// SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note /* * * (C) COPYRIGHT 2022-2023 ARM Limited. All rights reserved. * * This program is free software and is provided to you under the terms of the * GNU General Public License version 2 as published by the Free Software * Foundation, and any use by you of this program is subject to the terms * of such GNU license. * * This program is distributed in the hope that 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, you can access it online at * http://www.gnu.org/licenses/gpl-2.0.html. * */ /** * DOC: Base kernel page migration implementation. */ #include #include #include #include /* Global integer used to determine if module parameter value has been * provided and if page migration feature is enabled. * Feature is disabled on all platforms by default. */ #if !IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT) /* If page migration support is explicitly compiled out, there should be no way to change * this int. Its value is automatically 0 as a global. */ const int kbase_page_migration_enabled; /* module_param is not called so this value cannot be changed at insmod when compiled * without support for page migration. */ #else /* -1 as default, 0 when manually set as off and 1 when manually set as on */ int kbase_page_migration_enabled = -1; module_param(kbase_page_migration_enabled, int, 0444); MODULE_PARM_DESC(kbase_page_migration_enabled, "Explicitly enable or disable page migration with 1 or 0 respectively."); #endif /* !IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT) */ KBASE_EXPORT_TEST_API(kbase_page_migration_enabled); bool kbase_is_page_migration_enabled(void) { /* Handle uninitialised int case */ if (kbase_page_migration_enabled < 0) return false; return IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT) && kbase_page_migration_enabled; } KBASE_EXPORT_SYMBOL(kbase_is_page_migration_enabled); #if (KERNEL_VERSION(6, 0, 0) <= LINUX_VERSION_CODE) static const struct movable_operations movable_ops; #endif bool kbase_alloc_page_metadata(struct kbase_device *kbdev, struct page *p, dma_addr_t dma_addr, u8 group_id) { struct kbase_page_metadata *page_md; /* A check for kbase_page_migration_enabled would help here too but it's already being * checked in the only caller of this function. */ if (!IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT)) return false; page_md = kzalloc(sizeof(struct kbase_page_metadata), GFP_KERNEL); if (!page_md) return false; SetPagePrivate(p); set_page_private(p, (unsigned long)page_md); page_md->dma_addr = dma_addr; page_md->status = PAGE_STATUS_SET(page_md->status, (u8)ALLOCATE_IN_PROGRESS); page_md->vmap_count = 0; page_md->group_id = group_id; spin_lock_init(&page_md->migrate_lock); lock_page(p); #if (KERNEL_VERSION(6, 0, 0) <= LINUX_VERSION_CODE) __SetPageMovable(p, &movable_ops); page_md->status = PAGE_MOVABLE_SET(page_md->status); #else /* In some corner cases, the driver may attempt to allocate memory pages * even before the device file is open and the mapping for address space * operations is created. In that case, it is impossible to assign address * space operations to memory pages: simply pretend that they are movable, * even if they are not. * * The page will go through all state transitions but it will never be * actually considered movable by the kernel. This is due to the fact that * the page cannot be marked as NOT_MOVABLE upon creation, otherwise the * memory pool will always refuse to add it to the pool and schedule * a worker thread to free it later. * * Page metadata may seem redundant in this case, but they are not, * because memory pools expect metadata to be present when page migration * is enabled and because the pages may always return to memory pools and * gain the movable property later on in their life cycle. */ if (kbdev->mem_migrate.inode && kbdev->mem_migrate.inode->i_mapping) { __SetPageMovable(p, kbdev->mem_migrate.inode->i_mapping); page_md->status = PAGE_MOVABLE_SET(page_md->status); } #endif unlock_page(p); return true; } static void kbase_free_page_metadata(struct kbase_device *kbdev, struct page *p, u8 *group_id) { struct device *const dev = kbdev->dev; struct kbase_page_metadata *page_md; dma_addr_t dma_addr; if (!IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT)) return; page_md = kbase_page_private(p); if (!page_md) return; if (group_id) *group_id = page_md->group_id; dma_addr = kbase_dma_addr(p); dma_unmap_page(dev, dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL); kfree(page_md); set_page_private(p, 0); ClearPagePrivate(p); } #if IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT) /* This function is only called when page migration * support is not explicitly compiled out. */ static void kbase_free_pages_worker(struct work_struct *work) { struct kbase_mem_migrate *mem_migrate = container_of(work, struct kbase_mem_migrate, free_pages_work); struct kbase_device *kbdev = container_of(mem_migrate, struct kbase_device, mem_migrate); struct page *p, *tmp; struct kbase_page_metadata *page_md; LIST_HEAD(free_list); spin_lock(&mem_migrate->free_pages_lock); list_splice_init(&mem_migrate->free_pages_list, &free_list); spin_unlock(&mem_migrate->free_pages_lock); list_for_each_entry_safe(p, tmp, &free_list, lru) { u8 group_id = 0; list_del_init(&p->lru); lock_page(p); page_md = kbase_page_private(p); if (page_md && IS_PAGE_MOVABLE(page_md->status)) { __ClearPageMovable(p); page_md->status = PAGE_MOVABLE_CLEAR(page_md->status); } unlock_page(p); kbase_free_page_metadata(kbdev, p, &group_id); kbdev->mgm_dev->ops.mgm_free_page(kbdev->mgm_dev, group_id, p, 0); } } #endif void kbase_free_page_later(struct kbase_device *kbdev, struct page *p) { struct kbase_mem_migrate *mem_migrate = &kbdev->mem_migrate; if (!IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT)) return; spin_lock(&mem_migrate->free_pages_lock); list_add(&p->lru, &mem_migrate->free_pages_list); spin_unlock(&mem_migrate->free_pages_lock); } /** * kbasep_migrate_page_pt_mapped - Migrate a memory page that is mapped * in a PGD of kbase_mmu_table. * * @old_page: Existing PGD page to remove * @new_page: Destination for migrating the existing PGD page to * * Replace an existing PGD page with a new page by migrating its content. More specifically: * the new page shall replace the existing PGD page in the MMU page table. Before returning, * the new page shall be set as movable and not isolated, while the old page shall lose * the movable property. The meta data attached to the PGD page is transferred to the * new (replacement) page. * * This function returns early with an error if called when not compiled with * CONFIG_PAGE_MIGRATION_SUPPORT. * * Return: 0 on migration success, or -EAGAIN for a later retry. Otherwise it's a failure * and the migration is aborted. */ static int kbasep_migrate_page_pt_mapped(struct page *old_page, struct page *new_page) { struct kbase_page_metadata *page_md = kbase_page_private(old_page); struct kbase_context *kctx = page_md->data.pt_mapped.mmut->kctx; struct kbase_device *kbdev = kctx->kbdev; dma_addr_t old_dma_addr = page_md->dma_addr; dma_addr_t new_dma_addr; int ret; if (!IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT)) return -EINVAL; /* Create a new dma map for the new page */ new_dma_addr = dma_map_page(kbdev->dev, new_page, 0, PAGE_SIZE, DMA_BIDIRECTIONAL); if (dma_mapping_error(kbdev->dev, new_dma_addr)) return -ENOMEM; /* Lock context to protect access to the page in physical allocation. * This blocks the CPU page fault handler from remapping pages. * Only MCU's mmut is device wide, i.e. no corresponding kctx. */ kbase_gpu_vm_lock_with_pmode_sync(kctx); ret = kbase_mmu_migrate_page( as_tagged(page_to_phys(old_page)), as_tagged(page_to_phys(new_page)), old_dma_addr, new_dma_addr, PGD_VPFN_LEVEL_GET_LEVEL(page_md->data.pt_mapped.pgd_vpfn_level)); if (ret == 0) { dma_unmap_page(kbdev->dev, old_dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL); __ClearPageMovable(old_page); ClearPagePrivate(old_page); put_page(old_page); #if (KERNEL_VERSION(6, 0, 0) <= LINUX_VERSION_CODE) __SetPageMovable(new_page, &movable_ops); page_md->status = PAGE_MOVABLE_SET(page_md->status); #else if (kbdev->mem_migrate.inode->i_mapping) { __SetPageMovable(new_page, kbdev->mem_migrate.inode->i_mapping); page_md->status = PAGE_MOVABLE_SET(page_md->status); } #endif SetPagePrivate(new_page); get_page(new_page); } else dma_unmap_page(kbdev->dev, new_dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL); /* Page fault handler for CPU mapping unblocked. */ kbase_gpu_vm_unlock_with_pmode_sync(kctx); return ret; } /* * kbasep_migrate_page_allocated_mapped - Migrate a memory page that is both * allocated and mapped. * * @old_page: Page to remove. * @new_page: Page to add. * * Replace an old page with a new page by migrating its content and all its * CPU and GPU mappings. More specifically: the new page shall replace the * old page in the MMU page table, as well as in the page array of the physical * allocation, which is used to create CPU mappings. Before returning, the new * page shall be set as movable and not isolated, while the old page shall lose * the movable property. * * This function returns early with an error if called when not compiled with * CONFIG_PAGE_MIGRATION_SUPPORT. */ static int kbasep_migrate_page_allocated_mapped(struct page *old_page, struct page *new_page) { struct kbase_page_metadata *page_md = kbase_page_private(old_page); struct kbase_context *kctx = page_md->data.mapped.mmut->kctx; dma_addr_t old_dma_addr, new_dma_addr; int ret; if (!IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT)) return -EINVAL; old_dma_addr = page_md->dma_addr; new_dma_addr = dma_map_page(kctx->kbdev->dev, new_page, 0, PAGE_SIZE, DMA_BIDIRECTIONAL); if (dma_mapping_error(kctx->kbdev->dev, new_dma_addr)) return -ENOMEM; /* Lock context to protect access to array of pages in physical allocation. * This blocks the CPU page fault handler from remapping pages. */ kbase_gpu_vm_lock_with_pmode_sync(kctx); /* Unmap the old physical range. */ unmap_mapping_range(kctx->kfile->filp->f_inode->i_mapping, page_md->data.mapped.vpfn << PAGE_SHIFT, PAGE_SIZE, 1); ret = kbase_mmu_migrate_page(as_tagged(page_to_phys(old_page)), as_tagged(page_to_phys(new_page)), old_dma_addr, new_dma_addr, MIDGARD_MMU_BOTTOMLEVEL); if (ret == 0) { dma_unmap_page(kctx->kbdev->dev, old_dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL); SetPagePrivate(new_page); get_page(new_page); /* Clear PG_movable from the old page and release reference. */ ClearPagePrivate(old_page); __ClearPageMovable(old_page); put_page(old_page); /* Set PG_movable to the new page. */ #if (KERNEL_VERSION(6, 0, 0) <= LINUX_VERSION_CODE) __SetPageMovable(new_page, &movable_ops); page_md->status = PAGE_MOVABLE_SET(page_md->status); #else if (kctx->kbdev->mem_migrate.inode->i_mapping) { __SetPageMovable(new_page, kctx->kbdev->mem_migrate.inode->i_mapping); page_md->status = PAGE_MOVABLE_SET(page_md->status); } #endif } else dma_unmap_page(kctx->kbdev->dev, new_dma_addr, PAGE_SIZE, DMA_BIDIRECTIONAL); /* Page fault handler for CPU mapping unblocked. */ kbase_gpu_vm_unlock_with_pmode_sync(kctx); return ret; } /** * kbase_page_isolate - Isolate a page for migration. * * @p: Pointer of the page struct of page to isolate. * @mode: LRU Isolation modes. * * Callback function for Linux to isolate a page and prepare it for migration. * This callback is not registered if compiled without CONFIG_PAGE_MIGRATION_SUPPORT. * * Return: true on success, false otherwise. */ static bool kbase_page_isolate(struct page *p, isolate_mode_t mode) { bool status_mem_pool = false; struct kbase_mem_pool *mem_pool = NULL; struct kbase_page_metadata *page_md = kbase_page_private(p); if (!IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT)) return false; CSTD_UNUSED(mode); if (!page_md || !IS_PAGE_MOVABLE(page_md->status)) return false; if (!spin_trylock(&page_md->migrate_lock)) return false; if (WARN_ON(IS_PAGE_ISOLATED(page_md->status))) { spin_unlock(&page_md->migrate_lock); return false; } switch (PAGE_STATUS_GET(page_md->status)) { case MEM_POOL: /* Prepare to remove page from memory pool later only if pool is not * in the process of termination. */ mem_pool = page_md->data.mem_pool.pool; status_mem_pool = true; preempt_disable(); atomic_inc(&mem_pool->isolation_in_progress_cnt); break; case ALLOCATED_MAPPED: /* Mark the page into isolated state, but only if it has no * kernel CPU mappings */ if (page_md->vmap_count == 0) page_md->status = PAGE_ISOLATE_SET(page_md->status, 1); break; case PT_MAPPED: /* Mark the page into isolated state. */ page_md->status = PAGE_ISOLATE_SET(page_md->status, 1); break; case SPILL_IN_PROGRESS: case ALLOCATE_IN_PROGRESS: case FREE_IN_PROGRESS: break; case NOT_MOVABLE: /* Opportunistically clear the movable property for these pages */ __ClearPageMovable(p); page_md->status = PAGE_MOVABLE_CLEAR(page_md->status); break; default: /* State should always fall in one of the previous cases! * Also notice that FREE_ISOLATED_IN_PROGRESS or * FREE_PT_ISOLATED_IN_PROGRESS is impossible because * that state only applies to pages that are already isolated. */ page_md->status = PAGE_ISOLATE_SET(page_md->status, 0); break; } spin_unlock(&page_md->migrate_lock); /* If the page is still in the memory pool: try to remove it. This will fail * if pool lock is taken which could mean page no longer exists in pool. */ if (status_mem_pool) { if (!spin_trylock(&mem_pool->pool_lock)) { atomic_dec(&mem_pool->isolation_in_progress_cnt); preempt_enable(); return false; } spin_lock(&page_md->migrate_lock); /* Check status again to ensure page has not been removed from memory pool. */ if (PAGE_STATUS_GET(page_md->status) == MEM_POOL) { page_md->status = PAGE_ISOLATE_SET(page_md->status, 1); list_del_init(&p->lru); mem_pool->cur_size--; } spin_unlock(&page_md->migrate_lock); spin_unlock(&mem_pool->pool_lock); atomic_dec(&mem_pool->isolation_in_progress_cnt); preempt_enable(); } return IS_PAGE_ISOLATED(page_md->status); } /** * kbase_page_migrate - Migrate content of old page to new page provided. * * @mapping: Pointer to address_space struct associated with pages. * @new_page: Pointer to the page struct of new page. * @old_page: Pointer to the page struct of old page. * @mode: Mode to determine if migration will be synchronised. * * Callback function for Linux to migrate the content of the old page to the * new page provided. * This callback is not registered if compiled without CONFIG_PAGE_MIGRATION_SUPPORT. * * Return: 0 on success, error code otherwise. */ #if (KERNEL_VERSION(6, 0, 0) > LINUX_VERSION_CODE) static int kbase_page_migrate(struct address_space *mapping, struct page *new_page, struct page *old_page, enum migrate_mode mode) #else static int kbase_page_migrate(struct page *new_page, struct page *old_page, enum migrate_mode mode) #endif { int err = 0; bool status_mem_pool = false; bool status_free_pt_isolated_in_progress = false; bool status_free_isolated_in_progress = false; bool status_pt_mapped = false; bool status_mapped = false; bool status_not_movable = false; struct kbase_page_metadata *page_md = kbase_page_private(old_page); struct kbase_device *kbdev = NULL; #if (KERNEL_VERSION(6, 0, 0) > LINUX_VERSION_CODE) CSTD_UNUSED(mapping); #endif CSTD_UNUSED(mode); if (!kbase_is_page_migration_enabled() || !page_md || !IS_PAGE_MOVABLE(page_md->status)) return -EINVAL; if (!spin_trylock(&page_md->migrate_lock)) return -EAGAIN; if (WARN_ON(!IS_PAGE_ISOLATED(page_md->status))) { spin_unlock(&page_md->migrate_lock); return -EINVAL; } switch (PAGE_STATUS_GET(page_md->status)) { case MEM_POOL: status_mem_pool = true; kbdev = page_md->data.mem_pool.kbdev; break; case ALLOCATED_MAPPED: status_mapped = true; break; case PT_MAPPED: status_pt_mapped = true; break; case FREE_ISOLATED_IN_PROGRESS: status_free_isolated_in_progress = true; kbdev = page_md->data.free_isolated.kbdev; break; case FREE_PT_ISOLATED_IN_PROGRESS: status_free_pt_isolated_in_progress = true; kbdev = page_md->data.free_pt_isolated.kbdev; break; case NOT_MOVABLE: status_not_movable = true; break; default: /* State should always fall in one of the previous cases! */ err = -EAGAIN; break; } spin_unlock(&page_md->migrate_lock); if (status_mem_pool || status_free_isolated_in_progress || status_free_pt_isolated_in_progress) { struct kbase_mem_migrate *mem_migrate = &kbdev->mem_migrate; kbase_free_page_metadata(kbdev, old_page, NULL); __ClearPageMovable(old_page); put_page(old_page); /* Just free new page to avoid lock contention. */ INIT_LIST_HEAD(&new_page->lru); get_page(new_page); set_page_private(new_page, 0); kbase_free_page_later(kbdev, new_page); queue_work(mem_migrate->free_pages_workq, &mem_migrate->free_pages_work); } else if (status_not_movable) { err = -EINVAL; } else if (status_mapped) { err = kbasep_migrate_page_allocated_mapped(old_page, new_page); } else if (status_pt_mapped) { err = kbasep_migrate_page_pt_mapped(old_page, new_page); } /* While we want to preserve the movability of pages for which we return * EAGAIN, according to the kernel docs, movable pages for which a critical * error is returned are called putback on, which may not be what we * expect. */ if (err < 0 && err != -EAGAIN) { __ClearPageMovable(old_page); page_md->status = PAGE_MOVABLE_CLEAR(page_md->status); } return err; } /** * kbase_page_putback - Return isolated page back to kbase. * * @p: Pointer of the page struct of page. * * Callback function for Linux to return isolated page back to kbase. This * will only be called for a page that has been isolated but failed to * migrate. This function will put back the given page to the state it was * in before it was isolated. * This callback is not registered if compiled without CONFIG_PAGE_MIGRATION_SUPPORT. */ static void kbase_page_putback(struct page *p) { bool status_mem_pool = false; bool status_free_isolated_in_progress = false; bool status_free_pt_isolated_in_progress = false; struct kbase_page_metadata *page_md = kbase_page_private(p); struct kbase_device *kbdev = NULL; if (!IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT)) return; /* If we don't have page metadata, the page may not belong to the * driver or may already have been freed, and there's nothing we can do */ if (!page_md) return; spin_lock(&page_md->migrate_lock); if (WARN_ON(!IS_PAGE_ISOLATED(page_md->status))) { spin_unlock(&page_md->migrate_lock); return; } switch (PAGE_STATUS_GET(page_md->status)) { case MEM_POOL: status_mem_pool = true; kbdev = page_md->data.mem_pool.kbdev; break; case ALLOCATED_MAPPED: page_md->status = PAGE_ISOLATE_SET(page_md->status, 0); break; case PT_MAPPED: case NOT_MOVABLE: /* Pages should no longer be isolated if they are in a stable state * and used by the driver. */ page_md->status = PAGE_ISOLATE_SET(page_md->status, 0); break; case FREE_ISOLATED_IN_PROGRESS: status_free_isolated_in_progress = true; kbdev = page_md->data.free_isolated.kbdev; break; case FREE_PT_ISOLATED_IN_PROGRESS: status_free_pt_isolated_in_progress = true; kbdev = page_md->data.free_pt_isolated.kbdev; break; default: /* State should always fall in one of the previous cases! */ break; } spin_unlock(&page_md->migrate_lock); /* If page was in a memory pool then just free it to avoid lock contention. The * same is also true to status_free_pt_isolated_in_progress. */ if (status_mem_pool || status_free_isolated_in_progress || status_free_pt_isolated_in_progress) { __ClearPageMovable(p); page_md->status = PAGE_MOVABLE_CLEAR(page_md->status); if (!WARN_ON_ONCE(!kbdev)) { struct kbase_mem_migrate *mem_migrate = &kbdev->mem_migrate; kbase_free_page_later(kbdev, p); queue_work(mem_migrate->free_pages_workq, &mem_migrate->free_pages_work); } } } #if (KERNEL_VERSION(6, 0, 0) <= LINUX_VERSION_CODE) static const struct movable_operations movable_ops = { .isolate_page = kbase_page_isolate, .migrate_page = kbase_page_migrate, .putback_page = kbase_page_putback, }; #else static const struct address_space_operations kbase_address_space_ops = { .isolate_page = kbase_page_isolate, .migratepage = kbase_page_migrate, .putback_page = kbase_page_putback, }; #endif #if (KERNEL_VERSION(6, 0, 0) > LINUX_VERSION_CODE) void kbase_mem_migrate_set_address_space_ops(struct kbase_device *kbdev, struct file *const filp) { if (!kbase_is_page_migration_enabled()) return; mutex_lock(&kbdev->fw_load_lock); if (filp) { filp->f_inode->i_mapping->a_ops = &kbase_address_space_ops; if (!kbdev->mem_migrate.inode) { kbdev->mem_migrate.inode = filp->f_inode; /* This reference count increment is balanced by iput() * upon termination. */ atomic_inc(&filp->f_inode->i_count); } else { WARN_ON(kbdev->mem_migrate.inode != filp->f_inode); } } mutex_unlock(&kbdev->fw_load_lock); } #endif void kbase_mem_migrate_init(struct kbase_device *kbdev) { #if !IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT) /* Page migration explicitly disabled at compile time - do nothing */ return; #else struct kbase_mem_migrate *mem_migrate = &kbdev->mem_migrate; /* Page migration support compiled in, either explicitly or * by default, so the default behaviour is to follow the choice * of large pages if not selected at insmod. Check insmod parameter * integer for a negative value to see if insmod parameter was * passed in at all (it will override the default negative value). */ if (kbase_page_migration_enabled < 0) kbase_page_migration_enabled = kbdev->pagesize_2mb ? 1 : 0; else dev_info(kbdev->dev, "Page migration support explicitly %s at insmod.", kbase_page_migration_enabled ? "enabled" : "disabled"); spin_lock_init(&mem_migrate->free_pages_lock); INIT_LIST_HEAD(&mem_migrate->free_pages_list); #if (KERNEL_VERSION(6, 0, 0) > LINUX_VERSION_CODE) mem_migrate->inode = NULL; #endif mem_migrate->free_pages_workq = alloc_workqueue("free_pages_workq", WQ_UNBOUND | WQ_MEM_RECLAIM, 1); INIT_WORK(&mem_migrate->free_pages_work, kbase_free_pages_worker); #endif } void kbase_mem_migrate_term(struct kbase_device *kbdev) { struct kbase_mem_migrate *mem_migrate = &kbdev->mem_migrate; #if !IS_ENABLED(CONFIG_PAGE_MIGRATION_SUPPORT) /* Page migration explicitly disabled at compile time - do nothing */ return; #endif if (mem_migrate->free_pages_workq) destroy_workqueue(mem_migrate->free_pages_workq); #if (KERNEL_VERSION(6, 0, 0) > LINUX_VERSION_CODE) iput(mem_migrate->inode); #endif }