powerpc: use the generic dma coherent remap allocator
authorChristoph Hellwig <hch@lst.de>
Wed, 14 Aug 2019 13:22:30 +0000 (15:22 +0200)
committerMichael Ellerman <mpe@ellerman.id.au>
Wed, 28 Aug 2019 13:19:34 +0000 (23:19 +1000)
This switches to using common code for the DMA allocations, including
potential use of the CMA allocator if configured.

Switching to the generic code enables DMA allocations from atomic
context, which is required by the DMA API documentation, and also
adds various other minor features drivers start relying upon.  It
also makes sure we have on tested code base for all architectures
that require uncached pte bits for coherent DMA allocations.

Another advantage is that consistent memory allocations now share
the general vmalloc pool instead of needing an explicit careout
from it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Christophe Leroy <christophe.leroy@c-s.fr> # tested on 8xx
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20190814132230.31874-2-hch@lst.de
arch/powerpc/Kconfig
arch/powerpc/include/asm/book3s/32/pgtable.h
arch/powerpc/include/asm/nohash/32/pgtable.h
arch/powerpc/mm/dma-noncoherent.c
arch/powerpc/mm/mem.c
arch/powerpc/mm/ptdump/ptdump.c
arch/powerpc/platforms/Kconfig.cputype

index cfb21856559b6e184390706501947a849380c892..edc19363a72906ecbbc7d115636c7190d41ec627 100644 (file)
@@ -1141,18 +1141,6 @@ config TASK_SIZE
        default "0x80000000" if PPC_8xx
        default "0xc0000000"
 
-config CONSISTENT_SIZE_BOOL
-       bool "Set custom consistent memory pool size"
-       depends on ADVANCED_OPTIONS && NOT_COHERENT_CACHE
-       help
-         This option allows you to set the size of the
-         consistent memory pool.  This pool of virtual memory
-         is used to make consistent memory allocations.
-
-config CONSISTENT_SIZE
-       hex "Size of consistent memory pool" if CONSISTENT_SIZE_BOOL
-       default "0x00200000" if NOT_COHERENT_CACHE
-
 config PIN_TLB
        bool "Pinned Kernel TLBs (860 ONLY)"
        depends on ADVANCED_OPTIONS && PPC_8xx && \
index 331a29a501a1b5447a4580d6564d27743faf14ff..0796533d37dd52c4d7b2be34c9c1b5f0d27e63dd 100644 (file)
@@ -148,21 +148,15 @@ int map_kernel_page(unsigned long va, phys_addr_t pa, pgprot_t prot);
  */
 #include <asm/fixmap.h>
 
-#ifdef CONFIG_HIGHMEM
-#define KVIRT_TOP      PKMAP_BASE
-#else
-#define KVIRT_TOP      FIXADDR_START
-#endif
-
 /*
  * ioremap_bot starts at that address. Early ioremaps move down from there,
  * until mem_init() at which point this becomes the top of the vmalloc
  * and ioremap space
  */
-#ifdef CONFIG_NOT_COHERENT_CACHE
-#define IOREMAP_TOP    ((KVIRT_TOP - CONFIG_CONSISTENT_SIZE) & PAGE_MASK)
+#ifdef CONFIG_HIGHMEM
+#define IOREMAP_TOP    PKMAP_BASE
 #else
-#define IOREMAP_TOP    KVIRT_TOP
+#define IOREMAP_TOP    FIXADDR_START
 #endif
 
 /* PPC32 shares vmalloc area with ioremap */
index 3e1a4c1e40f048c553f30d13c5dd5aff4fcee20c..552b96eef0c8ed8b517d6fc662d07640a2f43172 100644 (file)
@@ -76,21 +76,15 @@ int map_kernel_page(unsigned long va, phys_addr_t pa, pgprot_t prot);
  */
 #include <asm/fixmap.h>
 
-#ifdef CONFIG_HIGHMEM
-#define KVIRT_TOP      PKMAP_BASE
-#else
-#define KVIRT_TOP      FIXADDR_START
-#endif
-
 /*
  * ioremap_bot starts at that address. Early ioremaps move down from there,
  * until mem_init() at which point this becomes the top of the vmalloc
  * and ioremap space
  */
-#ifdef CONFIG_NOT_COHERENT_CACHE
-#define IOREMAP_TOP    ((KVIRT_TOP - CONFIG_CONSISTENT_SIZE) & PAGE_MASK)
+#ifdef CONFIG_HIGHMEM
+#define IOREMAP_TOP    PKMAP_BASE
 #else
-#define IOREMAP_TOP    KVIRT_TOP
+#define IOREMAP_TOP    FIXADDR_START
 #endif
 
 /* PPC32 shares vmalloc area with ioremap */
index c617282d5b2aead7023cc27b47cbab0d4a5c4586..4272ca5e8159abd6f2f814b9ca4186d73806e65c 100644 (file)
  *    Copyright (C) 2001 Dan Malek (dmalek@jlc.net)
  *
  *  Copyright (C) 2000 Russell King
- *
- * Consistent memory allocators.  Used for DMA devices that want to
- * share uncached memory with the processor core.  The function return
- * is the virtual address and 'dma_handle' is the physical address.
- * Mostly stolen from the ARM port, with some changes for PowerPC.
- *                                             -- Dan
- *
- * Reorganized to get rid of the arch-specific consistent_* functions
- * and provide non-coherent implementations for the DMA API. -Matt
- *
- * Added in_interrupt() safe dma_alloc_coherent()/dma_free_coherent()
- * implementation. This is pulled straight from ARM and barely
- * modified. -Matt
  */
 
-#include <linux/sched.h>
-#include <linux/slab.h>
 #include <linux/kernel.h>
 #include <linux/errno.h>
-#include <linux/string.h>
 #include <linux/types.h>
 #include <linux/highmem.h>
 #include <linux/dma-direct.h>
 #include <linux/dma-noncoherent.h>
-#include <linux/export.h>
 
 #include <asm/tlbflush.h>
 #include <asm/dma.h>
 
-#include <mm/mmu_decl.h>
-
-/*
- * This address range defaults to a value that is safe for all
- * platforms which currently set CONFIG_NOT_COHERENT_CACHE. It
- * can be further configured for specific applications under
- * the "Advanced Setup" menu. -Matt
- */
-#define CONSISTENT_BASE                (IOREMAP_TOP)
-#define CONSISTENT_END                 (CONSISTENT_BASE + CONFIG_CONSISTENT_SIZE)
-#define CONSISTENT_OFFSET(x)   (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
-
-/*
- * This is the page table (2MB) covering uncached, DMA consistent allocations
- */
-static DEFINE_SPINLOCK(consistent_lock);
-
-/*
- * VM region handling support.
- *
- * This should become something generic, handling VM region allocations for
- * vmalloc and similar (ioremap, module space, etc).
- *
- * I envisage vmalloc()'s supporting vm_struct becoming:
- *
- *  struct vm_struct {
- *    struct vm_region region;
- *    unsigned long    flags;
- *    struct page      **pages;
- *    unsigned int     nr_pages;
- *    unsigned long    phys_addr;
- *  };
- *
- * get_vm_area() would then call vm_region_alloc with an appropriate
- * struct vm_region head (eg):
- *
- *  struct vm_region vmalloc_head = {
- *     .vm_list        = LIST_HEAD_INIT(vmalloc_head.vm_list),
- *     .vm_start       = VMALLOC_START,
- *     .vm_end         = VMALLOC_END,
- *  };
- *
- * However, vmalloc_head.vm_start is variable (typically, it is dependent on
- * the amount of RAM found at boot time.)  I would imagine that get_vm_area()
- * would have to initialise this each time prior to calling vm_region_alloc().
- */
-struct ppc_vm_region {
-       struct list_head        vm_list;
-       unsigned long           vm_start;
-       unsigned long           vm_end;
-};
-
-static struct ppc_vm_region consistent_head = {
-       .vm_list        = LIST_HEAD_INIT(consistent_head.vm_list),
-       .vm_start       = CONSISTENT_BASE,
-       .vm_end         = CONSISTENT_END,
-};
-
-static struct ppc_vm_region *
-ppc_vm_region_alloc(struct ppc_vm_region *head, size_t size, gfp_t gfp)
-{
-       unsigned long addr = head->vm_start, end = head->vm_end - size;
-       unsigned long flags;
-       struct ppc_vm_region *c, *new;
-
-       new = kmalloc(sizeof(struct ppc_vm_region), gfp);
-       if (!new)
-               goto out;
-
-       spin_lock_irqsave(&consistent_lock, flags);
-
-       list_for_each_entry(c, &head->vm_list, vm_list) {
-               if ((addr + size) < addr)
-                       goto nospc;
-               if ((addr + size) <= c->vm_start)
-                       goto found;
-               addr = c->vm_end;
-               if (addr > end)
-                       goto nospc;
-       }
-
- found:
-       /*
-        * Insert this entry _before_ the one we found.
-        */
-       list_add_tail(&new->vm_list, &c->vm_list);
-       new->vm_start = addr;
-       new->vm_end = addr + size;
-
-       spin_unlock_irqrestore(&consistent_lock, flags);
-       return new;
-
- nospc:
-       spin_unlock_irqrestore(&consistent_lock, flags);
-       kfree(new);
- out:
-       return NULL;
-}
-
-static struct ppc_vm_region *ppc_vm_region_find(struct ppc_vm_region *head, unsigned long addr)
-{
-       struct ppc_vm_region *c;
-
-       list_for_each_entry(c, &head->vm_list, vm_list) {
-               if (c->vm_start == addr)
-                       goto out;
-       }
-       c = NULL;
- out:
-       return c;
-}
-
-/*
- * Allocate DMA-coherent memory space and return both the kernel remapped
- * virtual and bus address for that space.
- */
-void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
-               gfp_t gfp, unsigned long attrs)
-{
-       struct page *page;
-       struct ppc_vm_region *c;
-       unsigned long order;
-       u64 mask = ISA_DMA_THRESHOLD, limit;
-
-       if (dev) {
-               mask = dev->coherent_dma_mask;
-
-               /*
-                * Sanity check the DMA mask - it must be non-zero, and
-                * must be able to be satisfied by a DMA allocation.
-                */
-               if (mask == 0) {
-                       dev_warn(dev, "coherent DMA mask is unset\n");
-                       goto no_page;
-               }
-
-               if ((~mask) & ISA_DMA_THRESHOLD) {
-                       dev_warn(dev, "coherent DMA mask %#llx is smaller "
-                                "than system GFP_DMA mask %#llx\n",
-                                mask, (unsigned long long)ISA_DMA_THRESHOLD);
-                       goto no_page;
-               }
-       }
-
-
-       size = PAGE_ALIGN(size);
-       limit = (mask + 1) & ~mask;
-       if ((limit && size >= limit) ||
-           size >= (CONSISTENT_END - CONSISTENT_BASE)) {
-               printk(KERN_WARNING "coherent allocation too big (requested %#x mask %#Lx)\n",
-                      size, mask);
-               return NULL;
-       }
-
-       order = get_order(size);
-
-       /* Might be useful if we ever have a real legacy DMA zone... */
-       if (mask != 0xffffffff)
-               gfp |= GFP_DMA;
-
-       page = alloc_pages(gfp, order);
-       if (!page)
-               goto no_page;
-
-       /*
-        * Invalidate any data that might be lurking in the
-        * kernel direct-mapped region for device DMA.
-        */
-       {
-               unsigned long kaddr = (unsigned long)page_address(page);
-               memset(page_address(page), 0, size);
-               flush_dcache_range(kaddr, kaddr + size);
-       }
-
-       /*
-        * Allocate a virtual address in the consistent mapping region.
-        */
-       c = ppc_vm_region_alloc(&consistent_head, size,
-                           gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
-       if (c) {
-               unsigned long vaddr = c->vm_start;
-               struct page *end = page + (1 << order);
-
-               split_page(page, order);
-
-               /*
-                * Set the "dma handle"
-                */
-               *dma_handle = phys_to_dma(dev, page_to_phys(page));
-
-               do {
-                       SetPageReserved(page);
-                       map_kernel_page(vaddr, page_to_phys(page),
-                                       pgprot_noncached(PAGE_KERNEL));
-                       page++;
-                       vaddr += PAGE_SIZE;
-               } while (size -= PAGE_SIZE);
-
-               /*
-                * Free the otherwise unused pages.
-                */
-               while (page < end) {
-                       __free_page(page);
-                       page++;
-               }
-
-               return (void *)c->vm_start;
-       }
-
-       if (page)
-               __free_pages(page, order);
- no_page:
-       return NULL;
-}
-
-/*
- * free a page as defined by the above mapping.
- */
-void arch_dma_free(struct device *dev, size_t size, void *vaddr,
-               dma_addr_t dma_handle, unsigned long attrs)
-{
-       struct ppc_vm_region *c;
-       unsigned long flags, addr;
-       
-       size = PAGE_ALIGN(size);
-
-       spin_lock_irqsave(&consistent_lock, flags);
-
-       c = ppc_vm_region_find(&consistent_head, (unsigned long)vaddr);
-       if (!c)
-               goto no_area;
-
-       if ((c->vm_end - c->vm_start) != size) {
-               printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
-                      __func__, c->vm_end - c->vm_start, size);
-               dump_stack();
-               size = c->vm_end - c->vm_start;
-       }
-
-       addr = c->vm_start;
-       do {
-               pte_t *ptep;
-               unsigned long pfn;
-
-               ptep = pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(addr),
-                                                              addr),
-                                                   addr),
-                                        addr);
-               if (!pte_none(*ptep) && pte_present(*ptep)) {
-                       pfn = pte_pfn(*ptep);
-                       pte_clear(&init_mm, addr, ptep);
-                       if (pfn_valid(pfn)) {
-                               struct page *page = pfn_to_page(pfn);
-                               __free_reserved_page(page);
-                       }
-               }
-               addr += PAGE_SIZE;
-       } while (size -= PAGE_SIZE);
-
-       flush_tlb_kernel_range(c->vm_start, c->vm_end);
-
-       list_del(&c->vm_list);
-
-       spin_unlock_irqrestore(&consistent_lock, flags);
-
-       kfree(c);
-       return;
-
- no_area:
-       spin_unlock_irqrestore(&consistent_lock, flags);
-       printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
-              __func__, vaddr);
-       dump_stack();
-}
-
 /*
  * make an area consistent.
  */
@@ -408,23 +116,15 @@ void arch_sync_dma_for_cpu(struct device *dev, phys_addr_t paddr,
        __dma_sync_page(paddr, size, dir);
 }
 
-/*
- * Return the PFN for a given cpu virtual address returned by arch_dma_alloc.
- */
-long arch_dma_coherent_to_pfn(struct device *dev, void *vaddr,
-               dma_addr_t dma_addr)
+void arch_dma_prep_coherent(struct page *page, size_t size)
 {
-       /* This should always be populated, so we don't test every
-        * level. If that fails, we'll have a nice crash which
-        * will be as good as a BUG_ON()
-        */
-       unsigned long cpu_addr = (unsigned long)vaddr;
-       pgd_t *pgd = pgd_offset_k(cpu_addr);
-       pud_t *pud = pud_offset(pgd, cpu_addr);
-       pmd_t *pmd = pmd_offset(pud, cpu_addr);
-       pte_t *ptep = pte_offset_kernel(pmd, cpu_addr);
+       unsigned long kaddr = (unsigned long)page_address(page);
 
-       if (pte_none(*ptep) || !pte_present(*ptep))
-               return 0;
-       return pte_pfn(*ptep);
+       flush_dcache_range(kaddr, kaddr + size);
+}
+
+static int __init atomic_pool_init(void)
+{
+       return dma_atomic_pool_init(GFP_KERNEL, pgprot_noncached(PAGE_KERNEL));
 }
+postcore_initcall(atomic_pool_init);
index 69f99128a8d6888673cd57dd8d1e5500f80e0635..be941d382c8d10d036a46210ac5e1e69e6907505 100644 (file)
@@ -302,10 +302,6 @@ void __init mem_init(void)
        pr_info("  * 0x%08lx..0x%08lx  : highmem PTEs\n",
                PKMAP_BASE, PKMAP_ADDR(LAST_PKMAP));
 #endif /* CONFIG_HIGHMEM */
-#ifdef CONFIG_NOT_COHERENT_CACHE
-       pr_info("  * 0x%08lx..0x%08lx  : consistent mem\n",
-               IOREMAP_TOP, IOREMAP_TOP + CONFIG_CONSISTENT_SIZE);
-#endif /* CONFIG_NOT_COHERENT_CACHE */
        if (ioremap_bot != IOREMAP_TOP)
                pr_info("  * 0x%08lx..0x%08lx  : early ioremap\n",
                        ioremap_bot, IOREMAP_TOP);
index ab6a572202b482f9b1a4c356a57553d10144b22f..2f9ddc29c5355454bd0b1003b09ca52c15b2ffcf 100644 (file)
@@ -84,10 +84,6 @@ static struct addr_marker address_markers[] = {
 #else
        { 0,    "Early I/O remap start" },
        { 0,    "Early I/O remap end" },
-#ifdef CONFIG_NOT_COHERENT_CACHE
-       { 0,    "Consistent mem start" },
-       { 0,    "Consistent mem end" },
-#endif
 #ifdef CONFIG_HIGHMEM
        { 0,    "Highmem PTEs start" },
        { 0,    "Highmem PTEs end" },
@@ -335,11 +331,6 @@ static void populate_markers(void)
 #else /* !CONFIG_PPC64 */
        address_markers[i++].start_address = ioremap_bot;
        address_markers[i++].start_address = IOREMAP_TOP;
-#ifdef CONFIG_NOT_COHERENT_CACHE
-       address_markers[i++].start_address = IOREMAP_TOP;
-       address_markers[i++].start_address = IOREMAP_TOP +
-                                            CONFIG_CONSISTENT_SIZE;
-#endif
 #ifdef CONFIG_HIGHMEM
        address_markers[i++].start_address = PKMAP_BASE;
        address_markers[i++].start_address = PKMAP_ADDR(LAST_PKMAP);
index 68c5cc075bfcceeb889ef4ab6daf55145826e2ef..12543e53fa96a507236ccad32f978ed20661bfa4 100644 (file)
@@ -460,8 +460,10 @@ config NOT_COHERENT_CACHE
        depends on 4xx || PPC_8xx || E200 || PPC_MPC512x || \
                GAMECUBE_COMMON || AMIGAONE
        select ARCH_HAS_DMA_COHERENT_TO_PFN
+       select ARCH_HAS_DMA_PREP_COHERENT
        select ARCH_HAS_SYNC_DMA_FOR_DEVICE
        select ARCH_HAS_SYNC_DMA_FOR_CPU
+       select DMA_DIRECT_REMAP
        default n if PPC_47x
        default y