tilegx pci: support I/O to arbitrarily-cached pages

The tilegx PCI root complex support (currently only in linux-next)
is limited to pages that are homed on cached in the default manner,
i.e. "hash-for-home".  This change supports delivery of I/O data to
pages that are cached in other ways (locally on a particular core,
uncached, user-managed incoherent, etc.).

A large part of the change is supporting flushing pages from cache
on particular homes so that we can transition the data that we are
delivering to or from the device appropriately.  The new homecache_finv*
routines handle this.

Some changes to page_table_range_init() were also required to make
the fixmap code work correctly on tilegx; it hadn't been used there
before.

We also remove some stub mark_caches_evicted_*() routines that
were just no-ops anyway.

Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>

2 files changed, 99 insertions, 116 deletions

diff --git a/arch/tile/mm/homecache.c b/arch/tile/mm/homecache.c
index dbcbdf7b8aa..5f7868dcd6d 100644
--- a/arch/tile/mm/homecache.c
+++ b/arch/tile/mm/homecache.c
@@ -64,10 +64,6 @@ early_param("noallocl2", set_noallocl2);
 
 #endif
 
-/* Provide no-op versions of these routines to keep flush_remote() cleaner. */
-#define mark_caches_evicted_start() 0
-#define mark_caches_evicted_finish(mask, timestamp) do {} while (0)
-
 
 /*
  * Update the irq_stat for cpus that we are going to interrupt
@@ -107,7 +103,6 @@ static void hv_flush_update(const struct cpumask *cache_cpumask,
  *    there's never any good reason for hv_flush_remote() to fail.
  *  - Accepts a 32-bit PFN rather than a 64-bit PA, which generally
  *    is the type that Linux wants to pass around anyway.
- *  - Centralizes the mark_caches_evicted() handling.
  *  - Canonicalizes that lengths of zero make cpumasks NULL.
  *  - Handles deferring TLB flushes for dataplane tiles.
  *  - Tracks remote interrupts in the per-cpu irq_cpustat_t.
@@ -126,7 +121,6 @@ void flush_remote(unsigned long cache_pfn, unsigned long cache_control,
 		  HV_Remote_ASID *asids, int asidcount)
 {
 	int rc;
-	int timestamp = 0;  /* happy compiler */
 	struct cpumask cache_cpumask_copy, tlb_cpumask_copy;
 	struct cpumask *cache_cpumask, *tlb_cpumask;
 	HV_PhysAddr cache_pa;
@@ -157,15 +151,11 @@ void flush_remote(unsigned long cache_pfn, unsigned long cache_control,
 	hv_flush_update(cache_cpumask, tlb_cpumask, tlb_va, tlb_length,
 			asids, asidcount);
 	cache_pa = (HV_PhysAddr)cache_pfn << PAGE_SHIFT;
-	if (cache_control & HV_FLUSH_EVICT_L2)
-		timestamp = mark_caches_evicted_start();
 	rc = hv_flush_remote(cache_pa, cache_control,
 			     cpumask_bits(cache_cpumask),
 			     tlb_va, tlb_length, tlb_pgsize,
 			     cpumask_bits(tlb_cpumask),
 			     asids, asidcount);
-	if (cache_control & HV_FLUSH_EVICT_L2)
-		mark_caches_evicted_finish(cache_cpumask, timestamp);
 	if (rc == 0)
 		return;
 	cpumask_scnprintf(cache_buf, sizeof(cache_buf), &cache_cpumask_copy);
@@ -180,85 +170,86 @@ void flush_remote(unsigned long cache_pfn, unsigned long cache_control,
 	panic("Unsafe to continue.");
 }
 
-void flush_remote_page(struct page *page, int order)
+static void homecache_finv_page_va(void* va, int home)
 {
-	int i, pages = (1 << order);
-	for (i = 0; i < pages; ++i, ++page) {
-		void *p = kmap_atomic(page);
-		int hfh = 0;
-		int home = page_home(page);
-#if CHIP_HAS_CBOX_HOME_MAP()
-		if (home == PAGE_HOME_HASH)
-			hfh = 1;
-		else
-#endif
-			BUG_ON(home < 0 || home >= NR_CPUS);
-		finv_buffer_remote(p, PAGE_SIZE, hfh);
-		kunmap_atomic(p);
+	if (home == smp_processor_id()) {
+		finv_buffer_local(va, PAGE_SIZE);
+	} else if (home == PAGE_HOME_HASH) {
+		finv_buffer_remote(va, PAGE_SIZE, 1);
+	} else {
+		BUG_ON(home < 0 || home >= NR_CPUS);
+		finv_buffer_remote(va, PAGE_SIZE, 0);
 	}
 }
 
-void homecache_evict(const struct cpumask *mask)
+void homecache_finv_map_page(struct page *page, int home)
 {
-	flush_remote(0, HV_FLUSH_EVICT_L2, mask, 0, 0, 0, NULL, NULL, 0);
+	unsigned long flags;
+	unsigned long va;
+	pte_t *ptep;
+	pte_t pte;
+
+	if (home == PAGE_HOME_UNCACHED)
+		return;
+	local_irq_save(flags);
+#ifdef CONFIG_HIGHMEM
+	va = __fix_to_virt(FIX_KMAP_BEGIN + kmap_atomic_idx_push() +
+			   (KM_TYPE_NR * smp_processor_id()));
+#else
+	va = __fix_to_virt(FIX_HOMECACHE_BEGIN + smp_processor_id());
+#endif
+	ptep = virt_to_pte(NULL, (unsigned long)va);
+	pte = pfn_pte(page_to_pfn(page), PAGE_KERNEL);
+	__set_pte(ptep, pte_set_home(pte, home));
+	homecache_finv_page_va((void *)va, home);
+	__pte_clear(ptep);
+	hv_flush_page(va, PAGE_SIZE);
+#ifdef CONFIG_HIGHMEM
+	kmap_atomic_idx_pop();
+#endif
+	local_irq_restore(flags);
 }
 
-/*
- * Return a mask of the cpus whose caches currently own these pages.
- * The return value is whether the pages are all coherently cached
- * (i.e. none are immutable, incoherent, or uncached).
- */
-static int homecache_mask(struct page *page, int pages,
-			  struct cpumask *home_mask)
+static void homecache_finv_page_home(struct page *page, int home)
 {
-	int i;
-	int cached_coherently = 1;
-	cpumask_clear(home_mask);
-	for (i = 0; i < pages; ++i) {
-		int home = page_home(&page[i]);
-		if (home == PAGE_HOME_IMMUTABLE ||
-		    home == PAGE_HOME_INCOHERENT) {
-			cpumask_copy(home_mask, cpu_possible_mask);
-			return 0;
-		}
-#if CHIP_HAS_CBOX_HOME_MAP()
-		if (home == PAGE_HOME_HASH) {
-			cpumask_or(home_mask, home_mask, &hash_for_home_map);
-			continue;
-		}
-#endif
-		if (home == PAGE_HOME_UNCACHED) {
-			cached_coherently = 0;
-			continue;
-		}
-		BUG_ON(home < 0 || home >= NR_CPUS);
-		cpumask_set_cpu(home, home_mask);
-	}
-	return cached_coherently;
+	if (!PageHighMem(page) && home == page_home(page))
+		homecache_finv_page_va(page_address(page), home);
+	else
+		homecache_finv_map_page(page, home);
 }
 
-/*
- * Return the passed length, or zero if it's long enough that we
- * believe we should evict the whole L2 cache.
- */
-static unsigned long cache_flush_length(unsigned long length)
+static inline bool incoherent_home(int home)
 {
-	return (length >= CHIP_L2_CACHE_SIZE()) ? HV_FLUSH_EVICT_L2 : length;
+	return home == PAGE_HOME_IMMUTABLE || home == PAGE_HOME_INCOHERENT;
 }
 
-/* Flush a page out of whatever cache(s) it is in. */
-void homecache_flush_cache(struct page *page, int order)
+static void homecache_finv_page_internal(struct page *page, int force_map)
 {
-	int pages = 1 << order;
-	int length = cache_flush_length(pages * PAGE_SIZE);
-	unsigned long pfn = page_to_pfn(page);
-	struct cpumask home_mask;
-
-	homecache_mask(page, pages, &home_mask);
-	flush_remote(pfn, length, &home_mask, 0, 0, 0, NULL, NULL, 0);
-	sim_validate_lines_evicted(PFN_PHYS(pfn), pages * PAGE_SIZE);
+	int home = page_home(page);
+	if (home == PAGE_HOME_UNCACHED)
+		return;
+	if (incoherent_home(home)) {
+		int cpu;
+		for_each_cpu(cpu, &cpu_cacheable_map)
+			homecache_finv_map_page(page, cpu);
+	} else if (force_map) {
+		/* Force if, e.g., the normal mapping is migrating. */
+		homecache_finv_map_page(page, home);
+	} else {
+		homecache_finv_page_home(page, home);
+	}
+	sim_validate_lines_evicted(PFN_PHYS(page_to_pfn(page)), PAGE_SIZE);
 }
 
+void homecache_finv_page(struct page *page)
+{
+	homecache_finv_page_internal(page, 0);
+}
+
+void homecache_evict(const struct cpumask *mask)
+{
+	flush_remote(0, HV_FLUSH_EVICT_L2, mask, 0, 0, 0, NULL, NULL, 0);
+}
 
 /* Report the home corresponding to a given PTE. */
 static int pte_to_home(pte_t pte)
@@ -441,15 +432,8 @@ struct page *homecache_alloc_pages_node(int nid, gfp_t gfp_mask,
 	return page;
 }
 
-void homecache_free_pages(unsigned long addr, unsigned int order)
+void __homecache_free_pages(struct page *page, unsigned int order)
 {
-	struct page *page;
-
-	if (addr == 0)
-		return;
-
-	VM_BUG_ON(!virt_addr_valid((void *)addr));
-	page = virt_to_page((void *)addr);
 	if (put_page_testzero(page)) {
 		homecache_change_page_home(page, order, initial_page_home());
 		if (order == 0) {
@@ -460,3 +444,13 @@ void homecache_free_pages(unsigned long addr, unsigned int order)
 		}
 	}
 }
+EXPORT_SYMBOL(__homecache_free_pages);
+
+void homecache_free_pages(unsigned long addr, unsigned int order)
+{
+	if (addr != 0) {
+		VM_BUG_ON(!virt_addr_valid((void *)addr));
+		__homecache_free_pages(virt_to_page((void *)addr), order);
+	}
+}
+EXPORT_SYMBOL(homecache_free_pages);
diff --git a/arch/tile/mm/init.c b/arch/tile/mm/init.c
index 630dd2ce2af..a2417a0a822 100644
--- a/arch/tile/mm/init.c
+++ b/arch/tile/mm/init.c
@@ -150,7 +150,21 @@ void __init shatter_pmd(pmd_t *pmd)
 	assign_pte(pmd, pte);
 }
 
-#ifdef CONFIG_HIGHMEM
+#ifdef __tilegx__
+static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va)
+{
+	pud_t *pud = pud_offset(&pgtables[pgd_index(va)], va);
+	if (pud_none(*pud))
+		assign_pmd(pud, alloc_pmd());
+	return pmd_offset(pud, va);
+}
+#else
+static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va)
+{
+	return pmd_offset(pud_offset(&pgtables[pgd_index(va)], va), va);
+}
+#endif
+
 /*
  * This function initializes a certain range of kernel virtual memory
  * with new bootmem page tables, everywhere page tables are missing in
@@ -163,24 +177,17 @@ void __init shatter_pmd(pmd_t *pmd)
  * checking the pgd every time.
  */
 static void __init page_table_range_init(unsigned long start,
-					 unsigned long end, pgd_t *pgd_base)
+					 unsigned long end, pgd_t *pgd)
 {
-	pgd_t *pgd;
-	int pgd_idx;
 	unsigned long vaddr;
-
-	vaddr = start;
-	pgd_idx = pgd_index(vaddr);
-	pgd = pgd_base + pgd_idx;
-
-	for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
-		pmd_t *pmd = pmd_offset(pud_offset(pgd, vaddr), vaddr);
+	start = round_down(start, PMD_SIZE);
+	end = round_up(end, PMD_SIZE);
+	for (vaddr = start; vaddr < end; vaddr += PMD_SIZE) {
+		pmd_t *pmd = get_pmd(pgd, vaddr);
 		if (pmd_none(*pmd))
 			assign_pte(pmd, alloc_pte());
-		vaddr += PMD_SIZE;
 	}
 }
-#endif /* CONFIG_HIGHMEM */
 
 
 #if CHIP_HAS_CBOX_HOME_MAP()
@@ -404,21 +411,6 @@ static inline pgprot_t ktext_set_nocache(pgprot_t prot)
 	return prot;
 }
 
-#ifndef __tilegx__
-static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va)
-{
-	return pmd_offset(pud_offset(&pgtables[pgd_index(va)], va), va);
-}
-#else
-static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va)
-{
-	pud_t *pud = pud_offset(&pgtables[pgd_index(va)], va);
-	if (pud_none(*pud))
-		assign_pmd(pud, alloc_pmd());
-	return pmd_offset(pud, va);
-}
-#endif
-
 /* Temporary page table we use for staging. */
 static pgd_t pgtables[PTRS_PER_PGD]
  __attribute__((aligned(HV_PAGE_TABLE_ALIGN)));
@@ -779,9 +771,6 @@ static void __init set_non_bootmem_pages_init(void)
  */
 void __init paging_init(void)
 {
-#ifdef CONFIG_HIGHMEM
-	unsigned long vaddr, end;
-#endif
 #ifdef __tilegx__
 	pud_t *pud;
 #endif
@@ -789,14 +778,14 @@ void __init paging_init(void)
 
 	kernel_physical_mapping_init(pgd_base);
 
-#ifdef CONFIG_HIGHMEM
 	/*
 	 * Fixed mappings, only the page table structure has to be
 	 * created - mappings will be set by set_fixmap():
 	 */
-	vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
-	end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
-	page_table_range_init(vaddr, end, pgd_base);
+	page_table_range_init(fix_to_virt(__end_of_fixed_addresses - 1),
+			      FIXADDR_TOP, pgd_base);
+
+#ifdef CONFIG_HIGHMEM
 	permanent_kmaps_init(pgd_base);
 #endif