From: Alan Cox alan@linux.intel.com
This fits alongside the GEM support to manage our resources on the card itself. It's not actually clear we need to configure the MMU at all. Further research is needed before removing it entirely. For now we suck it in (slightly abused) from the old semi-free driver.
Signed-off-by: Alan Cox alan@linux.intel.com ---
drivers/gpu/drm/gma500/gtt.c | 500 ++++++++++++++++++++++++ drivers/gpu/drm/gma500/gtt.h | 61 +++ drivers/gpu/drm/gma500/mmu.c | 858 ++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 1419 insertions(+), 0 deletions(-) create mode 100644 drivers/gpu/drm/gma500/gtt.c create mode 100644 drivers/gpu/drm/gma500/gtt.h create mode 100644 drivers/gpu/drm/gma500/mmu.c
diff --git a/drivers/gpu/drm/gma500/gtt.c b/drivers/gpu/drm/gma500/gtt.c new file mode 100644 index 0000000..461ead2 --- /dev/null +++ b/drivers/gpu/drm/gma500/gtt.c @@ -0,0 +1,500 @@ +/* + * Copyright (c) 2007, Intel Corporation. + * All Rights Reserved. + * + * This program is free software; you can redistribute it and/or modify it + * under the terms and conditions of the GNU General Public License, + * version 2, as published by the Free Software Foundation. + * + * This program is distributed in the hope it will be useful, but WITHOUT + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for + * more details. + * + * You should have received a copy of the GNU General Public License along with + * this program; if not, write to the Free Software Foundation, Inc., + * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA. + * + * Authors: Thomas Hellstrom <thomas-at-tungstengraphics.com> + * Alan Cox alan@linux.intel.com + */ + +#include <drm/drmP.h> +#include "psb_drv.h" + + +/* + * GTT resource allocator - manage page mappings in GTT space + */ + +/** + * psb_gtt_mask_pte - generate GTT pte entry + * @pfn: page number to encode + * @type: type of memory in the GTT + * + * Set the GTT entry for the appropriate memory type. + */ +static inline uint32_t psb_gtt_mask_pte(uint32_t pfn, int type) +{ + uint32_t mask = PSB_PTE_VALID; + + if (type & PSB_MMU_CACHED_MEMORY) + mask |= PSB_PTE_CACHED; + if (type & PSB_MMU_RO_MEMORY) + mask |= PSB_PTE_RO; + if (type & PSB_MMU_WO_MEMORY) + mask |= PSB_PTE_WO; + + return (pfn << PAGE_SHIFT) | mask; +} + +/** + * psb_gtt_entry - find the GTT entries for a gtt_range + * @dev: our DRM device + * @r: our GTT range + * + * Given a gtt_range object return the GTT offset of the page table + * entries for this gtt_range + */ +u32 *psb_gtt_entry(struct drm_device *dev, struct gtt_range *r) +{ + struct drm_psb_private *dev_priv = dev->dev_private; + unsigned long offset; + + offset = r->resource.start - dev_priv->gtt_mem->start; + + return dev_priv->gtt_map + (offset >> PAGE_SHIFT); +} + +/** + * psb_gtt_insert - put an object into the GTT + * @dev: our DRM device + * @r: our GTT range + * + * Take our preallocated GTT range and insert the GEM object into + * the GTT. + * + * FIXME: gtt lock ? + */ +static int psb_gtt_insert(struct drm_device *dev, struct gtt_range *r) +{ + u32 *gtt_slot, pte; + struct page **pages; + int i; + + if (r->pages == NULL) { + WARN_ON(1); + return -EINVAL; + } + + WARN_ON(r->stolen); /* refcount these maybe ? */ + + gtt_slot = psb_gtt_entry(dev, r); + pages = r->pages; + + /* Make sure changes are visible to the GPU */ + set_pages_array_uc(pages, r->npage); + + /* Write our page entries into the GTT itself */ + for (i = 0; i < r->npage; i++) { + pte = psb_gtt_mask_pte(page_to_pfn(*pages++), 0/*type*/); + iowrite32(pte, gtt_slot++); + } + /* Make sure all the entries are set before we return */ + ioread32(gtt_slot - 1); + return 0; +} + +/** + * psb_gtt_remove - remove an object from the GTT + * @dev: our DRM device + * @r: our GTT range + * + * Remove a preallocated GTT range from the GTT. Overwrite all the + * page table entries with the dummy page + */ + +static void psb_gtt_remove(struct drm_device *dev, struct gtt_range *r) +{ + struct drm_psb_private *dev_priv = dev->dev_private; + u32 *gtt_slot, pte; + int i; + + WARN_ON(r->stolen); + + gtt_slot = psb_gtt_entry(dev, r); + pte = psb_gtt_mask_pte(page_to_pfn(dev_priv->scratch_page), 0); + + for (i = 0; i < r->npage; i++) + iowrite32(pte, gtt_slot++); + ioread32(gtt_slot - 1); + set_pages_array_wb(r->pages, r->npage); +} + +/** + * psb_gtt_attach_pages - attach and pin GEM pages + * @gt: the gtt range + * + * Pin and build an in kernel list of the pages that back our GEM object. + * While we hold this the pages cannot be swapped out + */ +static int psb_gtt_attach_pages(struct gtt_range *gt) +{ + struct inode *inode; + struct address_space *mapping; + int i; + struct page *p; + int pages = gt->gem.size / PAGE_SIZE; + + WARN_ON(gt->pages); + + /* This is the shared memory object that backs the GEM resource */ + inode = gt->gem.filp->f_path.dentry->d_inode; + mapping = inode->i_mapping; + + gt->pages = kmalloc(pages * sizeof(struct page *), GFP_KERNEL); + if (gt->pages == NULL) + return -ENOMEM; + gt->npage = pages; + + for (i = 0; i < pages; i++) { + /* FIXME: review flags later */ + p = read_cache_page_gfp(mapping, i, + __GFP_COLD | GFP_KERNEL); + if (IS_ERR(p)) + goto err; + gt->pages[i] = p; + } + return 0; + +err: + while (i--) + page_cache_release(gt->pages[i]); + kfree(gt->pages); + gt->pages = NULL; + return PTR_ERR(p); +} + +/** + * psb_gtt_detach_pages - attach and pin GEM pages + * @gt: the gtt range + * + * Undo the effect of psb_gtt_attach_pages. At this point the pages + * must have been removed from the GTT as they could now be paged out + * and move bus address. + */ +static void psb_gtt_detach_pages(struct gtt_range *gt) +{ + int i; + for (i = 0; i < gt->npage; i++) { + /* FIXME: do we need to force dirty */ + set_page_dirty(gt->pages[i]); + page_cache_release(gt->pages[i]); + } + kfree(gt->pages); + gt->pages = NULL; +} + +/** + * psb_gtt_pin - pin pages into the GTT + * @gt: range to pin + * + * Pin a set of pages into the GTT. The pins are refcounted so that + * multiple pins need multiple unpins to undo. + * + * Non GEM backed objects treat this as a no-op as they are always GTT + * backed objects. + */ +int psb_gtt_pin(struct gtt_range *gt) +{ + int ret = 0; + struct drm_device *dev = gt->gem.dev; + struct drm_psb_private *dev_priv = dev->dev_private; + + mutex_lock(&dev_priv->gtt_mutex); + + if (gt->in_gart == 0 && gt->stolen == 0) { + ret = psb_gtt_attach_pages(gt); + if (ret < 0) + goto out; + ret = psb_gtt_insert(dev, gt); + if (ret < 0) { + psb_gtt_detach_pages(gt); + goto out; + } + } + gt->in_gart++; +out: + mutex_unlock(&dev_priv->gtt_mutex); + return ret; +} + +/** + * psb_gtt_unpin - Drop a GTT pin requirement + * @gt: range to pin + * + * Undoes the effect of psb_gtt_pin. On the last drop the GEM object + * will be removed from the GTT which will also drop the page references + * and allow the VM to clean up or page stuff. + * + * Non GEM backed objects treat this as a no-op as they are always GTT + * backed objects. + */ +void psb_gtt_unpin(struct gtt_range *gt) +{ + struct drm_device *dev = gt->gem.dev; + struct drm_psb_private *dev_priv = dev->dev_private; + + mutex_lock(&dev_priv->gtt_mutex); + + WARN_ON(!gt->in_gart); + + gt->in_gart--; + if (gt->in_gart == 0 && gt->stolen == 0) { + psb_gtt_remove(dev, gt); + psb_gtt_detach_pages(gt); + } + mutex_unlock(&dev_priv->gtt_mutex); +} + +/* + * GTT resource allocator - allocate and manage GTT address space + */ + +/** + * psb_gtt_alloc_range - allocate GTT address space + * @dev: Our DRM device + * @len: length (bytes) of address space required + * @name: resource name + * @backed: resource should be backed by stolen pages + * + * Ask the kernel core to find us a suitable range of addresses + * to use for a GTT mapping. + * + * Returns a gtt_range structure describing the object, or NULL on + * error. On successful return the resource is both allocated and marked + * as in use. + */ +struct gtt_range *psb_gtt_alloc_range(struct drm_device *dev, int len, + const char *name, int backed) +{ + struct drm_psb_private *dev_priv = dev->dev_private; + struct gtt_range *gt; + struct resource *r = dev_priv->gtt_mem; + int ret; + unsigned long start, end; + + if (backed) { + /* The start of the GTT is the stolen pages */ + start = r->start; + end = r->start + dev_priv->gtt.stolen_size - 1; + } else { + /* The rest we will use for GEM backed objects */ + start = r->start + dev_priv->gtt.stolen_size; + end = r->end; + } + + gt = kzalloc(sizeof(struct gtt_range), GFP_KERNEL); + if (gt == NULL) + return NULL; + gt->resource.name = name; + gt->stolen = backed; + gt->in_gart = backed; + /* Ensure this is set for non GEM objects */ + gt->gem.dev = dev; + ret = allocate_resource(dev_priv->gtt_mem, >->resource, + len, start, end, PAGE_SIZE, NULL, NULL); + if (ret == 0) { + gt->offset = gt->resource.start - r->start; + return gt; + } + kfree(gt); + return NULL; +} + +/** + * psb_gtt_free_range - release GTT address space + * @dev: our DRM device + * @gt: a mapping created with psb_gtt_alloc_range + * + * Release a resource that was allocated with psb_gtt_alloc_range. If the + * object has been pinned by mmap users we clean this up here currently. + */ +void psb_gtt_free_range(struct drm_device *dev, struct gtt_range *gt) +{ + /* Undo the mmap pin if we are destroying the object */ + if (gt->mmapping) { + psb_gtt_unpin(gt); + gt->mmapping = 0; + } + WARN_ON(gt->in_gart && !gt->stolen); + release_resource(>->resource); + kfree(gt); +} + +void psb_gtt_alloc(struct drm_device *dev) +{ + struct drm_psb_private *dev_priv = dev->dev_private; + init_rwsem(&dev_priv->gtt.sem); +} + +void psb_gtt_takedown(struct drm_device *dev) +{ + struct drm_psb_private *dev_priv = dev->dev_private; + + if (dev_priv->gtt_map) { + iounmap(dev_priv->gtt_map); + dev_priv->gtt_map = NULL; + } + if (dev_priv->gtt_initialized) { + pci_write_config_word(dev->pdev, PSB_GMCH_CTRL, + dev_priv->gmch_ctrl); + PSB_WVDC32(dev_priv->pge_ctl, PSB_PGETBL_CTL); + (void) PSB_RVDC32(PSB_PGETBL_CTL); + } + if (dev_priv->vram_addr) + iounmap(dev_priv->gtt_map); +} + +int psb_gtt_init(struct drm_device *dev, int resume) +{ + struct drm_psb_private *dev_priv = dev->dev_private; + unsigned gtt_pages; + unsigned long stolen_size, vram_stolen_size; + unsigned i, num_pages; + unsigned pfn_base; + uint32_t vram_pages; + uint32_t dvmt_mode = 0; + struct psb_gtt *pg; + + int ret = 0; + uint32_t pte; + + mutex_init(&dev_priv->gtt_mutex); + + psb_gtt_alloc(dev); + pg = &dev_priv->gtt; + + /* Enable the GTT */ + pci_read_config_word(dev->pdev, PSB_GMCH_CTRL, &dev_priv->gmch_ctrl); + pci_write_config_word(dev->pdev, PSB_GMCH_CTRL, + dev_priv->gmch_ctrl | _PSB_GMCH_ENABLED); + + dev_priv->pge_ctl = PSB_RVDC32(PSB_PGETBL_CTL); + PSB_WVDC32(dev_priv->pge_ctl | _PSB_PGETBL_ENABLED, PSB_PGETBL_CTL); + (void) PSB_RVDC32(PSB_PGETBL_CTL); + + /* The root resource we allocate address space from */ + dev_priv->gtt_initialized = 1; + + pg->gtt_phys_start = dev_priv->pge_ctl & PAGE_MASK; + + /* + * FIXME: video mmu has hw bug to access 0x0D0000000, + * then make gatt start at 0x0e000,0000 + */ + pg->mmu_gatt_start = 0xE0000000; + + pg->gtt_start = pci_resource_start(dev->pdev, PSB_GTT_RESOURCE); + gtt_pages = pci_resource_len(dev->pdev, PSB_GTT_RESOURCE) + >> PAGE_SHIFT; + /* CDV workaround */ + if (pg->gtt_start == 0 || gtt_pages == 0) { + dev_err(dev->dev, "GTT PCI BAR not initialized.\n"); + gtt_pages = 64; + pg->gtt_start = dev_priv->pge_ctl; + } + + pg->gatt_start = pci_resource_start(dev->pdev, PSB_GATT_RESOURCE); + pg->gatt_pages = pci_resource_len(dev->pdev, PSB_GATT_RESOURCE) + >> PAGE_SHIFT; + dev_priv->gtt_mem = &dev->pdev->resource[PSB_GATT_RESOURCE]; + + if (pg->gatt_pages == 0 || pg->gatt_start == 0) { + static struct resource fudge; /* Preferably peppermint */ + + /* This can occur on CDV SDV systems. Fudge it in this case. + We really don't care what imaginary space is being allocated + at this point */ + dev_err(dev->dev, "GATT PCI BAR not initialized.\n"); + pg->gatt_start = 0x40000000; + pg->gatt_pages = (128 * 1024 * 1024) >> PAGE_SHIFT; + fudge.start = 0x40000000; + fudge.end = 0x40000000 + 128 * 1024 * 1024 - 1; + fudge.name = "fudge"; + fudge.flags = IORESOURCE_MEM; + dev_priv->gtt_mem = &fudge; + } + + pci_read_config_dword(dev->pdev, PSB_BSM, &dev_priv->stolen_base); + vram_stolen_size = pg->gtt_phys_start - dev_priv->stolen_base + - PAGE_SIZE; + + stolen_size = vram_stolen_size; + + printk(KERN_INFO "Stolen memory information\n"); + printk(KERN_INFO " base in RAM: 0x%x\n", dev_priv->stolen_base); + printk(KERN_INFO " size: %luK, calculated by (GTT RAM base) - (Stolen base), seems wrong\n", + vram_stolen_size/1024); + dvmt_mode = (dev_priv->gmch_ctrl >> 4) & 0x7; + printk(KERN_INFO " the correct size should be: %dM(dvmt mode=%d)\n", + (dvmt_mode == 1) ? 1 : (2 << (dvmt_mode - 1)), dvmt_mode); + + if (resume && (gtt_pages != pg->gtt_pages) && + (stolen_size != pg->stolen_size)) { + dev_err(dev->dev, "GTT resume error.\n"); + ret = -EINVAL; + goto out_err; + } + + pg->gtt_pages = gtt_pages; + pg->stolen_size = stolen_size; + dev_priv->vram_stolen_size = vram_stolen_size; + + /* + * Map the GTT and the stolen memory area + */ + dev_priv->gtt_map = ioremap_nocache(pg->gtt_phys_start, + gtt_pages << PAGE_SHIFT); + if (!dev_priv->gtt_map) { + dev_err(dev->dev, "Failure to map gtt.\n"); + ret = -ENOMEM; + goto out_err; + } + + dev_priv->vram_addr = ioremap_wc(dev_priv->stolen_base, stolen_size); + if (!dev_priv->vram_addr) { + dev_err(dev->dev, "Failure to map stolen base.\n"); + ret = -ENOMEM; + goto out_err; + } + + /* + * Insert vram stolen pages into the GTT + */ + + pfn_base = dev_priv->stolen_base >> PAGE_SHIFT; + vram_pages = num_pages = vram_stolen_size >> PAGE_SHIFT; + printk(KERN_INFO"Set up %d stolen pages starting at 0x%08x, GTT offset %dK\n", + num_pages, pfn_base << PAGE_SHIFT, 0); + for (i = 0; i < num_pages; ++i) { + pte = psb_gtt_mask_pte(pfn_base + i, 0); + iowrite32(pte, dev_priv->gtt_map + i); + } + + /* + * Init rest of GTT to the scratch page to avoid accidents or scribbles + */ + + pfn_base = page_to_pfn(dev_priv->scratch_page); + pte = psb_gtt_mask_pte(pfn_base, 0); + for (; i < gtt_pages; ++i) + iowrite32(pte, dev_priv->gtt_map + i); + + (void) ioread32(dev_priv->gtt_map + i - 1); + return 0; + +out_err: + psb_gtt_takedown(dev); + return ret; +} diff --git a/drivers/gpu/drm/gma500/gtt.h b/drivers/gpu/drm/gma500/gtt.h new file mode 100644 index 0000000..e0e1cb6 --- /dev/null +++ b/drivers/gpu/drm/gma500/gtt.h @@ -0,0 +1,61 @@ +/************************************************************************** + * Copyright (c) 2007-2008, Intel Corporation. + * All Rights Reserved. + * + * This program is free software; you can redistribute it and/or modify it + * under the terms and conditions of the GNU General Public License, + * version 2, as published by the Free Software Foundation. + * + * This program is distributed in the hope it will be useful, but WITHOUT + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for + * more details. + * + * You should have received a copy of the GNU General Public License along with + * this program; if not, write to the Free Software Foundation, Inc., + * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA. + * + **************************************************************************/ + +#ifndef _PSB_GTT_H_ +#define _PSB_GTT_H_ + +#include <drm/drmP.h> + +/* This wants cleaning up with respect to the psb_dev and un-needed stuff */ +struct psb_gtt { + uint32_t gatt_start; + uint32_t mmu_gatt_start; + uint32_t gtt_start; + uint32_t gtt_phys_start; + unsigned gtt_pages; + unsigned gatt_pages; + unsigned long stolen_size; + unsigned long vram_stolen_size; + struct rw_semaphore sem; +}; + +/* Exported functions */ +extern int psb_gtt_init(struct drm_device *dev, int resume); +extern void psb_gtt_takedown(struct drm_device *dev); + +/* Each gtt_range describes an allocation in the GTT area */ +struct gtt_range { + struct resource resource; /* Resource for our allocation */ + u32 offset; /* GTT offset of our object */ + struct drm_gem_object gem; /* GEM high level stuff */ + int in_gart; /* Currently in the GART (ref ct) */ + bool stolen; /* Backed from stolen RAM */ + bool mmapping; /* Is mmappable */ + struct page **pages; /* Backing pages if present */ + int npage; /* Number of backing pages */ +}; + +extern struct gtt_range *psb_gtt_alloc_range(struct drm_device *dev, int len, + const char *name, int backed); +extern void psb_gtt_kref_put(struct gtt_range *gt); +extern void psb_gtt_free_range(struct drm_device *dev, struct gtt_range *gt); +extern int psb_gtt_pin(struct gtt_range *gt); +extern void psb_gtt_unpin(struct gtt_range *gt); + +#endif diff --git a/drivers/gpu/drm/gma500/mmu.c b/drivers/gpu/drm/gma500/mmu.c new file mode 100644 index 0000000..c904d73 --- /dev/null +++ b/drivers/gpu/drm/gma500/mmu.c @@ -0,0 +1,858 @@ +/************************************************************************** + * Copyright (c) 2007, Intel Corporation. + * + * This program is free software; you can redistribute it and/or modify it + * under the terms and conditions of the GNU General Public License, + * version 2, as published by the Free Software Foundation. + * + * This program is distributed in the hope it will be useful, but WITHOUT + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for + * more details. + * + * You should have received a copy of the GNU General Public License along with + * this program; if not, write to the Free Software Foundation, Inc., + * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA. + * + **************************************************************************/ +#include <drm/drmP.h> +#include "psb_drv.h" +#include "psb_reg.h" + +/* + * Code for the SGX MMU: + */ + +/* + * clflush on one processor only: + * clflush should apparently flush the cache line on all processors in an + * SMP system. + */ + +/* + * kmap atomic: + * The usage of the slots must be completely encapsulated within a spinlock, and + * no other functions that may be using the locks for other purposed may be + * called from within the locked region. + * Since the slots are per processor, this will guarantee that we are the only + * user. + */ + +/* + * TODO: Inserting ptes from an interrupt handler: + * This may be desirable for some SGX functionality where the GPU can fault in + * needed pages. For that, we need to make an atomic insert_pages function, that + * may fail. + * If it fails, the caller need to insert the page using a workqueue function, + * but on average it should be fast. + */ + +struct psb_mmu_driver { + /* protects driver- and pd structures. Always take in read mode + * before taking the page table spinlock. + */ + struct rw_semaphore sem; + + /* protects page tables, directory tables and pt tables. + * and pt structures. + */ + spinlock_t lock; + + atomic_t needs_tlbflush; + + uint8_t __iomem *register_map; + struct psb_mmu_pd *default_pd; + /*uint32_t bif_ctrl;*/ + int has_clflush; + int clflush_add; + unsigned long clflush_mask; + + struct drm_psb_private *dev_priv; +}; + +struct psb_mmu_pd; + +struct psb_mmu_pt { + struct psb_mmu_pd *pd; + uint32_t index; + uint32_t count; + struct page *p; + uint32_t *v; +}; + +struct psb_mmu_pd { + struct psb_mmu_driver *driver; + int hw_context; + struct psb_mmu_pt **tables; + struct page *p; + struct page *dummy_pt; + struct page *dummy_page; + uint32_t pd_mask; + uint32_t invalid_pde; + uint32_t invalid_pte; +}; + +static inline uint32_t psb_mmu_pt_index(uint32_t offset) +{ + return (offset >> PSB_PTE_SHIFT) & 0x3FF; +} + +static inline uint32_t psb_mmu_pd_index(uint32_t offset) +{ + return offset >> PSB_PDE_SHIFT; +} + +static inline void psb_clflush(void *addr) +{ + __asm__ __volatile__("clflush (%0)\n" : : "r"(addr) : "memory"); +} + +static inline void psb_mmu_clflush(struct psb_mmu_driver *driver, + void *addr) +{ + if (!driver->has_clflush) + return; + + mb(); + psb_clflush(addr); + mb(); +} + +static void psb_page_clflush(struct psb_mmu_driver *driver, struct page* page) +{ + uint32_t clflush_add = driver->clflush_add >> PAGE_SHIFT; + uint32_t clflush_count = PAGE_SIZE / clflush_add; + int i; + uint8_t *clf; + + clf = kmap_atomic(page, KM_USER0); + mb(); + for (i = 0; i < clflush_count; ++i) { + psb_clflush(clf); + clf += clflush_add; + } + mb(); + kunmap_atomic(clf, KM_USER0); +} + +static void psb_pages_clflush(struct psb_mmu_driver *driver, + struct page *page[], unsigned long num_pages) +{ + int i; + + if (!driver->has_clflush) + return ; + + for (i = 0; i < num_pages; i++) + psb_page_clflush(driver, *page++); +} + +static void psb_mmu_flush_pd_locked(struct psb_mmu_driver *driver, + int force) +{ + atomic_set(&driver->needs_tlbflush, 0); +} + +static void psb_mmu_flush_pd(struct psb_mmu_driver *driver, int force) +{ + down_write(&driver->sem); + psb_mmu_flush_pd_locked(driver, force); + up_write(&driver->sem); +} + +void psb_mmu_flush(struct psb_mmu_driver *driver, int rc_prot) +{ + if (rc_prot) + down_write(&driver->sem); + if (rc_prot) + up_write(&driver->sem); +} + +void psb_mmu_set_pd_context(struct psb_mmu_pd *pd, int hw_context) +{ + /*ttm_tt_cache_flush(&pd->p, 1);*/ + psb_pages_clflush(pd->driver, &pd->p, 1); + down_write(&pd->driver->sem); + wmb(); + psb_mmu_flush_pd_locked(pd->driver, 1); + pd->hw_context = hw_context; + up_write(&pd->driver->sem); + +} + +static inline unsigned long psb_pd_addr_end(unsigned long addr, + unsigned long end) +{ + + addr = (addr + PSB_PDE_MASK + 1) & ~PSB_PDE_MASK; + return (addr < end) ? addr : end; +} + +static inline uint32_t psb_mmu_mask_pte(uint32_t pfn, int type) +{ + uint32_t mask = PSB_PTE_VALID; + + if (type & PSB_MMU_CACHED_MEMORY) + mask |= PSB_PTE_CACHED; + if (type & PSB_MMU_RO_MEMORY) + mask |= PSB_PTE_RO; + if (type & PSB_MMU_WO_MEMORY) + mask |= PSB_PTE_WO; + + return (pfn << PAGE_SHIFT) | mask; +} + +struct psb_mmu_pd *psb_mmu_alloc_pd(struct psb_mmu_driver *driver, + int trap_pagefaults, int invalid_type) +{ + struct psb_mmu_pd *pd = kmalloc(sizeof(*pd), GFP_KERNEL); + uint32_t *v; + int i; + + if (!pd) + return NULL; + + pd->p = alloc_page(GFP_DMA32); + if (!pd->p) + goto out_err1; + pd->dummy_pt = alloc_page(GFP_DMA32); + if (!pd->dummy_pt) + goto out_err2; + pd->dummy_page = alloc_page(GFP_DMA32); + if (!pd->dummy_page) + goto out_err3; + + if (!trap_pagefaults) { + pd->invalid_pde = + psb_mmu_mask_pte(page_to_pfn(pd->dummy_pt), + invalid_type); + pd->invalid_pte = + psb_mmu_mask_pte(page_to_pfn(pd->dummy_page), + invalid_type); + } else { + pd->invalid_pde = 0; + pd->invalid_pte = 0; + } + + v = kmap(pd->dummy_pt); + for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i) + v[i] = pd->invalid_pte; + + kunmap(pd->dummy_pt); + + v = kmap(pd->p); + for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i) + v[i] = pd->invalid_pde; + + kunmap(pd->p); + + clear_page(kmap(pd->dummy_page)); + kunmap(pd->dummy_page); + + pd->tables = vmalloc_user(sizeof(struct psb_mmu_pt *) * 1024); + if (!pd->tables) + goto out_err4; + + pd->hw_context = -1; + pd->pd_mask = PSB_PTE_VALID; + pd->driver = driver; + + return pd; + +out_err4: + __free_page(pd->dummy_page); +out_err3: + __free_page(pd->dummy_pt); +out_err2: + __free_page(pd->p); +out_err1: + kfree(pd); + return NULL; +} + +void psb_mmu_free_pt(struct psb_mmu_pt *pt) +{ + __free_page(pt->p); + kfree(pt); +} + +void psb_mmu_free_pagedir(struct psb_mmu_pd *pd) +{ + struct psb_mmu_driver *driver = pd->driver; + struct psb_mmu_pt *pt; + int i; + + down_write(&driver->sem); + if (pd->hw_context != -1) + psb_mmu_flush_pd_locked(driver, 1); + + /* Should take the spinlock here, but we don't need to do that + since we have the semaphore in write mode. */ + + for (i = 0; i < 1024; ++i) { + pt = pd->tables[i]; + if (pt) + psb_mmu_free_pt(pt); + } + + vfree(pd->tables); + __free_page(pd->dummy_page); + __free_page(pd->dummy_pt); + __free_page(pd->p); + kfree(pd); + up_write(&driver->sem); +} + +static struct psb_mmu_pt *psb_mmu_alloc_pt(struct psb_mmu_pd *pd) +{ + struct psb_mmu_pt *pt = kmalloc(sizeof(*pt), GFP_KERNEL); + void *v; + uint32_t clflush_add = pd->driver->clflush_add >> PAGE_SHIFT; + uint32_t clflush_count = PAGE_SIZE / clflush_add; + spinlock_t *lock = &pd->driver->lock; + uint8_t *clf; + uint32_t *ptes; + int i; + + if (!pt) + return NULL; + + pt->p = alloc_page(GFP_DMA32); + if (!pt->p) { + kfree(pt); + return NULL; + } + + spin_lock(lock); + + v = kmap_atomic(pt->p, KM_USER0); + clf = (uint8_t *) v; + ptes = (uint32_t *) v; + for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i) + *ptes++ = pd->invalid_pte; + + + if (pd->driver->has_clflush && pd->hw_context != -1) { + mb(); + for (i = 0; i < clflush_count; ++i) { + psb_clflush(clf); + clf += clflush_add; + } + mb(); + } + + kunmap_atomic(v, KM_USER0); + spin_unlock(lock); + + pt->count = 0; + pt->pd = pd; + pt->index = 0; + + return pt; +} + +struct psb_mmu_pt *psb_mmu_pt_alloc_map_lock(struct psb_mmu_pd *pd, + unsigned long addr) +{ + uint32_t index = psb_mmu_pd_index(addr); + struct psb_mmu_pt *pt; + uint32_t *v; + spinlock_t *lock = &pd->driver->lock; + + spin_lock(lock); + pt = pd->tables[index]; + while (!pt) { + spin_unlock(lock); + pt = psb_mmu_alloc_pt(pd); + if (!pt) + return NULL; + spin_lock(lock); + + if (pd->tables[index]) { + spin_unlock(lock); + psb_mmu_free_pt(pt); + spin_lock(lock); + pt = pd->tables[index]; + continue; + } + + v = kmap_atomic(pd->p, KM_USER0); + pd->tables[index] = pt; + v[index] = (page_to_pfn(pt->p) << 12) | pd->pd_mask; + pt->index = index; + kunmap_atomic((void *) v, KM_USER0); + + if (pd->hw_context != -1) { + psb_mmu_clflush(pd->driver, (void *) &v[index]); + atomic_set(&pd->driver->needs_tlbflush, 1); + } + } + pt->v = kmap_atomic(pt->p, KM_USER0); + return pt; +} + +static struct psb_mmu_pt *psb_mmu_pt_map_lock(struct psb_mmu_pd *pd, + unsigned long addr) +{ + uint32_t index = psb_mmu_pd_index(addr); + struct psb_mmu_pt *pt; + spinlock_t *lock = &pd->driver->lock; + + spin_lock(lock); + pt = pd->tables[index]; + if (!pt) { + spin_unlock(lock); + return NULL; + } + pt->v = kmap_atomic(pt->p, KM_USER0); + return pt; +} + +static void psb_mmu_pt_unmap_unlock(struct psb_mmu_pt *pt) +{ + struct psb_mmu_pd *pd = pt->pd; + uint32_t *v; + + kunmap_atomic(pt->v, KM_USER0); + if (pt->count == 0) { + v = kmap_atomic(pd->p, KM_USER0); + v[pt->index] = pd->invalid_pde; + pd->tables[pt->index] = NULL; + + if (pd->hw_context != -1) { + psb_mmu_clflush(pd->driver, + (void *) &v[pt->index]); + atomic_set(&pd->driver->needs_tlbflush, 1); + } + kunmap_atomic(pt->v, KM_USER0); + spin_unlock(&pd->driver->lock); + psb_mmu_free_pt(pt); + return; + } + spin_unlock(&pd->driver->lock); +} + +static inline void psb_mmu_set_pte(struct psb_mmu_pt *pt, + unsigned long addr, uint32_t pte) +{ + pt->v[psb_mmu_pt_index(addr)] = pte; +} + +static inline void psb_mmu_invalidate_pte(struct psb_mmu_pt *pt, + unsigned long addr) +{ + pt->v[psb_mmu_pt_index(addr)] = pt->pd->invalid_pte; +} + + +void psb_mmu_mirror_gtt(struct psb_mmu_pd *pd, + uint32_t mmu_offset, uint32_t gtt_start, + uint32_t gtt_pages) +{ + uint32_t *v; + uint32_t start = psb_mmu_pd_index(mmu_offset); + struct psb_mmu_driver *driver = pd->driver; + int num_pages = gtt_pages; + + down_read(&driver->sem); + spin_lock(&driver->lock); + + v = kmap_atomic(pd->p, KM_USER0); + v += start; + + while (gtt_pages--) { + *v++ = gtt_start | pd->pd_mask; + gtt_start += PAGE_SIZE; + } + + /*ttm_tt_cache_flush(&pd->p, num_pages);*/ + psb_pages_clflush(pd->driver, &pd->p, num_pages); + kunmap_atomic(v, KM_USER0); + spin_unlock(&driver->lock); + + if (pd->hw_context != -1) + atomic_set(&pd->driver->needs_tlbflush, 1); + + up_read(&pd->driver->sem); + psb_mmu_flush_pd(pd->driver, 0); +} + +struct psb_mmu_pd *psb_mmu_get_default_pd(struct psb_mmu_driver *driver) +{ + struct psb_mmu_pd *pd; + + /* down_read(&driver->sem); */ + pd = driver->default_pd; + /* up_read(&driver->sem); */ + + return pd; +} + +/* Returns the physical address of the PD shared by sgx/msvdx */ +uint32_t psb_get_default_pd_addr(struct psb_mmu_driver *driver) +{ + struct psb_mmu_pd *pd; + + pd = psb_mmu_get_default_pd(driver); + return page_to_pfn(pd->p) << PAGE_SHIFT; +} + +void psb_mmu_driver_takedown(struct psb_mmu_driver *driver) +{ + psb_mmu_free_pagedir(driver->default_pd); + kfree(driver); +} + +struct psb_mmu_driver *psb_mmu_driver_init(uint8_t __iomem * registers, + int trap_pagefaults, + int invalid_type, + struct drm_psb_private *dev_priv) +{ + struct psb_mmu_driver *driver; + + driver = kmalloc(sizeof(*driver), GFP_KERNEL); + + if (!driver) + return NULL; + driver->dev_priv = dev_priv; + + driver->default_pd = psb_mmu_alloc_pd(driver, trap_pagefaults, + invalid_type); + if (!driver->default_pd) + goto out_err1; + + spin_lock_init(&driver->lock); + init_rwsem(&driver->sem); + down_write(&driver->sem); + driver->register_map = registers; + atomic_set(&driver->needs_tlbflush, 1); + + driver->has_clflush = 0; + + if (boot_cpu_has(X86_FEATURE_CLFLSH)) { + uint32_t tfms, misc, cap0, cap4, clflush_size; + + /* + * clflush size is determined at kernel setup for x86_64 + * but not for i386. We have to do it here. + */ + + cpuid(0x00000001, &tfms, &misc, &cap0, &cap4); + clflush_size = ((misc >> 8) & 0xff) * 8; + driver->has_clflush = 1; + driver->clflush_add = + PAGE_SIZE * clflush_size / sizeof(uint32_t); + driver->clflush_mask = driver->clflush_add - 1; + driver->clflush_mask = ~driver->clflush_mask; + } + + up_write(&driver->sem); + return driver; + +out_err1: + kfree(driver); + return NULL; +} + +static void psb_mmu_flush_ptes(struct psb_mmu_pd *pd, + unsigned long address, uint32_t num_pages, + uint32_t desired_tile_stride, + uint32_t hw_tile_stride) +{ + struct psb_mmu_pt *pt; + uint32_t rows = 1; + uint32_t i; + unsigned long addr; + unsigned long end; + unsigned long next; + unsigned long add; + unsigned long row_add; + unsigned long clflush_add = pd->driver->clflush_add; + unsigned long clflush_mask = pd->driver->clflush_mask; + + if (!pd->driver->has_clflush) { + /*ttm_tt_cache_flush(&pd->p, num_pages);*/ + psb_pages_clflush(pd->driver, &pd->p, num_pages); + return; + } + + if (hw_tile_stride) + rows = num_pages / desired_tile_stride; + else + desired_tile_stride = num_pages; + + add = desired_tile_stride << PAGE_SHIFT; + row_add = hw_tile_stride << PAGE_SHIFT; + mb(); + for (i = 0; i < rows; ++i) { + + addr = address; + end = addr + add; + + do { + next = psb_pd_addr_end(addr, end); + pt = psb_mmu_pt_map_lock(pd, addr); + if (!pt) + continue; + do { + psb_clflush(&pt->v + [psb_mmu_pt_index(addr)]); + } while (addr += + clflush_add, + (addr & clflush_mask) < next); + + psb_mmu_pt_unmap_unlock(pt); + } while (addr = next, next != end); + address += row_add; + } + mb(); +} + +void psb_mmu_remove_pfn_sequence(struct psb_mmu_pd *pd, + unsigned long address, uint32_t num_pages) +{ + struct psb_mmu_pt *pt; + unsigned long addr; + unsigned long end; + unsigned long next; + unsigned long f_address = address; + + down_read(&pd->driver->sem); + + addr = address; + end = addr + (num_pages << PAGE_SHIFT); + + do { + next = psb_pd_addr_end(addr, end); + pt = psb_mmu_pt_alloc_map_lock(pd, addr); + if (!pt) + goto out; + do { + psb_mmu_invalidate_pte(pt, addr); + --pt->count; + } while (addr += PAGE_SIZE, addr < next); + psb_mmu_pt_unmap_unlock(pt); + + } while (addr = next, next != end); + +out: + if (pd->hw_context != -1) + psb_mmu_flush_ptes(pd, f_address, num_pages, 1, 1); + + up_read(&pd->driver->sem); + + if (pd->hw_context != -1) + psb_mmu_flush(pd->driver, 0); + + return; +} + +void psb_mmu_remove_pages(struct psb_mmu_pd *pd, unsigned long address, + uint32_t num_pages, uint32_t desired_tile_stride, + uint32_t hw_tile_stride) +{ + struct psb_mmu_pt *pt; + uint32_t rows = 1; + uint32_t i; + unsigned long addr; + unsigned long end; + unsigned long next; + unsigned long add; + unsigned long row_add; + unsigned long f_address = address; + + if (hw_tile_stride) + rows = num_pages / desired_tile_stride; + else + desired_tile_stride = num_pages; + + add = desired_tile_stride << PAGE_SHIFT; + row_add = hw_tile_stride << PAGE_SHIFT; + + /* down_read(&pd->driver->sem); */ + + /* Make sure we only need to flush this processor's cache */ + + for (i = 0; i < rows; ++i) { + + addr = address; + end = addr + add; + + do { + next = psb_pd_addr_end(addr, end); + pt = psb_mmu_pt_map_lock(pd, addr); + if (!pt) + continue; + do { + psb_mmu_invalidate_pte(pt, addr); + --pt->count; + + } while (addr += PAGE_SIZE, addr < next); + psb_mmu_pt_unmap_unlock(pt); + + } while (addr = next, next != end); + address += row_add; + } + if (pd->hw_context != -1) + psb_mmu_flush_ptes(pd, f_address, num_pages, + desired_tile_stride, hw_tile_stride); + + /* up_read(&pd->driver->sem); */ + + if (pd->hw_context != -1) + psb_mmu_flush(pd->driver, 0); +} + +int psb_mmu_insert_pfn_sequence(struct psb_mmu_pd *pd, uint32_t start_pfn, + unsigned long address, uint32_t num_pages, + int type) +{ + struct psb_mmu_pt *pt; + uint32_t pte; + unsigned long addr; + unsigned long end; + unsigned long next; + unsigned long f_address = address; + int ret = 0; + + down_read(&pd->driver->sem); + + addr = address; + end = addr + (num_pages << PAGE_SHIFT); + + do { + next = psb_pd_addr_end(addr, end); + pt = psb_mmu_pt_alloc_map_lock(pd, addr); + if (!pt) { + ret = -ENOMEM; + goto out; + } + do { + pte = psb_mmu_mask_pte(start_pfn++, type); + psb_mmu_set_pte(pt, addr, pte); + pt->count++; + } while (addr += PAGE_SIZE, addr < next); + psb_mmu_pt_unmap_unlock(pt); + + } while (addr = next, next != end); + +out: + if (pd->hw_context != -1) + psb_mmu_flush_ptes(pd, f_address, num_pages, 1, 1); + + up_read(&pd->driver->sem); + + if (pd->hw_context != -1) + psb_mmu_flush(pd->driver, 1); + + return ret; +} + +int psb_mmu_insert_pages(struct psb_mmu_pd *pd, struct page **pages, + unsigned long address, uint32_t num_pages, + uint32_t desired_tile_stride, + uint32_t hw_tile_stride, int type) +{ + struct psb_mmu_pt *pt; + uint32_t rows = 1; + uint32_t i; + uint32_t pte; + unsigned long addr; + unsigned long end; + unsigned long next; + unsigned long add; + unsigned long row_add; + unsigned long f_address = address; + int ret = 0; + + if (hw_tile_stride) { + if (num_pages % desired_tile_stride != 0) + return -EINVAL; + rows = num_pages / desired_tile_stride; + } else { + desired_tile_stride = num_pages; + } + + add = desired_tile_stride << PAGE_SHIFT; + row_add = hw_tile_stride << PAGE_SHIFT; + + down_read(&pd->driver->sem); + + for (i = 0; i < rows; ++i) { + + addr = address; + end = addr + add; + + do { + next = psb_pd_addr_end(addr, end); + pt = psb_mmu_pt_alloc_map_lock(pd, addr); + if (!pt) { + ret = -ENOMEM; + goto out; + } + do { + pte = + psb_mmu_mask_pte(page_to_pfn(*pages++), + type); + psb_mmu_set_pte(pt, addr, pte); + pt->count++; + } while (addr += PAGE_SIZE, addr < next); + psb_mmu_pt_unmap_unlock(pt); + + } while (addr = next, next != end); + + address += row_add; + } +out: + if (pd->hw_context != -1) + psb_mmu_flush_ptes(pd, f_address, num_pages, + desired_tile_stride, hw_tile_stride); + + up_read(&pd->driver->sem); + + if (pd->hw_context != -1) + psb_mmu_flush(pd->driver, 1); + + return ret; +} + +int psb_mmu_virtual_to_pfn(struct psb_mmu_pd *pd, uint32_t virtual, + unsigned long *pfn) +{ + int ret; + struct psb_mmu_pt *pt; + uint32_t tmp; + spinlock_t *lock = &pd->driver->lock; + + down_read(&pd->driver->sem); + pt = psb_mmu_pt_map_lock(pd, virtual); + if (!pt) { + uint32_t *v; + + spin_lock(lock); + v = kmap_atomic(pd->p, KM_USER0); + tmp = v[psb_mmu_pd_index(virtual)]; + kunmap_atomic(v, KM_USER0); + spin_unlock(lock); + + if (tmp != pd->invalid_pde || !(tmp & PSB_PTE_VALID) || + !(pd->invalid_pte & PSB_PTE_VALID)) { + ret = -EINVAL; + goto out; + } + ret = 0; + *pfn = pd->invalid_pte >> PAGE_SHIFT; + goto out; + } + tmp = pt->v[psb_mmu_pt_index(virtual)]; + if (!(tmp & PSB_PTE_VALID)) { + ret = -EINVAL; + } else { + ret = 0; + *pfn = tmp >> PAGE_SHIFT; + } + psb_mmu_pt_unmap_unlock(pt); +out: + up_read(&pd->driver->sem); + return ret; +}