Skip to contentSkip to conversation

raid6check.c add page size check and repair

PATCH In review Piergiorgio Sartor 3 participants 7 messages Jan 18, 2014 – Jan 23, 2014 MD RAID

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 raid6check.c add page size check and repair

Hi Neil,

I prepared a patch, against mdadm 3.3, to reduce the check error report and the (automatic) repair, from stripe to page (4K) size.
The patch is fully untested, it compiles, but I would like someone to review the code.

Next patch will be about verbosity control.

bye,

--- --- ---

raid6check.c +73 -33
diff -uNr a/raid6check.c b/raid6check.c
@@ -27,6 +27,9 @@
 #include <signal.h>
 #include <sys/mman.h>
 
+#define CHECK_PAGE_BITS (12)
+#define CHECK_PAGE_SIZE (1 << CHECK_PAGE_BITS)
+
 enum repair {
 	NO_REPAIR = 0,
 	MANUAL_REPAIR,
@@ -73,15 +76,15 @@
 	}
 }
 
-/* Try to find out if a specific disk has problems */
-int raid6_stats(int *results, int raid_disks, int chunk_size)
+/* Try to find out if a specific disk has problems in a CHECK_PAGE_SIZE page size */
+int raid6_stats_blk(int *results, int raid_disks)
 {
 	int i;
 	int curr_broken_disk = -255;
 	int prev_broken_disk = -255;
 	int broken_status = 0;
 
-	for(i = 0; i < chunk_size; i++) {
+	for(i = 0; i < CHECK_PAGE_SIZE; i++) {
 
 		if(results[i] != -255)
 			curr_broken_disk = results[i];
@@ -112,6 +115,16 @@
 	return curr_broken_disk;
 }
 
+/* Collect disks status for a strip in CHECK_PAGE_SIZE page size blocks */
+void raid6_stats(int *disk, int *results, int raid_disks, int chunk_size)
+{
+	int i, j;
+
+	for(i = 0, j = 0; i < chunk_size; i += CHECK_PAGE_SIZE, j++) {
+		disk[j] = raid6_stats_blk(&results[i], raid_disks);
+	}
+}
+
 int lock_stripe(struct mdinfo *info, unsigned long long start,
 		int chunk_size, int data_disks, sighandler_t *sig) {
 	int rv;
@@ -152,13 +165,14 @@
 	char *stripe_buf = xmalloc(raid_disks * chunk_size);
 	char **stripes = xmalloc(raid_disks * sizeof(char*));
 	char **blocks = xmalloc(raid_disks * sizeof(char*));
+	char **blocks_page = xmalloc(raid_disks * sizeof(char*));
 	int *block_index_for_slot = xmalloc(raid_disks * sizeof(int));
 	uint8_t *p = xmalloc(chunk_size);
 	uint8_t *q = xmalloc(chunk_size);
 	int *results = xmalloc(chunk_size * sizeof(int));
 	sighandler_t *sig = xmalloc(3 * sizeof(sighandler_t));
 
-	int i;
+	int i, j;
 	int diskP, diskQ;
 	int data_disks = raid_disks - 2;
 	int err = 0;
@@ -172,7 +186,7 @@
 		stripes[i] = stripe_buf + i * chunk_size;
 
 	while (length > 0) {
-		int disk;
+		int disk[chunk_size >> CHECK_PAGE_BITS];
 
 		printf("pos --> %llu\n", start);
 
@@ -217,26 +231,31 @@
 		block_index_for_slot[diskP] = data_disks;
 		blocks[data_disks+1] = stripes[diskQ];
 		block_index_for_slot[diskQ] = data_disks+1;
-
+/* Do we really need the code below? */
+#if 0
 		if (memcmp(p, stripes[diskP], chunk_size) != 0) {
 			printf("P(%d) wrong at %llu\n", diskP, start);
 		}
 		if (memcmp(q, stripes[diskQ], chunk_size) != 0) {
 			printf("Q(%d) wrong at %llu\n", diskQ, start);
 		}
+#endif
 		raid6_collect(chunk_size, p, q, stripes[diskP], stripes[diskQ], results);
-		disk = raid6_stats(results, raid_disks, chunk_size);
+		raid6_stats(disk, results, raid_disks, chunk_size);
 
-		if(disk >= -2) {
-			disk = geo_map(disk, start, raid_disks, level, layout);
-		}
-		if(disk >= 0) {
-			printf("Error detected at %llu: possible failed disk slot: %d --> %s\n",
-				start, disk, name[disk]);
-		}
-		if(disk == -65535) {
-			printf("Error detected at %llu: disk slot unknown\n", start);
+		for(j = 0; i < (chunk_size >> CHECK_PAGE_BITS); j++) {
+			if(disk[j] >= -2) {
+				disk[j] = geo_map(disk[j], start, raid_disks, level, layout);
+			}
+			if(disk[j] >= 0) {
+				printf("Error detected at %llu, page %d: possible failed disk slot: %d --> %s\n",
+					start, j, disk[j], name[disk[j]]);
+			}
+			if(disk[j] == -65535) {
+				printf("Error detected at %llu, page %d: disk slot unknown\n", start, j);
+			}
 		}
+
 		if(repair == MANUAL_REPAIR) {
 			printf("Repairing stripe %llu\n", start);
 			printf("Assuming slots %d (%s) and %d (%s) are incorrect\n",
@@ -325,21 +344,37 @@
 				goto exitCheck;
 			}
 
-		} else if (disk >= 0 && repair == AUTO_REPAIR) {
-			printf("Auto-repairing slot %d (%s)\n", disk, name[disk]);
-			if (disk == diskQ) {
-				qsyndrome(p, (uint8_t*)stripes[diskQ], (uint8_t**)blocks, data_disks, chunk_size);
+		}
+
+		int pages_to_write_count = 0;
+		int page_to_write[chunk_size >> CHECK_PAGE_BITS];
+		for(j = 0; i < (chunk_size >> CHECK_PAGE_BITS); j++) {
+			if (disk[j] >= 0 && repair == AUTO_REPAIR) {
+				printf("Auto-repairing slot %d (%s)\n", disk[j], name[disk[j]]);
+				pages_to_write_count++;
+				page_to_write[j] = 1;
+				for(i = 0; i < raid_disks; i++) {
+					blocks_page[i] = blocks[i] + j * CHECK_PAGE_SIZE;
+				}
+				if (disk[j] == diskQ) {
+					qsyndrome(p, (uint8_t*)stripes[diskQ], (uint8_t**)blocks_page, data_disks, CHECK_PAGE_SIZE);
+				} else {
+					char *all_but_failed_blocks[data_disks];
+					int failed_block_index = block_index_for_slot[disk[j]];
+					for (i=0; i < data_disks; i++)
+						if (failed_block_index == i)
+							all_but_failed_blocks[i] = stripes[diskP] + j * CHECK_PAGE_SIZE;
+						else
+							all_but_failed_blocks[i] = blocks_page[i];
+					xor_blocks(stripes[disk[j]] + j * CHECK_PAGE_SIZE,
+					all_but_failed_blocks, data_disks, CHECK_PAGE_SIZE);
+				}
 			} else {
-				char *all_but_failed_blocks[data_disks];
-				int failed_block_index = block_index_for_slot[disk];
-				for (i=0; i < data_disks; i++)
-					if (failed_block_index == i)
-						all_but_failed_blocks[i] = stripes[diskP];
-					else
-						all_but_failed_blocks[i] = blocks[i];
-				xor_blocks(stripes[disk],
-					all_but_failed_blocks, data_disks, chunk_size);
+				page_to_write[j] = 0;
 			}
+		}
+
+		if(pages_to_write_count > 0) {
 
 			err = lock_stripe(info, start, chunk_size, data_disks, sig);
 			if(err != 0) {
@@ -348,14 +383,19 @@
 				goto exitCheck;
 			}
 
-			lseek64(source[disk], offsets[disk] + start * chunk_size, 0);
-			int write_res = write(source[disk], stripes[disk], chunk_size);
+			int write_res = 0;
+			for(j = 0; j < (chunk_size >> CHECK_PAGE_BITS); j++) {
+				if(page_to_write[j] == 1) {
+					lseek64(source[disk[j]], offsets[disk[j]] + start * chunk_size + j * CHECK_PAGE_SIZE, 0);
+					write_res += write(source[disk[j]], stripes[disk[j]] + j * CHECK_PAGE_SIZE, CHECK_PAGE_SIZE);
+				}
+			}
 
 			err = unlock_all_stripes(info, sig);
-			if(err != 0 || write_res != chunk_size)
+			if (err != 0 || write_res != (CHECK_PAGE_SIZE * pages_to_write_count))
 				goto exitCheck;
 
-			if (write_res != chunk_size) {
+			if (write_res != (CHECK_PAGE_SIZE * pages_to_write_count)) {
 				fprintf(stderr, "Failed to write a full chunk.\n");
 				goto exitCheck;
 			}
Signature · 1 line
piergiorgio
#1 lore raw

Hello Piergiorgio,

I'm going to review it in more detail later on this week and I think I also still have an outstanding patch. But on quick glance, I think your patch introduces a memory leak.

raid6check.c line 173

Quoted 10 lines
raid6check.c +1 -0
@@ -152,13 +165,14 @@
  	char *stripe_buf = xmalloc(raid_disks * chunk_size);
  	char **stripes = xmalloc(raid_disks * sizeof(char*));
  	char **blocks = xmalloc(raid_disks * sizeof(char*));
+	char **blocks_page = xmalloc(raid_disks * sizeof(char*));
  	int *block_index_for_slot = xmalloc(raid_disks * sizeof(int));
  	uint8_t *p = xmalloc(chunk_size);
  	uint8_t *q = xmalloc(chunk_size);
  	int *results = xmalloc(chunk_size * sizeof(int));
  	sighandler_t *sig = xmalloc(3 * sizeof(sighandler_t));

I don't see free(blocks_page).

Btw, clangs static code checker is rather good to find such issues. Actually, it reports several issues with mdadm, I just didn't have the time to look into it in more detail.

Btw2: Any chance you could at least add the '-p' argument to your diff command? I think if you just would use git, it would do that automatically.

Thanks,
Bernd

#2 lore raw
Quoted 19 lines · Bernd Schubert

Hello Piergiorgio,

I'm going to review it in more detail later on this week and I think I also still have an outstanding patch. But on quick glance, I think your patch introduces a memory leak.

Review line 173

Quoted 10 lines
(quoted diff) +1 -0
@@ -152,13 +165,14 @@
 	char *stripe_buf = xmalloc(raid_disks * chunk_size);
 	char **stripes = xmalloc(raid_disks * sizeof(char*));
 	char **blocks = xmalloc(raid_disks * sizeof(char*));
+	char **blocks_page = xmalloc(raid_disks * sizeof(char*));
 	int *block_index_for_slot = xmalloc(raid_disks * sizeof(int));
 	uint8_t *p = xmalloc(chunk_size);
 	uint8_t *q = xmalloc(chunk_size);
 	int *results = xmalloc(chunk_size * sizeof(int));
 	sighandler_t *sig = xmalloc(3 * sizeof(sighandler_t));

I don't see free(blocks_page).

Quoted 7 lines

Btw, clangs static code checker is rather good to find such issues. Actually, it reports several issues with mdadm, I just didn't have the time to look into it in more detail.

Btw2: Any chance you could at least add the '-p' argument to your diff command? I think if you just would use git, it would do that automatically.

Hi Bernd,

good catch, I'm pretty sure there are even more :-)

Here the re-patch, i.e. the same patch, with the free(). This was done, with "diff -uNpr", I hope it fits your needs. BTW, could you point me to some instruction on how to get the git tree of mdadm?

--- --- ---

raid6check.c +74 -33
diff -uNrp a/raid6check.c b/raid6check.c
@@ -27,6 +27,9 @@
 #include <signal.h>
 #include <sys/mman.h>
 
+#define CHECK_PAGE_BITS (12)
+#define CHECK_PAGE_SIZE (1 << CHECK_PAGE_BITS)
+
 enum repair {
 	NO_REPAIR = 0,
 	MANUAL_REPAIR,
@@ -73,15 +76,15 @@ void raid6_collect(int chunk_size, uint8
 	}
 }
 
-/* Try to find out if a specific disk has problems */
-int raid6_stats(int *results, int raid_disks, int chunk_size)
+/* Try to find out if a specific disk has problems in a CHECK_PAGE_SIZE page size */
+int raid6_stats_blk(int *results, int raid_disks)
 {
 	int i;
 	int curr_broken_disk = -255;
 	int prev_broken_disk = -255;
 	int broken_status = 0;
 
-	for(i = 0; i < chunk_size; i++) {
+	for(i = 0; i < CHECK_PAGE_SIZE; i++) {
 
 		if(results[i] != -255)
 			curr_broken_disk = results[i];
@@ -112,6 +115,16 @@ int raid6_stats(int *results, int raid_d
 	return curr_broken_disk;
 }
 
+/* Collect disks status for a strip in CHECK_PAGE_SIZE page size blocks */
+void raid6_stats(int *disk, int *results, int raid_disks, int chunk_size)
+{
+	int i, j;
+
+	for(i = 0, j = 0; i < chunk_size; i += CHECK_PAGE_SIZE, j++) {
+		disk[j] = raid6_stats_blk(&results[i], raid_disks);
+	}
+}
+
 int lock_stripe(struct mdinfo *info, unsigned long long start,
 		int chunk_size, int data_disks, sighandler_t *sig) {
 	int rv;
@@ -152,13 +165,14 @@ int check_stripes(struct mdinfo *info, i
 	char *stripe_buf = xmalloc(raid_disks * chunk_size);
 	char **stripes = xmalloc(raid_disks * sizeof(char*));
 	char **blocks = xmalloc(raid_disks * sizeof(char*));
+	char **blocks_page = xmalloc(raid_disks * sizeof(char*));
 	int *block_index_for_slot = xmalloc(raid_disks * sizeof(int));
 	uint8_t *p = xmalloc(chunk_size);
 	uint8_t *q = xmalloc(chunk_size);
 	int *results = xmalloc(chunk_size * sizeof(int));
 	sighandler_t *sig = xmalloc(3 * sizeof(sighandler_t));
 
-	int i;
+	int i, j;
 	int diskP, diskQ;
 	int data_disks = raid_disks - 2;
 	int err = 0;
@@ -172,7 +186,7 @@ int check_stripes(struct mdinfo *info, i
 		stripes[i] = stripe_buf + i * chunk_size;
 
 	while (length > 0) {
-		int disk;
+		int disk[chunk_size >> CHECK_PAGE_BITS];
 
 		printf("pos --> %llu\n", start);
 
@@ -217,26 +231,31 @@ int check_stripes(struct mdinfo *info, i
 		block_index_for_slot[diskP] = data_disks;
 		blocks[data_disks+1] = stripes[diskQ];
 		block_index_for_slot[diskQ] = data_disks+1;
-
+/* Do we really need the code below? */
+#if 0
 		if (memcmp(p, stripes[diskP], chunk_size) != 0) {
 			printf("P(%d) wrong at %llu\n", diskP, start);
 		}
 		if (memcmp(q, stripes[diskQ], chunk_size) != 0) {
 			printf("Q(%d) wrong at %llu\n", diskQ, start);
 		}
+#endif
 		raid6_collect(chunk_size, p, q, stripes[diskP], stripes[diskQ], results);
-		disk = raid6_stats(results, raid_disks, chunk_size);
+		raid6_stats(disk, results, raid_disks, chunk_size);
 
-		if(disk >= -2) {
-			disk = geo_map(disk, start, raid_disks, level, layout);
-		}
-		if(disk >= 0) {
-			printf("Error detected at %llu: possible failed disk slot: %d --> %s\n",
-				start, disk, name[disk]);
-		}
-		if(disk == -65535) {
-			printf("Error detected at %llu: disk slot unknown\n", start);
+		for(j = 0; i < (chunk_size >> CHECK_PAGE_BITS); j++) {
+			if(disk[j] >= -2) {
+				disk[j] = geo_map(disk[j], start, raid_disks, level, layout);
+			}
+			if(disk[j] >= 0) {
+				printf("Error detected at %llu, page %d: possible failed disk slot: %d --> %s\n",
+					start, j, disk[j], name[disk[j]]);
+			}
+			if(disk[j] == -65535) {
+				printf("Error detected at %llu, page %d: disk slot unknown\n", start, j);
+			}
 		}
+
 		if(repair == MANUAL_REPAIR) {
 			printf("Repairing stripe %llu\n", start);
 			printf("Assuming slots %d (%s) and %d (%s) are incorrect\n",
@@ -325,21 +344,37 @@ int check_stripes(struct mdinfo *info, i
 				goto exitCheck;
 			}
 
-		} else if (disk >= 0 && repair == AUTO_REPAIR) {
-			printf("Auto-repairing slot %d (%s)\n", disk, name[disk]);
-			if (disk == diskQ) {
-				qsyndrome(p, (uint8_t*)stripes[diskQ], (uint8_t**)blocks, data_disks, chunk_size);
+		}
+
+		int pages_to_write_count = 0;
+		int page_to_write[chunk_size >> CHECK_PAGE_BITS];
+		for(j = 0; i < (chunk_size >> CHECK_PAGE_BITS); j++) {
+			if (disk[j] >= 0 && repair == AUTO_REPAIR) {
+				printf("Auto-repairing slot %d (%s)\n", disk[j], name[disk[j]]);
+				pages_to_write_count++;
+				page_to_write[j] = 1;
+				for(i = 0; i < raid_disks; i++) {
+					blocks_page[i] = blocks[i] + j * CHECK_PAGE_SIZE;
+				}
+				if (disk[j] == diskQ) {
+					qsyndrome(p, (uint8_t*)stripes[diskQ], (uint8_t**)blocks_page, data_disks, CHECK_PAGE_SIZE);
+				} else {
+					char *all_but_failed_blocks[data_disks];
+					int failed_block_index = block_index_for_slot[disk[j]];
+					for (i=0; i < data_disks; i++)
+						if (failed_block_index == i)
+							all_but_failed_blocks[i] = stripes[diskP] + j * CHECK_PAGE_SIZE;
+						else
+							all_but_failed_blocks[i] = blocks_page[i];
+					xor_blocks(stripes[disk[j]] + j * CHECK_PAGE_SIZE,
+					all_but_failed_blocks, data_disks, CHECK_PAGE_SIZE);
+				}
 			} else {
-				char *all_but_failed_blocks[data_disks];
-				int failed_block_index = block_index_for_slot[disk];
-				for (i=0; i < data_disks; i++)
-					if (failed_block_index == i)
-						all_but_failed_blocks[i] = stripes[diskP];
-					else
-						all_but_failed_blocks[i] = blocks[i];
-				xor_blocks(stripes[disk],
-					all_but_failed_blocks, data_disks, chunk_size);
+				page_to_write[j] = 0;
 			}
+		}
+
+		if(pages_to_write_count > 0) {
 
 			err = lock_stripe(info, start, chunk_size, data_disks, sig);
 			if(err != 0) {
@@ -348,14 +383,19 @@ int check_stripes(struct mdinfo *info, i
 				goto exitCheck;
 			}
 
-			lseek64(source[disk], offsets[disk] + start * chunk_size, 0);
-			int write_res = write(source[disk], stripes[disk], chunk_size);
+			int write_res = 0;
+			for(j = 0; j < (chunk_size >> CHECK_PAGE_BITS); j++) {
+				if(page_to_write[j] == 1) {
+					lseek64(source[disk[j]], offsets[disk[j]] + start * chunk_size + j * CHECK_PAGE_SIZE, 0);
+					write_res += write(source[disk[j]], stripes[disk[j]] + j * CHECK_PAGE_SIZE, CHECK_PAGE_SIZE);
+				}
+			}
 
 			err = unlock_all_stripes(info, sig);
-			if(err != 0 || write_res != chunk_size)
+			if (err != 0 || write_res != (CHECK_PAGE_SIZE * pages_to_write_count))
 				goto exitCheck;
 
-			if (write_res != chunk_size) {
+			if (write_res != (CHECK_PAGE_SIZE * pages_to_write_count)) {
 				fprintf(stderr, "Failed to write a full chunk.\n");
 				goto exitCheck;
 			}
@@ -370,6 +410,7 @@ exitCheck:
 	free(stripe_buf);
 	free(stripes);
 	free(blocks);
+	free(blocks_page);
 	free(block_index_for_slot);
 	free(p);
 	free(q);
Signature · 1 line
piergiorgio
#3 lore raw

Hi again,

I'm polluting the mailing list now... Sorry!

I found a couple of really stupid bugs, so this re-re-patch should fix them.

Sorry for the noise.

--- --- --

Binary files a/dlink.o and b/dlink.o differ Binary files a/lib.o and b/lib.o differ Binary files a/maps.o and b/maps.o differ Binary files a/raid6check and b/raid6check differ

raid6check.c +74 -33
diff -uNrp a/raid6check.c b/raid6check.c
@@ -27,6 +27,9 @@
 #include <signal.h>
 #include <sys/mman.h>
 
+#define CHECK_PAGE_BITS (12)
+#define CHECK_PAGE_SIZE (1 << CHECK_PAGE_BITS)
+
 enum repair {
 	NO_REPAIR = 0,
 	MANUAL_REPAIR,
@@ -73,15 +76,15 @@ void raid6_collect(int chunk_size, uint8
 	}
 }
 
-/* Try to find out if a specific disk has problems */
-int raid6_stats(int *results, int raid_disks, int chunk_size)
+/* Try to find out if a specific disk has problems in a CHECK_PAGE_SIZE page size */
+int raid6_stats_blk(int *results, int raid_disks)
 {
 	int i;
 	int curr_broken_disk = -255;
 	int prev_broken_disk = -255;
 	int broken_status = 0;
 
-	for(i = 0; i < chunk_size; i++) {
+	for(i = 0; i < CHECK_PAGE_SIZE; i++) {
 
 		if(results[i] != -255)
 			curr_broken_disk = results[i];
@@ -112,6 +115,16 @@ int raid6_stats(int *results, int raid_d
 	return curr_broken_disk;
 }
 
+/* Collect disks status for a strip in CHECK_PAGE_SIZE page size blocks */
+void raid6_stats(int *disk, int *results, int raid_disks, int chunk_size)
+{
+	int i, j;
+
+	for(i = 0, j = 0; i < chunk_size; i += CHECK_PAGE_SIZE, j++) {
+		disk[j] = raid6_stats_blk(&results[i], raid_disks);
+	}
+}
+
 int lock_stripe(struct mdinfo *info, unsigned long long start,
 		int chunk_size, int data_disks, sighandler_t *sig) {
 	int rv;
@@ -152,13 +165,14 @@ int check_stripes(struct mdinfo *info, i
 	char *stripe_buf = xmalloc(raid_disks * chunk_size);
 	char **stripes = xmalloc(raid_disks * sizeof(char*));
 	char **blocks = xmalloc(raid_disks * sizeof(char*));
+	char **blocks_page = xmalloc(raid_disks * sizeof(char*));
 	int *block_index_for_slot = xmalloc(raid_disks * sizeof(int));
 	uint8_t *p = xmalloc(chunk_size);
 	uint8_t *q = xmalloc(chunk_size);
 	int *results = xmalloc(chunk_size * sizeof(int));
 	sighandler_t *sig = xmalloc(3 * sizeof(sighandler_t));
 
-	int i;
+	int i, j;
 	int diskP, diskQ;
 	int data_disks = raid_disks - 2;
 	int err = 0;
@@ -172,7 +186,7 @@ int check_stripes(struct mdinfo *info, i
 		stripes[i] = stripe_buf + i * chunk_size;
 
 	while (length > 0) {
-		int disk;
+		int disk[chunk_size >> CHECK_PAGE_BITS];
 
 		printf("pos --> %llu\n", start);
 
@@ -217,26 +231,31 @@ int check_stripes(struct mdinfo *info, i
 		block_index_for_slot[diskP] = data_disks;
 		blocks[data_disks+1] = stripes[diskQ];
 		block_index_for_slot[diskQ] = data_disks+1;
-
+/* Do we really need the code below? */
+#if 0
 		if (memcmp(p, stripes[diskP], chunk_size) != 0) {
 			printf("P(%d) wrong at %llu\n", diskP, start);
 		}
 		if (memcmp(q, stripes[diskQ], chunk_size) != 0) {
 			printf("Q(%d) wrong at %llu\n", diskQ, start);
 		}
+#endif
 		raid6_collect(chunk_size, p, q, stripes[diskP], stripes[diskQ], results);
-		disk = raid6_stats(results, raid_disks, chunk_size);
+		raid6_stats(disk, results, raid_disks, chunk_size);
 
-		if(disk >= -2) {
-			disk = geo_map(disk, start, raid_disks, level, layout);
-		}
-		if(disk >= 0) {
-			printf("Error detected at %llu: possible failed disk slot: %d --> %s\n",
-				start, disk, name[disk]);
-		}
-		if(disk == -65535) {
-			printf("Error detected at %llu: disk slot unknown\n", start);
+		for(j = 0; j < (chunk_size >> CHECK_PAGE_BITS); j++) {
+			if(disk[j] >= -2) {
+				disk[j] = geo_map(disk[j], start, raid_disks, level, layout);
+			}
+			if(disk[j] >= 0) {
+				printf("Error detected at %llu, page %d: possible failed disk slot: %d --> %s\n",
+					start, j, disk[j], name[disk[j]]);
+			}
+			if(disk[j] == -65535) {
+				printf("Error detected at %llu, page %d: disk slot unknown\n", start, j);
+			}
 		}
+
 		if(repair == MANUAL_REPAIR) {
 			printf("Repairing stripe %llu\n", start);
 			printf("Assuming slots %d (%s) and %d (%s) are incorrect\n",
@@ -325,21 +344,37 @@ int check_stripes(struct mdinfo *info, i
 				goto exitCheck;
 			}
 
-		} else if (disk >= 0 && repair == AUTO_REPAIR) {
-			printf("Auto-repairing slot %d (%s)\n", disk, name[disk]);
-			if (disk == diskQ) {
-				qsyndrome(p, (uint8_t*)stripes[diskQ], (uint8_t**)blocks, data_disks, chunk_size);
+		}
+
+		int pages_to_write_count = 0;
+		int page_to_write[chunk_size >> CHECK_PAGE_BITS];
+		for(j = 0; j < (chunk_size >> CHECK_PAGE_BITS); j++) {
+			if (disk[j] >= 0 && repair == AUTO_REPAIR) {
+				printf("Auto-repairing slot %d (%s)\n", disk[j], name[disk[j]]);
+				pages_to_write_count++;
+				page_to_write[j] = 1;
+				for(i = 0; i < raid_disks; i++) {
+					blocks_page[i] = blocks[i] + j * CHECK_PAGE_SIZE;
+				}
+				if (disk[j] == diskQ) {
+					qsyndrome(p, (uint8_t*)stripes[diskQ], (uint8_t**)blocks_page, data_disks, CHECK_PAGE_SIZE);
+				} else {
+					char *all_but_failed_blocks[data_disks];
+					int failed_block_index = block_index_for_slot[disk[j]];
+					for (i=0; i < data_disks; i++)
+						if (failed_block_index == i)
+							all_but_failed_blocks[i] = stripes[diskP] + j * CHECK_PAGE_SIZE;
+						else
+							all_but_failed_blocks[i] = blocks_page[i];
+					xor_blocks(stripes[disk[j]] + j * CHECK_PAGE_SIZE,
+					all_but_failed_blocks, data_disks, CHECK_PAGE_SIZE);
+				}
 			} else {
-				char *all_but_failed_blocks[data_disks];
-				int failed_block_index = block_index_for_slot[disk];
-				for (i=0; i < data_disks; i++)
-					if (failed_block_index == i)
-						all_but_failed_blocks[i] = stripes[diskP];
-					else
-						all_but_failed_blocks[i] = blocks[i];
-				xor_blocks(stripes[disk],
-					all_but_failed_blocks, data_disks, chunk_size);
+				page_to_write[j] = 0;
 			}
+		}
+
+		if(pages_to_write_count > 0) {
 
 			err = lock_stripe(info, start, chunk_size, data_disks, sig);
 			if(err != 0) {
@@ -348,14 +383,19 @@ int check_stripes(struct mdinfo *info, i
 				goto exitCheck;
 			}
 
-			lseek64(source[disk], offsets[disk] + start * chunk_size, 0);
-			int write_res = write(source[disk], stripes[disk], chunk_size);
+			int write_res = 0;
+			for(j = 0; j < (chunk_size >> CHECK_PAGE_BITS); j++) {
+				if(page_to_write[j] == 1) {
+					lseek64(source[disk[j]], offsets[disk[j]] + start * chunk_size + j * CHECK_PAGE_SIZE, 0);
+					write_res += write(source[disk[j]], stripes[disk[j]] + j * CHECK_PAGE_SIZE, CHECK_PAGE_SIZE);
+				}
+			}
 
 			err = unlock_all_stripes(info, sig);
-			if(err != 0 || write_res != chunk_size)
+			if (err != 0 || write_res != (CHECK_PAGE_SIZE * pages_to_write_count))
 				goto exitCheck;
 
-			if (write_res != chunk_size) {
+			if (write_res != (CHECK_PAGE_SIZE * pages_to_write_count)) {
 				fprintf(stderr, "Failed to write a full chunk.\n");
 				goto exitCheck;
 			}
@@ -370,6 +410,7 @@ exitCheck:
 	free(stripe_buf);
 	free(stripes);
 	free(blocks);
+	free(blocks_page);
 	free(block_index_for_slot);
 	free(p);
 	free(q);
Signature · 1 line
piergiorgio
#4 lore raw
Quoted 9 lines · On Mon, 20 Jan 2014 20:10:22 +0100 Piergiorgio Sartor…

Hi again,

I'm polluting the mailing list now... Sorry!

I found a couple of really stupid bugs, so this
re-re-patch should fix them.

Sorry for the noise.

Hi,
I've applied that patch - it looks believable.

However:
1/ If you could try putting a description of what the patch does, and why, at the top, that would help a lot.
See http://www.ozlabs.org/~akpm/stuff/tpp.txt

2/ I'm not really happy about:

Quoted 4 lines
raid6check.c +0 -0
@@ -217,26 +231,31 @@ int check_stripes(struct mdinfo *info, i
 		block_index_for_slot[diskP] = data_disks;
 		blocks[data_disks+1] = stripes[diskQ];
 		block_index_for_slot[diskQ] = data_disks+1;

raid6check.c int check_stripes(struct mdinfo *info, i · line 242

Quoted 10 lines
raid6check.c +3 -1
-
+/* Do we really need the code below? */
+#if 0
 		if (memcmp(p, stripes[diskP], chunk_size) != 0) {
 			printf("P(%d) wrong at %llu\n", diskP, start);
 		}
 		if (memcmp(q, stripes[diskQ], chunk_size) != 0) {
 			printf("Q(%d) wrong at %llu\n", diskQ, start);
 		}
+#endif

If you think it should be removed, then delete it completely. It will still exist in the git history so they is no really loss of code.
If you aren't sure, then don't delete it at all. Ask. Explain your thinking and seek opinions.
In either case it should be in a separate patch.
I've left the "#if 0" etc for now but if you could push through to a resolution and send a patch I would appreciate it.

Thanks,
NeilBrown

#5 lore raw
Quoted 38 lines · NeilBrown
Quoted 9 lines · On Mon, 20 Jan 2014 20:10:22 +0100 Piergiorgio Sartor…

Hi again,

I'm polluting the mailing list now... Sorry!

I found a couple of really stupid bugs, so this
re-re-patch should fix them.

Sorry for the noise.

Hi,
I've applied that patch - it looks believable.

However:
1/ If you could try putting a description of what the patch does, and why, at the top, that would help a lot.
See http://www.ozlabs.org/~akpm/stuff/tpp.txt

2/ I'm not really happy about:

Quoted 4 lines
raid6check.c +0 -0
@@ -217,26 +231,31 @@ int check_stripes(struct mdinfo *info, i
 		block_index_for_slot[diskP] = data_disks;
 		blocks[data_disks+1] = stripes[diskQ];
 		block_index_for_slot[diskQ] = data_disks+1;

raid6check.c int check_stripes(struct mdinfo *info, i · line 242

Quoted 10 lines
raid6check.c +3 -1
-
+/* Do we really need the code below? */
+#if 0
 		if (memcmp(p, stripes[diskP], chunk_size) != 0) {
 			printf("P(%d) wrong at %llu\n", diskP, start);
 		}
 		if (memcmp(q, stripes[diskQ], chunk_size) != 0) {
 			printf("Q(%d) wrong at %llu\n", diskQ, start);
 		}
+#endif
Quoted 7 lines

If you think it should be removed, then delete it completely. It will still exist in the git history so they is no really loss of code.
If you aren't sure, then don't delete it at all. Ask. Explain your thinking and seek opinions.
In either case it should be in a separate patch.
I've left the "#if 0" etc for now but if you could push through to a resolution and send a patch I would appreciate it.

Hi Neil,

thanks for your review and comments.

I'll send a patch (separate email)
removing the code.

bye,

Quoted 3 lines

Thanks,
NeilBrown

Signature · 1 line
piergiorgio
#6 lore raw
In reply to Piergiorgio Sartor

Hello Piergiorgio,

Quoted 6 lines · Piergiorgio Sartor

Hi Bernd,

good catch, I'm pretty sure there are even more :-)

Here the re-patch, i.e. the same patch, with the free(). This was done, with "diff -uNpr", I hope it fits your needs.

Thanks! Much better :)

Quoted 2 lines

BTW, could you point me to some instruction on how to get the git tree of mdadm?

git clone git://git.neil.brown.name/mdadm.git

For patch management I then use 'stg' (stacked-git). But without stg, so directly with git and patch rebasing/stacking also works.

Hope it helps,
Bernd