[PATCH v3] Improve the performance of --num-threads -d 31

[PATCH v3] Improve the performance of --num-threads -d 31

[Zhou Wenjian]

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@ti.jp.nec.com>
Signed-off-by: Zhou Wenjian <zhouwj-fnst@cn.fujitsu.com>
---
 makedumpfile.c | 300 +++++++++++++++++++++++++++++++++++----------------------
 makedumpfile.h |  33 ++++---
 2 files changed, 203 insertions(+), 130 deletions(-)

diff --git a/makedumpfile.c b/makedumpfile.c
index fa0b779..d75f7f7 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)
-				continue;
-
-			if (page_data_buf[index].ready != 0)
+			if (page_flag_buf->ready == FLAG_READY)
 				continue;
 
-			pthread_mutex_lock(&page_data_buf[index].mutex);
-
-			if (page_data_buf[index].ready != 0)
-				goto unlock;
+			/* 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;
 
-			buf_ready = TRUE;
+			page_flag_buf->pfn = pfn;
+			page_flag_buf->ready = FLAG_FILLING;
+			pthread_mutex_unlock(&info->current_pfn_mutex);
 
-			page_data_buf[index].pfn = pfn;
-			page_data_buf[index].ready = 1;
 
-			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 +7363,22 @@ 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);
 
-	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 +7391,94 @@ 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.
+		 */
+		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;
+			}
+
+			/*
+			 * If all the threads have reached the end, we will finish writing.
+			 */
+			if (end_count >= info->num_threads)
+				goto finish;
+
+			/*
+			 * Since it has the probabilty that there is no page_flag_buf being ready,
+			 * we should recheck if it happens.
+			 */
+			if (check_count == 0)
+				continue;
+
+			/*
+			 * 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;
+			}
 
-		/* 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 (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 +7494,12 @@ 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 +7537,8 @@ out:
 		}
 	}
 
-	if (page_data_buf != NULL) {
-		for (i = 0; i < page_data_num; i++) {
-			pthread_mutex_destroy(&page_data_buf[i].mutex);
-		}
-	}
-
 	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 +7631,7 @@ 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..5865827 100644
--- a/makedumpfile.h
+++ b/makedumpfile.h
@@ -977,7 +977,7 @@ 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 +985,33 @@ 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 +1300,11 @@ struct DumpInfo {
 	pthread_t **threads;
 	struct thread_args *kdump_thread_args;
 	struct page_data *page_data_buf;
+	struct page_flag **page_flag_buf;
 	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@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH v3] Improve the performance of --num-threads -d 31

["Zhou, Wenjian/周文剑"]

Hello Minfei and Minoru,

This is a reply to the previous letters.
Sorry for so late.
And it is also an introduction of the patch v3.

About performance:
I modified the patch after reading Minoru's code.
Thanks Minoru, it has better performance now.

About bugs:
There are two bugs in the patch v2.
First is reported by Minfei.
Minoru's explanation and solution are right.
But I used another solution.
+			if (pfn >= kdump_thread_args->end_pfn) {
+				page_data_buf[index].used = FALSE;
+				page_flag_buf->ready = FLAG_READY;
+				break;
+			}
The bug is caused by the above code.
Just remove "page_flag_buf->ready = FLAG_READY;" is OK.

Second bug is what I met.
If run makedumpfile a lot of times, a few pages' data will not be correct.
I fix it by using "volatile" and info->page_data_mutex.
So I delete the original mutex in page_data_buf. It can also use less memory.
I have run it 500 times and haven't got any wrong page data.

-- 
Thanks
Zhou

On 03/04/2016 04:35 PM, Zhou Wenjian wrote:
> 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@ti.jp.nec.com>
> Signed-off-by: Zhou Wenjian <zhouwj-fnst@cn.fujitsu.com>
> ---
>   makedumpfile.c | 300 +++++++++++++++++++++++++++++++++++----------------------
>   makedumpfile.h |  33 ++++---
>   2 files changed, 203 insertions(+), 130 deletions(-)
>
> diff --git a/makedumpfile.c b/makedumpfile.c
> index fa0b779..d75f7f7 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)
> -				continue;
> -
> -			if (page_data_buf[index].ready != 0)
> +			if (page_flag_buf->ready == FLAG_READY)
>   				continue;
>
> -			pthread_mutex_lock(&page_data_buf[index].mutex);
> -
> -			if (page_data_buf[index].ready != 0)
> -				goto unlock;
> +			/* 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;
>
> -			buf_ready = TRUE;
> +			page_flag_buf->pfn = pfn;
> +			page_flag_buf->ready = FLAG_FILLING;
> +			pthread_mutex_unlock(&info->current_pfn_mutex);
>
> -			page_data_buf[index].pfn = pfn;
> -			page_data_buf[index].ready = 1;
>
> -			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 +7363,22 @@ 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);
>
> -	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 +7391,94 @@ 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.
> +		 */
> +		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;
> +			}
> +
> +			/*
> +			 * If all the threads have reached the end, we will finish writing.
> +			 */
> +			if (end_count >= info->num_threads)
> +				goto finish;
> +
> +			/*
> +			 * Since it has the probabilty that there is no page_flag_buf being ready,
> +			 * we should recheck if it happens.
> +			 */
> +			if (check_count == 0)
> +				continue;
> +
> +			/*
> +			 * 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;
> +			}
>
> -		/* 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 (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 +7494,12 @@ 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 +7537,8 @@ out:
>   		}
>   	}
>
> -	if (page_data_buf != NULL) {
> -		for (i = 0; i < page_data_num; i++) {
> -			pthread_mutex_destroy(&page_data_buf[i].mutex);
> -		}
> -	}
> -
>   	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 +7631,7 @@ 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..5865827 100644
> --- a/makedumpfile.h
> +++ b/makedumpfile.h
> @@ -977,7 +977,7 @@ 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 +985,33 @@ 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 +1300,11 @@ struct DumpInfo {
>   	pthread_t **threads;
>   	struct thread_args *kdump_thread_args;
>   	struct page_data *page_data_buf;
> +	struct page_flag **page_flag_buf;
>   	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;
>



_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH v3] Improve the performance of --num-threads -d 31

["Zhou, Wenjian/周文剑"]

On 03/04/2016 05:07 PM, "Zhou, Wenjian/周文剑" wrote:
> Hello Minfei and Minoru,
>
> This is a reply to the previous letters.
> Sorry for so late.
> And it is also an introduction of the patch v3.
>
> About performance:
> I modified the patch after reading Minoru's code.
> Thanks Minoru, it has better performance now.
>
> About bugs:
> There are two bugs in the patch v2.
> First is reported by Minfei.
> Minoru's explanation and solution are right.
> But I used another solution.
> +            if (pfn >= kdump_thread_args->end_pfn) {
> +                page_data_buf[index].used = FALSE;
> +                page_flag_buf->ready = FLAG_READY;
> +                break;
> +            }
> The bug is caused by the above code.
> Just remove "page_flag_buf->ready = FLAG_READY;" is OK.
>
> Second bug is what I met.
> If run makedumpfile a lot of times, a few pages' data will not be correct.
> I fix it by using "volatile" and info->page_data_mutex.
> So I delete the original mutex in page_data_buf. It can also use less memory.
> I have run it 500 times and haven't got any wrong page data.
>

Sorry, the second bug still exists.


-- 
Thanks
Zhou



_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH v3] Improve the performance of --num-threads -d 31

[Minfei Huang]

On 03/07/16 at 08:17am, "Zhou, Wenjian/周文剑" wrote:
> On 03/04/2016 05:07 PM, "Zhou, Wenjian/周文剑" wrote:
> >Second bug is what I met.
> >If run makedumpfile a lot of times, a few pages' data will not be correct.
> >I fix it by using "volatile" and info->page_data_mutex.
> >So I delete the original mutex in page_data_buf. It can also use less memory.
> >I have run it 500 times and haven't got any wrong page data.
> >
> 
> Sorry, the second bug still exists.

Hi, Wenjian.

Due to the unresolved issue, I will hang on the test and wait for your
new update.

Thanks
Minfei

_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH v3] Improve the performance of --num-threads -d 31

[Minfei Huang]

Hi, Wenjian.

Since your patch applied Minoru Usui's draft patch, I have vierified it.
And it works well under 4T memory.

applied makedumpfile with option -d 31
real    3m48.681s

applied makedumpfile with option -d 31 --num-threads 32
real    3m3.533s

original makedumpfile with option -d 31
real    3m57.163s

Now with or without option --num-thread, it takes almost same time.

Thanks
Minfei

On 03/04/16 at 04:35pm, Zhou Wenjian wrote:
> 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@ti.jp.nec.com>
> Signed-off-by: Zhou Wenjian <zhouwj-fnst@cn.fujitsu.com>
> ---
>  makedumpfile.c | 300 +++++++++++++++++++++++++++++++++++----------------------
>  makedumpfile.h |  33 ++++---
>  2 files changed, 203 insertions(+), 130 deletions(-)
> 
> diff --git a/makedumpfile.c b/makedumpfile.c
> index fa0b779..d75f7f7 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)
> -				continue;
> -
> -			if (page_data_buf[index].ready != 0)
> +			if (page_flag_buf->ready == FLAG_READY)
>  				continue;
>  
> -			pthread_mutex_lock(&page_data_buf[index].mutex);
> -
> -			if (page_data_buf[index].ready != 0)
> -				goto unlock;
> +			/* 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;
>  
> -			buf_ready = TRUE;
> +			page_flag_buf->pfn = pfn;
> +			page_flag_buf->ready = FLAG_FILLING;
> +			pthread_mutex_unlock(&info->current_pfn_mutex);
>  
> -			page_data_buf[index].pfn = pfn;
> -			page_data_buf[index].ready = 1;
>  
> -			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 +7363,22 @@ 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);
>  
> -	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 +7391,94 @@ 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.
> +		 */
> +		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;
> +			}
> +
> +			/*
> +			 * If all the threads have reached the end, we will finish writing.
> +			 */
> +			if (end_count >= info->num_threads)
> +				goto finish;
> +
> +			/*
> +			 * Since it has the probabilty that there is no page_flag_buf being ready,
> +			 * we should recheck if it happens.
> +			 */
> +			if (check_count == 0)
> +				continue;
> +
> +			/*
> +			 * 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;
> +			}
>  
> -		/* 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 (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 +7494,12 @@ 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 +7537,8 @@ out:
>  		}
>  	}
>  
> -	if (page_data_buf != NULL) {
> -		for (i = 0; i < page_data_num; i++) {
> -			pthread_mutex_destroy(&page_data_buf[i].mutex);
> -		}
> -	}
> -
>  	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 +7631,7 @@ 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..5865827 100644
> --- a/makedumpfile.h
> +++ b/makedumpfile.h
> @@ -977,7 +977,7 @@ 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 +985,33 @@ 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 +1300,11 @@ struct DumpInfo {
>  	pthread_t **threads;
>  	struct thread_args *kdump_thread_args;
>  	struct page_data *page_data_buf;
> +	struct page_flag **page_flag_buf;
>  	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@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/kexec

_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH v3] Improve the performance of --num-threads -d 31

[Minoru Usui]

Hi, Minfei, Zhou

> -----Original Message-----
> From: kexec [mailto:kexec-bounces@lists.infradead.org] On Behalf Of Minfei Huang
> Sent: Monday, March 07, 2016 4:05 PM
> To: Zhou Wenjian <zhouwj-fnst@cn.fujitsu.com>
> Cc: kexec@lists.infradead.org
> Subject: Re: [PATCH v3] Improve the performance of --num-threads -d 31
> 
> Hi, Wenjian.
> 
> Since your patch applied Minoru Usui's draft patch, I have vierified it.
> And it works well under 4T memory.
> 
> applied makedumpfile with option -d 31
> real    3m48.681s
> 
> applied makedumpfile with option -d 31 --num-threads 32
> real    3m3.533s
> 
> original makedumpfile with option -d 31
> real    3m57.163s
> 
> Now with or without option --num-thread, it takes almost same time.

Thank you for your test on big memory machine.
I am relieved to see your test results.

Zhou, my patch may not be sophisticated, 
but getting mutex logic is robust, it works well and its performance is not bad.
I think it reaches the goal of the original purpose.
I understand you don't like getting mutex, but if you are debugging now, could you apply my approach?

Thanks,
Minoru Usui


> Thanks
> Minfei
> 
> On 03/04/16 at 04:35pm, Zhou Wenjian wrote:
> > 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@ti.jp.nec.com>
> > Signed-off-by: Zhou Wenjian <zhouwj-fnst@cn.fujitsu.com>
> > ---
> >  makedumpfile.c | 300 +++++++++++++++++++++++++++++++++++----------------------
> >  makedumpfile.h |  33 ++++---
> >  2 files changed, 203 insertions(+), 130 deletions(-)
> >
> > diff --git a/makedumpfile.c b/makedumpfile.c
> > index fa0b779..d75f7f7 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)
> > -				continue;
> > -
> > -			if (page_data_buf[index].ready != 0)
> > +			if (page_flag_buf->ready == FLAG_READY)
> >  				continue;
> >
> > -			pthread_mutex_lock(&page_data_buf[index].mutex);
> > -
> > -			if (page_data_buf[index].ready != 0)
> > -				goto unlock;
> > +			/* 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;
> >
> > -			buf_ready = TRUE;
> > +			page_flag_buf->pfn = pfn;
> > +			page_flag_buf->ready = FLAG_FILLING;
> > +			pthread_mutex_unlock(&info->current_pfn_mutex);
> >
> > -			page_data_buf[index].pfn = pfn;
> > -			page_data_buf[index].ready = 1;
> >
> > -			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 +7363,22 @@ 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);
> >
> > -	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 +7391,94 @@ 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.
> > +		 */
> > +		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;
> > +			}
> > +
> > +			/*
> > +			 * If all the threads have reached the end, we will finish writing.
> > +			 */
> > +			if (end_count >= info->num_threads)
> > +				goto finish;
> > +
> > +			/*
> > +			 * Since it has the probabilty that there is no page_flag_buf being ready,
> > +			 * we should recheck if it happens.
> > +			 */
> > +			if (check_count == 0)
> > +				continue;
> > +
> > +			/*
> > +			 * 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;
> > +			}
> >
> > -		/* 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 (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 +7494,12 @@ 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 +7537,8 @@ out:
> >  		}
> >  	}
> >
> > -	if (page_data_buf != NULL) {
> > -		for (i = 0; i < page_data_num; i++) {
> > -			pthread_mutex_destroy(&page_data_buf[i].mutex);
> > -		}
> > -	}
> > -
> >  	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 +7631,7 @@ 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..5865827 100644
> > --- a/makedumpfile.h
> > +++ b/makedumpfile.h
> > @@ -977,7 +977,7 @@ 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 +985,33 @@ 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 +1300,11 @@ struct DumpInfo {
> >  	pthread_t **threads;
> >  	struct thread_args *kdump_thread_args;
> >  	struct page_data *page_data_buf;
> > +	struct page_flag **page_flag_buf;
> >  	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@lists.infradead.org
> > http://lists.infradead.org/mailman/listinfo/kexec
> 
> _______________________________________________
> kexec mailing list
> kexec@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/kexec

_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH v3] Improve the performance of --num-threads -d 31

[Minoru Usui]

Hi, Minfei

> -----Original Message-----
> From: kexec [mailto:kexec-bounces@lists.infradead.org] On Behalf Of Minoru Usui
> Sent: Monday, March 07, 2016 4:37 PM
> To: Minfei Huang <mhuang@redhat.com>; Zhou Wenjian <zhouwj-fnst@cn.fujitsu.com>
> Cc: kexec@lists.infradead.org
> Subject: Re: [PATCH v3] Improve the performance of --num-threads -d 31
> 
> Hi, Minfei, Zhou
> 
> > -----Original Message-----
> > From: kexec [mailto:kexec-bounces@lists.infradead.org] On Behalf Of Minfei Huang
> > Sent: Monday, March 07, 2016 4:05 PM
> > To: Zhou Wenjian <zhouwj-fnst@cn.fujitsu.com>
> > Cc: kexec@lists.infradead.org
> > Subject: Re: [PATCH v3] Improve the performance of --num-threads -d 31
> >
> > Hi, Wenjian.
> >
> > Since your patch applied Minoru Usui's draft patch, I have vierified it.
> > And it works well under 4T memory.
> >
> > applied makedumpfile with option -d 31
> > real    3m48.681s
> >
> > applied makedumpfile with option -d 31 --num-threads 32
> > real    3m3.533s
> >
> > original makedumpfile with option -d 31
> > real    3m57.163s
> >
> > Now with or without option --num-thread, it takes almost same time.
> 
> Thank you for your test on big memory machine.
> I am relieved to see your test results.

I'm sorry, I may be misunderstood.
Is above result is Zhou's v3 patch or v2 + my patch?

If above result is Zhou's v3 patch, could you test v2 + my patch?

Thanks
Minoru Usui


> Zhou, my patch may not be sophisticated,
> but getting mutex logic is robust, it works well and its performance is not bad.
> I think it reaches the goal of the original purpose.
> I understand you don't like getting mutex, but if you are debugging now, could you apply my approach?
> 
> Thanks,
> Minoru Usui
> 
> 
> > Thanks
> > Minfei
> >
> > On 03/04/16 at 04:35pm, Zhou Wenjian wrote:
> > > 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@ti.jp.nec.com>
> > > Signed-off-by: Zhou Wenjian <zhouwj-fnst@cn.fujitsu.com>
> > > ---
> > >  makedumpfile.c | 300 +++++++++++++++++++++++++++++++++++----------------------
> > >  makedumpfile.h |  33 ++++---
> > >  2 files changed, 203 insertions(+), 130 deletions(-)
> > >
> > > diff --git a/makedumpfile.c b/makedumpfile.c
> > > index fa0b779..d75f7f7 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)
> > > -				continue;
> > > -
> > > -			if (page_data_buf[index].ready != 0)
> > > +			if (page_flag_buf->ready == FLAG_READY)
> > >  				continue;
> > >
> > > -			pthread_mutex_lock(&page_data_buf[index].mutex);
> > > -
> > > -			if (page_data_buf[index].ready != 0)
> > > -				goto unlock;
> > > +			/* 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;
> > >
> > > -			buf_ready = TRUE;
> > > +			page_flag_buf->pfn = pfn;
> > > +			page_flag_buf->ready = FLAG_FILLING;
> > > +			pthread_mutex_unlock(&info->current_pfn_mutex);
> > >
> > > -			page_data_buf[index].pfn = pfn;
> > > -			page_data_buf[index].ready = 1;
> > >
> > > -			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 +7363,22 @@ 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);
> > >
> > > -	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 +7391,94 @@ 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.
> > > +		 */
> > > +		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;
> > > +			}
> > > +
> > > +			/*
> > > +			 * If all the threads have reached the end, we will finish writing.
> > > +			 */
> > > +			if (end_count >= info->num_threads)
> > > +				goto finish;
> > > +
> > > +			/*
> > > +			 * Since it has the probabilty that there is no page_flag_buf being ready,
> > > +			 * we should recheck if it happens.
> > > +			 */
> > > +			if (check_count == 0)
> > > +				continue;
> > > +
> > > +			/*
> > > +			 * 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;
> > > +			}
> > >
> > > -		/* 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 (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 +7494,12 @@ 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 +7537,8 @@ out:
> > >  		}
> > >  	}
> > >
> > > -	if (page_data_buf != NULL) {
> > > -		for (i = 0; i < page_data_num; i++) {
> > > -			pthread_mutex_destroy(&page_data_buf[i].mutex);
> > > -		}
> > > -	}
> > > -
> > >  	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 +7631,7 @@ 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..5865827 100644
> > > --- a/makedumpfile.h
> > > +++ b/makedumpfile.h
> > > @@ -977,7 +977,7 @@ 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 +985,33 @@ 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 +1300,11 @@ struct DumpInfo {
> > >  	pthread_t **threads;
> > >  	struct thread_args *kdump_thread_args;
> > >  	struct page_data *page_data_buf;
> > > +	struct page_flag **page_flag_buf;
> > >  	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@lists.infradead.org
> > > http://lists.infradead.org/mailman/listinfo/kexec
> >
> > _______________________________________________
> > kexec mailing list
> > kexec@lists.infradead.org
> > http://lists.infradead.org/mailman/listinfo/kexec
> 
> _______________________________________________
> kexec mailing list
> kexec@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/kexec

_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH v3] Improve the performance of --num-threads -d 31

[Minfei Huang]

On 03/07/16 at 08:16am, Minoru Usui wrote:
> Hi, Minfei
> > >
> > > Hi, Wenjian.
> > >
> > > Since your patch applied Minoru Usui's draft patch, I have vierified it.
> > > And it works well under 4T memory.
> > >
> > > applied makedumpfile with option -d 31
> > > real    3m48.681s
> > >
> > > applied makedumpfile with option -d 31 --num-threads 32
> > > real    3m3.533s
> > >
> > > original makedumpfile with option -d 31
> > > real    3m57.163s
> > >
> > > Now with or without option --num-thread, it takes almost same time.
> > 
> > Thank you for your test on big memory machine.
> > I am relieved to see your test results.
> 
> I'm sorry, I may be misunderstood.
> Is above result is Zhou's v3 patch or v2 + my patch?

I have applied Zhou's v3 patch.

> 
> If above result is Zhou's v3 patch, could you test v2 + my patch?

Your patch fails to be applied based on Zhou's v2 patch. There is a rej
file to be generated during applying.

Thanks
Minfei

_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH v3] Improve the performance of --num-threads -d 31

["Zhou, Wenjian/周文剑"]

Hi,

I did the performance test.
v2 + Minoru's patch and v3 almost have the same performance in -d 31.

-- 
Thanks
Zhou

On 03/07/2016 04:39 PM, Minfei Huang wrote:
> On 03/07/16 at 08:16am, Minoru Usui wrote:
>> Hi, Minfei
>>>>
>>>> Hi, Wenjian.
>>>>
>>>> Since your patch applied Minoru Usui's draft patch, I have vierified it.
>>>> And it works well under 4T memory.
>>>>
>>>> applied makedumpfile with option -d 31
>>>> real    3m48.681s
>>>>
>>>> applied makedumpfile with option -d 31 --num-threads 32
>>>> real    3m3.533s
>>>>
>>>> original makedumpfile with option -d 31
>>>> real    3m57.163s
>>>>
>>>> Now with or without option --num-thread, it takes almost same time.
>>>
>>> Thank you for your test on big memory machine.
>>> I am relieved to see your test results.
>>
>> I'm sorry, I may be misunderstood.
>> Is above result is Zhou's v3 patch or v2 + my patch?
>
> I have applied Zhou's v3 patch.
>
>>
>> If above result is Zhou's v3 patch, could you test v2 + my patch?
>
> Your patch fails to be applied based on Zhou's v2 patch. There is a rej
> file to be generated during applying.
>
> Thanks
> Minfei
>
>



_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH v3] Improve the performance of --num-threads -d 31

["Zhou, Wenjian/周文?"]

On 03/07/2016 03:36 PM, Minoru Usui wrote:
> Hi, Minfei, Zhou
> 
>> -----Original Message-----
>> From: kexec [mailto:kexec-bounces@lists.infradead.org] On Behalf Of Minfei Huang
>> Sent: Monday, March 07, 2016 4:05 PM
>> To: Zhou Wenjian <zhouwj-fnst@cn.fujitsu.com>
>> Cc: kexec@lists.infradead.org
>> Subject: Re: [PATCH v3] Improve the performance of --num-threads -d 31
>>
>> Hi, Wenjian.
>>
>> Since your patch applied Minoru Usui's draft patch, I have vierified it.
>> And it works well under 4T memory.
>>
>> applied makedumpfile with option -d 31
>> real    3m48.681s
>>
>> applied makedumpfile with option -d 31 --num-threads 32
>> real    3m3.533s
>>
>> original makedumpfile with option -d 31
>> real    3m57.163s
>>
>> Now with or without option --num-thread, it takes almost same time.
> 
> Thank you for your test on big memory machine.
> I am relieved to see your test results.
> 
> Zhou, my patch may not be sophisticated,
> but getting mutex logic is robust, it works well and its performance is not bad.
> I think it reaches the goal of the original purpose.
> I understand you don't like getting mutex, but if you are debugging now, could you apply my approach?
> 
> Thanks,
> Minoru Usui
> 
> 

Hi Minoru,

I have test your patch since you post.
I can get a failure easily.
If you run it several times, the dumpfile it produces may be different from the correct one.

+			info->current_pfn = pfn + 1;
+			pthread_mutex_unlock(&info->current_pfn_mutex);
+			page_flag_buf->pfn = pfn;
+			page_flag_buf->ready = FLAG_FILLING;
The above code should be changed to the following:

+			info->current_pfn = pfn + 1;
+			page_flag_buf->pfn = pfn;
+			page_flag_buf->ready = FLAG_FILLING;
+			pthread_mutex_unlock(&info->current_pfn_mutex);

I haven't run your code with the bug fixed.
I'll try now.

By the way, with my patch v3, I can't get an error dumpfile until run it by 5000 times.

-- 
Thanks
Zhou

>> Thanks
>> Minfei
>>
>> On 03/04/16 at 04:35pm, Zhou Wenjian wrote:
>>> 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@ti.jp.nec.com>
>>> Signed-off-by: Zhou Wenjian <zhouwj-fnst@cn.fujitsu.com>
>>> ---
>>>   makedumpfile.c | 300 +++++++++++++++++++++++++++++++++++----------------------
>>>   makedumpfile.h |  33 ++++---
>>>   2 files changed, 203 insertions(+), 130 deletions(-)
>>>
>>> diff --git a/makedumpfile.c b/makedumpfile.c
>>> index fa0b779..d75f7f7 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)
>>> -				continue;
>>> -
>>> -			if (page_data_buf[index].ready != 0)
>>> +			if (page_flag_buf->ready == FLAG_READY)
>>>   				continue;
>>>
>>> -			pthread_mutex_lock(&page_data_buf[index].mutex);
>>> -
>>> -			if (page_data_buf[index].ready != 0)
>>> -				goto unlock;
>>> +			/* 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;
>>>
>>> -			buf_ready = TRUE;
>>> +			page_flag_buf->pfn = pfn;
>>> +			page_flag_buf->ready = FLAG_FILLING;
>>> +			pthread_mutex_unlock(&info->current_pfn_mutex);
>>>
>>> -			page_data_buf[index].pfn = pfn;
>>> -			page_data_buf[index].ready = 1;
>>>
>>> -			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 +7363,22 @@ 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);
>>>
>>> -	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 +7391,94 @@ 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.
>>> +		 */
>>> +		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;
>>> +			}
>>> +
>>> +			/*
>>> +			 * If all the threads have reached the end, we will finish writing.
>>> +			 */
>>> +			if (end_count >= info->num_threads)
>>> +				goto finish;
>>> +
>>> +			/*
>>> +			 * Since it has the probabilty that there is no page_flag_buf being ready,
>>> +			 * we should recheck if it happens.
>>> +			 */
>>> +			if (check_count == 0)
>>> +				continue;
>>> +
>>> +			/*
>>> +			 * 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;
>>> +			}
>>>
>>> -		/* 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 (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 +7494,12 @@ 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 +7537,8 @@ out:
>>>   		}
>>>   	}
>>>
>>> -	if (page_data_buf != NULL) {
>>> -		for (i = 0; i < page_data_num; i++) {
>>> -			pthread_mutex_destroy(&page_data_buf[i].mutex);
>>> -		}
>>> -	}
>>> -
>>>   	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 +7631,7 @@ 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..5865827 100644
>>> --- a/makedumpfile.h
>>> +++ b/makedumpfile.h
>>> @@ -977,7 +977,7 @@ 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 +985,33 @@ 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 +1300,11 @@ struct DumpInfo {
>>>   	pthread_t **threads;
>>>   	struct thread_args *kdump_thread_args;
>>>   	struct page_data *page_data_buf;
>>> +	struct page_flag **page_flag_buf;
>>>   	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@lists.infradead.org
>>> http://lists.infradead.org/mailman/listinfo/kexec
>>
>> _______________________________________________
>> kexec mailing list
>> kexec@lists.infradead.org
>> http://lists.infradead.org/mailman/listinfo/kexec




_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH v3] Improve the performance of --num-threads -d 31

["Zhou, Wenjian/周文剑"]

Hi Minoru,

I got two error dumpfiles by running it less than 100 times.

-- 
Thanks
Zhou

On 03/07/2016 04:22 PM, "Zhou, Wenjian/周文?" wrote:
> Hi Minoru,
>
> I have test your patch since you post.
> I can get a failure easily.
> If you run it several times, the dumpfile it produces may be different from the correct one.
>
> +			info->current_pfn = pfn + 1;
> +			pthread_mutex_unlock(&info->current_pfn_mutex);
> +			page_flag_buf->pfn = pfn;
> +			page_flag_buf->ready = FLAG_FILLING;
> The above code should be changed to the following:
>
> +			info->current_pfn = pfn + 1;
> +			page_flag_buf->pfn = pfn;
> +			page_flag_buf->ready = FLAG_FILLING;
> +			pthread_mutex_unlock(&info->current_pfn_mutex);
>
> I haven't run your code with the bug fixed.
> I'll try now.
>
> By the way, with my patch v3, I can't get an error dumpfile until run it by 5000 times.



_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

RE: [PATCH v3] Improve the performance of --num-threads -d 31

[Minoru Usui]

Hi, Zhou

> -----Original Message-----
> From: kexec [mailto:kexec-bounces@lists.infradead.org] On Behalf Of "Zhou, Wenjian/周文?"
> Sent: Monday, March 07, 2016 5:43 PM
> To: Usui Minoru(碓井 成) <min-usui@ti.jp.nec.com>
> Cc: Minfei Huang <mhuang@redhat.com>; kexec@lists.infradead.org
> Subject: Re: [PATCH v3] Improve the performance of --num-threads -d 31
> 
> Hi Minoru,
> 
> I got two error dumpfiles by running it less than 100 times.

Oh, I'm sorry.
Thank you for your test.

I'm waiting for your v4 patch.

Thanks
Minoru Usui



> --
> Thanks
> Zhou
> 
> On 03/07/2016 04:22 PM, "Zhou, Wenjian/周文?" wrote:
> > Hi Minoru,
> >
> > I have test your patch since you post.
> > I can get a failure easily.
> > If you run it several times, the dumpfile it produces may be different from the correct one.
> >
> > +			info->current_pfn = pfn + 1;
> > +			pthread_mutex_unlock(&info->current_pfn_mutex);
> > +			page_flag_buf->pfn = pfn;
> > +			page_flag_buf->ready = FLAG_FILLING;
> > The above code should be changed to the following:
> >
> > +			info->current_pfn = pfn + 1;
> > +			page_flag_buf->pfn = pfn;
> > +			page_flag_buf->ready = FLAG_FILLING;
> > +			pthread_mutex_unlock(&info->current_pfn_mutex);
> >
> > I haven't run your code with the bug fixed.
> > I'll try now.
> >
> > By the way, with my patch v3, I can't get an error dumpfile until run it by 5000 times.
> 
> 
> 
> _______________________________________________
> kexec mailing list
> kexec@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/kexec
_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: [PATCH v3] Improve the performance of --num-threads -d 31

["Zhou, Wenjian/周文剑"]

Hi Minoru,

I guess if you change the part of locking page_data_buf to the following,
it will be more correct.
But I'm not sure if you will get the same problem as which in my patch v3.

+		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);

-- 
Thanks
Zhou

On 03/07/2016 04:52 PM, Minoru Usui wrote:
> Hi, Zhou
>
>> -----Original Message-----
>> From: kexec [mailto:kexec-bounces@lists.infradead.org] On Behalf Of "Zhou, Wenjian/周文?"
>> Sent: Monday, March 07, 2016 5:43 PM
>> To: Usui Minoru(碓井 成) <min-usui@ti.jp.nec.com>
>> Cc: Minfei Huang <mhuang@redhat.com>; kexec@lists.infradead.org
>> Subject: Re: [PATCH v3] Improve the performance of --num-threads -d 31
>>
>> Hi Minoru,
>>
>> I got two error dumpfiles by running it less than 100 times.
>
> Oh, I'm sorry.
> Thank you for your test.
>
> I'm waiting for your v4 patch.
>
> Thanks
> Minoru Usui
>
>
>
>> --
>> Thanks
>> Zhou
>>
>> On 03/07/2016 04:22 PM, "Zhou, Wenjian/周文?" wrote:
>>> Hi Minoru,
>>>
>>> I have test your patch since you post.
>>> I can get a failure easily.
>>> If you run it several times, the dumpfile it produces may be different from the correct one.
>>>
>>> +			info->current_pfn = pfn + 1;
>>> +			pthread_mutex_unlock(&info->current_pfn_mutex);
>>> +			page_flag_buf->pfn = pfn;
>>> +			page_flag_buf->ready = FLAG_FILLING;
>>> The above code should be changed to the following:
>>>
>>> +			info->current_pfn = pfn + 1;
>>> +			page_flag_buf->pfn = pfn;
>>> +			page_flag_buf->ready = FLAG_FILLING;
>>> +			pthread_mutex_unlock(&info->current_pfn_mutex);
>>>
>>> I haven't run your code with the bug fixed.
>>> I'll try now.
>>>
>>> By the way, with my patch v3, I can't get an error dumpfile until run it by 5000 times.
>>
>>
>>
>> _______________________________________________
>> kexec mailing list
>> kexec@lists.infradead.org
>> http://lists.infradead.org/mailman/listinfo/kexec




_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec