| Message ID | 20240202215332.118728-13-david@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | libvhost-user: support more memslots and cleanup memslot handling code | expand |
On Fri, Feb 2, 2024 at 4:55 PM David Hildenbrand <david@redhat.com> wrote: > > In the past, QEMU would create memory regions that could partially cover > hugetlb pages, making mmap() fail if we would use the mmap_offset as an > fd_offset. For that reason, we never used the mmap_offset as an offset into > the fd and instead always mapped the fd from the very start. > > However, that can easily result in us mmap'ing a lot of unnecessary > parts of an fd, possibly repeatedly. > > QEMU nowadays does not create memory regions that partially cover huge > pages -- it never really worked with postcopy. QEMU handles merging of > regions that partially cover huge pages (due to holes in boot memory) since > 2018 in c1ece84e7c93 ("vhost: Huge page align and merge"). > > Let's be a bit careful and not unconditionally convert the > mmap_offset into an fd_offset. Instead, let's simply detect the hugetlb > size and pass as much as we can as fd_offset, making sure that we call > mmap() with a properly aligned offset. > > With QEMU and a virtio-mem device that is fully plugged (50GiB using 50 > memslots) the qemu-storage daemon process consumes in the VA space > 1281GiB before this change and 58GiB after this change. > > Example debug output: > ================ Vhost user message ================ > Request: VHOST_USER_ADD_MEM_REG (37) > Flags: 0x9 > Size: 40 > Fds: 59 > Adding region 50 > guest_phys_addr: 0x0000000d80000000 > memory_size: 0x0000000040000000 > userspace_addr 0x00007f54ebffe000 > mmap_offset 0x0000000c00000000 > fd_offset: 0x0000000c00000000 > new mmap_offset: 0x0000000000000000 > mmap_addr: 0x00007f7ecc000000 > Successfully added new region > ================ Vhost user message ================ > Request: VHOST_USER_ADD_MEM_REG (37) > Flags: 0x9 > Size: 40 > Fds: 59 > Adding region 51 > guest_phys_addr: 0x0000000dc0000000 > memory_size: 0x0000000040000000 > userspace_addr 0x00007f552bffe000 > mmap_offset 0x0000000c40000000 > fd_offset: 0x0000000c40000000 > new mmap_offset: 0x0000000000000000 > mmap_addr: 0x00007f7e8c000000 > Successfully added new region > > Signed-off-by: David Hildenbrand <david@redhat.com> Reviewed-by: Raphael Norwitz <raphael@enfabrica.net> > --- > subprojects/libvhost-user/libvhost-user.c | 50 ++++++++++++++++++++--- > 1 file changed, 45 insertions(+), 5 deletions(-) > > diff --git a/subprojects/libvhost-user/libvhost-user.c b/subprojects/libvhost-user/libvhost-user.c > index 75e47b7bb3..7d8293dc84 100644 > --- a/subprojects/libvhost-user/libvhost-user.c > +++ b/subprojects/libvhost-user/libvhost-user.c > @@ -43,6 +43,8 @@ > #include <fcntl.h> > #include <sys/ioctl.h> > #include <linux/vhost.h> > +#include <sys/vfs.h> > +#include <linux/magic.h> > > #ifdef __NR_userfaultfd > #include <linux/userfaultfd.h> > @@ -281,12 +283,36 @@ vu_remove_all_mem_regs(VuDev *dev) > dev->nregions = 0; > } > > +static size_t > +get_fd_pagesize(int fd) > +{ > + static size_t pagesize; > +#if defined(__linux__) > + struct statfs fs; > + int ret; > + > + do { > + ret = fstatfs(fd, &fs); > + } while (ret != 0 && errno == EINTR); > + > + if (!ret && fs.f_type == HUGETLBFS_MAGIC) { > + return fs.f_bsize; > + } > +#endif > + > + if (!pagesize) { > + pagesize = getpagesize(); > + } > + return pagesize; > +} > + > static void > _vu_add_mem_reg(VuDev *dev, VhostUserMemoryRegion *msg_region, int fd) > { > const uint64_t start_gpa = msg_region->guest_phys_addr; > const uint64_t end_gpa = start_gpa + msg_region->memory_size; > int prot = PROT_READ | PROT_WRITE; > + uint64_t mmap_offset, fd_offset; > VuDevRegion *r; > void *mmap_addr; > int low = 0; > @@ -335,11 +361,25 @@ _vu_add_mem_reg(VuDev *dev, VhostUserMemoryRegion *msg_region, int fd) > idx = low; > > /* > - * We don't use offset argument of mmap() since the mapped address has > - * to be page aligned, and we use huge pages. > + * Convert most of msg_region->mmap_offset to fd_offset. In almost all > + * cases, this will leave us with mmap_offset == 0, mmap()'ing only > + * what we really need. Only if a memory region would partially cover > + * hugetlb pages, we'd get mmap_offset != 0, which usually doesn't happen > + * anymore (i.e., modern QEMU). > + * > + * Note that mmap() with hugetlb would fail if the offset into the file > + * is not aligned to the huge page size. > */ > - mmap_addr = mmap(0, msg_region->memory_size + msg_region->mmap_offset, > - prot, MAP_SHARED | MAP_NORESERVE, fd, 0); > + fd_offset = ALIGN_DOWN(msg_region->mmap_offset, get_fd_pagesize(fd)); > + mmap_offset = msg_region->mmap_offset - fd_offset; > + > + DPRINT(" fd_offset: 0x%016"PRIx64"\n", > + fd_offset); > + DPRINT(" adj mmap_offset: 0x%016"PRIx64"\n", > + mmap_offset); > + > + mmap_addr = mmap(0, msg_region->memory_size + mmap_offset, > + prot, MAP_SHARED | MAP_NORESERVE, fd, fd_offset); > if (mmap_addr == MAP_FAILED) { > vu_panic(dev, "region mmap error: %s", strerror(errno)); > return; > @@ -354,7 +394,7 @@ _vu_add_mem_reg(VuDev *dev, VhostUserMemoryRegion *msg_region, int fd) > r->size = msg_region->memory_size; > r->qva = msg_region->userspace_addr; > r->mmap_addr = (uint64_t)(uintptr_t)mmap_addr; > - r->mmap_offset = msg_region->mmap_offset; > + r->mmap_offset = mmap_offset; > dev->nregions++; > > if (dev->postcopy_listening) { > -- > 2.43.0 > >
diff --git a/subprojects/libvhost-user/libvhost-user.c b/subprojects/libvhost-user/libvhost-user.c index 75e47b7bb3..7d8293dc84 100644 --- a/subprojects/libvhost-user/libvhost-user.c +++ b/subprojects/libvhost-user/libvhost-user.c @@ -43,6 +43,8 @@ #include <fcntl.h> #include <sys/ioctl.h> #include <linux/vhost.h> +#include <sys/vfs.h> +#include <linux/magic.h> #ifdef __NR_userfaultfd #include <linux/userfaultfd.h> @@ -281,12 +283,36 @@ vu_remove_all_mem_regs(VuDev *dev) dev->nregions = 0; } +static size_t +get_fd_pagesize(int fd) +{ + static size_t pagesize; +#if defined(__linux__) + struct statfs fs; + int ret; + + do { + ret = fstatfs(fd, &fs); + } while (ret != 0 && errno == EINTR); + + if (!ret && fs.f_type == HUGETLBFS_MAGIC) { + return fs.f_bsize; + } +#endif + + if (!pagesize) { + pagesize = getpagesize(); + } + return pagesize; +} + static void _vu_add_mem_reg(VuDev *dev, VhostUserMemoryRegion *msg_region, int fd) { const uint64_t start_gpa = msg_region->guest_phys_addr; const uint64_t end_gpa = start_gpa + msg_region->memory_size; int prot = PROT_READ | PROT_WRITE; + uint64_t mmap_offset, fd_offset; VuDevRegion *r; void *mmap_addr; int low = 0; @@ -335,11 +361,25 @@ _vu_add_mem_reg(VuDev *dev, VhostUserMemoryRegion *msg_region, int fd) idx = low; /* - * We don't use offset argument of mmap() since the mapped address has - * to be page aligned, and we use huge pages. + * Convert most of msg_region->mmap_offset to fd_offset. In almost all + * cases, this will leave us with mmap_offset == 0, mmap()'ing only + * what we really need. Only if a memory region would partially cover + * hugetlb pages, we'd get mmap_offset != 0, which usually doesn't happen + * anymore (i.e., modern QEMU). + * + * Note that mmap() with hugetlb would fail if the offset into the file + * is not aligned to the huge page size. */ - mmap_addr = mmap(0, msg_region->memory_size + msg_region->mmap_offset, - prot, MAP_SHARED | MAP_NORESERVE, fd, 0); + fd_offset = ALIGN_DOWN(msg_region->mmap_offset, get_fd_pagesize(fd)); + mmap_offset = msg_region->mmap_offset - fd_offset; + + DPRINT(" fd_offset: 0x%016"PRIx64"\n", + fd_offset); + DPRINT(" adj mmap_offset: 0x%016"PRIx64"\n", + mmap_offset); + + mmap_addr = mmap(0, msg_region->memory_size + mmap_offset, + prot, MAP_SHARED | MAP_NORESERVE, fd, fd_offset); if (mmap_addr == MAP_FAILED) { vu_panic(dev, "region mmap error: %s", strerror(errno)); return; @@ -354,7 +394,7 @@ _vu_add_mem_reg(VuDev *dev, VhostUserMemoryRegion *msg_region, int fd) r->size = msg_region->memory_size; r->qva = msg_region->userspace_addr; r->mmap_addr = (uint64_t)(uintptr_t)mmap_addr; - r->mmap_offset = msg_region->mmap_offset; + r->mmap_offset = mmap_offset; dev->nregions++; if (dev->postcopy_listening) {
In the past, QEMU would create memory regions that could partially cover hugetlb pages, making mmap() fail if we would use the mmap_offset as an fd_offset. For that reason, we never used the mmap_offset as an offset into the fd and instead always mapped the fd from the very start. However, that can easily result in us mmap'ing a lot of unnecessary parts of an fd, possibly repeatedly. QEMU nowadays does not create memory regions that partially cover huge pages -- it never really worked with postcopy. QEMU handles merging of regions that partially cover huge pages (due to holes in boot memory) since 2018 in c1ece84e7c93 ("vhost: Huge page align and merge"). Let's be a bit careful and not unconditionally convert the mmap_offset into an fd_offset. Instead, let's simply detect the hugetlb size and pass as much as we can as fd_offset, making sure that we call mmap() with a properly aligned offset. With QEMU and a virtio-mem device that is fully plugged (50GiB using 50 memslots) the qemu-storage daemon process consumes in the VA space 1281GiB before this change and 58GiB after this change. Example debug output: ================ Vhost user message ================ Request: VHOST_USER_ADD_MEM_REG (37) Flags: 0x9 Size: 40 Fds: 59 Adding region 50 guest_phys_addr: 0x0000000d80000000 memory_size: 0x0000000040000000 userspace_addr 0x00007f54ebffe000 mmap_offset 0x0000000c00000000 fd_offset: 0x0000000c00000000 new mmap_offset: 0x0000000000000000 mmap_addr: 0x00007f7ecc000000 Successfully added new region ================ Vhost user message ================ Request: VHOST_USER_ADD_MEM_REG (37) Flags: 0x9 Size: 40 Fds: 59 Adding region 51 guest_phys_addr: 0x0000000dc0000000 memory_size: 0x0000000040000000 userspace_addr 0x00007f552bffe000 mmap_offset 0x0000000c40000000 fd_offset: 0x0000000c40000000 new mmap_offset: 0x0000000000000000 mmap_addr: 0x00007f7e8c000000 Successfully added new region Signed-off-by: David Hildenbrand <david@redhat.com> --- subprojects/libvhost-user/libvhost-user.c | 50 ++++++++++++++++++++--- 1 file changed, 45 insertions(+), 5 deletions(-)