From: Zhou Guanghui zhouguanghui1@huawei.com

ascend inclusion category: feature bugzilla: https://gitee.com/openeuler/kernel/issues/I931MI CVE: NA

-----------------------------------------------------------------

A range is reserved from the user address space to ensure that the process cannot apply for the address range.

Signed-off-by: Zhou Guanghui zhouguanghui1@huawei.com --- fs/hugetlbfs/inode.c | 13 ++++++++ include/linux/mman.h | 43 ++++++++++++++++++++++++++ include/uapi/asm-generic/mman-common.h | 1 + mm/mmap.c | 33 ++++++++++++++++++++ 4 files changed, 90 insertions(+)

diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c index f001dfe52a14..f151370afdc8 100644 --- a/fs/hugetlbfs/inode.c +++ b/fs/hugetlbfs/inode.c @@ -197,6 +197,9 @@ hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr, info.high_limit = arch_get_mmap_end(addr, len, flags); info.align_mask = PAGE_MASK & ~huge_page_mask(h); info.align_offset = 0; + + svsp_mmap_get_area(&info, flags); + return vm_unmapped_area(&info); }

@@ -213,6 +216,9 @@ hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr, info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base); info.align_mask = PAGE_MASK & ~huge_page_mask(h); info.align_offset = 0; + + svsp_mmap_get_area(&info, flags); + addr = vm_unmapped_area(&info);

/* @@ -226,6 +232,9 @@ hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr, info.flags = 0; info.low_limit = current->mm->mmap_base; info.high_limit = arch_get_mmap_end(addr, len, flags); + + svsp_mmap_get_area(&info, flags); + addr = vm_unmapped_area(&info); }

@@ -255,6 +264,10 @@ generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,

if (addr) { addr = ALIGN(addr, huge_page_size(h)); + + if (svsp_mmap_check(addr, len, flags)) + return -ENOMEM; + vma = find_vma(mm, addr); if (mmap_end - len >= addr && (!vma || addr + len <= vm_start_gap(vma))) diff --git a/include/linux/mman.h b/include/linux/mman.h index 40d94411d492..9f0250993ef7 100644 --- a/include/linux/mman.h +++ b/include/linux/mman.h @@ -8,6 +8,49 @@ #include <linux/atomic.h> #include <uapi/linux/mman.h>

+extern int enable_mmap_svsp; + +#ifdef CONFIG_ASCEND_SVSP +#define SVSP_MMAP_BASE 0x180000000000UL +#define SVSP_MMAP_SIZE 0x080000000000UL + +static inline int svsp_mmap_check(unsigned long addr, unsigned long len, + unsigned long flags) +{ + if (enable_mmap_svsp && (flags & MAP_SVSP) && + (addr < (SVSP_MMAP_BASE + SVSP_MMAP_SIZE)) && + (addr > SVSP_MMAP_BASE)) + return -EINVAL; + else + return 0; +} + +static inline void svsp_mmap_get_area(struct vm_unmapped_area_info *info, + unsigned long flags) +{ + if (enable_mmap_svsp && (flags & MAP_SVSP)) { + info->low_limit = SVSP_MMAP_BASE; + info->high_limit = SVSP_MMAP_BASE + SVSP_MMAP_SIZE; + } else { + info->low_limit = max(info->low_limit, TASK_UNMAPPED_BASE); + info->high_limit = min(info->high_limit, SVSP_MMAP_BASE); + } +} + +#else +#define SVSP_MMAP_BASE 0 +#define SVSP_MMAP_SIZE 0 +static inline int svsp_mmap_check(unsigned long addr, unsigned long len, + unsigned long flags) +{ + return 0; +} + +static inline void svsp_mmap_get_area(struct vm_unmapped_area_info *info, + unsigned long flags) +{ } +#endif + /* * Arrange for legacy / undefined architecture specific flags to be * ignored by mmap handling code. diff --git a/include/uapi/asm-generic/mman-common.h b/include/uapi/asm-generic/mman-common.h index 14e5498efd7a..096c9018d2d5 100644 --- a/include/uapi/asm-generic/mman-common.h +++ b/include/uapi/asm-generic/mman-common.h @@ -30,6 +30,7 @@ #define MAP_SYNC 0x080000 /* perform synchronous page faults for the mapping */ #define MAP_FIXED_NOREPLACE 0x100000 /* MAP_FIXED which doesn't unmap underlying mapping */

+#define MAP_SVSP 0x400000 #define MAP_UNINITIALIZED 0x4000000 /* For anonymous mmap, memory could be * uninitialized */

diff --git a/mm/mmap.c b/mm/mmap.c index 5d78ebaeed60..94077c5356eb 100644 --- a/mm/mmap.c +++ b/mm/mmap.c @@ -1728,6 +1728,10 @@ generic_get_unmapped_area(struct file *filp, unsigned long addr,

if (addr) { addr = PAGE_ALIGN(addr); + + if (svsp_mmap_check(addr, len, flags)) + return -ENOMEM; + vma = find_vma_prev(mm, addr, &prev); if (mmap_end - len >= addr && addr >= mmap_min_addr && (!vma || addr + len <= vm_start_gap(vma)) && @@ -1741,6 +1745,9 @@ generic_get_unmapped_area(struct file *filp, unsigned long addr, info.high_limit = mmap_end; info.align_mask = 0; info.align_offset = 0; + + svsp_mmap_get_area(&info, flags); + return vm_unmapped_area(&info); }

@@ -1781,6 +1788,10 @@ generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr, /* requesting a specific address */ if (addr) { addr = PAGE_ALIGN(addr); + + if (svsp_mmap_check(addr, len, flags)) + return -ENOMEM; + vma = find_vma_prev(mm, addr, &prev); if (mmap_end - len >= addr && addr >= mmap_min_addr && (!vma || addr + len <= vm_start_gap(vma)) && @@ -1794,6 +1805,9 @@ generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr, info.high_limit = arch_get_mmap_base(addr, mm->mmap_base); info.align_mask = 0; info.align_offset = 0; + + svsp_mmap_get_area(&info, flags); + addr = vm_unmapped_area(&info);

/* @@ -1807,6 +1821,9 @@ generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr, info.flags = 0; info.low_limit = TASK_UNMAPPED_BASE; info.high_limit = mmap_end; + + svsp_mmap_get_area(&info, flags); + addr = vm_unmapped_area(&info); }

@@ -3946,3 +3963,19 @@ static int __meminit init_reserve_notifier(void) return 0; } subsys_initcall(init_reserve_notifier); + +int enable_mmap_svsp __read_mostly; + +#ifdef CONFIG_ASCEND_SVSP + +static int __init ascend_enable_mmap_svsp(char *s) +{ + enable_mmap_svsp = 1; + + pr_info("Ascend enable svsp mmap features\n"); + + return 1; +} +__setup("enable_mmap_svsp", ascend_enable_mmap_svsp); + +#endif

From: Jian Zhang zhangjian210@huawei.com

ascend inclusion category: feature bugzilla: https://gitee.com/openeuler/kernel/issues/I931MI CVE: NA

----------------------------------

When using svm_svsp_bind_core, the expected pasid is the pasid of the origin mm with its highest bit set, so we need to check and get the expected pasid instead of allocating a new one.

Signed-off-by: Jian Zhang zhangjian210@huawei.com Signed-off-by: Yuan Can yuancan@huawei.com --- drivers/iommu/iommu-sva.c | 4 ++++ drivers/iommu/iommu.c | 36 +++++++++++++++++++++++++++++++++++- include/linux/iommu.h | 3 +++ 3 files changed, 42 insertions(+), 1 deletion(-)

diff --git a/drivers/iommu/iommu-sva.c b/drivers/iommu/iommu-sva.c index b78671a8a914..99b2977ff708 100644 --- a/drivers/iommu/iommu-sva.c +++ b/drivers/iommu/iommu-sva.c @@ -25,6 +25,10 @@ static int iommu_sva_alloc_pasid(struct mm_struct *mm, struct device *dev) if (mm_valid_pasid(mm)) { if (mm->pasid >= dev->iommu->max_pasids) ret = -EOVERFLOW; +#ifdef CONFIG_ASCEND_SVSP + else + ret = iommu_reserve_svsp_pasid(mm, dev); +#endif goto out; }

diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c index 32f0f6c461b3..594d9e984d57 100644 --- a/drivers/iommu/iommu.c +++ b/drivers/iommu/iommu.c @@ -3697,20 +3697,54 @@ struct iommu_domain *iommu_sva_domain_alloc(struct device *dev, return domain; }

+#ifdef CONFIG_ASCEND_SVSP +int iommu_reserve_svsp_pasid(struct mm_struct *mm, struct device *dev) +{ + int ret; + int max_pasids = dev->iommu->max_pasids; + + /* This should never happen. */ + if (!max_pasids) + return -ENOSPC; + + /* Do nothing if the given PASID not in SVSP pasid range. */ + if (mm->pasid < max_pasids >> 1) + return 0; + + /* Reserve the given SVSP pasid. */ + ret = ida_alloc_range(&iommu_global_pasid_ida, mm->pasid, mm->pasid, + GFP_KERNEL); + if (ret == -ENOMEM) + return -ENOSPC; + + return 0; +} +#endif + ioasid_t iommu_alloc_global_pasid(struct device *dev) { int ret; + int max_pasids;

/* max_pasids == 0 means that the device does not support PASID */ if (!dev->iommu->max_pasids) return IOMMU_PASID_INVALID;

+ max_pasids = dev->iommu->max_pasids - 1; +#ifdef CONFIG_ASCEND_SVSP + /* + * When SVSP feature is enabled, the pasid space is divided into two + * equal parts, the normal pasids are distributed in the range with + * smaller value. + */ + max_pasids = max_pasids >> 1; +#endif /* * max_pasids is set up by vendor driver based on number of PASID bits * supported but the IDA allocation is inclusive. */ ret = ida_alloc_range(&iommu_global_pasid_ida, IOMMU_FIRST_GLOBAL_PASID, - dev->iommu->max_pasids - 1, GFP_KERNEL); + max_pasids, GFP_KERNEL); return ret < 0 ? IOMMU_PASID_INVALID : ret; } EXPORT_SYMBOL_GPL(iommu_alloc_global_pasid); diff --git a/include/linux/iommu.h b/include/linux/iommu.h index a1be18a687b6..e6f22dfa0205 100644 --- a/include/linux/iommu.h +++ b/include/linux/iommu.h @@ -833,6 +833,9 @@ iommu_get_domain_for_dev_pasid(struct device *dev, ioasid_t pasid, unsigned int type); ioasid_t iommu_alloc_global_pasid(struct device *dev); void iommu_free_global_pasid(ioasid_t pasid); +#ifdef CONFIG_ASCEND_SVSP +int iommu_reserve_svsp_pasid(struct mm_struct *mm, struct device *dev); +#endif #else /* CONFIG_IOMMU_API */

struct iommu_ops {};

driver/svm: 这个标题是不是要改下

在 2024/2/22 21:25, Yuan Can 写道:

From: Zhou Guanghui zhouguanghui1@huawei.com

ascend inclusion category: feature bugzilla: https://gitee.com/openeuler/kernel/issues/I931MI CVE: NA

---

In the SVA scenario, the master device of the SMMU can access the address space of the process by the operation of binding process. The page table of the process is shared by the MMU and SMMU, In this case, if some addrss range cannot be accessed through the MMU, access to this address range needs to be isolated from the process page table.

If the default domain of the SMMU is used, the isolation can be achieved. However, the address space created by default domain do not like the address space of the process which can be shared by multiple SMMUs.

Therefore, we add a shadow virtual space of process. This address space is used only by the SMMU.

Signed-off-by: Zhou Guanghui zhouguanghui1@huawei.com

arch/arm64/Kconfig | 15 +++++++++++++++ kernel/fork.c | 1 + mm/mmap.c | 1 + 3 files changed, 17 insertions(+)

diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig index 67bc2ad13453..3e84b2572827 100644 --- a/arch/arm64/Kconfig +++ b/arch/arm64/Kconfig @@ -2455,6 +2455,21 @@ config DMI However, even with this option, the resultant kernel should continue to boot on existing non-UEFI platforms.

+config ASCEND_SVSP

• bool "Enable support for SVSP"

• default n

• depends on ARM_SMMU_V3_SVA

• help

•  SVSP means Shadow Virtual Space of Process, which is an IO space used for
  

•  master device of SMMU.
  

•  This space is only used for the master device, since the target address
  

•  after traslation by MMU is not an accessible physical address, but can
  

•  accessible by SMMU.
  

•  Therefore, in the sva scenario, an svsp is created for each target process
  

•  address space.
  

• endmenu # "Boot options"

  config COMPAT

diff --git a/kernel/fork.c b/kernel/fork.c index 00e5886c9e04..735bc670b49e 100644 --- a/kernel/fork.c +++ b/kernel/fork.c @@ -1373,6 +1373,7 @@ struct mm_struct *mm_alloc(void) memset(mm, 0, sizeof(*mm)); return mm_init(mm, current, current_user_ns()); } +EXPORT_SYMBOL_GPL(mm_alloc);

static inline void __mmput(struct mm_struct *mm) { diff --git a/mm/mmap.c b/mm/mmap.c index 21d1fc39bf21..5d78ebaeed60 100644 --- a/mm/mmap.c +++ b/mm/mmap.c @@ -2684,6 +2684,7 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,

return do_vmi_munmap(&vmi, mm, start, len, uf, false); } +EXPORT_SYMBOL(do_munmap);

static unsigned long __mmap_region(struct mm_struct *mm, struct file *file, unsigned long addr,