Currently, the TDVF cannot work with gmem in-place conversion because current implementation of KVM_TDX_INIT_MEM_REGION in KVM requires gmem of TDVF to be valid for both shared and private at the same time.
To workaround it, explicitly not enable in-place conversion for internal MemoryRegion with gmem. So that TDVF doesn't use in-place conversion gmem and KVM_TDX_INIT_MEM_REGION will initialize the gmem with the separate shared memory. To make in-place conversion work with TDX's initial memory, the one possible solution and flow would be as below and it requires KVM change: - QEMU create gmem as shared; - QEMU mmap the gmem and load TDVF binary into it; - QEMU convert gmem to private with the content preserved[1]; - QEMU invokes KVM_TDX_INIT_MEM_REGION without valid src, so that KVM knows to fetch the content in-place and use in-place PAGE.ADD for TDX. [1] https://lore.kernel.org/all/ag0pnijvpl0cz...@google.com/ Signed-off-by: Xiaoyao Li <xiaoyao...@intel.com> --- include/system/memory.h | 3 +++ system/memory.c | 2 +- system/physmem.c | 8 +++++--- 3 files changed, 9 insertions(+), 4 deletions(-) diff --git a/include/system/memory.h b/include/system/memory.h index f14fbf65805d..89d6449cef70 100644 --- a/include/system/memory.h +++ b/include/system/memory.h @@ -256,6 +256,9 @@ typedef struct IOMMUTLBEvent { */ #define RAM_PRIVATE (1 << 13) +/* Don't use enable in-place conversion for the guest mmefd backend */ +#define RAM_GUEST_MEMFD_NO_INPLACE (1 << 14) + static inline void iommu_notifier_init(IOMMUNotifier *n, IOMMUNotify fn, IOMMUNotifierFlag flags, hwaddr start, hwaddr end, diff --git a/system/memory.c b/system/memory.c index 6870a41629ef..c1b73abc4c94 100644 --- a/system/memory.c +++ b/system/memory.c @@ -3702,7 +3702,7 @@ bool memory_region_init_ram_guest_memfd(MemoryRegion *mr, DeviceState *owner_dev; if (!memory_region_init_ram_flags_nomigrate(mr, owner, name, size, - RAM_GUEST_MEMFD, errp)) { + RAM_GUEST_MEMFD | RAM_GUEST_MEMFD_NO_INPLACE, errp)) { return false; } /* This will assert if owner is neither NULL nor a DeviceState. diff --git a/system/physmem.c b/system/physmem.c index ea1c27ea2b99..c23379082f38 100644 --- a/system/physmem.c +++ b/system/physmem.c @@ -1916,7 +1916,8 @@ static void ram_block_add(RAMBlock *new_block, Error **errp) if (new_block->flags & RAM_GUEST_MEMFD) { int ret; - bool in_place = kvm_guest_memfd_inplace_supported; + bool in_place = !(new_block->flags & RAM_GUEST_MEMFD_NO_INPLACE) && + kvm_guest_memfd_inplace_supported; new_block->guest_memfd_flags = 0; @@ -2230,7 +2231,8 @@ RAMBlock *qemu_ram_alloc_internal(ram_addr_t size, ram_addr_t max_size, ram_flags &= ~RAM_PRIVATE; assert((ram_flags & ~(RAM_SHARED | RAM_RESIZEABLE | RAM_PREALLOC | - RAM_NORESERVE | RAM_GUEST_MEMFD)) == 0); + RAM_NORESERVE | RAM_GUEST_MEMFD | + RAM_GUEST_MEMFD_NO_INPLACE)) == 0); assert(!host ^ (ram_flags & RAM_PREALLOC)); assert(max_size >= size); @@ -2314,7 +2316,7 @@ RAMBlock *qemu_ram_alloc(ram_addr_t size, uint32_t ram_flags, MemoryRegion *mr, Error **errp) { assert((ram_flags & ~(RAM_SHARED | RAM_NORESERVE | RAM_GUEST_MEMFD | - RAM_PRIVATE)) == 0); + RAM_PRIVATE | RAM_GUEST_MEMFD_NO_INPLACE)) == 0); return qemu_ram_alloc_internal(size, size, NULL, NULL, ram_flags, mr, errp); } -- 2.43.0
