/* * Copyright 2008 Advanced Micro Devices, Inc. * Copyright 2008 Red Hat Inc. * Copyright 2009 Jerome Glisse. * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included in * all copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR * OTHER DEALINGS IN THE SOFTWARE. * * Authors: Dave Airlie * Alex Deucher * Jerome Glisse
*/
/* * GPUVM * GPUVM is similar to the legacy gart on older asics, however * rather than there being a single global gart table * for the entire GPU, there are multiple VM page tables active * at any given time. The VM page tables can contain a mix * vram pages and system memory pages and system memory pages * can be mapped as snooped (cached system pages) or unsnooped * (uncached system pages). * Each VM has an ID associated with it and there is a page table * associated with each VMID. When execting a command buffer, * the kernel tells the ring what VMID to use for that command * buffer. VMIDs are allocated dynamically as commands are submitted. * The userspace drivers maintain their own address space and the kernel * sets up their pages tables accordingly when they submit their * command buffers and a VMID is assigned. * Cayman/Trinity support up to 8 active VMs at any given time; * SI supports 16.
*/
/** * radeon_vm_num_pdes - return the number of page directory entries * * @rdev: radeon_device pointer * * Calculate the number of page directory entries (cayman+).
*/ staticunsigned radeon_vm_num_pdes(struct radeon_device *rdev)
{ return rdev->vm_manager.max_pfn >> radeon_vm_block_size;
}
/** * radeon_vm_directory_size - returns the size of the page directory in bytes * * @rdev: radeon_device pointer * * Calculate the size of the page directory in bytes (cayman+).
*/ staticunsigned radeon_vm_directory_size(struct radeon_device *rdev)
{ return RADEON_GPU_PAGE_ALIGN(radeon_vm_num_pdes(rdev) * 8);
}
/** * radeon_vm_manager_init - init the vm manager * * @rdev: radeon_device pointer * * Init the vm manager (cayman+). * Returns 0 for success, error for failure.
*/ int radeon_vm_manager_init(struct radeon_device *rdev)
{ int r;
if (!rdev->vm_manager.enabled) {
r = radeon_asic_vm_init(rdev); if (r) return r;
rdev->vm_manager.enabled = true;
} return 0;
}
/** * radeon_vm_manager_fini - tear down the vm manager * * @rdev: radeon_device pointer * * Tear down the VM manager (cayman+).
*/ void radeon_vm_manager_fini(struct radeon_device *rdev)
{ int i;
if (!rdev->vm_manager.enabled) return;
for (i = 0; i < RADEON_NUM_VM; ++i)
radeon_fence_unref(&rdev->vm_manager.active[i]);
radeon_asic_vm_fini(rdev);
rdev->vm_manager.enabled = false;
}
/** * radeon_vm_get_bos - add the vm BOs to a validation list * * @rdev: radeon_device pointer * @vm: vm providing the BOs * @head: head of validation list * * Add the page directory to the list of BOs to * validate for command submission (cayman+).
*/ struct radeon_bo_list *radeon_vm_get_bos(struct radeon_device *rdev, struct radeon_vm *vm, struct list_head *head)
{ struct radeon_bo_list *list; unsigned i, idx;
list = kvmalloc_array(vm->max_pde_used + 2, sizeof(struct radeon_bo_list), GFP_KERNEL); if (!list) return NULL;
/* add the vm page table to the list */
list[0].robj = vm->page_directory;
list[0].preferred_domains = RADEON_GEM_DOMAIN_VRAM;
list[0].allowed_domains = RADEON_GEM_DOMAIN_VRAM;
list[0].shared = true;
list[0].tiling_flags = 0;
list_add(&list[0].list, head);
for (i = 0, idx = 1; i <= vm->max_pde_used; i++) { if (!vm->page_tables[i].bo) continue;
/** * radeon_vm_grab_id - allocate the next free VMID * * @rdev: radeon_device pointer * @vm: vm to allocate id for * @ring: ring we want to submit job to * * Allocate an id for the vm (cayman+). * Returns the fence we need to sync to (if any). * * Global and local mutex must be locked!
*/ struct radeon_fence *radeon_vm_grab_id(struct radeon_device *rdev, struct radeon_vm *vm, int ring)
{ struct radeon_fence *best[RADEON_NUM_RINGS] = {}; struct radeon_vm_id *vm_id = &vm->ids[ring];
unsigned choices[2] = {}; unsigned i;
/* check if the id is still valid */ if (vm_id->id && vm_id->last_id_use &&
vm_id->last_id_use == rdev->vm_manager.active[vm_id->id]) return NULL;
/* we definitely need to flush */
vm_id->pd_gpu_addr = ~0ll;
/* skip over VMID 0, since it is the system VM */ for (i = 1; i < rdev->vm_manager.nvm; ++i) { struct radeon_fence *fence = rdev->vm_manager.active[i];
if (fence == NULL) { /* found a free one */
vm_id->id = i;
trace_radeon_vm_grab_id(i, ring); return NULL;
}
if (radeon_fence_is_earlier(fence, best[fence->ring])) {
best[fence->ring] = fence;
choices[fence->ring == ring ? 0 : 1] = i;
}
}
for (i = 0; i < 2; ++i) { if (choices[i]) {
vm_id->id = choices[i];
trace_radeon_vm_grab_id(choices[i], ring); return rdev->vm_manager.active[choices[i]];
}
}
/* should never happen */
BUG(); return NULL;
}
/** * radeon_vm_flush - hardware flush the vm * * @rdev: radeon_device pointer * @vm: vm we want to flush * @ring: ring to use for flush * @updates: last vm update that is waited for * * Flush the vm (cayman+). * * Global and local mutex must be locked!
*/ void radeon_vm_flush(struct radeon_device *rdev, struct radeon_vm *vm, int ring, struct radeon_fence *updates)
{
uint64_t pd_addr = radeon_bo_gpu_offset(vm->page_directory); struct radeon_vm_id *vm_id = &vm->ids[ring];
if (pd_addr != vm_id->pd_gpu_addr || !vm_id->flushed_updates ||
radeon_fence_is_earlier(vm_id->flushed_updates, updates)) {
/** * radeon_vm_fence - remember fence for vm * * @rdev: radeon_device pointer * @vm: vm we want to fence * @fence: fence to remember * * Fence the vm (cayman+). * Set the fence used to protect page table and id. * * Global and local mutex must be locked!
*/ void radeon_vm_fence(struct radeon_device *rdev, struct radeon_vm *vm, struct radeon_fence *fence)
{ unsigned vm_id = vm->ids[fence->ring].id;
/** * radeon_vm_bo_find - find the bo_va for a specific vm & bo * * @vm: requested vm * @bo: requested buffer object * * Find @bo inside the requested vm (cayman+). * Search inside the @bos vm list for the requested vm * Returns the found bo_va or NULL if none is found * * Object has to be reserved!
*/ struct radeon_bo_va *radeon_vm_bo_find(struct radeon_vm *vm, struct radeon_bo *bo)
{ struct radeon_bo_va *bo_va;
list_for_each_entry(bo_va, &bo->va, bo_list) { if (bo_va->vm == vm) return bo_va;
} return NULL;
}
/** * radeon_vm_bo_add - add a bo to a specific vm * * @rdev: radeon_device pointer * @vm: requested vm * @bo: radeon buffer object * * Add @bo into the requested vm (cayman+). * Add @bo to the list of bos associated with the vm * Returns newly added bo_va or NULL for failure * * Object has to be reserved!
*/ struct radeon_bo_va *radeon_vm_bo_add(struct radeon_device *rdev, struct radeon_vm *vm, struct radeon_bo *bo)
{ struct radeon_bo_va *bo_va;
/** * radeon_vm_set_pages - helper to call the right asic function * * @rdev: radeon_device pointer * @ib: indirect buffer to fill with commands * @pe: addr of the page entry * @addr: dst addr to write into pe * @count: number of page entries to update * @incr: increase next addr by incr bytes * @flags: hw access flags * * Traces the parameters and calls the right asic functions * to setup the page table using the DMA.
*/ staticvoid radeon_vm_set_pages(struct radeon_device *rdev, struct radeon_ib *ib,
uint64_t pe,
uint64_t addr, unsigned count,
uint32_t incr, uint32_t flags)
{
trace_radeon_vm_set_page(pe, addr, count, incr, flags);
/** * radeon_vm_bo_set_addr - set bos virtual address inside a vm * * @rdev: radeon_device pointer * @bo_va: bo_va to store the address * @soffset: requested offset of the buffer in the VM address space * @flags: attributes of pages (read/write/valid/etc.) * * Set offset of @bo_va (cayman+). * Validate and set the offset requested within the vm address space. * Returns 0 for success, error for failure. * * Object has to be reserved and gets unreserved by this function!
*/ int radeon_vm_bo_set_addr(struct radeon_device *rdev, struct radeon_bo_va *bo_va,
uint64_t soffset,
uint32_t flags)
{
uint64_t size = radeon_bo_size(bo_va->bo); struct radeon_vm *vm = bo_va->vm; unsigned last_pfn, pt_idx;
uint64_t eoffset; int r;
if (soffset) { /* make sure object fit at this offset */
eoffset = soffset + size - 1; if (soffset >= eoffset) {
r = -EINVAL; goto error_unreserve;
}
if (eoffset > vm->max_pde_used)
vm->max_pde_used = eoffset;
radeon_bo_unreserve(bo_va->bo);
/* walk over the address space and allocate the page tables */ for (pt_idx = soffset; pt_idx <= eoffset; ++pt_idx) { struct radeon_bo *pt;
if (vm->page_tables[pt_idx].bo) continue;
/* drop mutex to allocate and clear page table */
mutex_unlock(&vm->mutex);
r = radeon_bo_create(rdev, RADEON_VM_PTE_COUNT * 8,
RADEON_GPU_PAGE_SIZE, true,
RADEON_GEM_DOMAIN_VRAM, 0,
NULL, NULL, &pt); if (r) return r;
r = radeon_vm_clear_bo(rdev, pt); if (r) {
radeon_bo_unref(&pt); return r;
}
/* aquire mutex again */
mutex_lock(&vm->mutex); if (vm->page_tables[pt_idx].bo) { /* someone else allocated the pt in the meantime */
mutex_unlock(&vm->mutex);
radeon_bo_unref(&pt);
mutex_lock(&vm->mutex); continue;
}
/** * radeon_vm_map_gart - get the physical address of a gart page * * @rdev: radeon_device pointer * @addr: the unmapped addr * * Look up the physical address of the page that the pte resolves * to (cayman+). * Returns the physical address of the page.
*/
uint64_t radeon_vm_map_gart(struct radeon_device *rdev, uint64_t addr)
{
uint64_t result;
/* page table offset */
result = rdev->gart.pages_entry[addr >> RADEON_GPU_PAGE_SHIFT];
result &= ~RADEON_GPU_PAGE_MASK;
return result;
}
/** * radeon_vm_page_flags - translate page flags to what the hw uses * * @flags: flags comming from userspace * * Translate the flags the userspace ABI uses to hw flags.
*/ static uint32_t radeon_vm_page_flags(uint32_t flags)
{
uint32_t hw_flags = 0;
/** * radeon_vm_update_page_directory - make sure that page directory is valid * * @rdev: radeon_device pointer * @vm: requested vm * * Allocates new page tables if necessary * and updates the page directory (cayman+). * Returns 0 for success, error for failure. * * Global and local mutex must be locked!
*/ int radeon_vm_update_page_directory(struct radeon_device *rdev, struct radeon_vm *vm)
{ struct radeon_bo *pd = vm->page_directory;
uint64_t pd_addr = radeon_bo_gpu_offset(pd);
uint32_t incr = RADEON_VM_PTE_COUNT * 8;
uint64_t last_pde = ~0, last_pt = ~0; unsigned count = 0, pt_idx, ndw; struct radeon_ib ib; int r;
/* padding, etc. */
ndw = 64;
/* assume the worst case */
ndw += vm->max_pde_used * 6;
/* update too big for an IB */ if (ndw > 0xfffff) return -ENOMEM;
r = radeon_ib_get(rdev, R600_RING_TYPE_DMA_INDEX, &ib, NULL, ndw * 4); if (r) return r;
ib.length_dw = 0;
/* walk over the address space and update the page directory */ for (pt_idx = 0; pt_idx <= vm->max_pde_used; ++pt_idx) { struct radeon_bo *bo = vm->page_tables[pt_idx].bo;
uint64_t pde, pt;
/** * radeon_vm_frag_ptes - add fragment information to PTEs * * @rdev: radeon_device pointer * @ib: IB for the update * @pe_start: first PTE to handle * @pe_end: last PTE to handle * @addr: addr those PTEs should point to * @flags: hw mapping flags * * Global and local mutex must be locked!
*/ staticvoid radeon_vm_frag_ptes(struct radeon_device *rdev, struct radeon_ib *ib,
uint64_t pe_start, uint64_t pe_end,
uint64_t addr, uint32_t flags)
{ /** * The MC L1 TLB supports variable sized pages, based on a fragment * field in the PTE. When this field is set to a non-zero value, page * granularity is increased from 4KB to (1 << (12 + frag)). The PTE * flags are considered valid for all PTEs within the fragment range * and corresponding mappings are assumed to be physically contiguous. * * The L1 TLB can store a single PTE for the whole fragment, * significantly increasing the space available for translation * caching. This leads to large improvements in throughput when the * TLB is under pressure. * * The L2 TLB distributes small and large fragments into two * asymmetric partitions. The large fragment cache is significantly * larger. Thus, we try to use large fragments wherever possible. * Userspace can support this by aligning virtual base address and * allocation size to the fragment size.
*/
/* NI is optimized for 256KB fragments, SI and newer for 64KB */
uint64_t frag_flags = ((rdev->family == CHIP_CAYMAN) ||
(rdev->family == CHIP_ARUBA)) ?
R600_PTE_FRAG_256KB : R600_PTE_FRAG_64KB;
uint64_t frag_align = ((rdev->family == CHIP_CAYMAN) ||
(rdev->family == CHIP_ARUBA)) ? 0x200 : 0x80;
/* handle the 4K area at the beginning */ if (pe_start != frag_start) {
count = (frag_start - pe_start) / 8;
radeon_vm_set_pages(rdev, ib, pe_start, addr, count,
RADEON_GPU_PAGE_SIZE, flags);
addr += RADEON_GPU_PAGE_SIZE * count;
}
/* handle the area in the middle */
count = (frag_end - frag_start) / 8;
radeon_vm_set_pages(rdev, ib, frag_start, addr, count,
RADEON_GPU_PAGE_SIZE, flags | frag_flags);
/* handle the 4K area at the end */ if (frag_end != pe_end) {
addr += RADEON_GPU_PAGE_SIZE * count;
count = (pe_end - frag_end) / 8;
radeon_vm_set_pages(rdev, ib, frag_end, addr, count,
RADEON_GPU_PAGE_SIZE, flags);
}
}
/** * radeon_vm_update_ptes - make sure that page tables are valid * * @rdev: radeon_device pointer * @vm: requested vm * @ib: indirect buffer to use for the update * @start: start of GPU address range * @end: end of GPU address range * @dst: destination address to map to * @flags: mapping flags * * Update the page tables in the range @start - @end (cayman+). * * Global and local mutex must be locked!
*/ staticint radeon_vm_update_ptes(struct radeon_device *rdev, struct radeon_vm *vm, struct radeon_ib *ib,
uint64_t start, uint64_t end,
uint64_t dst, uint32_t flags)
{
uint64_t mask = RADEON_VM_PTE_COUNT - 1;
uint64_t last_pte = ~0, last_dst = ~0; unsigned count = 0;
uint64_t addr;
/* walk over the address space and update the page tables */ for (addr = start; addr < end; ) {
uint64_t pt_idx = addr >> radeon_vm_block_size; struct radeon_bo *pt = vm->page_tables[pt_idx].bo; unsigned nptes;
uint64_t pte; int r;
radeon_sync_resv(rdev, &ib->sync, pt->tbo.base.resv, true);
r = dma_resv_reserve_fences(pt->tbo.base.resv, 1); if (r) return r;
/** * radeon_vm_fence_pts - fence page tables after an update * * @vm: requested vm * @start: start of GPU address range * @end: end of GPU address range * @fence: fence to use * * Fence the page tables in the range @start - @end (cayman+). * * Global and local mutex must be locked!
*/ staticvoid radeon_vm_fence_pts(struct radeon_vm *vm,
uint64_t start, uint64_t end, struct radeon_fence *fence)
{ unsigned i;
start >>= radeon_vm_block_size;
end = (end - 1) >> radeon_vm_block_size;
for (i = start; i <= end; ++i)
radeon_bo_fence(vm->page_tables[i].bo, fence, true);
}
/** * radeon_vm_bo_update - map a bo into the vm page table * * @rdev: radeon_device pointer * @bo_va: radeon buffer virtual address object * @mem: ttm mem * * Fill in the page table entries for @bo (cayman+). * Returns 0 for success, -EINVAL for failure. * * Object have to be reserved and mutex must be locked!
*/ int radeon_vm_bo_update(struct radeon_device *rdev, struct radeon_bo_va *bo_va, struct ttm_resource *mem)
{ struct radeon_vm *vm = bo_va->vm; struct radeon_ib ib; unsigned nptes, ncmds, ndw;
uint64_t addr;
uint32_t flags; int r;
if (!bo_va->it.start) {
dev_err(rdev->dev, "bo %p don't has a mapping in vm %p\n",
bo_va->bo, vm); return -EINVAL;
}
spin_lock(&vm->status_lock); if (mem) { if (list_empty(&bo_va->vm_status)) {
spin_unlock(&vm->status_lock); return 0;
}
list_del_init(&bo_va->vm_status);
} else {
list_del(&bo_va->vm_status);
list_add(&bo_va->vm_status, &vm->cleared);
}
spin_unlock(&vm->status_lock);
/* reserve space for one command every (1 << BLOCK_SIZE) entries
or 2k dwords (whatever is smaller) */
ncmds = (nptes >> min(radeon_vm_block_size, 11)) + 1;
/** * radeon_vm_clear_freed - clear freed BOs in the PT * * @rdev: radeon_device pointer * @vm: requested vm * * Make sure all freed BOs are cleared in the PT. * Returns 0 for success. * * PTs have to be reserved and mutex must be locked!
*/ int radeon_vm_clear_freed(struct radeon_device *rdev, struct radeon_vm *vm)
{ struct radeon_bo_va *bo_va; int r = 0;
r = radeon_vm_bo_update(rdev, bo_va, NULL);
radeon_bo_unref(&bo_va->bo);
radeon_fence_unref(&bo_va->last_pt_update);
spin_lock(&vm->status_lock);
list_del(&bo_va->vm_status);
kfree(bo_va); if (r) break;
}
spin_unlock(&vm->status_lock); return r;
}
/** * radeon_vm_clear_invalids - clear invalidated BOs in the PT * * @rdev: radeon_device pointer * @vm: requested vm * * Make sure all invalidated BOs are cleared in the PT. * Returns 0 for success. * * PTs have to be reserved and mutex must be locked!
*/ int radeon_vm_clear_invalids(struct radeon_device *rdev, struct radeon_vm *vm)
{ struct radeon_bo_va *bo_va; int r;
/** * radeon_vm_bo_rmv - remove a bo to a specific vm * * @rdev: radeon_device pointer * @bo_va: requested bo_va * * Remove @bo_va->bo from the requested vm (cayman+). * * Object have to be reserved!
*/ void radeon_vm_bo_rmv(struct radeon_device *rdev, struct radeon_bo_va *bo_va)
{ struct radeon_vm *vm = bo_va->vm;
list_del(&bo_va->bo_list);
mutex_lock(&vm->mutex); if (bo_va->it.start || bo_va->it.last)
interval_tree_remove(&bo_va->it, &vm->va);
r = radeon_bo_create(rdev, pd_size, align, true,
RADEON_GEM_DOMAIN_VRAM, 0, NULL,
NULL, &vm->page_directory); if (r) {
kfree(vm->page_tables);
vm->page_tables = NULL; return r;
}
r = radeon_vm_clear_bo(rdev, vm->page_directory); if (r) {
radeon_bo_unref(&vm->page_directory);
vm->page_directory = NULL;
kfree(vm->page_tables);
vm->page_tables = NULL; return r;
}
return 0;
}
/** * radeon_vm_fini - tear down a vm instance * * @rdev: radeon_device pointer * @vm: requested vm * * Tear down @vm (cayman+). * Unbind the VM and remove all bos from the vm bo list
*/ void radeon_vm_fini(struct radeon_device *rdev, struct radeon_vm *vm)
{ struct radeon_bo_va *bo_va, *tmp; int i, r;
if (!RB_EMPTY_ROOT(&vm->va.rb_root))
dev_err(rdev->dev, "still active bo inside vm\n");
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.