On Tue, Jan 15, 2019 at 10:49:45AM +0800, Yi Zhang wrote:
> On 2019-01-14 at 17:07:02 -0200, Eduardo Habkost wrote:
> > On Wed, Jan 02, 2019 at 01:26:15PM +0800, Zhang Yi wrote:
> > > When a file supporting DAX is used as vNVDIMM backend, mmap it with
> > > MAP_SYNC flag in addition which can ensure file system metadata
> > > synced in each guest writes to the backend file, without other QEMU
> > > actions (e.g., periodic fsync() by QEMU).
> > >
> > > Signed-off-by: Haozhong Zhang <haozhong.zh...@intel.com>
> > > Signed-off-by: Zhang Yi <yi.z.zh...@linux.intel.com>
> > > ---
> > > include/qemu/osdep.h | 16 ++++++++++++++++
> > > util/mmap-alloc.c | 12 +++++++++++-
> > > 2 files changed, 27 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/include/qemu/osdep.h b/include/qemu/osdep.h
> > > index 3bf48bc..bb1eba1 100644
> > > --- a/include/qemu/osdep.h
> > > +++ b/include/qemu/osdep.h
> > > @@ -410,6 +410,22 @@ void qemu_anon_ram_free(void *ptr, size_t size);
> > > # define QEMU_VMALLOC_ALIGN getpagesize()
> > > #endif
> > >
> > > +/*
> > > + * MAP_SHARED_VALIDATE and MAP_SYNC are introduced in Linux kernel
> > > + * 4.15, so they may not be defined when compiling on older kernels.
> > > + */
> > > +#ifdef CONFIG_LINUX
> > > +
> > > +#include <asm-generic/mman.h>
> > > +
> > > +#ifndef MAP_SYNC
> > > +#define MAP_SYNC 0x0
> > > +#endif
> > > +
> > > +#else /* !CONFIG_LINUX */
> > > +#define MAP_SYNC 0x0
> > > +#endif /* CONFIG_LINUX */
> > > +
> > > #ifdef CONFIG_POSIX
> > > struct qemu_signalfd_siginfo {
> > > uint32_t ssi_signo; /* Signal number */
> > > diff --git a/util/mmap-alloc.c b/util/mmap-alloc.c
> > > index 8f0a740..a9d5e56 100644
> > > --- a/util/mmap-alloc.c
> > > +++ b/util/mmap-alloc.c
> > > @@ -99,6 +99,8 @@ void *qemu_ram_mmap(int fd, size_t size, size_t align,
> > > uint32_t flags)
> > > void *ptr = mmap(0, total, PROT_NONE, MAP_ANONYMOUS | MAP_PRIVATE,
> > > -1, 0);
> > > #endif
> > > bool shared = flags & RAM_SHARED;
> > > + bool is_pmem = flags & RAM_PMEM;
> > > + int mmap_xflags = 0;
> > > size_t offset;
> > > void *ptr1;
> > >
> > > @@ -109,13 +111,21 @@ void *qemu_ram_mmap(int fd, size_t size, size_t
> > > align, uint32_t flags)
> > > assert(is_power_of_2(align));
> > > /* Always align to host page size */
> > > assert(align >= getpagesize());
> > > + if (shared && is_pmem) {
> > > + mmap_xflags |= MAP_SYNC;
> > > + }
> > >
> > > offset = QEMU_ALIGN_UP((uintptr_t)ptr, align) - (uintptr_t)ptr;
> > > + retry_mmap_fd:
> > > ptr1 = mmap(ptr + offset, size, PROT_READ | PROT_WRITE,
> > > MAP_FIXED |
> > > (fd == -1 ? MAP_ANONYMOUS : 0) |
> > > - (shared ? MAP_SHARED : MAP_PRIVATE),
> > > + (shared ? MAP_SHARED : MAP_PRIVATE) | mmap_xflags,
> > > fd, 0);
> > > + if ((ptr1 == MAP_FAILED) && (mmap_xflags & MAP_SYNC)) {
> > > + mmap_xflags &= ~MAP_SYNC;
> > > + goto retry_mmap_fd;
> >
> > Do we have use cases where using pmem=on without MAP_SYNC isn't
> > going to cause problems? If not, shouldn't we at least print a
> Yes, we have a case that direct use dax device but not a files on
> dax aware file system, we prefer to don't set the MAP_SYNC if user
> haven't much knowledge about that. it may took some potencial
> performance issues with MAP_SYNC.
I think you will have to be quite a bit more specific.
If there's a performance / functionality tradeoff here
then hiding it behind an option with an inscrutable name
isn't a good idea. Neither is ignoring failures silently.
> > warning here? Otherwise, won't we still need an option for cases
> > that require MAP_SYNC to be working?
> >
> > > + }
> >
> > --
> > Eduardo
