On AMD based systems, mmap'ing a PCI MMIO region does not return proper
values. This is because the mmap_mem function does not consider the fact
that IO regions are not to be encrypted.
In the failing kernel, here's the output --
[root@foo]# ./memaccess 0xd0000000 -t pmem -l 32
0 (0 ) : ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff
10 (16 ) : ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff
I instrumented the kernel to print out the PTE value --
Jun 20 11:46:35 bur-e1-2l-303 kernel: pfn = 0xfffffff80f0866b2, vma->pgoff = 0xd
0000, flags = 0x5044471, prot = 0x8000800000000025
Jun 20 11:46:35 bur-e1-2l-303 kernel: pte = 0x80008000d0000235, pfn = 0xd0000
Note that 0x8000800000000025 -- bit 47 is set. It should not be set for a
MMIO region.
When I disable memory encryption (mem_encrypt=off command line), things work
as they should.
[root@foo]# ./memaccess 0xd0000000 -t pmem -l 32
0 (0 ) : 20 00 00 01 40 08 00 04 f1 00 00 14 0a 00 ff 07
10 (16 ) : 65 f6 70 02 c0 05 00 00 a0 04 00 00 0b 00 00 00
James Puthukattukaran(1):
drivers/char/mem.c: Disable encryption bit in page tables for MMIO
access
drivers/char/mem.c | 27 +++++++++++++++++++++++++++
1 file changed, 27 insertions(+)
Attempting to mmap to a memory mapped IO space returns -1s because the
memory encryption bit is set for these pages. According to the AMD spec,
this bit should not be set for non-DRAM space. The patch checks if this
is an memory IO region being accessed and decrypts accordingly.
Signed-off-by: James Puthukattukaran<[email protected]>
---
drivers/char/mem.c | 27 +++++++++++++++++++++++++++
1 file changed, 27 insertions(+)
diff --git a/drivers/char/mem.c b/drivers/char/mem.c
index ffeb60d..beaa374 100644
--- a/drivers/char/mem.c
+++ b/drivers/char/mem.c
@@ -355,10 +355,30 @@ static inline int private_mapping_ok(struct vm_area_struct *vma)
#endif
};
+static int check_flags(struct resource *res, void *arg)
+{
+ int *mmio = arg;
+ *mmio = (!((res->flags & IORESOURCE_SYSTEM_RAM) ==
+ IORESOURCE_SYSTEM_RAM) && (res->desc == IORES_DESC_NONE));
+ return *mmio;
+}
+
+static void check_iomem_region(phys_addr_t addr, size_t size,
+ int *mmio)
+{
+ u64 start, end;
+
+ start = (u64)addr;
+ end = start + size - 1;
+ *mmio = 0;
+ walk_mem_res(start, end, mmio, check_flags);
+}
+
static int mmap_mem(struct file *file, struct vm_area_struct *vma)
{
size_t size = vma->vm_end - vma->vm_start;
phys_addr_t offset = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
+ int mmio;
/* Does it even fit in phys_addr_t? */
if (offset >> PAGE_SHIFT != vma->vm_pgoff)
@@ -387,6 +407,13 @@ static int mmap_mem(struct file *file, struct vm_area_struct *vma)
vma->vm_ops = &mmap_mem_ops;
+ if (mem_encrypt_active()) {
+ check_iomem_region(vma->vm_pgoff, size, &mmio);
+ if (mmio)
+ vma->vm_page_prot =
+ pgprot_decrypted(vma->vm_page_prot);
+ }
+
/* Remap-pfn-range will mark the range VM_IO */
if (remap_pfn_range(vma,
vma->vm_start,