Re: [Qemu-devel] [PATCH V9 10/16] xen: Introduce the Xen mapcache

[Anthony Liguori <anthony@codemonkey.ws>]

On 01/25/2011 08:29 AM, anthony.perard@citrix.com wrote:
> From: Jun Nakajima<jun.nakajima@intel.com>
>
> On IA32 host or IA32 PAE host, at present, generally, we can't create
> an HVM guest with more than 2G memory, because generally it's almost
> impossible for Qemu to find a large enough and consecutive virtual
> address space to map an HVM guest's whole physical address space.
> The attached patch fixes this issue using dynamic mapping based on
> little blocks of memory.
>
> Each call to qemu_get_ram_ptr makes a call to qemu_map_cache with the
> lock option, so mapcache will not unmap these ram_ptr.
>
> Signed-off-by: Jun Nakajima<jun.nakajima@intel.com>
> Signed-off-by: Anthony PERARD<anthony.perard@citrix.com>
> Signed-off-by: Stefano Stabellini<stefano.stabellini@eu.citrix.com>
> ---
>   Makefile.target     |    3 +
>   configure           |    3 +
>   exec.c              |   40 ++++++-
>   hw/xen.h            |   13 +++
>   xen-all.c           |   64 +++++++++++
>   xen-mapcache-stub.c |   33 ++++++
>   xen-mapcache.c      |  301 +++++++++++++++++++++++++++++++++++++++++++++++++++
>   xen-mapcache.h      |   14 +++
>   xen-stub.c          |    4 +
>   9 files changed, 471 insertions(+), 4 deletions(-)
>   create mode 100644 xen-mapcache-stub.c
>   create mode 100644 xen-mapcache.c
>   create mode 100644 xen-mapcache.h
>
> diff --git a/Makefile.target b/Makefile.target
> index 8126da9..18b3959 100644
> --- a/Makefile.target
> +++ b/Makefile.target
> @@ -207,9 +207,12 @@ QEMU_CFLAGS += $(VNC_JPEG_CFLAGS)
>   QEMU_CFLAGS += $(VNC_PNG_CFLAGS)
>
>   # xen support
> +CONFIG_NO_XEN_MAPCACHE = $(if $(subst n,,$(CONFIG_XEN_MAPCACHE)),n,y)
>   obj-$(CONFIG_XEN) += xen_interfaces.o
>   obj-$(CONFIG_XEN) += xen-all.o
>   obj-$(CONFIG_NO_XEN) += xen-stub.o
> +obj-$(CONFIG_XEN_MAPCACHE) += xen-mapcache.o
> +obj-$(CONFIG_NO_XEN_MAPCACHE) += xen-mapcache-stub.o
>
>   # xen backend driver support
>   obj-$(CONFIG_XEN) += xen_backend.o xen_devconfig.o
> diff --git a/configure b/configure
> index fde9bad..c9a13e1 100755
> --- a/configure
> +++ b/configure
> @@ -3069,6 +3069,9 @@ case "$target_arch2" in
>         echo "CONFIG_XEN=y">>  $config_target_mak
>         echo "LIBS+=$xen_libs">>  $config_target_mak
>         echo "CONFIG_XEN_CTRL_INTERFACE_VERSION=$xen_ctrl_version">>  $config_target_mak
> +      if test "$cpu" = "i386" -o "$cpu" = "x86_64"; then
> +          echo "CONFIG_XEN_MAPCACHE=y">>  $config_target_mak
> +      fi
>       fi
>   esac
>   case "$target_arch2" in
> diff --git a/exec.c b/exec.c
> index e950df2..3b137dc 100644
> --- a/exec.c
> +++ b/exec.c
> @@ -32,6 +32,7 @@
>   #include "hw/qdev.h"
>   #include "osdep.h"
>   #include "kvm.h"
> +#include "hw/xen.h"
>   #include "qemu-timer.h"
>   #if defined(CONFIG_USER_ONLY)
>   #include<qemu.h>
> @@ -51,6 +52,8 @@
>   #include<libutil.h>
>   #endif
>   #endif
> +#else /* !CONFIG_USER_ONLY */
> +#include "xen-mapcache.h"
>   #endif
>
>   //#define DEBUG_TB_INVALIDATE
> @@ -2835,6 +2838,7 @@ ram_addr_t qemu_ram_alloc_from_ptr(DeviceState *dev, const char *name,
>           }
>       }
>
> +    new_block->offset = find_ram_offset(size);
>       if (host) {
>           new_block->host = host;
>       } else {
> @@ -2856,13 +2860,15 @@ ram_addr_t qemu_ram_alloc_from_ptr(DeviceState *dev, const char *name,
>                                      PROT_EXEC|PROT_READ|PROT_WRITE,
>                                      MAP_SHARED | MAP_ANONYMOUS, -1, 0);
>   #else
> -            new_block->host = qemu_vmalloc(size);
> +            if (xen_mapcache_enabled()) {
> +                xen_ram_alloc(new_block->offset, size);
> +            } else {
> +                new_block->host = qemu_vmalloc(size);
> +            }
>   #endif
>               qemu_madvise(new_block->host, size, QEMU_MADV_MERGEABLE);
>           }
>       }
> -
> -    new_block->offset = find_ram_offset(size);
>       new_block->length = size;
>
>       QLIST_INSERT_HEAD(&ram_list.blocks, new_block, next);
> @@ -2903,7 +2909,11 @@ void qemu_ram_free(ram_addr_t addr)
>   #if defined(TARGET_S390X)&&  defined(CONFIG_KVM)
>                   munmap(block->host, block->length);
>   #else
> -                qemu_vfree(block->host);
> +                if (xen_mapcache_enabled()) {
> +                    qemu_invalidate_entry(block->host);
> +                } else {
> +                    qemu_vfree(block->host);
> +                }
>   #endif
>               }
>               qemu_free(block);
> @@ -2929,6 +2939,15 @@ void *qemu_get_ram_ptr(ram_addr_t addr)
>           if (addr - block->offset<  block->length) {
>               QLIST_REMOVE(block, next);
>               QLIST_INSERT_HEAD(&ram_list.blocks, block, next);
> +            if (xen_mapcache_enabled()) {
> +                /* We need to check if the requested address is in the RAM
> +                 * because we don't want to map the entire memory in QEMU.
> +                 */
> +                if (block->offset == 0) {
> +                    return qemu_map_cache(addr, 0, 1);
> +                }
> +                block->host = qemu_map_cache(block->offset, block->length, 1);
> +            }
>               return block->host + (addr - block->offset);
>           }
>       }
> @@ -2964,11 +2983,21 @@ int qemu_ram_addr_from_host(void *ptr, ram_addr_t *ram_addr)
>       uint8_t *host = ptr;
>
>       QLIST_FOREACH(block,&ram_list.blocks, next) {
> +        /* This case append when the block is not mapped. */
> +        if (block->host == NULL) {
> +            continue;
> +        }
>           if (host - block->host<  block->length) {
>               *ram_addr = block->offset + (host - block->host);
>               return 0;
>           }
>       }
> +
> +    if (xen_mapcache_enabled()) {
> +        *ram_addr = qemu_ram_addr_from_mapcache(ptr);
> +        return 0;
> +    }
> +
>       return -1;
>   }
>
> @@ -3879,6 +3908,9 @@ void cpu_physical_memory_unmap(void *buffer, target_phys_addr_t len,
>       if (is_write) {
>           cpu_physical_memory_write(bounce.addr, bounce.buffer, access_len);
>       }
> +    if (xen_enabled()) {
> +        qemu_invalidate_entry(buffer);
> +    }
>       qemu_vfree(bounce.buffer);
>       bounce.buffer = NULL;
>       cpu_notify_map_clients();
> diff --git a/hw/xen.h b/hw/xen.h
> index 338cf76..dd3fb68 100644
> --- a/hw/xen.h
> +++ b/hw/xen.h
> @@ -31,10 +31,23 @@ static inline int xen_enabled(void)
>   #endif
>   }
>
> +static inline int xen_mapcache_enabled(void)
> +{
> +#ifdef CONFIG_XEN_MAPCACHE
> +    return xen_enabled();
> +#else
> +    return 0;
> +#endif
> +}
> +
>   int xen_pci_slot_get_pirq(PCIDevice *pci_dev, int irq_num);
>   void xen_piix3_set_irq(void *opaque, int irq_num, int level);
>   void xen_piix_pci_write_config_client(uint32_t address, uint32_t val, int len);
>
>   int xen_init(int smp_cpus);
>
> +#if defined(NEED_CPU_H)&&  !defined(CONFIG_USER_ONLY)
> +void xen_ram_alloc(ram_addr_t ram_addr, ram_addr_t size);
> +#endif
> +
>   #endif /* QEMU_HW_XEN_H */
> diff --git a/xen-all.c b/xen-all.c
> index 205cbc4..2b9e71c 100644
> --- a/xen-all.c
> +++ b/xen-all.c
> @@ -13,6 +13,8 @@
>   #include "hw/xen_backend.h"
>   #include "hw/xen_redirect.h"
>
> +#include "xen-mapcache.h"
> +
>   /* Xen specific function for piix pci */
>
>   int xen_pci_slot_get_pirq(PCIDevice *pci_dev, int irq_num)
> @@ -55,6 +57,64 @@ qemu_irq *i8259_xen_init(void)
>       return qemu_allocate_irqs(i8259_set_irq, NULL, 16);
>   }
>
> +
> +/* Memory Ops */
> +
> +static void xen_ram_init(ram_addr_t ram_size)
> +{
> +    RAMBlock *new_block;
> +    ram_addr_t below_4g_mem_size, above_4g_mem_size = 0;
> +
> +    new_block = qemu_mallocz(sizeof (*new_block));
> +    pstrcpy(new_block->idstr, sizeof (new_block->idstr), "xen.ram");
> +    new_block->host = NULL;
> +    new_block->offset = 0;
> +    new_block->length = ram_size;
> +
> +    QLIST_INSERT_HEAD(&ram_list.blocks, new_block, next);
> +
> +    ram_list.phys_dirty = qemu_realloc(ram_list.phys_dirty,
> +                                       new_block->length>>  TARGET_PAGE_BITS);
> +    memset(ram_list.phys_dirty + (new_block->offset>>  TARGET_PAGE_BITS),
> +           0xff, new_block->length>>  TARGET_PAGE_BITS);
> +
> +    if (ram_size>= 0xe0000000 ) {
> +        above_4g_mem_size = ram_size - 0xe0000000;
> +        below_4g_mem_size = 0xe0000000;
> +    } else {
> +        below_4g_mem_size = ram_size;
> +    }
> +
> +    cpu_register_physical_memory(0, below_4g_mem_size, new_block->offset);
> +#if TARGET_PHYS_ADDR_BITS>  32
> +    if (above_4g_mem_size>  0) {
> +        cpu_register_physical_memory(0x100000000ULL, above_4g_mem_size,
> +                                     new_block->offset + below_4g_mem_size);
> +    }
> +#endif
> +}
> +
> +void xen_ram_alloc(ram_addr_t ram_addr, ram_addr_t size)
> +{
> +    unsigned long nr_pfn;
> +    xen_pfn_t *pfn_list;
> +    int i;
> +
> +    nr_pfn = size>>  TARGET_PAGE_BITS;
> +    pfn_list = qemu_malloc(sizeof (*pfn_list) * nr_pfn);
> +
> +    for (i = 0; i<  nr_pfn; i++) {
> +        pfn_list[i] = (ram_addr>>  TARGET_PAGE_BITS) + i;
> +    }
> +
> +    if (xc_domain_populate_physmap_exact(xen_xc, xen_domid, nr_pfn, 0, 0, pfn_list)) {
> +        hw_error("xen: failed to populate ram at %lx", ram_addr);
> +    }
> +
> +    qemu_free(pfn_list);
> +}
> +
> +
>   /* Initialise Xen */
>
>   int xen_init(int smp_cpus)
> @@ -68,5 +128,9 @@ int xen_init(int smp_cpus)
>           return -1;
>       }
>
> +    /* Init RAM management */
> +    qemu_map_cache_init();
> +    xen_ram_init(ram_size);
> +
>       return 0;
>   }
> diff --git a/xen-mapcache-stub.c b/xen-mapcache-stub.c
> new file mode 100644
> index 0000000..69ce2e7
> --- /dev/null
> +++ b/xen-mapcache-stub.c
> @@ -0,0 +1,33 @@
> +#include "config.h"
> +
> +#include "exec-all.h"
> +#include "qemu-common.h"
> +#include "cpu-common.h"
> +#include "xen-mapcache.h"
> +
> +int qemu_map_cache_init(void)
> +{
> +    return 0;
> +}
> +
> +uint8_t *qemu_map_cache(target_phys_addr_t phys_addr, target_phys_addr_t size, uint8_t lock)
> +{
> +    return qemu_get_ram_ptr(phys_addr);
> +}
> +
> +void qemu_map_cache_unlock(void *buffer)
> +{
> +}
> +
> +ram_addr_t qemu_ram_addr_from_mapcache(void *ptr)
> +{
> +    return -1;
> +}
> +
> +void qemu_invalidate_map_cache(void)
> +{
> +}
> +
> +void qemu_invalidate_entry(uint8_t *buffer)
> +{
> +}
> diff --git a/xen-mapcache.c b/xen-mapcache.c
> new file mode 100644
> index 0000000..3e1cca9
> --- /dev/null
> +++ b/xen-mapcache.c
> @@ -0,0 +1,301 @@
>    

Needs a copyright.

> +#include "config.h"
> +
> +#include "hw/xen_backend.h"
> +#include "blockdev.h"
> +
> +#include<xen/hvm/params.h>
> +#include<sys/mman.h>
> +
> +#include "xen-mapcache.h"
> +
> +
> +//#define MAPCACHE_DEBUG
> +
> +#ifdef MAPCACHE_DEBUG
> +#  define DPRINTF(fmt, ...) do { \
> +    fprintf(stderr, "xen_mapcache: " fmt, ## __VA_ARGS__); \
> +} while (0)
> +#else
> +#  define DPRINTF(fmt, ...) do { } while (0)
> +#endif
> +
> +#if defined(__i386__)
> +#  define MAX_MCACHE_SIZE    0x40000000 /* 1GB max for x86 */
> +#  define MCACHE_BUCKET_SHIFT 16
> +#elif defined(__x86_64__)
> +#  define MAX_MCACHE_SIZE    0x1000000000 /* 64GB max for x86_64 */
> +#  define MCACHE_BUCKET_SHIFT 20
> +#endif
> +#define MCACHE_BUCKET_SIZE (1UL<<  MCACHE_BUCKET_SHIFT)
>    

This is pretty awful.  Set limits based on host address with, not based 
on i386 vs. x86_64.

> +#define BITS_PER_LONG (sizeof(long) * 8)
> +#define BITS_TO_LONGS(bits) (((bits) + BITS_PER_LONG - 1) / BITS_PER_LONG)
> +#define DECLARE_BITMAP(name, bits) unsigned long name[BITS_TO_LONGS(bits)]
> +#define test_bit(bit, map) \
> +    (!!((map)[(bit) / BITS_PER_LONG]&  (1UL<<  ((bit) % BITS_PER_LONG))))
> +
> +typedef struct MapCacheEntry {
> +    target_phys_addr_t paddr_index;
> +    uint8_t *vaddr_base;
> +    DECLARE_BITMAP(valid_mapping, MCACHE_BUCKET_SIZE>>  XC_PAGE_SHIFT);
> +    uint8_t lock;
> +    struct MapCacheEntry *next;
> +} MapCacheEntry;
> +
> +typedef struct MapCacheRev {
> +    uint8_t *vaddr_req;
> +    target_phys_addr_t paddr_index;
> +    QTAILQ_ENTRY(MapCacheRev) next;
> +} MapCacheRev;
> +
> +typedef struct MapCache {
> +    MapCacheEntry *entry;
> +    unsigned long nr_buckets;
> +    QTAILQ_HEAD(map_cache_head, MapCacheRev) locked_entries;
> +
> +    /* For most cases (>99.9%), the page address is the same. */
> +    target_phys_addr_t last_address_index;
> +    uint8_t *last_address_vaddr;
> +} MapCache;
> +
> +static MapCache *mapcache;
> +
> +
> +int qemu_map_cache_init(void)
> +{
> +    unsigned long size;
> +
> +    mapcache = qemu_mallocz(sizeof (MapCache));
> +
> +    QTAILQ_INIT(&mapcache->locked_entries);
> +    mapcache->last_address_index = -1;
> +
> +    mapcache->nr_buckets = (((MAX_MCACHE_SIZE>>  XC_PAGE_SHIFT) +
> +                   (1UL<<  (MCACHE_BUCKET_SHIFT - XC_PAGE_SHIFT)) - 1)>>
> +                  (MCACHE_BUCKET_SHIFT - XC_PAGE_SHIFT));
> +
> +    /*
> +     * Use mmap() directly: lets us allocate a big hash table with no up-front
> +     * cost in storage space. The OS will allocate memory only for the buckets
> +     * that we actually use. All others will contain all zeroes.
>    

You don't need to use mmap.  malloc() does the same thing.

> +     */
> +    size = mapcache->nr_buckets * sizeof (MapCacheEntry);
> +    size = (size + XC_PAGE_SIZE - 1)&  ~(XC_PAGE_SIZE - 1);
> +    DPRINTF("qemu_map_cache_init, nr_buckets = %lx size %lu\n", mapcache->nr_buckets, size);
> +    mapcache->entry = mmap(NULL, size, PROT_READ|PROT_WRITE,
> +                          MAP_SHARED|MAP_ANON, -1, 0);
> +    if (mapcache->entry == MAP_FAILED) {
> +        return -1;
> +    }
> +
> +    return 0;
> +}
> +
> +static void qemu_remap_bucket(MapCacheEntry *entry,
> +                              target_phys_addr_t size,
> +                              target_phys_addr_t address_index)
> +{
> +    uint8_t *vaddr_base;
> +    xen_pfn_t *pfns;
> +    int *err;
> +    unsigned int i, j;
> +    target_phys_addr_t nb_pfn = size>>  XC_PAGE_SHIFT;
> +
> +    pfns = qemu_mallocz(nb_pfn * sizeof (xen_pfn_t));
> +    err = qemu_mallocz(nb_pfn * sizeof (int));
> +
> +    if (entry->vaddr_base != NULL) {
> +        if (munmap(entry->vaddr_base, size) != 0) {
> +            perror("unmap fails");
> +            exit(-1);
> +        }
> +    }
> +
> +    for (i = 0; i<  nb_pfn; i++) {
> +        pfns[i] = (address_index<<  (MCACHE_BUCKET_SHIFT-XC_PAGE_SHIFT)) + i;
> +    }
> +
> +    vaddr_base = xc_map_foreign_bulk(xen_xc, xen_domid, PROT_READ|PROT_WRITE,
> +                                     pfns, err, nb_pfn);
> +    if (vaddr_base == NULL) {
> +        perror("xc_map_foreign_bulk");
> +        exit(-1);
> +    }
> +
> +    entry->vaddr_base = vaddr_base;
> +    entry->paddr_index = address_index;
> +
> +    for (i = 0; i<  nb_pfn; i += BITS_PER_LONG) {
> +        unsigned long word = 0;
> +        if ((i + BITS_PER_LONG)>  nb_pfn) {
> +            j = nb_pfn % BITS_PER_LONG;
> +        } else {
> +            j = BITS_PER_LONG;
> +        }
> +        while (j>  0) {
> +            word = (word<<  1) | !err[i + --j];
> +        }
> +        entry->valid_mapping[i / BITS_PER_LONG] = word;
> +    }
> +
> +    qemu_free(pfns);
> +    qemu_free(err);
> +}
> +
> +uint8_t *qemu_map_cache(target_phys_addr_t phys_addr, target_phys_addr_t size, uint8_t lock)
> +{
> +    MapCacheEntry *entry, *pentry = NULL;
> +    target_phys_addr_t address_index  = phys_addr>>  MCACHE_BUCKET_SHIFT;
> +    target_phys_addr_t address_offset = phys_addr&  (MCACHE_BUCKET_SIZE - 1);
> +
> +    if (address_index == mapcache->last_address_index&&  !lock) {
> +        return mapcache->last_address_vaddr + address_offset;
> +    }
> +
> +    entry =&mapcache->entry[address_index % mapcache->nr_buckets];
> +
> +    while (entry&&  entry->lock&&  entry->paddr_index != address_index&&  entry->vaddr_base) {
> +        pentry = entry;
> +        entry = entry->next;
> +    }
> +    if (!entry) {
> +        entry = qemu_mallocz(sizeof (MapCacheEntry));
> +        pentry->next = entry;
> +        qemu_remap_bucket(entry, size ? : MCACHE_BUCKET_SIZE, address_index);
> +    } else if (!entry->lock) {
> +        if (!entry->vaddr_base || entry->paddr_index != address_index ||
> +            !test_bit(address_offset>>  XC_PAGE_SHIFT, entry->valid_mapping)) {
> +            qemu_remap_bucket(entry, size ? : MCACHE_BUCKET_SIZE, address_index);
> +        }
> +    }
> +
> +    if (!test_bit(address_offset>>  XC_PAGE_SHIFT, entry->valid_mapping)) {
> +        mapcache->last_address_index = -1;
> +        return NULL;
> +    }
> +
> +    mapcache->last_address_index = address_index;
> +    mapcache->last_address_vaddr = entry->vaddr_base;
> +    if (lock) {
> +        MapCacheRev *reventry = qemu_mallocz(sizeof(MapCacheRev));
> +        entry->lock++;
> +        reventry->vaddr_req = mapcache->last_address_vaddr + address_offset;
> +        reventry->paddr_index = mapcache->last_address_index;
> +        QTAILQ_INSERT_TAIL(&mapcache->locked_entries, reventry, next);
> +    }
> +
> +    return mapcache->last_address_vaddr + address_offset;
> +}
> +
> +ram_addr_t qemu_ram_addr_from_mapcache(void *ptr)
> +{
> +    MapCacheRev *reventry;
> +    target_phys_addr_t paddr_index;
> +    int found = 0;
> +
> +    QTAILQ_FOREACH(reventry,&mapcache->locked_entries, next) {
> +        if (reventry->vaddr_req == ptr) {
> +            paddr_index = reventry->paddr_index;
> +            found = 1;
> +            break;
> +        }
> +    }
> +    if (!found) {
> +        fprintf(stderr, "qemu_ram_addr_from_mapcache, could not find %p\n", ptr);
> +        QTAILQ_FOREACH(reventry,&mapcache->locked_entries, next) {
> +            DPRINTF("   %lx ->  %p is present\n", reventry->paddr_index,
> +                    reventry->vaddr_req);
> +        }
> +        abort();
> +        return 0;
> +    }
> +
> +    return paddr_index<<  MCACHE_BUCKET_SHIFT;
> +}
> +
> +void qemu_invalidate_entry(uint8_t *buffer)
> +{
> +    MapCacheEntry *entry = NULL, *pentry = NULL;
> +    MapCacheRev *reventry;
> +    target_phys_addr_t paddr_index;
> +    int found = 0;
> +
> +    if (mapcache->last_address_vaddr == buffer) {
> +        mapcache->last_address_index = -1;
> +    }
> +
> +    QTAILQ_FOREACH(reventry,&mapcache->locked_entries, next) {
> +        if (reventry->vaddr_req == buffer) {
> +            paddr_index = reventry->paddr_index;
> +            found = 1;
> +            break;
> +        }
> +    }
> +    if (!found) {
> +        DPRINTF("qemu_invalidate_entry, could not find %p\n", buffer);
> +        QTAILQ_FOREACH(reventry,&mapcache->locked_entries, next) {
> +            DPRINTF("   %lx ->  %p is present\n", reventry->paddr_index, reventry->vaddr_req);
> +        }
> +        return;
> +    }
> +    QTAILQ_REMOVE(&mapcache->locked_entries, reventry, next);
> +    qemu_free(reventry);
> +
> +    entry =&mapcache->entry[paddr_index % mapcache->nr_buckets];
> +    while (entry&&  entry->paddr_index != paddr_index) {
> +        pentry = entry;
> +        entry = entry->next;
> +    }
> +    if (!entry) {
> +        DPRINTF("Trying to unmap address %p that is not in the mapcache!\n", buffer);
> +        return;
> +    }
> +    entry->lock--;
> +    if (entry->lock>  0 || pentry == NULL) {
> +        return;
> +    }
> +
> +    pentry->next = entry->next;
> +    if (munmap(entry->vaddr_base, MCACHE_BUCKET_SIZE) != 0) {
> +        perror("unmap fails");
> +        exit(-1);
> +    }
> +    qemu_free(entry);
> +}
> +
> +void qemu_invalidate_map_cache(void)
> +{
> +    unsigned long i;
> +    MapCacheRev *reventry;
> +
> +    qemu_aio_flush();
>    

This is bizarre?  Why is this needed?

Regards,

Anthony Liguori