2 * Copyright 2018 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
27 #include <linux/io-64-nonatomic-lo-hi.h>
29 #include <asm/hypervisor.h>
33 #include "amdgpu_gmc.h"
34 #include "amdgpu_ras.h"
35 #include "amdgpu_reset.h"
36 #include "amdgpu_xgmi.h"
38 #include <drm/drm_drv.h>
39 #include <drm/ttm/ttm_tt.h>
42 * amdgpu_gmc_pdb0_alloc - allocate vram for pdb0
44 * @adev: amdgpu_device pointer
46 * Allocate video memory for pdb0 and map it for CPU access
47 * Returns 0 for success, error for failure.
49 int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev)
52 struct amdgpu_bo_param bp;
53 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
54 uint32_t pde0_page_shift = adev->gmc.vmid0_page_table_block_size + 21;
55 uint32_t npdes = (vram_size + (1ULL << pde0_page_shift) - 1) >> pde0_page_shift;
57 memset(&bp, 0, sizeof(bp));
58 bp.size = PAGE_ALIGN((npdes + 1) * 8);
59 bp.byte_align = PAGE_SIZE;
60 bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
61 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
62 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
63 bp.type = ttm_bo_type_kernel;
65 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
67 r = amdgpu_bo_create(adev, &bp, &adev->gmc.pdb0_bo);
71 r = amdgpu_bo_reserve(adev->gmc.pdb0_bo, false);
73 goto bo_reserve_failure;
75 r = amdgpu_bo_pin(adev->gmc.pdb0_bo, AMDGPU_GEM_DOMAIN_VRAM);
78 r = amdgpu_bo_kmap(adev->gmc.pdb0_bo, &adev->gmc.ptr_pdb0);
82 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
86 amdgpu_bo_unpin(adev->gmc.pdb0_bo);
88 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
90 amdgpu_bo_unref(&adev->gmc.pdb0_bo);
95 * amdgpu_gmc_get_pde_for_bo - get the PDE for a BO
97 * @bo: the BO to get the PDE for
98 * @level: the level in the PD hirarchy
99 * @addr: resulting addr
100 * @flags: resulting flags
102 * Get the address and flags to be used for a PDE (Page Directory Entry).
104 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
105 uint64_t *addr, uint64_t *flags)
107 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
109 switch (bo->tbo.resource->mem_type) {
111 *addr = bo->tbo.ttm->dma_address[0];
114 *addr = amdgpu_bo_gpu_offset(bo);
120 *flags = amdgpu_ttm_tt_pde_flags(bo->tbo.ttm, bo->tbo.resource);
121 amdgpu_gmc_get_vm_pde(adev, level, addr, flags);
125 * amdgpu_gmc_pd_addr - return the address of the root directory
127 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo)
129 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
132 /* TODO: move that into ASIC specific code */
133 if (adev->asic_type >= CHIP_VEGA10) {
134 uint64_t flags = AMDGPU_PTE_VALID;
136 amdgpu_gmc_get_pde_for_bo(bo, -1, &pd_addr, &flags);
139 pd_addr = amdgpu_bo_gpu_offset(bo);
145 * amdgpu_gmc_set_pte_pde - update the page tables using CPU
147 * @adev: amdgpu_device pointer
148 * @cpu_pt_addr: cpu address of the page table
149 * @gpu_page_idx: entry in the page table to update
150 * @addr: dst addr to write into pte/pde
151 * @flags: access flags
153 * Update the page tables using CPU.
155 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
156 uint32_t gpu_page_idx, uint64_t addr,
159 void __iomem *ptr = (void *)cpu_pt_addr;
163 * The following is for PTE only. GART does not have PDEs.
165 value = addr & 0x0000FFFFFFFFF000ULL;
167 writeq(value, ptr + (gpu_page_idx * 8));
173 * amdgpu_gmc_agp_addr - return the address in the AGP address space
175 * @bo: TTM BO which needs the address, must be in GTT domain
177 * Tries to figure out how to access the BO through the AGP aperture. Returns
178 * AMDGPU_BO_INVALID_OFFSET if that is not possible.
180 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo)
182 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
185 return AMDGPU_BO_INVALID_OFFSET;
187 if (bo->ttm->num_pages != 1 || bo->ttm->caching == ttm_cached)
188 return AMDGPU_BO_INVALID_OFFSET;
190 if (bo->ttm->dma_address[0] + PAGE_SIZE >= adev->gmc.agp_size)
191 return AMDGPU_BO_INVALID_OFFSET;
193 return adev->gmc.agp_start + bo->ttm->dma_address[0];
197 * amdgpu_gmc_vram_location - try to find VRAM location
199 * @adev: amdgpu device structure holding all necessary information
200 * @mc: memory controller structure holding memory information
201 * @base: base address at which to put VRAM
203 * Function will try to place VRAM at base address provided
206 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
209 uint64_t vis_limit = (uint64_t)amdgpu_vis_vram_limit << 20;
210 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
212 mc->vram_start = base;
213 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
214 if (limit < mc->real_vram_size)
215 mc->real_vram_size = limit;
217 if (vis_limit && vis_limit < mc->visible_vram_size)
218 mc->visible_vram_size = vis_limit;
220 if (mc->real_vram_size < mc->visible_vram_size)
221 mc->visible_vram_size = mc->real_vram_size;
223 if (mc->xgmi.num_physical_nodes == 0) {
224 mc->fb_start = mc->vram_start;
225 mc->fb_end = mc->vram_end;
227 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
228 mc->mc_vram_size >> 20, mc->vram_start,
229 mc->vram_end, mc->real_vram_size >> 20);
232 /** amdgpu_gmc_sysvm_location - place vram and gart in sysvm aperture
234 * @adev: amdgpu device structure holding all necessary information
235 * @mc: memory controller structure holding memory information
237 * This function is only used if use GART for FB translation. In such
238 * case, we use sysvm aperture (vmid0 page tables) for both vram
239 * and gart (aka system memory) access.
241 * GPUVM (and our organization of vmid0 page tables) require sysvm
242 * aperture to be placed at a location aligned with 8 times of native
243 * page size. For example, if vm_context0_cntl.page_table_block_size
244 * is 12, then native page size is 8G (2M*2^12), sysvm should start
245 * with a 64G aligned address. For simplicity, we just put sysvm at
246 * address 0. So vram start at address 0 and gart is right after vram.
248 void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
250 u64 hive_vram_start = 0;
251 u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes - 1;
252 mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id;
253 mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1;
254 mc->gart_start = hive_vram_end + 1;
255 mc->gart_end = mc->gart_start + mc->gart_size - 1;
256 mc->fb_start = hive_vram_start;
257 mc->fb_end = hive_vram_end;
258 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
259 mc->mc_vram_size >> 20, mc->vram_start,
260 mc->vram_end, mc->real_vram_size >> 20);
261 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
262 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
266 * amdgpu_gmc_gart_location - try to find GART location
268 * @adev: amdgpu device structure holding all necessary information
269 * @mc: memory controller structure holding memory information
270 * @gart_placement: GART placement policy with respect to VRAM
272 * Function will place try to place GART before or after VRAM.
273 * If GART size is bigger than space left then we ajust GART size.
274 * Thus function will never fails.
276 void amdgpu_gmc_gart_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
277 enum amdgpu_gart_placement gart_placement)
279 const uint64_t four_gb = 0x100000000ULL;
280 u64 size_af, size_bf;
281 /*To avoid the hole, limit the max mc address to AMDGPU_GMC_HOLE_START*/
282 u64 max_mc_address = min(adev->gmc.mc_mask, AMDGPU_GMC_HOLE_START - 1);
284 /* VCE doesn't like it when BOs cross a 4GB segment, so align
285 * the GART base on a 4GB boundary as well.
287 size_bf = mc->fb_start;
288 size_af = max_mc_address + 1 - ALIGN(mc->fb_end + 1, four_gb);
290 if (mc->gart_size > max(size_bf, size_af)) {
291 dev_warn(adev->dev, "limiting GART\n");
292 mc->gart_size = max(size_bf, size_af);
295 switch (gart_placement) {
296 case AMDGPU_GART_PLACEMENT_HIGH:
297 mc->gart_start = max_mc_address - mc->gart_size + 1;
299 case AMDGPU_GART_PLACEMENT_LOW:
302 case AMDGPU_GART_PLACEMENT_BEST_FIT:
304 if ((size_bf >= mc->gart_size && size_bf < size_af) ||
305 (size_af < mc->gart_size))
308 mc->gart_start = max_mc_address - mc->gart_size + 1;
312 mc->gart_start &= ~(four_gb - 1);
313 mc->gart_end = mc->gart_start + mc->gart_size - 1;
314 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
315 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
319 * amdgpu_gmc_agp_location - try to find AGP location
320 * @adev: amdgpu device structure holding all necessary information
321 * @mc: memory controller structure holding memory information
323 * Function will place try to find a place for the AGP BAR in the MC address
326 * AGP BAR will be assigned the largest available hole in the address space.
327 * Should be called after VRAM and GART locations are setup.
329 void amdgpu_gmc_agp_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
331 const uint64_t sixteen_gb = 1ULL << 34;
332 const uint64_t sixteen_gb_mask = ~(sixteen_gb - 1);
333 u64 size_af, size_bf;
335 if (mc->fb_start > mc->gart_start) {
336 size_bf = (mc->fb_start & sixteen_gb_mask) -
337 ALIGN(mc->gart_end + 1, sixteen_gb);
338 size_af = mc->mc_mask + 1 - ALIGN(mc->fb_end + 1, sixteen_gb);
340 size_bf = mc->fb_start & sixteen_gb_mask;
341 size_af = (mc->gart_start & sixteen_gb_mask) -
342 ALIGN(mc->fb_end + 1, sixteen_gb);
345 if (size_bf > size_af) {
346 mc->agp_start = (mc->fb_start - size_bf) & sixteen_gb_mask;
347 mc->agp_size = size_bf;
349 mc->agp_start = ALIGN(mc->fb_end + 1, sixteen_gb);
350 mc->agp_size = size_af;
353 mc->agp_end = mc->agp_start + mc->agp_size - 1;
354 dev_info(adev->dev, "AGP: %lluM 0x%016llX - 0x%016llX\n",
355 mc->agp_size >> 20, mc->agp_start, mc->agp_end);
359 * amdgpu_gmc_set_agp_default - Set the default AGP aperture value.
360 * @adev: amdgpu device structure holding all necessary information
361 * @mc: memory controller structure holding memory information
363 * To disable the AGP aperture, you need to set the start to a larger
364 * value than the end. This function sets the default value which
365 * can then be overridden using amdgpu_gmc_agp_location() if you want
366 * to enable the AGP aperture on a specific chip.
369 void amdgpu_gmc_set_agp_default(struct amdgpu_device *adev,
370 struct amdgpu_gmc *mc)
372 mc->agp_start = 0xffffffffffff;
378 * amdgpu_gmc_fault_key - get hask key from vm fault address and pasid
380 * @addr: 48 bit physical address, page aligned (36 significant bits)
381 * @pasid: 16 bit process address space identifier
383 static inline uint64_t amdgpu_gmc_fault_key(uint64_t addr, uint16_t pasid)
385 return addr << 4 | pasid;
389 * amdgpu_gmc_filter_faults - filter VM faults
391 * @adev: amdgpu device structure
392 * @ih: interrupt ring that the fault received from
393 * @addr: address of the VM fault
394 * @pasid: PASID of the process causing the fault
395 * @timestamp: timestamp of the fault
398 * True if the fault was filtered and should not be processed further.
399 * False if the fault is a new one and needs to be handled.
401 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
402 struct amdgpu_ih_ring *ih, uint64_t addr,
403 uint16_t pasid, uint64_t timestamp)
405 struct amdgpu_gmc *gmc = &adev->gmc;
406 uint64_t stamp, key = amdgpu_gmc_fault_key(addr, pasid);
407 struct amdgpu_gmc_fault *fault;
410 /* Stale retry fault if timestamp goes backward */
411 if (amdgpu_ih_ts_after(timestamp, ih->processed_timestamp))
414 /* If we don't have space left in the ring buffer return immediately */
415 stamp = max(timestamp, AMDGPU_GMC_FAULT_TIMEOUT + 1) -
416 AMDGPU_GMC_FAULT_TIMEOUT;
417 if (gmc->fault_ring[gmc->last_fault].timestamp >= stamp)
420 /* Try to find the fault in the hash */
421 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
422 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
423 while (fault->timestamp >= stamp) {
426 if (atomic64_read(&fault->key) == key) {
428 * if we get a fault which is already present in
429 * the fault_ring and the timestamp of
430 * the fault is after the expired timestamp,
431 * then this is a new fault that needs to be added
432 * into the fault ring.
434 if (fault->timestamp_expiry != 0 &&
435 amdgpu_ih_ts_after(fault->timestamp_expiry,
442 tmp = fault->timestamp;
443 fault = &gmc->fault_ring[fault->next];
445 /* Check if the entry was reused */
446 if (fault->timestamp >= tmp)
450 /* Add the fault to the ring */
451 fault = &gmc->fault_ring[gmc->last_fault];
452 atomic64_set(&fault->key, key);
453 fault->timestamp = timestamp;
455 /* And update the hash */
456 fault->next = gmc->fault_hash[hash].idx;
457 gmc->fault_hash[hash].idx = gmc->last_fault++;
462 * amdgpu_gmc_filter_faults_remove - remove address from VM faults filter
464 * @adev: amdgpu device structure
465 * @addr: address of the VM fault
466 * @pasid: PASID of the process causing the fault
468 * Remove the address from fault filter, then future vm fault on this address
469 * will pass to retry fault handler to recover.
471 void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
474 struct amdgpu_gmc *gmc = &adev->gmc;
475 uint64_t key = amdgpu_gmc_fault_key(addr, pasid);
476 struct amdgpu_ih_ring *ih;
477 struct amdgpu_gmc_fault *fault;
483 if (adev->irq.retry_cam_enabled)
487 /* Get the WPTR of the last entry in IH ring */
488 last_wptr = amdgpu_ih_get_wptr(adev, ih);
489 /* Order wptr with ring data. */
491 /* Get the timetamp of the last entry in IH ring */
492 last_ts = amdgpu_ih_decode_iv_ts(adev, ih, last_wptr, -1);
494 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
495 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
497 if (atomic64_read(&fault->key) == key) {
499 * Update the timestamp when this fault
502 fault->timestamp_expiry = last_ts;
506 tmp = fault->timestamp;
507 fault = &gmc->fault_ring[fault->next];
508 } while (fault->timestamp < tmp);
511 int amdgpu_gmc_ras_sw_init(struct amdgpu_device *adev)
516 r = amdgpu_umc_ras_sw_init(adev);
520 /* mmhub ras block */
521 r = amdgpu_mmhub_ras_sw_init(adev);
526 r = amdgpu_hdp_ras_sw_init(adev);
530 /* mca.x ras block */
531 r = amdgpu_mca_mp0_ras_sw_init(adev);
535 r = amdgpu_mca_mp1_ras_sw_init(adev);
539 r = amdgpu_mca_mpio_ras_sw_init(adev);
544 r = amdgpu_xgmi_ras_sw_init(adev);
551 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
556 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev)
562 * The latest engine allocation on gfx9/10 is:
563 * Engine 2, 3: firmware
564 * Engine 0, 1, 4~16: amdgpu ring,
565 * subject to change when ring number changes
566 * Engine 17: Gart flushes
568 #define AMDGPU_VMHUB_INV_ENG_BITMAP 0x1FFF3
570 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev)
572 struct amdgpu_ring *ring;
573 unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] = {0};
575 unsigned vmhub, inv_eng;
577 /* init the vm inv eng for all vmhubs */
578 for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) {
579 vm_inv_engs[i] = AMDGPU_VMHUB_INV_ENG_BITMAP;
580 /* reserve engine 5 for firmware */
581 if (adev->enable_mes)
582 vm_inv_engs[i] &= ~(1 << 5);
583 /* reserve mmhub engine 3 for firmware */
584 if (adev->enable_umsch_mm)
585 vm_inv_engs[i] &= ~(1 << 3);
588 for (i = 0; i < adev->num_rings; ++i) {
589 ring = adev->rings[i];
590 vmhub = ring->vm_hub;
592 if (ring == &adev->mes.ring ||
593 ring == &adev->umsch_mm.ring)
596 inv_eng = ffs(vm_inv_engs[vmhub]);
598 dev_err(adev->dev, "no VM inv eng for ring %s\n",
603 ring->vm_inv_eng = inv_eng - 1;
604 vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng);
606 dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n",
607 ring->name, ring->vm_inv_eng, ring->vm_hub);
613 void amdgpu_gmc_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
614 uint32_t vmhub, uint32_t flush_type)
616 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
617 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
618 struct dma_fence *fence;
619 struct amdgpu_job *job;
622 if (!hub->sdma_invalidation_workaround || vmid ||
623 !adev->mman.buffer_funcs_enabled || !adev->ib_pool_ready ||
624 !ring->sched.ready) {
626 * A GPU reset should flush all TLBs anyway, so no need to do
627 * this while one is ongoing.
629 if (!down_read_trylock(&adev->reset_domain->sem))
632 if (adev->gmc.flush_tlb_needs_extra_type_2)
633 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid,
636 if (adev->gmc.flush_tlb_needs_extra_type_0 && flush_type == 2)
637 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid,
640 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid, vmhub,
642 up_read(&adev->reset_domain->sem);
646 /* The SDMA on Navi 1x has a bug which can theoretically result in memory
647 * corruption if an invalidation happens at the same time as an VA
648 * translation. Avoid this by doing the invalidation from the SDMA
649 * itself at least for GART.
651 mutex_lock(&adev->mman.gtt_window_lock);
652 r = amdgpu_job_alloc_with_ib(ring->adev, &adev->mman.high_pr,
653 AMDGPU_FENCE_OWNER_UNDEFINED,
654 16 * 4, AMDGPU_IB_POOL_IMMEDIATE,
659 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
660 job->vm_needs_flush = true;
661 job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop;
662 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
663 fence = amdgpu_job_submit(job);
664 mutex_unlock(&adev->mman.gtt_window_lock);
666 dma_fence_wait(fence, false);
667 dma_fence_put(fence);
672 mutex_unlock(&adev->mman.gtt_window_lock);
673 dev_err(adev->dev, "Error flushing GPU TLB using the SDMA (%d)!\n", r);
676 int amdgpu_gmc_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid,
677 uint32_t flush_type, bool all_hub,
680 u32 usec_timeout = amdgpu_sriov_vf(adev) ? SRIOV_USEC_TIMEOUT :
682 struct amdgpu_ring *ring = &adev->gfx.kiq[inst].ring;
683 struct amdgpu_kiq *kiq = &adev->gfx.kiq[inst];
689 * A GPU reset should flush all TLBs anyway, so no need to do
690 * this while one is ongoing.
692 if (!down_read_trylock(&adev->reset_domain->sem))
695 if (!adev->gmc.flush_pasid_uses_kiq || !ring->sched.ready) {
696 if (adev->gmc.flush_tlb_needs_extra_type_2)
697 adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid,
701 if (adev->gmc.flush_tlb_needs_extra_type_0 && flush_type == 2)
702 adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid,
706 adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid,
711 /* 2 dwords flush + 8 dwords fence */
712 ndw = kiq->pmf->invalidate_tlbs_size + 8;
714 if (adev->gmc.flush_tlb_needs_extra_type_2)
715 ndw += kiq->pmf->invalidate_tlbs_size;
717 if (adev->gmc.flush_tlb_needs_extra_type_0)
718 ndw += kiq->pmf->invalidate_tlbs_size;
720 spin_lock(&adev->gfx.kiq[inst].ring_lock);
721 r = amdgpu_ring_alloc(ring, ndw);
723 spin_unlock(&adev->gfx.kiq[inst].ring_lock);
724 goto error_unlock_reset;
726 if (adev->gmc.flush_tlb_needs_extra_type_2)
727 kiq->pmf->kiq_invalidate_tlbs(ring, pasid, 2, all_hub);
729 if (flush_type == 2 && adev->gmc.flush_tlb_needs_extra_type_0)
730 kiq->pmf->kiq_invalidate_tlbs(ring, pasid, 0, all_hub);
732 kiq->pmf->kiq_invalidate_tlbs(ring, pasid, flush_type, all_hub);
733 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
735 amdgpu_ring_undo(ring);
736 spin_unlock(&adev->gfx.kiq[inst].ring_lock);
737 goto error_unlock_reset;
740 amdgpu_ring_commit(ring);
741 spin_unlock(&adev->gfx.kiq[inst].ring_lock);
742 if (amdgpu_fence_wait_polling(ring, seq, usec_timeout) < 1) {
743 dev_err(adev->dev, "timeout waiting for kiq fence\n");
749 up_read(&adev->reset_domain->sem);
753 void amdgpu_gmc_fw_reg_write_reg_wait(struct amdgpu_device *adev,
754 uint32_t reg0, uint32_t reg1,
755 uint32_t ref, uint32_t mask,
758 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_inst];
759 struct amdgpu_ring *ring = &kiq->ring;
760 signed long r, cnt = 0;
764 if (adev->mes.ring.sched.ready) {
765 amdgpu_mes_reg_write_reg_wait(adev, reg0, reg1,
770 spin_lock_irqsave(&kiq->ring_lock, flags);
771 amdgpu_ring_alloc(ring, 32);
772 amdgpu_ring_emit_reg_write_reg_wait(ring, reg0, reg1,
774 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
778 amdgpu_ring_commit(ring);
779 spin_unlock_irqrestore(&kiq->ring_lock, flags);
781 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
783 /* don't wait anymore for IRQ context */
784 if (r < 1 && in_interrupt())
788 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
790 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
791 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
794 if (cnt > MAX_KIQ_REG_TRY)
800 amdgpu_ring_undo(ring);
801 spin_unlock_irqrestore(&kiq->ring_lock, flags);
803 dev_err(adev->dev, "failed to write reg %x wait reg %x\n", reg0, reg1);
807 * amdgpu_gmc_tmz_set -- check and set if a device supports TMZ
808 * @adev: amdgpu_device pointer
810 * Check and set if an the device @adev supports Trusted Memory
813 void amdgpu_gmc_tmz_set(struct amdgpu_device *adev)
815 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
817 case IP_VERSION(9, 2, 2):
818 case IP_VERSION(9, 1, 0):
819 /* RENOIR looks like RAVEN */
820 case IP_VERSION(9, 3, 0):
822 case IP_VERSION(10, 3, 7):
824 case IP_VERSION(11, 0, 1):
825 if (amdgpu_tmz == 0) {
826 adev->gmc.tmz_enabled = false;
828 "Trusted Memory Zone (TMZ) feature disabled (cmd line)\n");
830 adev->gmc.tmz_enabled = true;
832 "Trusted Memory Zone (TMZ) feature enabled\n");
835 case IP_VERSION(10, 1, 10):
836 case IP_VERSION(10, 1, 1):
837 case IP_VERSION(10, 1, 2):
838 case IP_VERSION(10, 1, 3):
839 case IP_VERSION(10, 3, 0):
840 case IP_VERSION(10, 3, 2):
841 case IP_VERSION(10, 3, 4):
842 case IP_VERSION(10, 3, 5):
843 case IP_VERSION(10, 3, 6):
845 case IP_VERSION(10, 3, 1):
847 case IP_VERSION(10, 3, 3):
848 case IP_VERSION(11, 0, 4):
849 case IP_VERSION(11, 5, 0):
850 case IP_VERSION(11, 5, 1):
851 case IP_VERSION(11, 5, 2):
852 /* Don't enable it by default yet.
854 if (amdgpu_tmz < 1) {
855 adev->gmc.tmz_enabled = false;
857 "Trusted Memory Zone (TMZ) feature disabled as experimental (default)\n");
859 adev->gmc.tmz_enabled = true;
861 "Trusted Memory Zone (TMZ) feature enabled as experimental (cmd line)\n");
865 adev->gmc.tmz_enabled = false;
867 "Trusted Memory Zone (TMZ) feature not supported\n");
873 * amdgpu_gmc_noretry_set -- set per asic noretry defaults
874 * @adev: amdgpu_device pointer
876 * Set a per asic default for the no-retry parameter.
879 void amdgpu_gmc_noretry_set(struct amdgpu_device *adev)
881 struct amdgpu_gmc *gmc = &adev->gmc;
882 uint32_t gc_ver = amdgpu_ip_version(adev, GC_HWIP, 0);
883 bool noretry_default = (gc_ver == IP_VERSION(9, 0, 1) ||
884 gc_ver == IP_VERSION(9, 4, 0) ||
885 gc_ver == IP_VERSION(9, 4, 1) ||
886 gc_ver == IP_VERSION(9, 4, 2) ||
887 gc_ver == IP_VERSION(9, 4, 3) ||
888 gc_ver == IP_VERSION(9, 4, 4) ||
889 gc_ver >= IP_VERSION(10, 3, 0));
891 if (!amdgpu_sriov_xnack_support(adev))
894 gmc->noretry = (amdgpu_noretry == -1) ? noretry_default : amdgpu_noretry;
897 void amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type,
900 struct amdgpu_vmhub *hub;
903 hub = &adev->vmhub[hub_type];
904 for (i = 0; i < 16; i++) {
905 reg = hub->vm_context0_cntl + hub->ctx_distance * i;
907 tmp = (hub_type == AMDGPU_GFXHUB(0)) ?
908 RREG32_SOC15_IP(GC, reg) :
909 RREG32_SOC15_IP(MMHUB, reg);
912 tmp |= hub->vm_cntx_cntl_vm_fault;
914 tmp &= ~hub->vm_cntx_cntl_vm_fault;
916 (hub_type == AMDGPU_GFXHUB(0)) ?
917 WREG32_SOC15_IP(GC, reg, tmp) :
918 WREG32_SOC15_IP(MMHUB, reg, tmp);
922 void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev)
927 * Some ASICs need to reserve a region of video memory to avoid access
930 adev->mman.stolen_reserved_offset = 0;
931 adev->mman.stolen_reserved_size = 0;
935 * Currently there is a bug where some memory client outside
936 * of the driver writes to first 8M of VRAM on S3 resume,
937 * this overrides GART which by default gets placed in first 8M and
938 * causes VM_FAULTS once GTT is accessed.
939 * Keep the stolen memory reservation until the while this is not solved.
941 switch (adev->asic_type) {
943 adev->mman.keep_stolen_vga_memory = true;
945 * VEGA10 SRIOV VF with MS_HYPERV host needs some firmware reserved area.
948 if (amdgpu_sriov_vf(adev) && hypervisor_is_type(X86_HYPER_MS_HYPERV)) {
949 adev->mman.stolen_reserved_offset = 0x500000;
950 adev->mman.stolen_reserved_size = 0x200000;
956 adev->mman.keep_stolen_vga_memory = true;
959 adev->mman.keep_stolen_vga_memory = false;
963 if (amdgpu_sriov_vf(adev) ||
964 !amdgpu_device_has_display_hardware(adev)) {
967 size = amdgpu_gmc_get_vbios_fb_size(adev);
969 if (adev->mman.keep_stolen_vga_memory)
970 size = max(size, (unsigned)AMDGPU_VBIOS_VGA_ALLOCATION);
973 /* set to 0 if the pre-OS buffer uses up most of vram */
974 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
977 if (size > AMDGPU_VBIOS_VGA_ALLOCATION) {
978 adev->mman.stolen_vga_size = AMDGPU_VBIOS_VGA_ALLOCATION;
979 adev->mman.stolen_extended_size = size - adev->mman.stolen_vga_size;
981 adev->mman.stolen_vga_size = size;
982 adev->mman.stolen_extended_size = 0;
987 * amdgpu_gmc_init_pdb0 - initialize PDB0
989 * @adev: amdgpu_device pointer
991 * This function is only used when GART page table is used
992 * for FB address translatioin. In such a case, we construct
993 * a 2-level system VM page table: PDB0->PTB, to cover both
994 * VRAM of the hive and system memory.
996 * PDB0 is static, initialized once on driver initialization.
997 * The first n entries of PDB0 are used as PTE by setting
998 * P bit to 1, pointing to VRAM. The n+1'th entry points
999 * to a big PTB covering system memory.
1002 void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev)
1005 uint64_t flags = adev->gart.gart_pte_flags; //TODO it is UC. explore NC/RW?
1006 /* Each PDE0 (used as PTE) covers (2^vmid0_page_table_block_size)*2M
1008 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
1009 u64 pde0_page_size = (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21;
1010 u64 vram_addr = adev->vm_manager.vram_base_offset -
1011 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1012 u64 vram_end = vram_addr + vram_size;
1013 u64 gart_ptb_gpu_pa = amdgpu_gmc_vram_pa(adev, adev->gart.bo);
1016 if (!drm_dev_enter(adev_to_drm(adev), &idx))
1019 flags |= AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE;
1020 flags |= AMDGPU_PTE_WRITEABLE;
1021 flags |= AMDGPU_PTE_SNOOPED;
1022 flags |= AMDGPU_PTE_FRAG((adev->gmc.vmid0_page_table_block_size + 9*1));
1023 flags |= AMDGPU_PDE_PTE_FLAG(adev);
1025 /* The first n PDE0 entries are used as PTE,
1028 for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size)
1029 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, vram_addr, flags);
1031 /* The n+1'th PDE0 entry points to a huge
1032 * PTB who has more than 512 entries each
1033 * pointing to a 4K system page
1035 flags = AMDGPU_PTE_VALID;
1036 flags |= AMDGPU_PTE_SNOOPED | AMDGPU_PDE_BFS_FLAG(adev, 0);
1037 /* Requires gart_ptb_gpu_pa to be 4K aligned */
1038 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, gart_ptb_gpu_pa, flags);
1043 * amdgpu_gmc_vram_mc2pa - calculate vram buffer's physical address from MC
1046 * @adev: amdgpu_device pointer
1047 * @mc_addr: MC address of buffer
1049 uint64_t amdgpu_gmc_vram_mc2pa(struct amdgpu_device *adev, uint64_t mc_addr)
1051 return mc_addr - adev->gmc.vram_start + adev->vm_manager.vram_base_offset;
1055 * amdgpu_gmc_vram_pa - calculate vram buffer object's physical address from
1058 * @adev: amdgpu_device pointer
1059 * @bo: amdgpu buffer object
1061 uint64_t amdgpu_gmc_vram_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo)
1063 return amdgpu_gmc_vram_mc2pa(adev, amdgpu_bo_gpu_offset(bo));
1067 * amdgpu_gmc_vram_cpu_pa - calculate vram buffer object's physical address
1070 * @adev: amdgpu_device pointer
1071 * @bo: amdgpu buffer object
1073 uint64_t amdgpu_gmc_vram_cpu_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo)
1075 return amdgpu_bo_gpu_offset(bo) - adev->gmc.vram_start + adev->gmc.aper_base;
1078 int amdgpu_gmc_vram_checking(struct amdgpu_device *adev)
1080 struct amdgpu_bo *vram_bo = NULL;
1081 uint64_t vram_gpu = 0;
1082 void *vram_ptr = NULL;
1084 int ret, size = 0x100000;
1087 ret = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE,
1088 AMDGPU_GEM_DOMAIN_VRAM,
1095 memset(vram_ptr, 0x86, size);
1096 memset(cptr, 0x86, 10);
1099 * Check the start, the mid, and the end of the memory if the content of
1100 * each byte is the pattern "0x86". If yes, we suppose the vram bo is
1103 * Note: If check the each byte of whole 1M bo, it will cost too many
1104 * seconds, so here, we just pick up three parts for emulation.
1106 ret = memcmp(vram_ptr, cptr, 10);
1109 goto release_buffer;
1112 ret = memcmp(vram_ptr + (size / 2), cptr, 10);
1115 goto release_buffer;
1118 ret = memcmp(vram_ptr + size - 10, cptr, 10);
1121 goto release_buffer;
1125 amdgpu_bo_free_kernel(&vram_bo, &vram_gpu,
1131 static ssize_t current_memory_partition_show(
1132 struct device *dev, struct device_attribute *addr, char *buf)
1134 struct drm_device *ddev = dev_get_drvdata(dev);
1135 struct amdgpu_device *adev = drm_to_adev(ddev);
1136 enum amdgpu_memory_partition mode;
1138 mode = adev->gmc.gmc_funcs->query_mem_partition_mode(adev);
1140 case AMDGPU_NPS1_PARTITION_MODE:
1141 return sysfs_emit(buf, "NPS1\n");
1142 case AMDGPU_NPS2_PARTITION_MODE:
1143 return sysfs_emit(buf, "NPS2\n");
1144 case AMDGPU_NPS3_PARTITION_MODE:
1145 return sysfs_emit(buf, "NPS3\n");
1146 case AMDGPU_NPS4_PARTITION_MODE:
1147 return sysfs_emit(buf, "NPS4\n");
1148 case AMDGPU_NPS6_PARTITION_MODE:
1149 return sysfs_emit(buf, "NPS6\n");
1150 case AMDGPU_NPS8_PARTITION_MODE:
1151 return sysfs_emit(buf, "NPS8\n");
1153 return sysfs_emit(buf, "UNKNOWN\n");
1157 static DEVICE_ATTR_RO(current_memory_partition);
1159 int amdgpu_gmc_sysfs_init(struct amdgpu_device *adev)
1161 if (!adev->gmc.gmc_funcs->query_mem_partition_mode)
1164 return device_create_file(adev->dev,
1165 &dev_attr_current_memory_partition);
1168 void amdgpu_gmc_sysfs_fini(struct amdgpu_device *adev)
1170 device_remove_file(adev->dev, &dev_attr_current_memory_partition);
1173 int amdgpu_gmc_get_nps_memranges(struct amdgpu_device *adev,
1174 struct amdgpu_mem_partition_info *mem_ranges,
1177 struct amdgpu_gmc_memrange *ranges;
1178 int range_cnt, ret, i, j;
1184 ret = amdgpu_discovery_get_nps_info(adev, &nps_type, &ranges,
1190 /* TODO: For now, expect ranges and partition count to be the same.
1191 * Adjust if there are holes expected in any NPS domain.
1193 if (range_cnt != exp_ranges) {
1196 "NPS config mismatch - expected ranges: %d discovery - nps mode: %d, nps ranges: %d",
1197 exp_ranges, nps_type, range_cnt);
1202 for (i = 0; i < exp_ranges; ++i) {
1203 if (ranges[i].base_address >= ranges[i].limit_address) {
1206 "Invalid NPS range - nps mode: %d, range[%d]: base: %llx limit: %llx",
1207 nps_type, i, ranges[i].base_address,
1208 ranges[i].limit_address);
1213 /* Check for overlaps, not expecting any now */
1214 for (j = i - 1; j >= 0; j--) {
1215 if (max(ranges[j].base_address,
1216 ranges[i].base_address) <=
1217 min(ranges[j].limit_address,
1218 ranges[i].limit_address)) {
1221 "overlapping ranges detected [ %llx - %llx ] | [%llx - %llx]",
1222 ranges[j].base_address,
1223 ranges[j].limit_address,
1224 ranges[i].base_address,
1225 ranges[i].limit_address);
1231 mem_ranges[i].range.fpfn =
1232 (ranges[i].base_address -
1233 adev->vm_manager.vram_base_offset) >>
1234 AMDGPU_GPU_PAGE_SHIFT;
1235 mem_ranges[i].range.lpfn =
1236 (ranges[i].limit_address -
1237 adev->vm_manager.vram_base_offset) >>
1238 AMDGPU_GPU_PAGE_SHIFT;
1239 mem_ranges[i].size =
1240 ranges[i].limit_address - ranges[i].base_address + 1;