Skip to contentSkip to conversation

parisc: Optimize per-pagetable spinlocks

PATCH New Helge Deller 1 participant 1 message Feb 12, 2021 PA-RISC architecture

lore
Share & tools
lore

Helge Deller

author maintainer
#0 lore raw

PATCH 1/1 parisc: Optimize per-pagetable spinlocks

On parisc a spinlock is stored in the next page behind the pgd which protects against parallel accesses to the pgd. That's why one additional page (PGD_ALLOC_ORDER) is allocated for the pgd.

Matthew Wilcox suggested that we instead should use a pointer in the struct page table for this spinlock and noted, that the comments for the PGD_ORDER and PMD_ORDER defines were wrong.

Both suggestions are addressed with this patch. Instead of having an own spinlock to protect the pgd, we now switch to use the existing page_table_lock. Additionally, beside loading the pgd into cr25 in switch_mm_irqs_off(), the physical address of this lock is loaded into cr28 (tr4), so that we can avoid implementing a complicated lookup in assembly for this lock in the TLB fault handlers.

The existing Hybrid L2/L3 page table scheme (where the pmd is adjacent to the pgd) has been dropped with this patch.

Remove the locking in set_pte() and the huge-page pte functions too. They trigger a spinlock recursion on 32bit machines and seem unnecessary.

Fixes b37d1c1898b2 ("parisc: Use per-pagetable spinlock")
---
 arch/parisc/Kconfig                   |  10 +++
 arch/parisc/include/asm/mmu_context.h |   7 ++
 arch/parisc/include/asm/page.h        |   2 +-
 arch/parisc/include/asm/pgalloc.h     |  76 ++++-------------
 arch/parisc/include/asm/pgtable.h     |  89 ++++----------------
 arch/parisc/kernel/asm-offsets.c      |   1 -
 arch/parisc/kernel/entry.S            | 116 +++++++++++---------------
 arch/parisc/mm/hugetlbpage.c          |  13 ---
 arch/parisc/mm/init.c                 |  10 +--
 9 files changed, 110 insertions(+), 214 deletions(-)
arch/parisc/Kconfig +10 -0
index 278462186ac4..8368a806a18f 100644
@@ -311,6 +311,16 @@ config IRQSTACKS
 	  for handling hard and soft interrupts.  This can help avoid
 	  overflowing the process kernel stacks.
 
+config TLB_PTLOCK
+	bool "Use page table locks in TLB fault handler"
+	depends on SMP
+	default n
+	help
+	  Select this option to enable page table locking in the TLB
+	  fault handler. This ensures that page table entries are
+	  updated consistently on SMP machines at the expense of some
+	  loss in performance.
+
 config HOTPLUG_CPU
 	bool
 	default y if SMP
arch/parisc/include/asm/mmu_context.h +7 -0
index 46f8c22c5977..726257648d9f 100644
@@ -5,6 +5,7 @@
 #include <linux/mm.h>
 #include <linux/sched.h>
 #include <linux/atomic.h>
+#include <linux/spinlock.h>
 #include <asm-generic/mm_hooks.h>
 
 /* on PA-RISC, we actually have enough contexts to justify an allocator
@@ -50,6 +51,12 @@ static inline void switch_mm_irqs_off(struct mm_struct *prev,
 		struct mm_struct *next, struct task_struct *tsk)
 {
 	if (prev != next) {
+#ifdef CONFIG_TLB_PTLOCK
+		/* put physical address of page_table_lock in cr28 (tr4)
+		   for TLB faults */
+		spinlock_t *pgd_lock = &next->page_table_lock;
+		mtctl(__pa(__ldcw_align(&pgd_lock->rlock.raw_lock)), 28);
+#endif
 		mtctl(__pa(next->pgd), 25);
 		load_context(next->context);
 	}
arch/parisc/include/asm/page.h +1 -1
index 6b3f6740a6a6..d00313d1274e 100644
@@ -112,7 +112,7 @@ extern int npmem_ranges;
 #else
 #define BITS_PER_PTE_ENTRY	2
 #define BITS_PER_PMD_ENTRY	2
-#define BITS_PER_PGD_ENTRY	BITS_PER_PMD_ENTRY
+#define BITS_PER_PGD_ENTRY	2
 #endif
 #define PGD_ENTRY_SIZE	(1UL << BITS_PER_PGD_ENTRY)
 #define PMD_ENTRY_SIZE	(1UL << BITS_PER_PMD_ENTRY)
arch/parisc/include/asm/pgalloc.h +18 -58

106 lines not on this page · raw message

arch/parisc/include/asm/pgtable.h +18 -71

170 lines not on this page · raw message

arch/parisc/kernel/asm-offsets.c +0 -1
index 305768a40773..cd2cc1b1648c 100644
@@ -268,7 +268,6 @@ int main(void)
 	DEFINE(ASM_BITS_PER_PGD, BITS_PER_PGD);
 	DEFINE(ASM_BITS_PER_PMD, BITS_PER_PMD);
 	DEFINE(ASM_BITS_PER_PTE, BITS_PER_PTE);
-	DEFINE(ASM_PGD_PMD_OFFSET, -(PAGE_SIZE << PGD_ORDER));
 	DEFINE(ASM_PMD_ENTRY, ((PAGE_OFFSET & PMD_MASK) >> PMD_SHIFT));
 	DEFINE(ASM_PGD_ENTRY, PAGE_OFFSET >> PGDIR_SHIFT);
 	DEFINE(ASM_PGD_ENTRY_SIZE, PGD_ENTRY_SIZE);
arch/parisc/kernel/entry.S +51 -65

342 lines not on this page · raw message

arch/parisc/mm/hugetlbpage.c +0 -13
index d7ba014a7fbb..43652de5f139 100644
@@ -142,24 +142,17 @@ static void __set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
 		     pte_t *ptep, pte_t entry)
 {
-	unsigned long flags;
-
-	spin_lock_irqsave(pgd_spinlock((mm)->pgd), flags);
 	__set_huge_pte_at(mm, addr, ptep, entry);
-	spin_unlock_irqrestore(pgd_spinlock((mm)->pgd), flags);
 }
 
 
 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
 			      pte_t *ptep)
 {
-	unsigned long flags;
 	pte_t entry;
 
-	spin_lock_irqsave(pgd_spinlock((mm)->pgd), flags);
 	entry = *ptep;
 	__set_huge_pte_at(mm, addr, ptep, __pte(0));
-	spin_unlock_irqrestore(pgd_spinlock((mm)->pgd), flags);
 
 	return entry;
 }
@@ -168,29 +161,23 @@ pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
 void huge_ptep_set_wrprotect(struct mm_struct *mm,
 				unsigned long addr, pte_t *ptep)
 {
-	unsigned long flags;
 	pte_t old_pte;
 
-	spin_lock_irqsave(pgd_spinlock((mm)->pgd), flags);
 	old_pte = *ptep;
 	__set_huge_pte_at(mm, addr, ptep, pte_wrprotect(old_pte));
-	spin_unlock_irqrestore(pgd_spinlock((mm)->pgd), flags);
 }
 
 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
 				unsigned long addr, pte_t *ptep,
 				pte_t pte, int dirty)
 {
-	unsigned long flags;
 	int changed;
 	struct mm_struct *mm = vma->vm_mm;
 
-	spin_lock_irqsave(pgd_spinlock((mm)->pgd), flags);
 	changed = !pte_same(*ptep, pte);
 	if (changed) {
 		__set_huge_pte_at(mm, addr, ptep, pte);
 	}
-	spin_unlock_irqrestore(pgd_spinlock((mm)->pgd), flags);
 	return changed;
 }
 
arch/parisc/mm/init.c +5 -5
index 3ec633b11b54..9ca4e4ff6895 100644
@@ -37,11 +37,6 @@ extern int  data_start;
 extern void parisc_kernel_start(void);	/* Kernel entry point in head.S */
 
 #if CONFIG_PGTABLE_LEVELS == 3
-/* NOTE: This layout exactly conforms to the hybrid L2/L3 page table layout
- * with the first pmd adjacent to the pgd and below it. gcc doesn't actually
- * guarantee that global objects will be laid out in memory in the same order
- * as the order of declaration, so put these in different sections and use
- * the linker script to order them. */
 pmd_t pmd0[PTRS_PER_PMD] __section(".data..vm0.pmd") __attribute__ ((aligned(PAGE_SIZE)));
 #endif
 
@@ -559,6 +554,11 @@ void __init mem_init(void)
 	BUILD_BUG_ON(PGD_ENTRY_SIZE != sizeof(pgd_t));
 	BUILD_BUG_ON(PAGE_SHIFT + BITS_PER_PTE + BITS_PER_PMD + BITS_PER_PGD
 			> BITS_PER_LONG);
+#if CONFIG_PGTABLE_LEVELS == 3
+	BUILD_BUG_ON(PT_INITIAL > PTRS_PER_PMD);
+#else
+	BUILD_BUG_ON(PT_INITIAL > PTRS_PER_PGD);
+#endif
 
 	high_memory = __va((max_pfn << PAGE_SHIFT));
 	set_max_mapnr(max_low_pfn);

--
2.29.2