Linux中DMA会使用硬件IOMMU如AMD IOMMU, INTEL VT-D， 也会使用软件的SWIOTLB
这篇梳理一下LINUX内核在有AMD IOMMU的情况下，是如何做DMA的，内容包括如下
1. struct iommu_ops amd_iommu_ops
2. struct dma_map_ops iommu_dma_ops
3. DMA struct dma_map_ops 与 struct iommu_ops的关系
Consistent, Streaming
4. struct io_pgtable_ops, 及与struct iommu_ops amd_iommu_ops的关系
1. 两处会设置struct iommu_ops amd_iommu_ops;
const struct iommu_ops amd_iommu_ops = {
.capable = amd_iommu_capable,
.domain_alloc = amd_iommu_domain_alloc, // 分配一个iommu_domain
.domain_free = amd_iommu_domain_free,
.attach_dev = amd_iommu_attach_device, //针对独立设备（即所在Group里只有自己），将设备所在Group与domain进行绑定
.detach_dev = amd_iommu_detach_device,
.map = amd_iommu_map, //用于映射domain内的iova，将长度为size以iova为起始地址的iova区域映射到以paddr为起始地址的物理地址。该函数只能用于UNMANAGED类型和DMA类型的domain
.iotlb_sync_map = amd_iommu_iotlb_sync_map,
.unmap = amd_iommu_unmap,
.iova_to_phys = amd_iommu_iova_to_phys, // 将iova转换成物理地址
.probe_device = amd_iommu_probe_device,
.release_device = amd_iommu_release_device,
.probe_finalize = amd_iommu_probe_finalize,
.device_group = amd_iommu_device_group,
.get_resv_regions = amd_iommu_get_resv_regions,
.put_resv_regions = generic_iommu_put_resv_regions,
.is_attach_deferred = amd_iommu_is_attach_deferred,
.pgsize_bitmap = AMD_IOMMU_PGSIZES,
.flush_iotlb_all = amd_iommu_flush_iotlb_all,
.iotlb_sync = amd_iommu_iotlb_sync,
.def_domain_type = amd_iommu_def_domain_type,
};
一处在struct iommu_device的iommu ops;
另一处在struct bus_type的iommu ops;
amd_iommu_init ->
iommu_go_to_state ->
state_next ->
amd_iommu_init_pci ->
iommu_init_pci ->
iommu_device_register
struct iommu_device {
struct list_head list;
const struct iommu_ops *ops;
struct fwnode_handle *fwnode;
struct device *dev;
};
/*
* Structure where we save information about one hardware AMD IOMMU in the
* system.
*/
struct amd_iommu {
...
struct iommu_device iommu;/* Handle for IOMMU core code */
...
}
amd_iommu_init ->
iommu_go_to_state ->
state_next ->
amd_iommu_init_pci ->
amd_iommu_init_api ->
bus_set_iommu -> //将自身挂入到 对应总线中
struct bus_type {
...
const struct iommu_ops *iommu_ops;
...
};
2. 设置struct dma_map_ops iommu_dma_ops
amd_iommu_ops.probe_finalize = amd_iommu_probe_finalize ->
iommu_setup_dma_ops->
struct device *dev->dma_ops = &iommu_dma_ops;
struct device {
...
const struct dma_map_ops *dma_ops; // DMA mapping operations for this device
...
};
static const struct dma_map_ops iommu_dma_ops = {
.alloc = iommu_dma_alloc,
.free = iommu_dma_free,
.alloc_pages = dma_common_alloc_pages,
.free_pages = dma_common_free_pages,
#ifdef CONFIG_DMA_REMAP
.alloc_noncontiguous = iommu_dma_alloc_noncontiguous,
.free_noncontiguous = iommu_dma_free_noncontiguous,
#endif
.mmap = iommu_dma_mmap,
.get_sgtable = iommu_dma_get_sgtable,
.map_page = iommu_dma_map_page,
.unmap_page = iommu_dma_unmap_page,
.map_sg = iommu_dma_map_sg,
.unmap_sg = iommu_dma_unmap_sg,
.sync_single_for_cpu = iommu_dma_sync_single_for_cpu,
.sync_single_for_device = iommu_dma_sync_single_for_device,
.sync_sg_for_cpu = iommu_dma_sync_sg_for_cpu,
.sync_sg_for_device = iommu_dma_sync_sg_for_device,
.map_resource = iommu_dma_map_resource,
.unmap_resource = iommu_dma_unmap_resource,
.get_merge_boundary = iommu_dma_get_merge_boundary,
};
iommu_dma_alloc ->
void *cpu_addr = iommu_dma_alloc_pages
dma_addr_t iova = __iommu_dma_map (-> iommu_dma_alloc_iova)
3. DMA -- struct dma_map_ops 与 struct iommu_ops
如果写过LINUX DMA驱动，会接触过以下几个函数：
Consistent
void * dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flag)
Streaming
dma_addr_t dma_map_single(struct device *dev, void *cpu_addr, size_t size, enum dma_data_direction direction)
dma_map_page
dma_map_sg
其中dma_alloc_coherent，的调用栈为：
dma_alloc_coherent ->
dma_alloc_attrs ->
struct dma_map_ops *ops->alloc == iommu_dma_alloc ->
__iommu_dma_map ->
iommu_map_atomic ->
_iommu_map->
__iommu_map->
__iommu_map_pages->
struct iommu_ops *ops->map_pages/map
这里的iommu_ops.map 在使用AMD IOMMU的时候，就是.map = amd_iommu_map,
#define dma_map_single(d, a, s, r) dma_map_single_attrs(d, a, s, r, 0)
#define dma_unmap_single(d, a, s, r) dma_unmap_single_attrs(d, a, s, r, 0)
#define dma_map_sg(d, s, n, r) dma_map_sg_attrs(d, s, n, r, 0)
#define dma_unmap_sg(d, s, n, r) dma_unmap_sg_attrs(d, s, n, r, 0)
#define dma_map_page(d, p, o, s, r) dma_map_page_attrs(d, p, o, s, r, 0)
#define dma_unmap_page(d, a, s, r) dma_unmap_page_attrs(d, a, s, r, 0)
对应到dma_map_single的调用栈为：
dma_map_single_attrs->
dma_map_page_attrs->
struct dma_map_ops *ops->map_page == iommu_dma_map_page ->
__iommu_dma_map_swiotlb ->
__iommu_dma_map ->
之后与dma_alloc_coherent相同
4. struct io_pgtable_ops
amd_iommu_map最终的实现在struct io_pgtable_ops
amd_iommu_map ->
struct io_pgtable_ops *ops->map
struct io_pgtable_ops {
int (*map)(struct io_pgtable_ops *ops, unsigned long iova,
phys_addr_t paddr, size_t size, int prot, gfp_t gfp);
int (*map_pages)(struct io_pgtable_ops *ops, unsigned long iova,
phys_addr_t paddr, size_t pgsize, size_t pgcount,
int prot, gfp_t gfp, size_t *mapped);
size_t (*unmap)(struct io_pgtable_ops *ops, unsigned long iova,
size_t size, struct iommu_iotlb_gather *gather);
size_t (*unmap_pages)(struct io_pgtable_ops *ops, unsigned long iova,
size_t pgsize, size_t pgcount,
struct iommu_iotlb_gather *gather);
phys_addr_t (*iova_to_phys)(struct io_pgtable_ops *ops,
unsigned long iova);
};
struct io_pgtable {
enum io_pgtable_fmt fmt;
void *cookie;
struct io_pgtable_cfg cfg;
struct io_pgtable_ops ops;
};
struct amd_io_pgtable {
struct io_pgtable_cfg pgtbl_cfg;
struct io_pgtable iop;
...
};
struct protection_domain {
...
struct iommu_domain domain; /* generic domain handle used by
iommu core code */
struct amd_io_pgtable iop;
...
};
struct amd_io_pgtable *pgtable
pgtable->iop.ops.map = iommu_v1_map_page; //maps a physical address into a DMA
address space. It allocates the page table pages if necessary, 建立DMA addr与paddr的page table
pgtable->iop.ops.unmap = iommu_v1_unmap_page;
pgtable->iop.ops.iova_to_phys = iommu_v1_iova_to_phys;
5. Summary
系统中的调用关系如下：
struct dma_map_ops iommu_dma_ops ->
struct iommu_ops amd_iommu_op ->
struct io_pgtable_ops
iommu是实现在dma mapping api下层的驱动，所以我们只需要使用dma mapping的相关api，不需要直接调用iommu接口
AMD IOMMU驱动实现了自己的struct io_pgtable_ops
类似在内核中的还有ARM SMMU与Apple DART等，详见include/linux/io-pgtable.h文件
Reference:
[1]
dma_map_ops 实现的三种方式_jason的笔记-CSDN博客_dma map
[2]
kernel是如何选择iommu的呢？_jason的笔记-CSDN博客[
[3]
iommu_dma_mmap + mmap - tycoon3 - 博客园 (cnblogs)
【4】
Documentation/core-api/dma-api-howto.rst
Documentation/core-api/dma-api.rst
Notes:
1. 处于同一个domain中的设备使用同一套映射做地址转换, 就是独立的页表
2. Group中default_domain和domain的概念：domain指group当前所在的domain，而default_domain指Group默认应该在的domain
进行attach操作时，会检查default_domain是否与domain相同，以此判断该Group是否已经attach到别的domain上了
如果Group有自己的default_domain，那么该函数iommu_detach_device在detach完成之后会重新attach到default_domain上
3. PCIE是一个点对点的协议，如果一个多function设备挂到了一个不支持ACS的bridge下，那么这两个function可以通过该bridge进行通信。这样的通信直接由bridge进行转发而无需通过Root Complex，自然也就无需通过IOMMU。这种情况下，这两个function的IOVA无法完全通过IOMMU隔离开，所以他们需要分到同一个Group中。同一个Group的设备应该是共用一个domain的
4. 每一个domain即代表一个iommu映射地址空间，即一个page table。一个Group逻辑上是需要与domain进行绑定的，即一个Group中的所有设备都位于一个domain中
Questions:
The difference between IOMMU_DOMAIN_UNMANAGED & IOMMU_DOMAIN_DMA?
/*
* This are the possible domain-types
*
* IOMMU_DOMAIN_BLOCKED - All DMA is blocked, can be used to isolate
* devices
* IOMMU_DOMAIN_IDENTITY - DMA addresses are system physical addresses
* IOMMU_DOMAIN_UNMANAGED - DMA mappings managed by IOMMU-API user, used
* for VMs
* IOMMU_DOMAIN_DMA - Internally used for DMA-API implementations.
* This flag allows IOMMU drivers to implement
* certain optimizations for these domains
* IOMMU_DOMAIN_DMA_FQ - As above, but definitely using batched TLB
* invalidation.
*/
更多推荐
AMD IOMMU与Linux (3) -- DMA
发布评论