tmakin/ack-tegra
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

ANDROID: KVM: arm64: Fix MMIO guard alignment for guests

Browse Source 
Currently pkvm_granule alignment is enforced for the start address of
the ioremap hook. This is a problem as this address doesn't have to be
aligned at all. At the same time, the size of the ioremap isn't verified
and we can simply overshoot the MMIO guard original request.

MMIO guard is solely here to indicate to the hypervisor where are the
MMIO regions. If we validate the RAM is aligned with the pkvm_granule,
we can safely overshoot the MMIO guard.

Bug: 381400679
Bug: 357781595
Fixes: 13c871aec2 ("ANDROID: KVM: arm64: Allow the pVM guest to boot with different granule")
Tested-by: Mostafa Saleh <smostafa@google.com>
Change-Id: I0ae27c1626fab17a3b58a6004b6b4f31c23c61a3
Signed-off-by: Vincent Donnefort <vdonnefort@google.com>
This commit is contained in:
Vincent Donnefort
2025-01-28 14:23:31 +00:00
parent 3ea0fa42a6
commit 4e27d44ab9
1 changed files with 48 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files
drivers/virt/coco/pkvm-guest/arm-pkvm-guest.c
+48 -15
View File
@@ -12,6 +12,7 @@
#include <linux/io.h>
#include <linux/mem_encrypt.h>
#include <linux/mem_relinquish.h>
#include <linux/memblock.h>
#include <linux/mm.h>
#include <linux/pgtable.h>
@@ -53,9 +54,6 @@ static int __arm_smccc_do_range(u32 func_id, phys_addr_t phys, int numgranules)
/*
* Apply func_id on the range [phys : phys + numpages * PAGE_SIZE)
*
* Use with cautious. Boundaries of the range will be aligned with pkvm_granule.
* This might lead to overshooting when pkvm_granule > PAGE_SIZE.
*/
static int arm_smccc_do_range(u32 func_id, phys_addr_t phys, int numpages,
bool func_has_range)
@@ -63,11 +61,13 @@ static int arm_smccc_do_range(u32 func_id, phys_addr_t phys, int numpages,
size_t size = numpages * PAGE_SIZE;
int numgranules;
if (!IS_ALIGNED(phys, pkvm_granule))
if (!IS_ALIGNED(phys, PAGE_SIZE))
return -EINVAL;
numgranules = DIV_ROUND_UP(size, pkvm_granule);
phys = ALIGN_DOWN(phys, pkvm_granule);
if (!IS_ALIGNED(phys | size, pkvm_granule))
return -EINVAL;
numgranules = size / pkvm_granule;
if (func_has_range)
return __arm_smccc_do_range(func_id, phys, numgranules);
@@ -77,18 +77,12 @@ static int arm_smccc_do_range(u32 func_id, phys_addr_t phys, int numpages,
static int pkvm_set_memory_encrypted(unsigned long addr, int numpages)
{
if (!IS_ALIGNED(numpages * PAGE_SIZE, pkvm_granule))
return -EINVAL;
return arm_smccc_do_range(ARM_SMCCC_VENDOR_HYP_KVM_MEM_UNSHARE_FUNC_ID,
virt_to_phys((void *)addr), numpages, pkvm_func_range);
}
static int pkvm_set_memory_decrypted(unsigned long addr, int numpages)
{
if (!IS_ALIGNED(numpages * PAGE_SIZE, pkvm_granule))
return -EINVAL;
return arm_smccc_do_range(ARM_SMCCC_VENDOR_HYP_KVM_MEM_SHARE_FUNC_ID,
virt_to_phys((void *)addr), numpages, pkvm_func_range);
}
@@ -105,6 +99,7 @@ static int mmio_guard_ioremap_hook(phys_addr_t phys, size_t size,
u32 func_id = pkvm_func_range ?
ARM_SMCCC_VENDOR_HYP_KVM_MMIO_RGUARD_MAP_FUNC_ID :
ARM_SMCCC_VENDOR_HYP_KVM_MMIO_GUARD_MAP_FUNC_ID;
phys_addr_t end;
/*
* We only expect MMIO emulation for regions mapped with device
@@ -113,9 +108,20 @@ static int mmio_guard_ioremap_hook(phys_addr_t phys, size_t size,
if (protval != PROT_DEVICE_nGnRE && protval != PROT_DEVICE_nGnRnE)
return 0;
WARN_ON_ONCE(arm_smccc_do_range(func_id, phys, DIV_ROUND_UP(size, PAGE_SIZE),
pkvm_func_range));
end = ALIGN(phys + size, PAGE_SIZE);
phys = ALIGN_DOWN(phys, PAGE_SIZE);
size = end - phys;
/*
* It is fine to overshoot MMIO guard requests. Its sole purpose is to
* indicate the hypervisor where are the MMIO regions and we have
* validated the alignment of the memory regions beforehand.
*/
end = ALIGN(phys + size, pkvm_granule);
phys = ALIGN_DOWN(phys, pkvm_granule);
WARN_ON_ONCE(arm_smccc_do_range(func_id, phys, (end - phys) >> PAGE_SHIFT,
pkvm_func_range));
return 0;
}
@@ -153,6 +159,32 @@ EXPORT_SYMBOL_GPL(page_relinquish);
#endif
static bool __dram_is_aligned(size_t pkvm_granule)
{
struct memblock_region *region, *prev = NULL;
for_each_mem_region(region) {
phys_addr_t prev_end;
if (!prev)
goto discontinuous;
prev_end = prev->base + prev->size;
if (prev_end == region->base)
goto contiguous;
if (!IS_ALIGNED(prev_end, pkvm_granule))
return false;
discontinuous:
if (!IS_ALIGNED(region->base, pkvm_granule))
return false;
contiguous:
prev = region;
}
return IS_ALIGNED(region->base + region->size, pkvm_granule);
}
void pkvm_init_hyp_services(void)
{
struct arm_smccc_res res;
@@ -172,7 +204,8 @@ void pkvm_init_hyp_services(void)
kvm_arm_hyp_service_available(ARM_SMCCC_KVM_FUNC_MEM_UNSHARE))
arm64_mem_crypt_ops_register(&pkvm_crypt_ops);
if (kvm_arm_hyp_service_available(ARM_SMCCC_KVM_FUNC_MMIO_GUARD_MAP))
if (kvm_arm_hyp_service_available(ARM_SMCCC_KVM_FUNC_MMIO_GUARD_MAP) &&
__dram_is_aligned(pkvm_granule))
arm64_ioremap_prot_hook_register(&mmio_guard_ioremap_hook);
#ifdef CONFIG_MEMORY_RELINQUISH