this is correct, and you do not e.g. need this to make some other use
> of _refcount, I would suggest to drop this patch.

The goal of this preparation patch is to consolidate s390's usage of
struct page fields so that struct ptdesc can be smaller. Its not particularly
mandatory; leaving _refcount in ptdesc only increases the struct by
8 bytes and can always be changed later.

However it is a little annoying since s390 is the only architecture
that egregiously uses space throughout struct page for their page
tables, rather than just the page table struct. For example, s390
gmap uses page->index which also aliases with pt_mm and
pt_frag_refcount. I'm not sure if/how gmap page tables interact
with s390 process page tables at all, but if it does that could
potentially cause problems with Hugh's patch as well :(

I can add _refcount to ptdesc if we would like, but I still
prefer if s390 could be simplified instead.
diff mbox series

Patch

diff --git a/arch/s390/mm/pgalloc.c b/arch/s390/mm/pgalloc.c
index 66ab68db9842..6b99932abc66 100644
--- a/arch/s390/mm/pgalloc.c
+++ b/arch/s390/mm/pgalloc.c
@@ -182,20 +182,17 @@  void page_table_free_pgste(struct page *page)
  * As follows from the above, no unallocated or fully allocated parent
  * pages are contained in mm_context_t::pgtable_list.
  *
- * The upper byte (bits 24-31) of the parent page _refcount is used
+ * The lower byte (bits 0-7) of the parent page pt_frag_refcount is used
  * for tracking contained 2KB-pgtables and has the following format:
  *
  *   PP  AA
- * 01234567    upper byte (bits 24-31) of struct page::_refcount
+ * 01234567    upper byte (bits 0-7) of struct page::pt_frag_refcount
  *   ||  ||
  *   ||  |+--- upper 2KB-pgtable is allocated
  *   ||  +---- lower 2KB-pgtable is allocated
  *   |+------- upper 2KB-pgtable is pending for removal
  *   +-------- lower 2KB-pgtable is pending for removal
  *
- * (See commit 620b4e903179 ("s390: use _refcount for pgtables") on why
- * using _refcount is possible).
- *
  * When 2KB-pgtable is allocated the corresponding AA bit is set to 1.
  * The parent page is either:
  *   - added to mm_context_t::pgtable_list in case the second half of the
@@ -243,11 +240,12 @@  unsigned long *page_table_alloc(struct mm_struct *mm)
 		if (!list_empty(&mm->context.pgtable_list)) {
 			page = list_first_entry(&mm->context.pgtable_list,
 						struct page, lru);
-			mask = atomic_read(&page->_refcount) >> 24;
+			mask = atomic_read(&page->pt_frag_refcount);
 			/*
 			 * The pending removal bits must also be checked.
 			 * Failure to do so might lead to an impossible
-			 * value of (i.e 0x13 or 0x23) written to _refcount.
+			 * value of (i.e 0x13 or 0x23) written to
+			 * pt_frag_refcount.
 			 * Such values violate the assumption that pending and
 			 * allocation bits are mutually exclusive, and the rest
 			 * of the code unrails as result. That could lead to
@@ -259,8 +257,8 @@  unsigned long *page_table_alloc(struct mm_struct *mm)
 				bit = mask & 1;		/* =1 -> second 2K */
 				if (bit)
 					table += PTRS_PER_PTE;
-				atomic_xor_bits(&page->_refcount,
-							0x01U << (bit + 24));
+				atomic_xor_bits(&page->pt_frag_refcount,
+							0x01U << bit);
 				list_del(&page->lru);
 			}
 		}
@@ -281,12 +279,12 @@  unsigned long *page_table_alloc(struct mm_struct *mm)
 	table = (unsigned long *) page_to_virt(page);
 	if (mm_alloc_pgste(mm)) {
 		/* Return 4K page table with PGSTEs */
-		atomic_xor_bits(&page->_refcount, 0x03U << 24);
+		atomic_xor_bits(&page->pt_frag_refcount, 0x03U);
 		memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
 		memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
 	} else {
 		/* Return the first 2K fragment of the page */
-		atomic_xor_bits(&page->_refcount, 0x01U << 24);
+		atomic_xor_bits(&page->pt_frag_refcount, 0x01U);
 		memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
 		spin_lock_bh(&mm->context.lock);
 		list_add(&page->lru, &mm->context.pgtable_list);
@@ -323,22 +321,19 @@  void page_table_free(struct mm_struct *mm, unsigned long *table)
 		 * will happen outside of the critical section from this
 		 * function or from __tlb_remove_table()
 		 */
-		mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
-		mask >>= 24;
+		mask = atomic_xor_bits(&page->pt_frag_refcount, 0x11U << bit);
 		if (mask & 0x03U)
 			list_add(&page->lru, &mm->context.pgtable_list);
 		else
 			list_del(&page->lru);
 		spin_unlock_bh(&mm->context.lock);
-		mask = atomic_xor_bits(&page->_refcount, 0x10U << (bit + 24));
-		mask >>= 24;
+		mask = atomic_xor_bits(&page->pt_frag_refcount, 0x10U << bit);
 		if (mask != 0x00U)
 			return;
 		half = 0x01U << bit;
 	} else {
 		half = 0x03U;
-		mask = atomic_xor_bits(&page->_refcount, 0x03U << 24);
-		mask >>= 24;
+		mask = atomic_xor_bits(&page->pt_frag_refcount, 0x03U);
 	}
 
 	page_table_release_check(page, table, half, mask);
@@ -368,8 +363,7 @@  void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
 	 * outside of the critical section from __tlb_remove_table() or from
 	 * page_table_free()
 	 */
-	mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
-	mask >>= 24;
+	mask = atomic_xor_bits(&page->pt_frag_refcount, 0x11U << bit);
 	if (mask & 0x03U)
 		list_add_tail(&page->lru, &mm->context.pgtable_list);
 	else
@@ -391,14 +385,12 @@  void __tlb_remove_table(void *_table)
 		return;
 	case 0x01U:	/* lower 2K of a 4K page table */
 	case 0x02U:	/* higher 2K of a 4K page table */
-		mask = atomic_xor_bits(&page->_refcount, mask << (4 + 24));
-		mask >>= 24;
+		mask = atomic_xor_bits(&page->pt_frag_refcount, mask << 4);
 		if (mask != 0x00U)
 			return;
 		break;
 	case 0x03U:	/* 4K page table with pgstes */
-		mask = atomic_xor_bits(&page->_refcount, 0x03U << 24);
-		mask >>= 24;
+		mask = atomic_xor_bits(&page->pt_frag_refcount, 0x03U);
 		break;
 	}