[PATCH v4] Improve the performance of --num-threads -d 31
"Zhou, Wenjian/周文剑"
zhouwj-fnst at cn.fujitsu.com
Thu Mar 17 22:48:40 PDT 2016
On 03/18/2016 12:16 PM, Minfei Huang wrote:
> On 03/18/16 at 10:46am, "Zhou, Wenjian/周文剑" wrote:
>> Hello Minfei,
>>
>> Since I can't produce the bug, I reviewed the patch and wrote an increment patch.
>> Though there are some bugs in the increment patch,
>> I wonder if the previous bug still exists with this patch.
>> Could you help me confirm it?
>
> Ok. I will help verify this increasing patch.
>
Thank you very much.
>>
>> And I have another question.
>> Did it only occur in patch v4?
>
> This issue doesn't exist in v3. I have pasted the test result with
> --num-thread 32 in that thread.
>
> applied makedumpfile with option -d 31 --num-threads 32
> real 3m3.533s
Oh, then the patch in the previous mail may not work.
I'm appreciated if you can also test the patch in this letter.
I introduced semaphore to fix the bug in the v3.
So I want to know if it is this which affects the result.
The attached patch is based on v4, used to remove semaohore.
--
Thanks
Zhou
>
> Thanks
> Minfei
>
>> The previous patches almost have the same logic.
>>
>> --
>> Thanks
>> Zhou
>>
>> On 03/15/2016 02:34 PM, Minfei Huang wrote:
>>> Hi, Zhou.
>>>
>>> I have applied this patch base on 1.5.9. There are several testcases I
>>> have tested.
>>>
>>> - makedumpfile --num-threads 64 -d 31
>>> real 0m0.010s
>>> user 0m0.002s
>>> sys 0m0.009s
>>>
>>> - makedumpfile --num-threads 31 -d 31
>>> real 2m40.915s
>>> user 10m50.900s
>>> sys 23m9.664s
>>>
>>> makedumpfile --num-threads 30 -d 31
>>> real 0m0.006s
>>> user 0m0.002s
>>> sys 0m0.004s
>>>
>>> makedumpfile --num-threads 32 -d 31
>>> real 0m0.007s
>>> user 0m0.002s
>>> sys 0m0.005s
>>>
>>> - makedumpfile --num-threads 8 -d 31
>>> real 2m32.692s
>>> user 7m4.630s
>>> sys 2m0.369s
>>>
>>> - makedumpfile --num-threads 1 -d 31
>>> real 4m42.423s
>>> user 7m27.153s
>>> sys 0m22.490s
>>>
>>> - makedumpfile.orig -d 31
>>> real 4m1.297s
>>> user 3m39.696s
>>> sys 0m15.200s
>>>
>>> This patch has a huge increment to the filter performance under 31. But
>>> it is not stable, since makedumpfile fails to dump vmcore intermittently.
>>> You can find the above test result, makedumpfile fails to dump vmcore
>>> with option --num-threads 64, also it may occur with option
>>> --number-threads 8.
>>>
>>> Thanks
>>> Minfei
>>>
>>> On 03/09/16 at 08:27am, Zhou Wenjian wrote:
>>>> v4:
>>>> 1. fix a bug caused by the logic
>>>> v3:
>>>> 1. remove some unused variables
>>>> 2. fix a bug caused by the wrong logic
>>>> 3. fix a bug caused by optimising
>>>> 4. improve more performance by using Minoru Usui's code
>>>>
>>>> multi-threads implementation will introduce extra cost when handling
>>>> each page. The origin implementation will also do the extra work for
>>>> filtered pages. So there is a big performance degradation in
>>>> --num-threads -d 31.
>>>> The new implementation won't do the extra work for filtered pages any
>>>> more. So the performance of -d 31 is close to that of serial processing.
>>>>
>>>> The new implementation is just like the following:
>>>> * The basic idea is producer producing page and consumer writing page.
>>>> * Each producer have a page_flag_buf list which is used for storing
>>>> page's description.
>>>> * The size of page_flag_buf is little so it won't take too much memory.
>>>> * And all producers will share a page_data_buf array which is
>>>> used for storing page's compressed data.
>>>> * The main thread is the consumer. It will find the next pfn and write
>>>> it into file.
>>>> * The next pfn is smallest pfn in all page_flag_buf.
>>>>
>>>> Signed-off-by: Minoru Usui <min-usui at ti.jp.nec.com>
>>>> Signed-off-by: Zhou Wenjian <zhouwj-fnst at cn.fujitsu.com>
>>>> ---
>>>> makedumpfile.c | 298 +++++++++++++++++++++++++++++++++++----------------------
>>>> makedumpfile.h | 35 ++++---
>>>> 2 files changed, 202 insertions(+), 131 deletions(-)
>>>>
>>>> diff --git a/makedumpfile.c b/makedumpfile.c
>>>> index fa0b779..2b0864a 100644
>>>> --- a/makedumpfile.c
>>>> +++ b/makedumpfile.c
>>>> @@ -3483,7 +3483,8 @@ initial_for_parallel()
>>>> unsigned long page_data_buf_size;
>>>> unsigned long limit_size;
>>>> int page_data_num;
>>>> - int i;
>>>> + struct page_flag *current;
>>>> + int i, j;
>>>>
>>>> len_buf_out = calculate_len_buf_out(info->page_size);
>>>>
>>>> @@ -3560,10 +3561,16 @@ initial_for_parallel()
>>>>
>>>> limit_size = (get_free_memory_size()
>>>> - MAP_REGION * info->num_threads) * 0.6;
>>>> + if (limit_size < 0) {
>>>> + MSG("Free memory is not enough for multi-threads\n");
>>>> + return FALSE;
>>>> + }
>>>>
>>>> page_data_num = limit_size / page_data_buf_size;
>>>> + info->num_buffers = 3 * info->num_threads;
>>>>
>>>> - info->num_buffers = MIN(NUM_BUFFERS, page_data_num);
>>>> + info->num_buffers = MAX(info->num_buffers, NUM_BUFFERS);
>>>> + info->num_buffers = MIN(info->num_buffers, page_data_num);
>>>>
>>>> DEBUG_MSG("Number of struct page_data for produce/consume: %d\n",
>>>> info->num_buffers);
>>>> @@ -3588,6 +3595,36 @@ initial_for_parallel()
>>>> }
>>>>
>>>> /*
>>>> + * initial page_flag for each thread
>>>> + */
>>>> + if ((info->page_flag_buf = malloc(sizeof(void *) * info->num_threads))
>>>> + == NULL) {
>>>> + MSG("Can't allocate memory for page_flag_buf. %s\n",
>>>> + strerror(errno));
>>>> + return FALSE;
>>>> + }
>>>> + memset(info->page_flag_buf, 0, sizeof(void *) * info->num_threads);
>>>> +
>>>> + for (i = 0; i < info->num_threads; i++) {
>>>> + if ((info->page_flag_buf[i] = calloc(1, sizeof(struct page_flag))) == NULL) {
>>>> + MSG("Can't allocate memory for page_flag. %s\n",
>>>> + strerror(errno));
>>>> + return FALSE;
>>>> + }
>>>> + current = info->page_flag_buf[i];
>>>> +
>>>> + for (j = 1; j < NUM_BUFFERS; j++) {
>>>> + if ((current->next = calloc(1, sizeof(struct page_flag))) == NULL) {
>>>> + MSG("Can't allocate memory for page_flag. %s\n",
>>>> + strerror(errno));
>>>> + return FALSE;
>>>> + }
>>>> + current = current->next;
>>>> + }
>>>> + current->next = info->page_flag_buf[i];
>>>> + }
>>>> +
>>>> + /*
>>>> * initial fd_memory for threads
>>>> */
>>>> for (i = 0; i < info->num_threads; i++) {
>>>> @@ -3612,7 +3649,8 @@ initial_for_parallel()
>>>> void
>>>> free_for_parallel()
>>>> {
>>>> - int i;
>>>> + int i, j;
>>>> + struct page_flag *current;
>>>>
>>>> if (info->threads != NULL) {
>>>> for (i = 0; i < info->num_threads; i++) {
>>>> @@ -3655,6 +3693,19 @@ free_for_parallel()
>>>> free(info->page_data_buf);
>>>> }
>>>>
>>>> + if (info->page_flag_buf != NULL) {
>>>> + for (i = 0; i < info->num_threads; i++) {
>>>> + for (j = 0; j < NUM_BUFFERS; j++) {
>>>> + if (info->page_flag_buf[i] != NULL) {
>>>> + current = info->page_flag_buf[i];
>>>> + info->page_flag_buf[i] = current->next;
>>>> + free(current);
>>>> + }
>>>> + }
>>>> + }
>>>> + free(info->page_flag_buf);
>>>> + }
>>>> +
>>>> if (info->parallel_info == NULL)
>>>> return;
>>>>
>>>> @@ -7075,11 +7126,11 @@ void *
>>>> kdump_thread_function_cyclic(void *arg) {
>>>> void *retval = PTHREAD_FAIL;
>>>> struct thread_args *kdump_thread_args = (struct thread_args *)arg;
>>>> - struct page_data *page_data_buf = kdump_thread_args->page_data_buf;
>>>> + volatile struct page_data *page_data_buf = kdump_thread_args->page_data_buf;
>>>> + volatile struct page_flag *page_flag_buf = kdump_thread_args->page_flag_buf;
>>>> struct cycle *cycle = kdump_thread_args->cycle;
>>>> - int page_data_num = kdump_thread_args->page_data_num;
>>>> - mdf_pfn_t pfn;
>>>> - int index;
>>>> + mdf_pfn_t pfn = cycle->start_pfn;
>>>> + int index = kdump_thread_args->thread_num;
>>>> int buf_ready;
>>>> int dumpable;
>>>> int fd_memory = 0;
>>>> @@ -7125,47 +7176,48 @@ kdump_thread_function_cyclic(void *arg) {
>>>> kdump_thread_args->thread_num);
>>>> }
>>>>
>>>> - while (1) {
>>>> - /* get next pfn */
>>>> - pthread_mutex_lock(&info->current_pfn_mutex);
>>>> - pfn = info->current_pfn;
>>>> - info->current_pfn++;
>>>> - pthread_mutex_unlock(&info->current_pfn_mutex);
>>>> -
>>>> - if (pfn >= kdump_thread_args->end_pfn)
>>>> - break;
>>>> -
>>>> - index = -1;
>>>> + /*
>>>> + * filtered page won't take anything
>>>> + * unfiltered zero page will only take a page_flag_buf
>>>> + * unfiltered non-zero page will take a page_flag_buf and a page_data_buf
>>>> + */
>>>> + while (pfn < cycle->end_pfn) {
>>>> buf_ready = FALSE;
>>>>
>>>> + pthread_mutex_lock(&info->page_data_mutex);
>>>> + while (page_data_buf[index].used != FALSE) {
>>>> + index = (index + 1) % info->num_buffers;
>>>> + }
>>>> + page_data_buf[index].used = TRUE;
>>>> + pthread_mutex_unlock(&info->page_data_mutex);
>>>> +
>>>> while (buf_ready == FALSE) {
>>>> pthread_testcancel();
>>>> -
>>>> - index = pfn % page_data_num;
>>>> -
>>>> - if (pfn - info->consumed_pfn > info->num_buffers)
>>>> + if (page_flag_buf->ready == FLAG_READY)
>>>> continue;
>>>>
>>>> - if (page_data_buf[index].ready != 0)
>>>> - continue;
>>>> -
>>>> - pthread_mutex_lock(&page_data_buf[index].mutex);
>>>> -
>>>> - if (page_data_buf[index].ready != 0)
>>>> - goto unlock;
>>>> -
>>>> - buf_ready = TRUE;
>>>> + /* get next dumpable pfn */
>>>> + pthread_mutex_lock(&info->current_pfn_mutex);
>>>> + for (pfn = info->current_pfn; pfn < cycle->end_pfn; pfn++) {
>>>> + dumpable = is_dumpable(
>>>> + info->fd_bitmap ? &bitmap_parallel : info->bitmap2,
>>>> + pfn,
>>>> + cycle);
>>>> + if (dumpable)
>>>> + break;
>>>> + }
>>>> + info->current_pfn = pfn + 1;
>>>>
>>>> - page_data_buf[index].pfn = pfn;
>>>> - page_data_buf[index].ready = 1;
>>>> + page_flag_buf->pfn = pfn;
>>>> + page_flag_buf->ready = FLAG_FILLING;
>>>> + pthread_mutex_unlock(&info->current_pfn_mutex);
>>>> + sem_post(&info->page_flag_buf_sem);
>>>>
>>>> - dumpable = is_dumpable(
>>>> - info->fd_bitmap ? &bitmap_parallel : info->bitmap2,
>>>> - pfn,
>>>> - cycle);
>>>> - page_data_buf[index].dumpable = dumpable;
>>>> - if (!dumpable)
>>>> - goto unlock;
>>>> + if (pfn >= cycle->end_pfn) {
>>>> + info->current_pfn = cycle->end_pfn;
>>>> + page_data_buf[index].used = FALSE;
>>>> + break;
>>>> + }
>>>>
>>>> if (!read_pfn_parallel(fd_memory, pfn, buf,
>>>> &bitmap_memory_parallel,
>>>> @@ -7178,11 +7230,11 @@ kdump_thread_function_cyclic(void *arg) {
>>>>
>>>> if ((info->dump_level & DL_EXCLUDE_ZERO)
>>>> && is_zero_page(buf, info->page_size)) {
>>>> - page_data_buf[index].zero = TRUE;
>>>> - goto unlock;
>>>> + page_flag_buf->zero = TRUE;
>>>> + goto next;
>>>> }
>>>>
>>>> - page_data_buf[index].zero = FALSE;
>>>> + page_flag_buf->zero = FALSE;
>>>>
>>>> /*
>>>> * Compress the page data.
>>>> @@ -7210,6 +7262,7 @@ kdump_thread_function_cyclic(void *arg) {
>>>> page_data_buf[index].flags =
>>>> DUMP_DH_COMPRESSED_LZO;
>>>> page_data_buf[index].size = size_out;
>>>> +
>>>> memcpy(page_data_buf[index].buf, buf_out, size_out);
>>>> #endif
>>>> #ifdef USESNAPPY
>>>> @@ -7232,12 +7285,14 @@ kdump_thread_function_cyclic(void *arg) {
>>>> page_data_buf[index].size = info->page_size;
>>>> memcpy(page_data_buf[index].buf, buf, info->page_size);
>>>> }
>>>> -unlock:
>>>> - pthread_mutex_unlock(&page_data_buf[index].mutex);
>>>> + page_flag_buf->index = index;
>>>> + buf_ready = TRUE;
>>>> +next:
>>>> + page_flag_buf->ready = FLAG_READY;
>>>> + page_flag_buf = page_flag_buf->next;
>>>>
>>>> }
>>>> }
>>>> -
>>>> retval = NULL;
>>>>
>>>> fail:
>>>> @@ -7265,14 +7320,15 @@ write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>>>> struct page_desc pd;
>>>> struct timeval tv_start;
>>>> struct timeval last, new;
>>>> - unsigned long long consuming_pfn;
>>>> pthread_t **threads = NULL;
>>>> struct thread_args *kdump_thread_args = NULL;
>>>> void *thread_result;
>>>> - int page_data_num;
>>>> + int page_buf_num;
>>>> struct page_data *page_data_buf = NULL;
>>>> int i;
>>>> int index;
>>>> + int end_count, consuming, check_count;
>>>> + mdf_pfn_t current_pfn, temp_pfn;
>>>>
>>>> if (info->flag_elf_dumpfile)
>>>> return FALSE;
>>>> @@ -7284,13 +7340,6 @@ write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>>>> goto out;
>>>> }
>>>>
>>>> - res = pthread_mutex_init(&info->consumed_pfn_mutex, NULL);
>>>> - if (res != 0) {
>>>> - ERRMSG("Can't initialize consumed_pfn_mutex. %s\n",
>>>> - strerror(res));
>>>> - goto out;
>>>> - }
>>>> -
>>>> res = pthread_mutex_init(&info->filter_mutex, NULL);
>>>> if (res != 0) {
>>>> ERRMSG("Can't initialize filter_mutex. %s\n", strerror(res));
>>>> @@ -7314,36 +7363,23 @@ write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>>>> end_pfn = cycle->end_pfn;
>>>>
>>>> info->current_pfn = start_pfn;
>>>> - info->consumed_pfn = start_pfn - 1;
>>>>
>>>> threads = info->threads;
>>>> kdump_thread_args = info->kdump_thread_args;
>>>>
>>>> - page_data_num = info->num_buffers;
>>>> + page_buf_num = info->num_buffers;
>>>> page_data_buf = info->page_data_buf;
>>>> + pthread_mutex_init(&info->page_data_mutex, NULL);
>>>> + sem_init(&info->page_flag_buf_sem, 0, 0);
>>>>
>>>> - for (i = 0; i < page_data_num; i++) {
>>>> - /*
>>>> - * producer will use pfn in page_data_buf to decide the
>>>> - * consumed pfn
>>>> - */
>>>> - page_data_buf[i].pfn = start_pfn - 1;
>>>> - page_data_buf[i].ready = 0;
>>>> - res = pthread_mutex_init(&page_data_buf[i].mutex, NULL);
>>>> - if (res != 0) {
>>>> - ERRMSG("Can't initialize mutex of page_data_buf. %s\n",
>>>> - strerror(res));
>>>> - goto out;
>>>> - }
>>>> - }
>>>> + for (i = 0; i < page_buf_num; i++)
>>>> + page_data_buf[i].used = FALSE;
>>>>
>>>> for (i = 0; i < info->num_threads; i++) {
>>>> kdump_thread_args[i].thread_num = i;
>>>> kdump_thread_args[i].len_buf_out = len_buf_out;
>>>> - kdump_thread_args[i].start_pfn = start_pfn;
>>>> - kdump_thread_args[i].end_pfn = end_pfn;
>>>> - kdump_thread_args[i].page_data_num = page_data_num;
>>>> kdump_thread_args[i].page_data_buf = page_data_buf;
>>>> + kdump_thread_args[i].page_flag_buf = info->page_flag_buf[i];
>>>> kdump_thread_args[i].cycle = cycle;
>>>>
>>>> res = pthread_create(threads[i], NULL,
>>>> @@ -7356,55 +7392,88 @@ write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>>>> }
>>>> }
>>>>
>>>> - consuming_pfn = start_pfn;
>>>> - index = -1;
>>>> + end_count = 0;
>>>> + while (1) {
>>>> + consuming = 0;
>>>> + check_count = 0;
>>>>
>>>> - gettimeofday(&last, NULL);
>>>> + /*
>>>> + * The basic idea is producer producing page and consumer writing page.
>>>> + * Each producer have a page_flag_buf list which is used for storing page's description.
>>>> + * The size of page_flag_buf is little so it won't take too much memory.
>>>> + * And all producers will share a page_data_buf array which is used for storing page's compressed data.
>>>> + * The main thread is the consumer. It will find the next pfn and write it into file.
>>>> + * The next pfn is smallest pfn in all page_flag_buf.
>>>> + */
>>>> + sem_wait(&info->page_flag_buf_sem);
>>>> + gettimeofday(&last, NULL);
>>>> + while (1) {
>>>> + current_pfn = end_pfn;
>>>>
>>>> - while (consuming_pfn < end_pfn) {
>>>> - index = consuming_pfn % page_data_num;
>>>> + /*
>>>> + * page_flag_buf is in circular linked list.
>>>> + * The array info->page_flag_buf[] records the current page_flag_buf in each thread's
>>>> + * page_flag_buf list.
>>>> + * consuming is used for recording in which thread the pfn is the smallest.
>>>> + * current_pfn is used for recording the value of pfn when checking the pfn.
>>>> + */
>>>> + for (i = 0; i < info->num_threads; i++) {
>>>> + if (info->page_flag_buf[i]->ready == FLAG_UNUSED)
>>>> + continue;
>>>> + temp_pfn = info->page_flag_buf[i]->pfn;
>>>>
>>>> - gettimeofday(&new, NULL);
>>>> - if (new.tv_sec - last.tv_sec > WAIT_TIME) {
>>>> - ERRMSG("Can't get data of pfn %llx.\n", consuming_pfn);
>>>> - goto out;
>>>> - }
>>>> + /*
>>>> + * count how many threads have reached the end.
>>>> + */
>>>> + if (temp_pfn >= end_pfn) {
>>>> + info->page_flag_buf[i]->ready = FLAG_UNUSED;
>>>> + end_count++;
>>>> + continue;
>>>> + }
>>>>
>>>> - /*
>>>> - * check pfn first without mutex locked to reduce the time
>>>> - * trying to lock the mutex
>>>> - */
>>>> - if (page_data_buf[index].pfn != consuming_pfn)
>>>> - continue;
>>>> + if (current_pfn < temp_pfn)
>>>> + continue;
>>>>
>>>> - if (pthread_mutex_trylock(&page_data_buf[index].mutex) != 0)
>>>> - continue;
>>>> + check_count++;
>>>> + consuming = i;
>>>> + current_pfn = temp_pfn;
>>>> + }
>>>>
>>>> - /* check whether the found one is ready to be consumed */
>>>> - if (page_data_buf[index].pfn != consuming_pfn ||
>>>> - page_data_buf[index].ready != 1) {
>>>> - goto unlock;
>>>> + /*
>>>> + * If all the threads have reached the end, we will finish writing.
>>>> + */
>>>> + if (end_count >= info->num_threads)
>>>> + goto finish;
>>>> +
>>>> + /*
>>>> + * If the page_flag_buf is not ready, the pfn recorded may be changed.
>>>> + * So we should recheck.
>>>> + */
>>>> + if (info->page_flag_buf[consuming]->ready != FLAG_READY) {
>>>> + gettimeofday(&new, NULL);
>>>> + if (new.tv_sec - last.tv_sec > WAIT_TIME) {
>>>> + ERRMSG("Can't get data of pfn.\n");
>>>> + goto out;
>>>> + }
>>>> + continue;
>>>> + }
>>>> +
>>>> + if (current_pfn == info->page_flag_buf[consuming]->pfn)
>>>> + break;
>>>> }
>>>>
>>>> if ((num_dumped % per) == 0)
>>>> print_progress(PROGRESS_COPY, num_dumped, info->num_dumpable);
>>>>
>>>> - /* next pfn is found, refresh last here */
>>>> - last = new;
>>>> - consuming_pfn++;
>>>> - info->consumed_pfn++;
>>>> - page_data_buf[index].ready = 0;
>>>> -
>>>> - if (page_data_buf[index].dumpable == FALSE)
>>>> - goto unlock;
>>>> -
>>>> num_dumped++;
>>>>
>>>> - if (page_data_buf[index].zero == TRUE) {
>>>> +
>>>> + if (info->page_flag_buf[consuming]->zero == TRUE) {
>>>> if (!write_cache(cd_header, pd_zero, sizeof(page_desc_t)))
>>>> goto out;
>>>> pfn_zero++;
>>>> } else {
>>>> + index = info->page_flag_buf[consuming]->index;
>>>> pd.flags = page_data_buf[index].flags;
>>>> pd.size = page_data_buf[index].size;
>>>> pd.page_flags = 0;
>>>> @@ -7420,12 +7489,12 @@ write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>>>> */
>>>> if (!write_cache(cd_page, page_data_buf[index].buf, pd.size))
>>>> goto out;
>>>> -
>>>> + page_data_buf[index].used = FALSE;
>>>> }
>>>> -unlock:
>>>> - pthread_mutex_unlock(&page_data_buf[index].mutex);
>>>> + info->page_flag_buf[consuming]->ready = FLAG_UNUSED;
>>>> + info->page_flag_buf[consuming] = info->page_flag_buf[consuming]->next;
>>>> }
>>>> -
>>>> +finish:
>>>> ret = TRUE;
>>>> /*
>>>> * print [100 %]
>>>> @@ -7463,15 +7532,9 @@ out:
>>>> }
>>>> }
>>>>
>>>> - if (page_data_buf != NULL) {
>>>> - for (i = 0; i < page_data_num; i++) {
>>>> - pthread_mutex_destroy(&page_data_buf[i].mutex);
>>>> - }
>>>> - }
>>>> -
>>>> + sem_destroy(&info->page_flag_buf_sem);
>>>> pthread_rwlock_destroy(&info->usemmap_rwlock);
>>>> pthread_mutex_destroy(&info->filter_mutex);
>>>> - pthread_mutex_destroy(&info->consumed_pfn_mutex);
>>>> pthread_mutex_destroy(&info->current_pfn_mutex);
>>>>
>>>> return ret;
>>>> @@ -7564,6 +7627,7 @@ write_kdump_pages_cyclic(struct cache_data *cd_header, struct cache_data *cd_pag
>>>> num_dumped++;
>>>> if (!read_pfn(pfn, buf))
>>>> goto out;
>>>> +
>>>> filter_data_buffer(buf, pfn_to_paddr(pfn), info->page_size);
>>>>
>>>> /*
>>>> diff --git a/makedumpfile.h b/makedumpfile.h
>>>> index e0b5bbf..4b315c0 100644
>>>> --- a/makedumpfile.h
>>>> +++ b/makedumpfile.h
>>>> @@ -44,6 +44,7 @@
>>>> #include "print_info.h"
>>>> #include "sadump_mod.h"
>>>> #include <pthread.h>
>>>> +#include <semaphore.h>
>>>>
>>>> /*
>>>> * Result of command
>>>> @@ -977,7 +978,7 @@ typedef unsigned long long int ulonglong;
>>>> #define PAGE_DATA_NUM (50)
>>>> #define WAIT_TIME (60 * 10)
>>>> #define PTHREAD_FAIL ((void *)-2)
>>>> -#define NUM_BUFFERS (50)
>>>> +#define NUM_BUFFERS (20)
>>>>
>>>> struct mmap_cache {
>>>> char *mmap_buf;
>>>> @@ -985,28 +986,33 @@ struct mmap_cache {
>>>> off_t mmap_end_offset;
>>>> };
>>>>
>>>> +enum {
>>>> + FLAG_UNUSED,
>>>> + FLAG_READY,
>>>> + FLAG_FILLING
>>>> +};
>>>> +struct page_flag {
>>>> + mdf_pfn_t pfn;
>>>> + char zero;
>>>> + char ready;
>>>> + short index;
>>>> + struct page_flag *next;
>>>> +};
>>>> +
>>>> struct page_data
>>>> {
>>>> - mdf_pfn_t pfn;
>>>> - int dumpable;
>>>> - int zero;
>>>> - unsigned int flags;
>>>> long size;
>>>> unsigned char *buf;
>>>> - pthread_mutex_t mutex;
>>>> - /*
>>>> - * whether the page_data is ready to be consumed
>>>> - */
>>>> - int ready;
>>>> + int flags;
>>>> + int used;
>>>> };
>>>>
>>>> struct thread_args {
>>>> int thread_num;
>>>> unsigned long len_buf_out;
>>>> - mdf_pfn_t start_pfn, end_pfn;
>>>> - int page_data_num;
>>>> struct cycle *cycle;
>>>> struct page_data *page_data_buf;
>>>> + struct page_flag *page_flag_buf;
>>>> };
>>>>
>>>> /*
>>>> @@ -1295,11 +1301,12 @@ struct DumpInfo {
>>>> pthread_t **threads;
>>>> struct thread_args *kdump_thread_args;
>>>> struct page_data *page_data_buf;
>>>> + struct page_flag **page_flag_buf;
>>>> + sem_t page_flag_buf_sem;
>>>> pthread_rwlock_t usemmap_rwlock;
>>>> mdf_pfn_t current_pfn;
>>>> pthread_mutex_t current_pfn_mutex;
>>>> - mdf_pfn_t consumed_pfn;
>>>> - pthread_mutex_t consumed_pfn_mutex;
>>>> + pthread_mutex_t page_data_mutex;
>>>> pthread_mutex_t filter_mutex;
>>>> };
>>>> extern struct DumpInfo *info;
>>>> --
>>>> 1.8.3.1
>>>>
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> kexec mailing list
>>>> kexec at lists.infradead.org
>>>> http://lists.infradead.org/mailman/listinfo/kexec
>>
>>
>>
>
>> From 1d7ad5dbbc29efe58171b1023ab0df09eb2815bf Mon Sep 17 00:00:00 2001
>> From: Zhou Wenjian <zhouwj-fnst at cn.fujitsu.com>
>> Date: Fri, 18 Mar 2016 10:35:35 +0800
>> Subject: [PATCH] increment
>>
>> ---
>> makedumpfile.c | 107 ++++++++++++++++++++++++++++++++++++++++++---------------
>> makedumpfile.h | 12 +++++--
>> 2 files changed, 89 insertions(+), 30 deletions(-)
>>
>> diff --git a/makedumpfile.c b/makedumpfile.c
>> index 2b0864a..a304a61 100644
>> --- a/makedumpfile.c
>> +++ b/makedumpfile.c
>> @@ -3477,6 +3477,42 @@ calibrate_machdep_info(void)
>> }
>>
>> int
>> +initial_parallel_area(int page_data_buf_size)
>> +{
>> + int i, j;
>> + struct page_flag *current;
>> + info->page_data_buf = info->parallel_area;
>> + void *page_data_buf = info->parallel_area + sizeof(struct page_data) * info->num_buffers;
>> + void *page_flag_list = page_data_buf + page_data_buf_size * info->num_buffers;
>> +
>> + for (i = 0; i < info->num_buffers; i++) {
>> + info->page_data_buf[i].buf = page_data_buf + page_data_buf_size * i;
>> + }
>> +
>> +
>> + if ((info->page_flag_list = malloc(sizeof(struct page_flag_list) * info->num_threads))
>> + == NULL) {
>> + MSG("Can't allocate memory for page_flag_buf. %s\n",
>> + strerror(errno));
>> + return FALSE;
>> + }
>> +
>> + for (i = 0; i < info->num_threads; i++) {
>> + info->page_flag_list[i].header = page_flag_list + sizeof(struct page_flag) * info->num_buffers * i;
>> + info->page_flag_list[i].current = 0;
>> + /*
>> + current = info->page_flag_buf[i];
>> + for (j = 1; j < info->num_buffers; j++) {
>> + current->next = current + sizeof(struct page_flag);
>> + current = current->next;
>> + }
>> + current->next = info->page_flag_buf[i];
>> +*/
>> + }
>> +
>> +}
>> +
>> +int
>> initial_for_parallel()
>> {
>> unsigned long len_buf_out;
>> @@ -3575,9 +3611,15 @@ initial_for_parallel()
>> DEBUG_MSG("Number of struct page_data for produce/consume: %d\n",
>> info->num_buffers);
>>
>> - /*
>> - * allocate memory for page_data
>> - */
>> + if ((info->parallel_area = calloc(info->num_buffers, sizeof(struct page_data)+page_data_buf_size + sizeof(struct page_flag)*info->num_threads))
>> + ==NULL) {
>> + MSG("Can't allocate memory for page_data_buf. %s\n",
>> + strerror(errno));
>> + return FALSE;
>> + }
>> +
>> + initial_parallel_area(page_data_buf_size);
>> +/*
>> if ((info->page_data_buf = malloc(sizeof(struct page_data) * info->num_buffers))
>> == NULL) {
>> MSG("Can't allocate memory for page_data_buf. %s\n",
>> @@ -3594,9 +3636,6 @@ initial_for_parallel()
>> }
>> }
>>
>> - /*
>> - * initial page_flag for each thread
>> - */
>> if ((info->page_flag_buf = malloc(sizeof(void *) * info->num_threads))
>> == NULL) {
>> MSG("Can't allocate memory for page_flag_buf. %s\n",
>> @@ -3623,7 +3662,7 @@ initial_for_parallel()
>> }
>> current->next = info->page_flag_buf[i];
>> }
>> -
>> +*/
>> /*
>> * initial fd_memory for threads
>> */
>> @@ -3685,6 +3724,15 @@ free_for_parallel()
>> if (info->kdump_thread_args != NULL)
>> free(info->kdump_thread_args);
>>
>> + if (info->page_flag_list != NULL) {
>> + free(info->page_flag_list);
>> + }
>> +
>> + if (info->parallel_area != NULL) {
>> + free(info->parallel_area);
>> + }
>> +
>> +/*
>> if (info->page_data_buf != NULL) {
>> for (i = 0; i < info->num_buffers; i++) {
>> if (info->page_data_buf[i].buf != NULL)
>> @@ -3705,7 +3753,7 @@ free_for_parallel()
>> }
>> free(info->page_flag_buf);
>> }
>> -
>> +*/
>> if (info->parallel_info == NULL)
>> return;
>>
>> @@ -7122,12 +7170,14 @@ int finalize_zlib(z_stream *stream)
>> return err;
>> }
>>
>> +#define CURRENT_FLAG page_flag_header[current_page_flag]
>> void *
>> kdump_thread_function_cyclic(void *arg) {
>> void *retval = PTHREAD_FAIL;
>> struct thread_args *kdump_thread_args = (struct thread_args *)arg;
>> volatile struct page_data *page_data_buf = kdump_thread_args->page_data_buf;
>> - volatile struct page_flag *page_flag_buf = kdump_thread_args->page_flag_buf;
>> + struct page_flag *page_flag_header = kdump_thread_args->page_flag_list->header;
>> + int current_page_flag = kdump_thread_args->page_flag_list->current;
>> struct cycle *cycle = kdump_thread_args->cycle;
>> mdf_pfn_t pfn = cycle->start_pfn;
>> int index = kdump_thread_args->thread_num;
>> @@ -7193,7 +7243,7 @@ kdump_thread_function_cyclic(void *arg) {
>>
>> while (buf_ready == FALSE) {
>> pthread_testcancel();
>> - if (page_flag_buf->ready == FLAG_READY)
>> + if (CURRENT_FLAG.ready == FLAG_READY)
>> continue;
>>
>> /* get next dumpable pfn */
>> @@ -7208,8 +7258,8 @@ kdump_thread_function_cyclic(void *arg) {
>> }
>> info->current_pfn = pfn + 1;
>>
>> - page_flag_buf->pfn = pfn;
>> - page_flag_buf->ready = FLAG_FILLING;
>> + CURRENT_FLAG.pfn = pfn;
>> + CURRENT_FLAG.ready = FLAG_FILLING;
>> pthread_mutex_unlock(&info->current_pfn_mutex);
>> sem_post(&info->page_flag_buf_sem);
>>
>> @@ -7230,11 +7280,11 @@ kdump_thread_function_cyclic(void *arg) {
>>
>> if ((info->dump_level & DL_EXCLUDE_ZERO)
>> && is_zero_page(buf, info->page_size)) {
>> - page_flag_buf->zero = TRUE;
>> + CURRENT_FLAG.zero = TRUE;
>> goto next;
>> }
>>
>> - page_flag_buf->zero = FALSE;
>> + CURRENT_FLAG.zero = FALSE;
>>
>> /*
>> * Compress the page data.
>> @@ -7285,11 +7335,11 @@ kdump_thread_function_cyclic(void *arg) {
>> page_data_buf[index].size = info->page_size;
>> memcpy(page_data_buf[index].buf, buf, info->page_size);
>> }
>> - page_flag_buf->index = index;
>> + CURRENT_FLAG.index = index;
>> buf_ready = TRUE;
>> next:
>> - page_flag_buf->ready = FLAG_READY;
>> - page_flag_buf = page_flag_buf->next;
>> + CURRENT_FLAG.ready = FLAG_READY;
>> + current_page_flag = (current_page_flag + 1) % info->num_buffers;
>>
>> }
>> }
>> @@ -7306,6 +7356,8 @@ fail:
>> pthread_exit(retval);
>> }
>>
>> +#define CURRENT_PAGE_FLAG(i) (info->page_flag_list[i].header)[info->page_flag_list[i].current]
>> +
>> int
>> write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>> struct cache_data *cd_page,
>> @@ -7379,7 +7431,7 @@ write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>> kdump_thread_args[i].thread_num = i;
>> kdump_thread_args[i].len_buf_out = len_buf_out;
>> kdump_thread_args[i].page_data_buf = page_data_buf;
>> - kdump_thread_args[i].page_flag_buf = info->page_flag_buf[i];
>> + kdump_thread_args[i].page_flag_list = &(info->page_flag_list[i]);
>> kdump_thread_args[i].cycle = cycle;
>>
>> res = pthread_create(threads[i], NULL,
>> @@ -7418,15 +7470,15 @@ write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>> * current_pfn is used for recording the value of pfn when checking the pfn.
>> */
>> for (i = 0; i < info->num_threads; i++) {
>> - if (info->page_flag_buf[i]->ready == FLAG_UNUSED)
>> + if (CURRENT_PAGE_FLAG(i).ready == FLAG_UNUSED)
>> continue;
>> - temp_pfn = info->page_flag_buf[i]->pfn;
>> + temp_pfn = CURRENT_PAGE_FLAG(i).pfn;
>>
>> /*
>> * count how many threads have reached the end.
>> */
>> if (temp_pfn >= end_pfn) {
>> - info->page_flag_buf[i]->ready = FLAG_UNUSED;
>> + CURRENT_PAGE_FLAG(i).ready = FLAG_UNUSED;
>> end_count++;
>> continue;
>> }
>> @@ -7449,7 +7501,7 @@ write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>> * If the page_flag_buf is not ready, the pfn recorded may be changed.
>> * So we should recheck.
>> */
>> - if (info->page_flag_buf[consuming]->ready != FLAG_READY) {
>> + if (CURRENT_PAGE_FLAG(consuming).ready != FLAG_READY) {
>> gettimeofday(&new, NULL);
>> if (new.tv_sec - last.tv_sec > WAIT_TIME) {
>> ERRMSG("Can't get data of pfn.\n");
>> @@ -7458,7 +7510,7 @@ write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>> continue;
>> }
>>
>> - if (current_pfn == info->page_flag_buf[consuming]->pfn)
>> + if (current_pfn == CURRENT_PAGE_FLAG(consuming).pfn)
>> break;
>> }
>>
>> @@ -7468,12 +7520,12 @@ write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>> num_dumped++;
>>
>>
>> - if (info->page_flag_buf[consuming]->zero == TRUE) {
>> + if (CURRENT_PAGE_FLAG(consuming).zero == TRUE) {
>> if (!write_cache(cd_header, pd_zero, sizeof(page_desc_t)))
>> goto out;
>> pfn_zero++;
>> } else {
>> - index = info->page_flag_buf[consuming]->index;
>> + index = CURRENT_PAGE_FLAG(consuming).index;
>> pd.flags = page_data_buf[index].flags;
>> pd.size = page_data_buf[index].size;
>> pd.page_flags = 0;
>> @@ -7491,8 +7543,9 @@ write_kdump_pages_parallel_cyclic(struct cache_data *cd_header,
>> goto out;
>> page_data_buf[index].used = FALSE;
>> }
>> - info->page_flag_buf[consuming]->ready = FLAG_UNUSED;
>> - info->page_flag_buf[consuming] = info->page_flag_buf[consuming]->next;
>> + CURRENT_PAGE_FLAG(consuming).ready = FLAG_UNUSED;
>> + info->page_flag_list[consuming].current += 1;
>> + info->page_flag_list[consuming].current %= info->num_buffers;
>> }
>> finish:
>> ret = TRUE;
>> diff --git a/makedumpfile.h b/makedumpfile.h
>> index 4b315c0..8c4bc1a 100644
>> --- a/makedumpfile.h
>> +++ b/makedumpfile.h
>> @@ -996,7 +996,12 @@ struct page_flag {
>> char zero;
>> char ready;
>> short index;
>> - struct page_flag *next;
>> +// struct page_flag *next;
>> +};
>> +
>> +struct page_flag_list {
>> + struct page_flag *header;
>> + int current;
>> };
>>
>> struct page_data
>> @@ -1012,7 +1017,7 @@ struct thread_args {
>> unsigned long len_buf_out;
>> struct cycle *cycle;
>> struct page_data *page_data_buf;
>> - struct page_flag *page_flag_buf;
>> + struct page_flag_list *page_flag_list;
>> };
>>
>> /*
>> @@ -1298,10 +1303,11 @@ struct DumpInfo {
>> */
>> int num_threads;
>> int num_buffers;
>> + void *parallel_area;
>> pthread_t **threads;
>> struct thread_args *kdump_thread_args;
>> struct page_data *page_data_buf;
>> - struct page_flag **page_flag_buf;
>> + struct page_flag_list *page_flag_list;
>> sem_t page_flag_buf_sem;
>> pthread_rwlock_t usemmap_rwlock;
>> mdf_pfn_t current_pfn;
>> --
>> 1.8.3.1
>>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-remove-sem.patch
Type: text/x-patch
Size: 2396 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/kexec/attachments/20160318/bf2e4d69/attachment-0001.bin>
More information about the kexec
mailing list