diff options
author | Jérôme Glisse <jglisse@redhat.com> | 2018-07-26 17:59:13 -0400 |
---|---|---|
committer | Ben Skeggs <bskeggs@redhat.com> | 2019-02-20 09:00:02 +1000 |
commit | 5be73b690875f7eb2d2defb54ccd7f2f12074984 (patch) | |
tree | a4a08ec3faeacfd2275c44545c240bf00d6d5c2d /drivers/gpu | |
parent | eeaf06ac1a5584e41cf289f8351e446bb131374b (diff) |
drm/nouveau/dmem: device memory helpers for SVM
Device memory can be use in SVM, in which case we do not have any of
the existing buffer object. This commit add infrastructure to allow
use of device memory without nouveau_bo. Again this is a temporary
solution until a rework of GPU memory management.
Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
Diffstat (limited to 'drivers/gpu')
-rw-r--r-- | drivers/gpu/drm/nouveau/Kbuild | 1 | ||||
-rw-r--r-- | drivers/gpu/drm/nouveau/Kconfig | 1 | ||||
-rw-r--r-- | drivers/gpu/drm/nouveau/nouveau_dmem.c | 912 | ||||
-rw-r--r-- | drivers/gpu/drm/nouveau/nouveau_dmem.h | 60 | ||||
-rw-r--r-- | drivers/gpu/drm/nouveau/nouveau_drm.c | 5 | ||||
-rw-r--r-- | drivers/gpu/drm/nouveau/nouveau_drv.h | 2 | ||||
-rw-r--r-- | drivers/gpu/drm/nouveau/nouveau_svm.c | 2 |
7 files changed, 983 insertions, 0 deletions
diff --git a/drivers/gpu/drm/nouveau/Kbuild b/drivers/gpu/drm/nouveau/Kbuild index 71bd48aafe64..581404e6544d 100644 --- a/drivers/gpu/drm/nouveau/Kbuild +++ b/drivers/gpu/drm/nouveau/Kbuild @@ -31,6 +31,7 @@ nouveau-y += nouveau_vga.o nouveau-y += nouveau_bo.o nouveau-y += nouveau_gem.o nouveau-$(CONFIG_DRM_NOUVEAU_SVM) += nouveau_svm.o +nouveau-$(CONFIG_DRM_NOUVEAU_SVM) += nouveau_dmem.o nouveau-y += nouveau_mem.o nouveau-y += nouveau_prime.o nouveau-y += nouveau_sgdma.o diff --git a/drivers/gpu/drm/nouveau/Kconfig b/drivers/gpu/drm/nouveau/Kconfig index e7b26a6f386f..00cd9ab8948d 100644 --- a/drivers/gpu/drm/nouveau/Kconfig +++ b/drivers/gpu/drm/nouveau/Kconfig @@ -78,6 +78,7 @@ config DRM_NOUVEAU_SVM depends on DRM_NOUVEAU depends on STAGING select HMM_MIRROR + select DEVICE_PRIVATE default n help Say Y here if you want to enable experimental support for diff --git a/drivers/gpu/drm/nouveau/nouveau_dmem.c b/drivers/gpu/drm/nouveau/nouveau_dmem.c new file mode 100644 index 000000000000..b8ba338df1fb --- /dev/null +++ b/drivers/gpu/drm/nouveau/nouveau_dmem.c @@ -0,0 +1,912 @@ +/* + * Copyright 2018 Red Hat Inc. + * + * 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. + */ +#include "nouveau_dmem.h" +#include "nouveau_drv.h" +#include "nouveau_chan.h" +#include "nouveau_dma.h" +#include "nouveau_mem.h" +#include "nouveau_bo.h" + +#include <nvif/class.h> +#include <nvif/object.h> +#include <nvif/if500b.h> +#include <nvif/if900b.h> + +#include <linux/sched/mm.h> +#include <linux/hmm.h> + +/* + * FIXME: this is ugly right now we are using TTM to allocate vram and we pin + * it in vram while in use. We likely want to overhaul memory management for + * nouveau to be more page like (not necessarily with system page size but a + * bigger page size) at lowest level and have some shim layer on top that would + * provide the same functionality as TTM. + */ +#define DMEM_CHUNK_SIZE (2UL << 20) +#define DMEM_CHUNK_NPAGES (DMEM_CHUNK_SIZE >> PAGE_SHIFT) + +struct nouveau_migrate; + +typedef int (*nouveau_migrate_copy_t)(struct nouveau_drm *drm, u64 npages, + u64 dst_addr, u64 src_addr); + +struct nouveau_dmem_chunk { + struct list_head list; + struct nouveau_bo *bo; + struct nouveau_drm *drm; + unsigned long pfn_first; + unsigned long callocated; + unsigned long bitmap[BITS_TO_LONGS(DMEM_CHUNK_NPAGES)]; + struct nvif_vma vma; + spinlock_t lock; +}; + +struct nouveau_dmem_migrate { + nouveau_migrate_copy_t copy_func; + struct nouveau_channel *chan; +}; + +struct nouveau_dmem { + struct hmm_devmem *devmem; + struct nouveau_dmem_migrate migrate; + struct list_head chunk_free; + struct list_head chunk_full; + struct list_head chunk_empty; + struct mutex mutex; +}; + +struct nouveau_migrate_hmem { + struct scatterlist *sg; + struct nouveau_mem mem; + unsigned long npages; + struct nvif_vma vma; +}; + +struct nouveau_dmem_fault { + struct nouveau_drm *drm; + struct nouveau_fence *fence; + struct nouveau_migrate_hmem hmem; +}; + +struct nouveau_migrate { + struct vm_area_struct *vma; + struct nouveau_drm *drm; + struct nouveau_fence *fence; + unsigned long npages; + struct nouveau_migrate_hmem hmem; +}; + +static void +nouveau_migrate_hmem_fini(struct nouveau_drm *drm, + struct nouveau_migrate_hmem *hmem) +{ + struct nvif_vmm *vmm = &drm->client.vmm.vmm; + + nouveau_mem_fini(&hmem->mem); + nvif_vmm_put(vmm, &hmem->vma); + + if (hmem->sg) { + dma_unmap_sg_attrs(drm->dev->dev, hmem->sg, + hmem->npages, DMA_BIDIRECTIONAL, + DMA_ATTR_SKIP_CPU_SYNC); + kfree(hmem->sg); + hmem->sg = NULL; + } +} + +static int +nouveau_migrate_hmem_init(struct nouveau_drm *drm, + struct nouveau_migrate_hmem *hmem, + unsigned long npages, + const unsigned long *pfns) +{ + struct nvif_vmm *vmm = &drm->client.vmm.vmm; + unsigned long i; + int ret; + + hmem->sg = kzalloc(npages * sizeof(*hmem->sg), GFP_KERNEL); + if (hmem->sg == NULL) + return -ENOMEM; + + for (i = 0, hmem->npages = 0; hmem->npages < npages; ++i) { + struct page *page; + + if (!pfns[i] || pfns[i] == MIGRATE_PFN_ERROR) + continue; + + page = migrate_pfn_to_page(pfns[i]); + if (page == NULL) { + ret = -EINVAL; + goto error; + } + + sg_set_page(&hmem->sg[hmem->npages], page, PAGE_SIZE, 0); + hmem->npages++; + } + sg_mark_end(&hmem->sg[hmem->npages - 1]); + + i = dma_map_sg_attrs(drm->dev->dev, hmem->sg, hmem->npages, + DMA_BIDIRECTIONAL, DMA_ATTR_SKIP_CPU_SYNC); + if (i != hmem->npages) { + ret = -ENOMEM; + goto error; + } + + ret = nouveau_mem_sgl(&hmem->mem, &drm->client, + hmem->npages, hmem->sg); + if (ret) + goto error; + + ret = nvif_vmm_get(vmm, LAZY, false, hmem->mem.mem.page, + 0, hmem->mem.mem.size, &hmem->vma); + if (ret) + goto error; + + ret = nouveau_mem_map(&hmem->mem, vmm, &hmem->vma); + if (ret) + goto error; + + return 0; + +error: + nouveau_migrate_hmem_fini(drm, hmem); + return ret; +} + + +static void +nouveau_dmem_free(struct hmm_devmem *devmem, struct page *page) +{ + struct nouveau_dmem_chunk *chunk; + struct nouveau_drm *drm; + unsigned long idx; + + chunk = (void *)hmm_devmem_page_get_drvdata(page); + idx = page_to_pfn(page) - chunk->pfn_first; + drm = chunk->drm; + + /* + * FIXME: + * + * This is really a bad example, we need to overhaul nouveau memory + * management to be more page focus and allow lighter locking scheme + * to be use in the process. + */ + spin_lock(&chunk->lock); + clear_bit(idx, chunk->bitmap); + WARN_ON(!chunk->callocated); + chunk->callocated--; + /* + * FIXME when chunk->callocated reach 0 we should add the chunk to + * a reclaim list so that it can be freed in case of memory pressure. + */ + spin_unlock(&chunk->lock); +} + +static void +nouveau_dmem_fault_alloc_and_copy(struct vm_area_struct *vma, + const unsigned long *src_pfns, + unsigned long *dst_pfns, + unsigned long start, + unsigned long end, + void *private) +{ + struct nouveau_dmem_fault *fault = private; + struct nouveau_drm *drm = fault->drm; + unsigned long addr, i, c, npages = 0; + nouveau_migrate_copy_t copy; + int ret; + + + /* First allocate new memory */ + for (addr = start, i = 0; addr < end; addr += PAGE_SIZE, i++) { + struct page *dpage, *spage; + + dst_pfns[i] = 0; + spage = migrate_pfn_to_page(src_pfns[i]); + if (!spage || !(src_pfns[i] & MIGRATE_PFN_MIGRATE)) + continue; + + dpage = hmm_vma_alloc_locked_page(vma, addr); + if (!dpage) { + dst_pfns[i] = MIGRATE_PFN_ERROR; + continue; + } + + dst_pfns[i] = migrate_pfn(page_to_pfn(dpage)) | + MIGRATE_PFN_LOCKED; + npages++; + } + + /* Create scatter list FIXME: get rid of scatter list */ + ret = nouveau_migrate_hmem_init(drm, &fault->hmem, npages, dst_pfns); + if (ret) + goto error; + + /* Copy things over */ + copy = drm->dmem->migrate.copy_func; + for (addr = start, i = c = 0; addr < end; addr += PAGE_SIZE, i++) { + struct nouveau_dmem_chunk *chunk; + struct page *spage, *dpage; + u64 src_addr, dst_addr; + + dpage = migrate_pfn_to_page(dst_pfns[i]); + if (!dpage || dst_pfns[i] == MIGRATE_PFN_ERROR) + continue; + + dst_addr = fault->hmem.vma.addr + (c << PAGE_SHIFT); + c++; + + spage = migrate_pfn_to_page(src_pfns[i]); + if (!spage || !(src_pfns[i] & MIGRATE_PFN_MIGRATE)) { + dst_pfns[i] = MIGRATE_PFN_ERROR; + __free_page(dpage); + continue; + } + + chunk = (void *)hmm_devmem_page_get_drvdata(spage); + src_addr = page_to_pfn(spage) - chunk->pfn_first; + src_addr = (src_addr << PAGE_SHIFT) + chunk->vma.addr; + + ret = copy(drm, 1, dst_addr, src_addr); + if (ret) { + dst_pfns[i] = MIGRATE_PFN_ERROR; + __free_page(dpage); + continue; + } + } + + nouveau_fence_new(drm->dmem->migrate.chan, false, &fault->fence); + + return; + +error: + for (addr = start, i = 0; addr < end; addr += PAGE_SIZE, ++i) { + struct page *page; + + if (!dst_pfns[i] || dst_pfns[i] == MIGRATE_PFN_ERROR) + continue; + + page = migrate_pfn_to_page(dst_pfns[i]); + dst_pfns[i] = MIGRATE_PFN_ERROR; + if (page == NULL) + continue; + + __free_page(page); + } +} + +void nouveau_dmem_fault_finalize_and_map(struct vm_area_struct *vma, + const unsigned long *src_pfns, + const unsigned long *dst_pfns, + unsigned long start, + unsigned long end, + void *private) +{ + struct nouveau_dmem_fault *fault = private; + struct nouveau_drm *drm = fault->drm; + + if (fault->fence) { + nouveau_fence_wait(fault->fence, true, false); + nouveau_fence_unref(&fault->fence); + } else { + /* + * FIXME wait for channel to be IDLE before calling finalizing + * the hmem object below (nouveau_migrate_hmem_fini()). + */ + } + nouveau_migrate_hmem_fini(drm, &fault->hmem); +} + +static const struct migrate_vma_ops nouveau_dmem_fault_migrate_ops = { + .alloc_and_copy = nouveau_dmem_fault_alloc_and_copy, + .finalize_and_map = nouveau_dmem_fault_finalize_and_map, +}; + +static int +nouveau_dmem_fault(struct hmm_devmem *devmem, + struct vm_area_struct *vma, + unsigned long addr, + const struct page *page, + unsigned int flags, + pmd_t *pmdp) +{ + struct drm_device *drm_dev = dev_get_drvdata(devmem->device); + unsigned long src[1] = {0}, dst[1] = {0}; + struct nouveau_dmem_fault fault = {0}; + int ret; + + + + /* + * FIXME what we really want is to find some heuristic to migrate more + * than just one page on CPU fault. When such fault happens it is very + * likely that more surrounding page will CPU fault too. + */ + fault.drm = nouveau_drm(drm_dev); + ret = migrate_vma(&nouveau_dmem_fault_migrate_ops, vma, addr, + addr + PAGE_SIZE, src, dst, &fault); + if (ret) + return VM_FAULT_SIGBUS; + + if (dst[0] == MIGRATE_PFN_ERROR) + return VM_FAULT_SIGBUS; + + return 0; +} + +static const struct hmm_devmem_ops +nouveau_dmem_devmem_ops = { + .free = nouveau_dmem_free, + .fault = nouveau_dmem_fault, +}; + +static int +nouveau_dmem_chunk_alloc(struct nouveau_drm *drm) +{ + struct nvif_vmm *vmm = &drm->client.vmm.vmm; + struct nouveau_dmem_chunk *chunk; + int ret; + + if (drm->dmem == NULL) + return -EINVAL; + + mutex_lock(&drm->dmem->mutex); + chunk = list_first_entry_or_null(&drm->dmem->chunk_empty, + struct nouveau_dmem_chunk, + list); + if (chunk == NULL) { + mutex_unlock(&drm->dmem->mutex); + return -ENOMEM; + } + + list_del(&chunk->list); + mutex_unlock(&drm->dmem->mutex); + + ret = nvif_vmm_get(vmm, LAZY, false, 12, 0, + DMEM_CHUNK_SIZE, &chunk->vma); + if (ret) + goto out; + + ret = nouveau_bo_new(&drm->client, DMEM_CHUNK_SIZE, 0, + TTM_PL_FLAG_VRAM, 0, 0, NULL, NULL, + &chunk->bo); + if (ret) + goto out; + + ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false); + if (ret) { + nouveau_bo_ref(NULL, &chunk->bo); + goto out; + } + + ret = nouveau_mem_map(nouveau_mem(&chunk->bo->bo.mem), vmm, &chunk->vma); + if (ret) { + nouveau_bo_unpin(chunk->bo); + nouveau_bo_ref(NULL, &chunk->bo); + goto out; + } + + bitmap_zero(chunk->bitmap, DMEM_CHUNK_NPAGES); + spin_lock_init(&chunk->lock); + +out: + mutex_lock(&drm->dmem->mutex); + if (chunk->bo) + list_add(&chunk->list, &drm->dmem->chunk_empty); + else + list_add_tail(&chunk->list, &drm->dmem->chunk_empty); + mutex_unlock(&drm->dmem->mutex); + + return ret; +} + +static struct nouveau_dmem_chunk * +nouveau_dmem_chunk_first_free_locked(struct nouveau_drm *drm) +{ + struct nouveau_dmem_chunk *chunk; + + chunk = list_first_entry_or_null(&drm->dmem->chunk_free, + struct nouveau_dmem_chunk, + list); + if (chunk) + return chunk; + + chunk = list_first_entry_or_null(&drm->dmem->chunk_empty, + struct nouveau_dmem_chunk, + list); + if (chunk->bo) + return chunk; + + return NULL; +} + +static int +nouveau_dmem_pages_alloc(struct nouveau_drm *drm, + unsigned long npages, + unsigned long *pages) +{ + struct nouveau_dmem_chunk *chunk; + unsigned long c; + int ret; + + memset(pages, 0xff, npages * sizeof(*pages)); + + mutex_lock(&drm->dmem->mutex); + for (c = 0; c < npages;) { + unsigned long i; + + chunk = nouveau_dmem_chunk_first_free_locked(drm); + if (chunk == NULL) { + mutex_unlock(&drm->dmem->mutex); + ret = nouveau_dmem_chunk_alloc(drm); + if (ret) { + if (c) + break; + return ret; + } + continue; + } + + spin_lock(&chunk->lock); + i = find_first_zero_bit(chunk->bitmap, DMEM_CHUNK_NPAGES); + while (i < DMEM_CHUNK_NPAGES && c < npages) { + pages[c] = chunk->pfn_first + i; + set_bit(i, chunk->bitmap); + chunk->callocated++; + c++; + + i = find_next_zero_bit(chunk->bitmap, + DMEM_CHUNK_NPAGES, i); + } + spin_unlock(&chunk->lock); + } + mutex_unlock(&drm->dmem->mutex); + + return 0; +} + +static struct page * +nouveau_dmem_page_alloc_locked(struct nouveau_drm *drm) +{ + unsigned long pfns[1]; + struct page *page; + int ret; + + /* FIXME stop all the miss-match API ... */ + ret = nouveau_dmem_pages_alloc(drm, 1, pfns); + if (ret) + return NULL; + + page = pfn_to_page(pfns[0]); + get_page(page); + lock_page(page); + return page; +} + +static void +nouveau_dmem_page_free_locked(struct nouveau_drm *drm, struct page *page) +{ + unlock_page(page); + put_page(page); +} + +void +nouveau_dmem_resume(struct nouveau_drm *drm) +{ + struct nouveau_dmem_chunk *chunk; + int ret; + + if (drm->dmem == NULL) + return; + + mutex_lock(&drm->dmem->mutex); + list_for_each_entry (chunk, &drm->dmem->chunk_free, list) { + ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false); + /* FIXME handle pin failure */ + WARN_ON(ret); + } + list_for_each_entry (chunk, &drm->dmem->chunk_full, list) { + ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false); + /* FIXME handle pin failure */ + WARN_ON(ret); + } + list_for_each_entry (chunk, &drm->dmem->chunk_empty, list) { + ret = nouveau_bo_pin(chunk->bo, TTM_PL_FLAG_VRAM, false); + /* FIXME handle pin failure */ + WARN_ON(ret); + } + mutex_unlock(&drm->dmem->mutex); +} + +void +nouveau_dmem_suspend(struct nouveau_drm *drm) +{ + struct nouveau_dmem_chunk *chunk; + + if (drm->dmem == NULL) + return; + + mutex_lock(&drm->dmem->mutex); + list_for_each_entry (chunk, &drm->dmem->chunk_free, list) { + nouveau_bo_unpin(chunk->bo); + } + list_for_each_entry (chunk, &drm->dmem->chunk_full, list) { + nouveau_bo_unpin(chunk->bo); + } + list_for_each_entry (chunk, &drm->dmem->chunk_empty, list) { + nouveau_bo_unpin(chunk->bo); + } + mutex_unlock(&drm->dmem->mutex); +} + +void +nouveau_dmem_fini(struct nouveau_drm *drm) +{ + struct nvif_vmm *vmm = &drm->client.vmm.vmm; + struct nouveau_dmem_chunk *chunk, *tmp; + + if (drm->dmem == NULL) + return; + + mutex_lock(&drm->dmem->mutex); + + WARN_ON(!list_empty(&drm->dmem->chunk_free)); + WARN_ON(!list_empty(&drm->dmem->chunk_full)); + + list_for_each_entry_safe (chunk, tmp, &drm->dmem->chunk_empty, list) { + if (chunk->bo) { + nouveau_bo_unpin(chunk->bo); + nouveau_bo_ref(NULL, &chunk->bo); + } + nvif_vmm_put(vmm, &chunk->vma); + list_del(&chunk->list); + kfree(chunk); + } + + mutex_unlock(&drm->dmem->mutex); +} + +static int +nvc0b5_migrate_copy(struct nouveau_drm *drm, u64 npages, + u64 dst_addr, u64 src_addr) +{ + struct nouveau_channel *chan = drm->dmem->migrate.chan; + int ret; + + ret = RING_SPACE(chan, 10); + if (ret) + return ret; + + BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8); + OUT_RING (chan, upper_32_bits(src_addr)); + OUT_RING (chan, lower_32_bits(src_addr)); + OUT_RING (chan, upper_32_bits(dst_addr)); + OUT_RING (chan, lower_32_bits(dst_addr)); + OUT_RING (chan, PAGE_SIZE); + OUT_RING (chan, PAGE_SIZE); + OUT_RING (chan, PAGE_SIZE); + OUT_RING (chan, npages); + BEGIN_IMC0(chan, NvSubCopy, 0x0300, 0x0386); + return 0; +} + +static int +nouveau_dmem_migrate_init(struct nouveau_drm *drm) +{ + switch (drm->ttm.copy.oclass) { + case PASCAL_DMA_COPY_A: + case PASCAL_DMA_COPY_B: + case VOLTA_DMA_COPY_A: + case TURING_DMA_COPY_A: + drm->dmem->migrate.copy_func = nvc0b5_migrate_copy; + drm->dmem->migrate.chan = drm->ttm.chan; + return 0; + default: + break; + } + return -ENODEV; +} + +void +nouveau_dmem_init(struct nouveau_drm *drm) +{ + struct device *device = drm->dev->dev; + unsigned long i, size; + int ret; + + /* This only make sense on PASCAL or newer */ + if (drm->client.device.info.family < NV_DEVICE_INFO_V0_PASCAL) + return; + + if (!(drm->dmem = kzalloc(sizeof(*drm->dmem), GFP_KERNEL))) + return; + + mutex_init(&drm->dmem->mutex); + INIT_LIST_HEAD(&drm->dmem->chunk_free); + INIT_LIST_HEAD(&drm->dmem->chunk_full); + INIT_LIST_HEAD(&drm->dmem->chunk_empty); + + size = ALIGN(drm->client.device.info.ram_user, DMEM_CHUNK_SIZE); + + /* Initialize migration dma helpers before registering memory */ + ret = nouveau_dmem_migrate_init(drm); + if (ret) { + kfree(drm->dmem); + drm->dmem = NULL; + return; + } + + /* + * FIXME we need some kind of policy to decide how much VRAM we + * want to register with HMM. For now just register everything + * and latter if we want to do thing like over commit then we + * could revisit this. + */ + drm->dmem->devmem = hmm_devmem_add(&nouveau_dmem_devmem_ops, + device, size); + if (drm->dmem->devmem == NULL) { + kfree(drm->dmem); + drm->dmem = NULL; + return; + } + + for (i = 0; i < (size / DMEM_CHUNK_SIZE); ++i) { + struct nouveau_dmem_chunk *chunk; + struct page *page; + unsigned long j; + + chunk = kzalloc(sizeof(*chunk), GFP_KERNEL); + if (chunk == NULL) { + nouveau_dmem_fini(drm); + return; + } + + chunk->drm = drm; + chunk->pfn_first = drm->dmem->devmem->pfn_first; + chunk->pfn_first += (i * DMEM_CHUNK_NPAGES); + list_add_tail(&chunk->list, &drm->dmem->chunk_empty); + + page = pfn_to_page(chunk->pfn_first); + for (j = 0; j < DMEM_CHUNK_NPAGES; ++j, ++page) { + hmm_devmem_page_set_drvdata(page, (long)chunk); + } + } + + NV_INFO(drm, "DMEM: registered %ldMB of device memory\n", size >> 20); +} + +static void +nouveau_dmem_migrate_alloc_and_copy(struct vm_area_struct *vma, + const unsigned long *src_pfns, + unsigned long *dst_pfns, + unsigned long start, + unsigned long end, + void *private) +{ + struct nouveau_migrate *migrate = private; + struct nouveau_drm *drm = migrate->drm; + unsigned long addr, i, c, npages = 0; + nouveau_migrate_copy_t copy; + int ret; + + /* First allocate new memory */ + for (addr = start, i = 0; addr < end; addr += PAGE_SIZE, i++) { + struct page *dpage, *spage; + + dst_pfns[i] = 0; + spage = migrate_pfn_to_page(src_pfns[i]); + if (!spage || !(src_pfns[i] & MIGRATE_PFN_MIGRATE)) + continue; + + dpage = nouveau_dmem_page_alloc_locked(drm); + if (!dpage) + continue; + + dst_pfns[i] = migrate_pfn(page_to_pfn(dpage)) | + MIGRATE_PFN_LOCKED | + MIGRATE_PFN_DEVICE; + npages++; + } + + if (!npages) + return; + + /* Create scatter list FIXME: get rid of scatter list */ + ret = nouveau_migrate_hmem_init(drm, &migrate->hmem, npages, src_pfns); + if (ret) + goto error; + + /* Copy things over */ + copy = drm->dmem->migrate.copy_func; + for (addr = start, i = c = 0; addr < end; addr += PAGE_SIZE, i++) { + struct nouveau_dmem_chunk *chunk; + struct page *spage, *dpage; + u64 src_addr, dst_addr; + + dpage = migrate_pfn_to_page(dst_pfns[i]); + if (!dpage || dst_pfns[i] == MIGRATE_PFN_ERROR) + continue; + + chunk = (void *)hmm_devmem_page_get_drvdata(dpage); + dst_addr = page_to_pfn(dpage) - chunk->pfn_first; + dst_addr = (dst_addr << PAGE_SHIFT) + chunk->vma.addr; + + spage = migrate_pfn_to_page(src_pfns[i]); + if (!spage || !(src_pfns[i] & MIGRATE_PFN_MIGRATE)) { + nouveau_dmem_page_free_locked(drm, dpage); + dst_pfns[i] = 0; + continue; + } + + src_addr = migrate->hmem.vma.addr + (c << PAGE_SHIFT); + c++; + + ret = copy(drm, 1, dst_addr, src_addr); + if (ret) { + nouveau_dmem_page_free_locked(drm, dpage); + dst_pfns[i] = 0; + continue; + } + } + + nouveau_fence_new(drm->dmem->migrate.chan, false, &migrate->fence); + + return; + +error: + for (addr = start, i = 0; addr < end; addr += PAGE_SIZE, ++i) { + struct page *page; + + if (!dst_pfns[i] || dst_pfns[i] == MIGRATE_PFN_ERROR) + continue; + + page = migrate_pfn_to_page(dst_pfns[i]); + dst_pfns[i] = MIGRATE_PFN_ERROR; + if (page == NULL) + continue; + + __free_page(page); + } +} + +void nouveau_dmem_migrate_finalize_and_map(struct vm_area_struct *vma, + const unsigned long *src_pfns, + const unsigned long *dst_pfns, + unsigned long start, + unsigned long end, + void *private) +{ + struct nouveau_migrate *migrate = private; + struct nouveau_drm *drm = migrate->drm; + + if (migrate->fence) { + nouveau_fence_wait(migrate->fence, true, false); + nouveau_fence_unref(&migrate->fence); + } else { + /* + * FIXME wait for channel to be IDLE before finalizing + * the hmem object below (nouveau_migrate_hmem_fini()) ? + */ + } + nouveau_migrate_hmem_fini(drm, &migrate->hmem); + + /* + * FIXME optimization: update GPU page table to point to newly + * migrated memory. + */ +} + +static const struct migrate_vma_ops nouveau_dmem_migrate_ops = { + .alloc_and_copy = nouveau_dmem_migrate_alloc_and_copy, + .finalize_and_map = nouveau_dmem_migrate_finalize_and_map, +}; + +int +nouveau_dmem_migrate_vma(struct nouveau_drm *drm, + struct vm_area_struct *vma, + unsigned long start, + unsigned long end) +{ + unsigned long *src_pfns, *dst_pfns, npages; + struct nouveau_migrate migrate = {0}; + unsigned long i, c, max; + int ret = 0; + + npages = (end - start) >> PAGE_SHIFT; + max = min(SG_MAX_SINGLE_ALLOC, npages); + src_pfns = kzalloc(sizeof(long) * max, GFP_KERNEL); + if (src_pfns == NULL) + return -ENOMEM; + dst_pfns = kzalloc(sizeof(long) * max, GFP_KERNEL); + if (dst_pfns == NULL) { + kfree(src_pfns); + return -ENOMEM; + } + + migrate.drm = drm; + migrate.vma = vma; + migrate.npages = npages; + for (i = 0; i < npages; i += c) { + unsigned long next; + + c = min(SG_MAX_SINGLE_ALLOC, npages); + next = start + (c << PAGE_SHIFT); + ret = migrate_vma(&nouveau_dmem_migrate_ops, vma, start, + next, src_pfns, dst_pfns, &migrate); + if (ret) + goto out; + start = next; + } + +out: + kfree(dst_pfns); + kfree(src_pfns); + return ret; +} + +static inline bool +nouveau_dmem_page(struct nouveau_drm *drm, struct page *page) +{ + if (!is_device_private_page(page)) + return false; + + if (drm->dmem->devmem != page->pgmap->data) + return false; + + return true; +} + +void +nouveau_dmem_convert_pfn(struct nouveau_drm *drm, + struct hmm_range *range) +{ + unsigned long i, npages; + + npages = (range->end - range->start) >> PAGE_SHIFT; + for (i = 0; i < npages; ++i) { + struct nouveau_dmem_chunk *chunk; + struct page *page; + uint64_t addr; + + page = hmm_pfn_to_page(range, range->pfns[i]); + if (page == NULL) + continue; + + if (!(range->pfns[i] & range->flags[HMM_PFN_DEVICE_PRIVATE])) { + continue; + } + + if (!nouveau_dmem_page(drm, page)) { + WARN(1, "Some unknown device memory !\n"); + range->pfns[i] = 0; + continue; + } + + chunk = (void *)hmm_devmem_page_get_drvdata(page); + addr = page_to_pfn(page) - chunk->pfn_first; + addr = (addr + chunk->bo->bo.mem.start) << PAGE_SHIFT; + + range->pfns[i] &= ((1UL << range->pfn_shift) - 1); + range->pfns[i] |= (addr >> PAGE_SHIFT) << range->pfn_shift; + } +} diff --git a/drivers/gpu/drm/nouveau/nouveau_dmem.h b/drivers/gpu/drm/nouveau/nouveau_dmem.h new file mode 100644 index 000000000000..9d97d756fb7d --- /dev/null +++ b/drivers/gpu/drm/nouveau/nouveau_dmem.h @@ -0,0 +1,60 @@ +/* + * Copyright 2018 Red Hat Inc. + * + * 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. + */ +#ifndef __NOUVEAU_DMEM_H__ +#define __NOUVEAU_DMEM_H__ +#include <nvif/os.h> +struct drm_device; +struct drm_file; +struct nouveau_drm; +struct hmm_range; + +#if IS_ENABLED(CONFIG_DRM_NOUVEAU_SVM) +void nouveau_dmem_init(struct nouveau_drm *); +void nouveau_dmem_fini(struct nouveau_drm *); +void nouveau_dmem_suspend(struct nouveau_drm *); +void nouveau_dmem_resume(struct nouveau_drm *); + +int nouveau_dmem_migrate_vma(struct nouveau_drm *drm, + struct vm_area_struct *vma, + unsigned long start, + unsigned long end); + +void nouveau_dmem_convert_pfn(struct nouveau_drm *drm, + struct hmm_range *range); +#else /* IS_ENABLED(CONFIG_DRM_NOUVEAU_SVM) */ +static inline void nouveau_dmem_init(struct nouveau_drm *drm) {} +static inline void nouveau_dmem_fini(struct nouveau_drm *drm) {} +static inline void nouveau_dmem_suspend(struct nouveau_drm *drm) {} +static inline void nouveau_dmem_resume(struct nouveau_drm *drm) {} + +static inline int nouveau_dmem_migrate_vma(struct nouveau_drm *drm, + struct vm_area_struct *vma, + unsigned long start, + unsigned long end) +{ + return 0; +} + +static inline void nouveau_dmem_convert_pfn(struct nouveau_drm *drm, + struct hmm_range *range) {} +#endif /* IS_ENABLED(CONFIG_DRM_NOUVEAU_SVM) */ +#endif diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c b/drivers/gpu/drm/nouveau/nouveau_drm.c index 53fdfa283ee2..ee2c4d498d7d 100644 --- a/drivers/gpu/drm/nouveau/nouveau_drm.c +++ b/drivers/gpu/drm/nouveau/nouveau_drm.c @@ -63,6 +63,7 @@ #include "nouveau_connector.h" #include "nouveau_platform.h" #include "nouveau_svm.h" +#include "nouveau_dmem.h" MODULE_PARM_DESC(config, "option string to pass to driver core"); static char *nouveau_config; @@ -551,6 +552,7 @@ nouveau_drm_device_init(struct drm_device *dev) nouveau_debugfs_init(drm); nouveau_hwmon_init(dev); nouveau_svm_init(drm); + nouveau_dmem_init(drm); nouveau_fbcon_init(dev); nouveau_led_init(dev); @@ -594,6 +596,7 @@ nouveau_drm_device_fini(struct drm_device *dev) nouveau_led_fini(dev); nouveau_fbcon_fini(dev); + nouveau_dmem_fini(drm); nouveau_svm_fini(drm); nouveau_hwmon_fini(dev); nouveau_debugfs_fini(drm); @@ -741,6 +744,7 @@ nouveau_do_suspend(struct drm_device *dev, bool runtime) int ret; nouveau_svm_suspend(drm); + nouveau_dmem_suspend(drm); nouveau_led_suspend(dev); if (dev->mode_config.num_crtc) { @@ -817,6 +821,7 @@ nouveau_do_resume(struct drm_device *dev, bool runtime) } nouveau_led_resume(dev); + nouveau_dmem_resume(drm); nouveau_svm_resume(drm); return 0; } diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h b/drivers/gpu/drm/nouveau/nouveau_drv.h index 1326b42f75e6..da847244479d 100644 --- a/drivers/gpu/drm/nouveau/nouveau_drv.h +++ b/drivers/gpu/drm/nouveau/nouveau_drv.h @@ -212,6 +212,8 @@ struct nouveau_drm { struct dev_pm_domain vga_pm_domain; struct nouveau_svm *svm; + + struct nouveau_dmem *dmem; }; static inline struct nouveau_drm * diff --git a/drivers/gpu/drm/nouveau/nouveau_svm.c b/drivers/gpu/drm/nouveau/nouveau_svm.c index 6158e99b1dc3..3ba980d80a1a 100644 --- a/drivers/gpu/drm/nouveau/nouveau_svm.c +++ b/drivers/gpu/drm/nouveau/nouveau_svm.c @@ -561,6 +561,8 @@ again: goto again; } + nouveau_dmem_convert_pfn(svm->drm, &range); + svmm->vmm->vmm.object.client->super = true; ret = nvif_object_ioctl(&svmm->vmm->vmm.object, &args, sizeof(args.i) + |