/* SPDX-License-Identifier: GPL-2.0-only */ /* Copyright (C) 2013--2026 Intel Corporation */ #ifndef IPU6_MMU_H #define IPU6_MMU_H #include <linux/list.h> #include <linux/spinlock_types.h> #include <linux/types.h> #include "ipu7-mmu-hw.h" struct device; struct ipu6_mmu; struct page; struct ipu6_mmu_info { struct device *dev; u32 *l1_pt; u32 l1_pt_dma; u32 **l2_pts; u32 *dummy_l2_pt; u32 dummy_l2_pteval; void *dummy_page; u32 dummy_page_pteval; dma_addr_t aperture_start; dma_addr_t aperture_end; unsigned long pgsize_bitmap; spinlock_t lock; /* Serialize access to users */ struct ipu6_dma_mapping *dmap; }; /* * MMU Invalidation HW bug workaround by ZLW mechanism * * Old IPU6 MMUV2 has a bug in the invalidation mechanism which might result in * wrong translation or replication of the translation. This will cause data * corruption. So we cannot directly use the MMU V2 invalidation registers * to invalidate the MMU. Instead, whenever an invalidate is called, we need to * clear the TLB by evicting all the valid translations by filling it with trash * buffer (which is guaranteed not to be used by any other processes). ZLW is * used to fill the L1 and L2 caches with the trash buffer translations. ZLW * or Zero length write, is pre-fetch mechanism to pre-fetch the pages in * advance to the L1 and L2 caches without triggering any memory operations. * * In MMU V2, L1 -> 16 streams and 64 blocks, maximum 16 blocks per stream * One L1 block has 16 entries, hence points to 16 * 4K pages * L2 -> 16 streams and 32 blocks. 2 blocks per streams * One L2 block maps to 1024 L1 entries, hence points to 4MB address range * 2 blocks per L2 stream means, 1 stream points to 8MB range * * As we need to clear the caches and 8MB being the biggest cache size, we need * to have trash buffer which points to 8MB address range. As these trash * buffers are not used for any memory transactions, we need only the least * amount of physical memory. So we reserve 8MB IOVA address range but only * one page is reserved from physical memory. Each of this 8MB IOVA address * range is then mapped to the same physical memory page. */ /* One L2 entry maps 1024 L1 entries and one L1 entry per page */ #define IPU6_MMUV2_L2_RANGE (1024 * PAGE_SIZE) /* Max L2 blocks per stream */ #define IPU6_MMUV2_MAX_L2_BLOCKS 2 /* Max L1 blocks per stream */ #define IPU6_MMUV2_MAX_L1_BLOCKS 16 #define IPU6_MMUV2_TRASH_RANGE (IPU6_MMUV2_L2_RANGE * IPU6_MMUV2_MAX_L2_BLOCKS) /* Entries per L1 block */ #define MMUV2_ENTRIES_PER_L1_BLOCK 16 #define MMUV2_TRASH_L1_BLOCK_OFFSET (MMUV2_ENTRIES_PER_L1_BLOCK * PAGE_SIZE) #define MMUV2_TRASH_L2_BLOCK_OFFSET IPU6_MMUV2_L2_RANGE /* * In some of the IPU6 MMUs, there is provision to configure L1 and L2 page * table caches. Both these L1 and L2 caches are divided into multiple sections * called streams. There is maximum 16 streams for both caches. Each of these * sections are subdivided into multiple blocks. When nr_l1streams = 0 and * nr_l2streams = 0, means the MMU is of type MMU_V1 and do not support * L1/L2 page table caches. * * L1 stream per block sizes are configurable and varies per usecase. * L2 has constant block sizes - 2 blocks per stream. * * MMU1 support pre-fetching of the pages to have less cache lookup misses. To * enable the pre-fetching, MMU1 AT (Address Translator) device registers * need to be configured. * * There are four types of memory accesses which requires ZLW configuration. * ZLW(Zero Length Write) is a mechanism to enable VT-d pre-fetching on IOMMU. * * 1. Sequential Access or 1D mode * Set ZLW_EN -> 1 * set ZLW_PAGE_CROSS_1D -> 1 * Set ZLW_N to "N" pages so that ZLW will be inserte N pages ahead where * N is pre-defined and hardcoded in the platform data * Set ZLW_2D -> 0 * * 2. ZLW 2D mode * Set ZLW_EN -> 1 * set ZLW_PAGE_CROSS_1D -> 1, * Set ZLW_N -> 0 * Set ZLW_2D -> 1 * * 3. ZLW Enable (no 1D or 2D mode) * Set ZLW_EN -> 1 * set ZLW_PAGE_CROSS_1D -> 0, * Set ZLW_N -> 0 * Set ZLW_2D -> 0 * * 4. ZLW disable * Set ZLW_EN -> 0 * set ZLW_PAGE_CROSS_1D -> 0, * Set ZLW_N -> 0 * Set ZLW_2D -> 0 * * To configure the ZLW for the above memory access, four registers are * available. Hence to track these four settings, we have the following entries * in the struct ipu6_mmu_hw. Each of these entries are per stream and * available only for the L1 streams. * * a. l1_zlw_en -> To track zlw enabled per stream (ZLW_EN) * b. l1_zlw_1d_mode -> Track 1D mode per stream. ZLW inserted at page boundary * c. l1_ins_zlw_ahead_pages -> to track how advance the ZLW need to be inserted * Insert ZLW request N pages ahead address. * d. l1_zlw_2d_mode -> To track 2D mode per stream (ZLW_2D) * * * Currently L1/L2 streams, blocks, AT ZLW configurations etc. are pre-defined * as per the usecase specific calculations. Any change to this pre-defined * table has to happen in sync with IPU6 FW. */ struct ipu6_mmu_hw { union { unsigned long offset; void __iomem *base; }; u32 info_bits; u8 nr_l1streams; /* * L1 has variable blocks per stream - total of 64 blocks and maximum of * 16 blocks per stream. Configurable by using the block start address * per stream. Block start address is calculated from the block size */ u8 l1_block_sz[IPU6_MMU_MAX_TLB_L1_STREAMS]; /* Is ZLW is enabled in each stream */ bool l1_zlw_en[IPU6_MMU_MAX_TLB_L1_STREAMS]; bool l1_zlw_1d_mode[IPU6_MMU_MAX_TLB_L1_STREAMS]; u8 l1_ins_zlw_ahead_pages[IPU6_MMU_MAX_TLB_L1_STREAMS]; bool l1_zlw_2d_mode[IPU6_MMU_MAX_TLB_L1_STREAMS]; u32 l1_stream_id_reg_offset; u32 l2_stream_id_reg_offset; u8 nr_l2streams; /* * L2 has fixed 2 blocks per stream. Block address is calculated * from the block size */ u8 l2_block_sz[IPU6_MMU_MAX_TLB_L2_STREAMS]; /* flag to track if WA is needed for successive invalidate HW bug */ bool insert_read_before_invalidate; }; struct ipu6_mmu_hw_ops { int (*init_hw_data)(struct ipu6_mmu *mmu, struct device *dev, void __iomem *base); int (*hw_init)(struct ipu6_mmu *mmu); void (*tlb_invalidate)(struct ipu6_mmu *mmu); }; struct ipu6_mmu { struct list_head node; union { struct ipu6_mmu_hw *ipu6_mmu_hw; struct ipu7_mmu_hw *ipu7_mmu_hw; }; unsigned int nr_mmus; unsigned int mmid; phys_addr_t pgtbl; struct device *dev; struct ipu6_dma_mapping *dmap; struct list_head vma_list; struct page *trash_page; dma_addr_t pci_trash_page; /* IOVA from PCI DMA services (parent) */ dma_addr_t iova_trash_page; /* IOVA for IPU6 child nodes to use */ bool ready; spinlock_t ready_lock; /* Serialize access to bool ready */ const struct ipu6_mmu_hw_ops *ops; }; extern const struct ipu6_mmu_hw_ops ipu6_mmu_ops; extern const struct ipu6_mmu_hw_ops ipu7_mmu_ops; struct ipu6_mmu *ipu6_mmu_init(struct device *dev, void __iomem *base, int mmid); void ipu6_mmu_cleanup(struct ipu6_mmu *mmu); int ipu6_mmu_hw_init(struct ipu6_mmu *mmu); void ipu6_mmu_hw_cleanup(struct ipu6_mmu *mmu); int ipu6_mmu_map(struct ipu6_mmu_info *mmu_info, unsigned long iova, phys_addr_t paddr, size_t size); void ipu6_mmu_unmap(struct ipu6_mmu_info *mmu_info, unsigned long iova, size_t size); phys_addr_t ipu6_mmu_iova_to_phys(struct ipu6_mmu_info *mmu_info, dma_addr_t iova); #endif