* Chuan Zheng (zhengch...@huawei.com) wrote:
> From: Zheng Chuan <zhengch...@huawei.com>
>
> Record hash results for each ramblock.
Please be careful when talking about 'ramblock' since we already use
that for a chunk of memory in QEMU.
> Signed-off-by: Zheng Chuan <zhengch...@huawei.com>
> Signed-off-by: YanYing Zhang <ann.zhuangyany...@huawei.com>
> ---
> migration/dirtyrate.c | 157
> ++++++++++++++++++++++++++++++++++++++++++++++++++
> migration/dirtyrate.h | 1 +
> 2 files changed, 158 insertions(+)
>
> diff --git a/migration/dirtyrate.c b/migration/dirtyrate.c
> index 6baf674..45cfc91 100644
> --- a/migration/dirtyrate.c
> +++ b/migration/dirtyrate.c
> @@ -10,12 +10,27 @@
> * See the COPYING file in the top-level directory.
> */
>
> +#include "qemu/osdep.h"
> +#include "qapi/error.h"
> +#include "crypto/hash.h"
> +#include "crypto/random.h"
> +#include "qemu/config-file.h"
> +#include "exec/memory.h"
> +#include "exec/ramblock.h"
> +#include "exec/target_page.h"
> +#include "qemu/rcu_queue.h"
> +#include "qapi/qapi-commands-migration.h"
> +#include "migration.h"
> #include "dirtyrate.h"
>
> static uint64_t sample_pages_per_gigabytes = DIRTYRATE_DEFAULT_SAMPLE_PAGES;
> static struct dirtyrate_statistics dirty_stat;
> CalculatingDirtyRateStage calculating_dirty_rate_stage = CAL_DIRTY_RATE_INIT;
>
> +#define RAMBLOCK_FOREACH_MIGRATABLE(block) \
> + INTERNAL_RAMBLOCK_FOREACH(block) \
> + if (!qemu_ram_is_migratable(block)) {} else
> +
Instead of redefining this here, please move the existing definition up
into migration/ram.h ?
> static void reset_dirtyrate_stat(void)
> {
> dirty_stat.total_dirty_samples = 0;
> @@ -44,6 +59,148 @@ static void update_dirtyrate(int64_t msec)
> dirty_stat.dirty_rate = dirty_rate;
> }
>
> +static int get_block_vfn_hash(struct block_dirty_info *info, unsigned long
> vfn,
> + uint8_t **md, size_t *hash_len)
> +{
> + struct iovec iov_array;
> + int ret = 0;
> + int nkey = 1;
> +
> + iov_array.iov_base = info->block_addr +
> + vfn * DIRTYRATE_SAMPLE_PAGE_SIZE;
> + iov_array.iov_len = DIRTYRATE_SAMPLE_PAGE_SIZE;
I'm a bit confused by how this is working; is 'vfn' an index
in SAMPLE_PAGE_SIZE's rather than individual pages? So this is
going to hash over something the size of a 'DIRTYRATE_SAMPLE_PAGE_SIZE'?
> + if (qcrypto_hash_bytesv(QCRYPTO_HASH_ALG_MD5,
> + &iov_array, nkey,
> + md, hash_len, NULL) < 0) {
> + ret = -1;
> + }
> +
> + return ret;
> +}
> +
> +static int save_block_hash(struct block_dirty_info *info)
> +{
> + unsigned long *rand_buf = NULL;
> + unsigned int sample_pages_count;
> + uint8_t *md = NULL;
> + size_t hash_len;
> + int i;
> + int ret = -1;
> +
> + sample_pages_count = info->sample_pages_count;
> + /* block size less than one page, return success to skip this block */
> + if (unlikely(info->block_pages == 0 || sample_pages_count == 0)) {
> + ret = 0;
> + goto out;
> + }
> +
> + /* use random bytes to pick sample page vfn */
> + rand_buf = g_malloc0_n(sample_pages_count, sizeof(unsigned long));
> + /* DEFAULT_READ_RANDOM_MAX_LIMIT 32M,
> + * can support 4T vm 1024 sample_pages_per_gigabytes
> + */
> + ret = qcrypto_random_bytes((unsigned char *)rand_buf,
> + sample_pages_count * sizeof(unsigned long),
> + NULL);
I think there may be a different way to do this without having to store
the rand_buf.
We already link to glib, and glib has a PRNG;
https://developer.gnome.org/glib/stable/glib-Random-Numbers.html
If you create a new prng is g_rand_new() at the start of day, and take a
copy with g_rand_copy(), then you can replay the sequence of random
numbers it generates without actually storing the sequence.
So g_rand_new(), g_rand_copy() and then before you traverse the set of
pages you go and take a copy again and use the copy; it'll give the same
sequence every time.
Note also, because you're allocating a potentially large array,
for this and the hash_result please use g_try_malloc0_n (or g_try_new0)
and fail properly if it returns NULL.
Note we have users with more than 4T of RAM in their VMs, although I
doubt in a single RAMBlock
> + if (ret) {
> + ret = -1;
> + goto out;
> + }
> +
> + hash_len = qcrypto_hash_digest_len(QCRYPTO_HASH_ALG_MD5);
> + info->hash_result = g_malloc0_n(sample_pages_count, sizeof(uint8_t) *
> hash_len);
> + info->sample_page_vfn = g_malloc0_n(sample_pages_count, sizeof(unsigned
> long));
> +
> + for (i = 0; i < sample_pages_count; i++) {
> + md = info->hash_result + i * hash_len;
> + info->sample_page_vfn[i] = rand_buf[i] % info->block_pages;
> + ret = get_block_vfn_hash(info, info->sample_page_vfn[i], &md,
> &hash_len);
> + if (ret < 0) {
> + goto out;
> + }
> + }
> + ret = 0;
> +out:
> + g_free(rand_buf);
> + return ret;
> +}
> +
> +static void get_block_dirty_info(RAMBlock *block, struct block_dirty_info
> *info,
> + struct dirtyrate_config *config)
> +{
> + uint64_t sample_pages_per_gigabytes = config->sample_pages_per_gigabytes;
> +
> + /* Right shift 30 bits to calc block size in GB */
> + info->sample_pages_count = (qemu_ram_get_used_length(block) *
> sample_pages_per_gigabytes) >> 30;
> +
> + info->block_pages = qemu_ram_get_used_length(block) >>
> DIRTYRATE_PAGE_SIZE_SHIFT;
> + info->block_addr = qemu_ram_get_host_addr(block);
> + strcpy(info->idstr, qemu_ram_get_idstr(block));
> +}
> +
> +static struct block_dirty_info *
> +alloc_block_dirty_info(int *block_index,
> + struct block_dirty_info *block_dinfo)
> +{
> + struct block_dirty_info *info = NULL;
> + int index = *block_index;
> +
> + if (!block_dinfo) {
> + block_dinfo = g_new(struct block_dirty_info, 1);
> + index = 0;
> + } else {
> + block_dinfo = g_realloc(block_dinfo, (index + 1) *
> + sizeof(struct block_dirty_info));
> + index++;
I think g_realloc works on a NULL pointer, so you might be able to
simplify.
> + }
> + info = &block_dinfo[index];
> + memset(info, 0, sizeof(struct block_dirty_info));
> +
> + *block_index = index;
> + return block_dinfo;
> +}
> +
> +static int ram_block_skip(RAMBlock *block)
> +{
> + if (!strstr(qemu_ram_get_idstr(block), "ram-node") &&
> + !strstr(qemu_ram_get_idstr(block), "memdimm")) {
> + if (strcmp(qemu_ram_get_idstr(block), "mach-virt.ram") ||
> + strcmp(block->idstr, "pc.ram")) {
> + return -1;
> + }
> + }
We can't tie this to the names you guess that RAMBlocks might have -
things like 'memdimm' are settable by the caller and can be anything,
so you need to picka different way of chosing which RAMBlocks to
use; I suggest anything larger than some cutoff size, that's really RAM.
> + return 0;
> +}
> +
> +static int record_block_hash_info(struct dirtyrate_config config,
> + struct block_dirty_info **block_dinfo, int
> *block_index)
> +{
> + struct block_dirty_info *info = NULL;
> + struct block_dirty_info *dinfo = NULL;
> + RAMBlock *block = NULL;
> + int index = 0;
> +
> + RAMBLOCK_FOREACH_MIGRATABLE(block) {
> + if (ram_block_skip(block) < 0) {
> + continue;
> + }
> + dinfo = alloc_block_dirty_info(&index, dinfo);
> + info = &dinfo[index];
> + get_block_dirty_info(block, info, &config);
> + if (save_block_hash(info) < 0) {
> + *block_dinfo = dinfo;
> + *block_index = index;
> + return -1;
> + }
> + }
> +
> + *block_dinfo = dinfo;
> + *block_index = index;
> +
> + return 0;
> +}
>
> static void calculate_dirtyrate(struct dirtyrate_config config, int64_t time)
> {
> diff --git a/migration/dirtyrate.h b/migration/dirtyrate.h
> index 2994535..4d9b3b8 100644
> --- a/migration/dirtyrate.h
> +++ b/migration/dirtyrate.h
> @@ -15,6 +15,7 @@
>
> /* take 256 pages per GB for cal dirty rate */
> #define DIRTYRATE_DEFAULT_SAMPLE_PAGES 256
> +#define DIRTYRATE_SAMPLE_PAGE_SIZE 4096
> #define DIRTYRATE_PAGE_SIZE_SHIFT 12
> #define BLOCK_INFO_MAX_LEN 256
> #define PAGE_SIZE_SHIFT 20
> --
> 1.8.3.1
>
--
Dr. David Alan Gilbert / dgilb...@redhat.com / Manchester, UK
