From: Joerg Roedel joro@8bytes.org
Sent: 26 March 2022 14:41
The io specific memcpy/memset functions use string mmio accesses to do their work. Under SEV the hypervisor can't emulate these instructions, because they read/write directly from/to encrypted memory.
KVM will inject a page fault exception into the guest when it is asked to emulate string mmio instructions for an SEV guest:
BUG: unable to handle page fault for address: ffffc90000065068 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 8000100000067 P4D 8000100000067 PUD 80001000fb067 PMD 80001000fc067 PTE 80000000fed40173 Oops: 0000 [#1] PREEMPT SMP NOPTI CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.17.0-rc7 #3
As string mmio for an SEV guest can not be supported by the hypervisor, unroll the instructions for CC_ATTR_GUEST_UNROLL_STRING_IO enabled kernels.
This issue appears when kernels are launched in recent libvirt-managed SEV virtual machines, because libvirt started to add a tpm-crb device to the guest by default.
The kernel driver for tpm-crb uses memcpy_to/from_io() functions to access MMIO memory, resulting in a page-fault injected by KVM and crashing the kernel at boot.
Cc: stable@vger.kernel.org #4.15+ Fixes: d8aa7eea78a1 ('x86/mm: Add Secure Encrypted Virtualization (SEV) support') Reviewed-by: Tom Lendacky thomas.lendacky@amd.com Signed-off-by: Joerg Roedel jroedel@suse.de
Changes v2->v3:
- Fix sparse warnings introduced by v2
arch/x86/lib/iomem.c | 65 ++++++++++++++++++++++++++++++++++++++------ 1 file changed, 57 insertions(+), 8 deletions(-)
diff --git a/arch/x86/lib/iomem.c b/arch/x86/lib/iomem.c index df50451d94ef..3e2f33fc33de 100644 --- a/arch/x86/lib/iomem.c +++ b/arch/x86/lib/iomem.c @@ -22,7 +22,7 @@ static __always_inline void rep_movs(void *to, const void *from, size_t n) : "memory"); }
-void memcpy_fromio(void *to, const volatile void __iomem *from, size_t n) +static void string_memcpy_fromio(void *to, const volatile void __iomem *from, size_t n) { if (unlikely(!n)) return; @@ -38,9 +38,8 @@ void memcpy_fromio(void *to, const volatile void __iomem *from, size_t n) } rep_movs(to, (const void *)from, n); } -EXPORT_SYMBOL(memcpy_fromio);
-void memcpy_toio(volatile void __iomem *to, const void *from, size_t n) +static void string_memcpy_toio(volatile void __iomem *to, const void *from, size_t n) { if (unlikely(!n)) return; @@ -56,14 +55,64 @@ void memcpy_toio(volatile void __iomem *to, const void *from, size_t n) } rep_movs((void *)to, (const void *) from, n); }
+static void unrolled_memcpy_fromio(void *to, const volatile void __iomem *from, size_t n) +{
- const volatile char __iomem *in = from;
- char *out = to;
- int i;
- for (i = 0; i < n; ++i)
out[i] = readb(&in[i]);+}
Wait a minute.... Aren't these functions supposed to be doing 'memory' copies? In which case they need to be using 64bit IO accesses where appropriate - otherwise the performance is horrid.
I thought the x86 memcpy_to/from_io() had been changed to always use a software loop rather than using whatever memcpy() ended up using. In particular the 'rep movsb' ERMS (EMRS?) copy that is fast (on some cpu) for memory-memory copies is always a byte copy on uncached locations typical for io addresses.
PIO reads from PCIe can be spectacularly slow. You really do want to use the largest register available.
David
- Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)