[LU-16948] sparse file migration broken

• Type: Bug
• Resolution: Unresolved
• Priority: Minor
• Affects Version/s: Lustre 2.16.0, Lustre 2.15.3
• Severity: 3
• Assignee: Etienne Aujames
• Reporter: Etienne Aujames

Description

On 2.15.3 sparse file migration does not keep the same number of blocks if the extents are non-aligned on the stripe_size (on master the extent must be > 64M):

Reproducer

[root@client client]# lfs setstripe -c2 test_simple
[root@client client]# dd if=/dev/urandom of=test_simple bs=512K count=$((128*2+1)) seek=1                                                                                                                                  
[root@client client]# dd if=/dev/urandom of=test_simple bs=512K count=$((128*2+1)) seek=$((512*2+1)) conv=fsync                          
                             
[root@client client]# filefrag -v test_simple                               
Filesystem type is: bd00bd0                                                 
File size of test_simple is 672137216 (656384 blocks of 1024 bytes)         
 ext:     device_logical:        physical_offset: length:  dev: flags:      
   0:        0..       3:    4063232..   4063235:      4: 0001: net         
   1:      512..   66559:    4063744..   4129791:  66048: 0001: net         
   2:   262656..  319487:    4137472..   4194303:  56832: 0001: net         
   3:   319488..  327679:     466944..    475135:   8192: 0001: net         
   4:   327680..  328703:     491520..    492543:   1024: 0001: net         
   5:        0..   65535:    2883584..   2949119:  65536: 0000: net         
   6:   262144..  327679:    2949120..   3014655:  65536: 0000: last,net    
test_simple: 5 extents found                                                

[root@client client]# du test_simple      
263176  test_simple                     

[root@client client]# strace -o report lfs migrate -c2 test_simple

[root@client client]# filefrag -v test_simple                              
Filesystem type is: bd00bd0                                                
File size of test_simple is 672137216 (656384 blocks of 1024 bytes)        
 ext:     device_logical:        physical_offset: length:  dev: flags:     
   0:        0..       3:     433676..    433679:      4: 0001: net        
   1:      512..    1023:     457244..    457755:    512: 0001: net        
   2:     1024..    1535:     461812..    462323:    512: 0001: net        
   3:     1536..    2047:     457756..    458267:    512: 0001: net        
   4:     2048..    3583:     462848..    464383:   1536: 0001: net        
   5:     3584..    4095:     397824..    398335:    512: 0001: net        
   6:     4096..   24575:     499712..    520191:  20480: 0001: net        
   7:    24576..   66559:    2629632..   2671615:  41984: 0001: net        
   8:   262656..  328703:    2679296..   2745343:  66048: 0001: net        
   9:        0..    1023:     473088..    474111:   1024: 0000: net        
  10:     1024..    1535:     466432..    466943:    512: 0000: net        
  11:     1536..    2047:     475136..    475647:    512: 0000: net        
  12:     2048..    3071:     474112..    475135:   1024: 0000: net        
  13:     3072..    3583:     475648..    476159:    512: 0000: net        
  14:     3584..   66047:    3014656..   3077119:  62464: 0000: net        
  15:   262144..  327679:    3088384..   3153919:  65536: 0000: last,net   
test_simple: 16 extents found                                              
                      
[root@client client]# du test_simple                           
263692  test_simple                                            

block_diff = 263692 - 263176 = 516kb

"lfs migrate" strace:

lseek(3, 4096, SEEK_DATA)               = 524288                                                                                                  
lseek(3, 524288, SEEK_HOLE)             = 135266304                                                                                               
pread64(3, "]\220\311\337\376 T@\3622\314\231\3\0100\202]\31j\246u\177\256\255\307\250\0\333\311\301\346\251"..., 1048576, 524288) = 1048576      
pwrite64(4, "]\220\311\337\376 T@\3622\314\231\3\0100\202]\31j\246u\177\256\255\307\250\0\333\311\301\346\251"..., 1048576, 524288) = 1048576 
...
pwrite64(4, "\266\270\20\357\243\366#jC\217\333\325\260V\314;\274b\204\250\4s\\\231\6\222\34\300\30n]\27"..., 1048576, 133693440) = 1048576  
pread64(3, "\325/\377m\316\251P\2723\27*%1\205\216\365; \tv$\241X\31>\342\331f\265td\353"..., 1048576, 134742016) = 1048576                  
pwrite64(4, "\325/\377m\316\251"..., 1048576, 134742016) = 1048576                     <------ 134742016 + 1048576 = 135790592 != seek_hole (135266304)
lseek(3, 135790592, SEEK_DATA)          = 537395200                                                                                          
lseek(3, 537395200, SEEK_HOLE)          = 672137216                                                                                          

Issue Links

• is related to: LU-16587 Make lfs migrate faster

Comments

[Etienne Aujames]

On master the block_diff should be something like this:
block_diff = buffer_size - (segment_size % buffer_size) = 64M - 512K

[Etienne Aujames]

This bug is linked to the buffer_size, on master the buffer_size is > 64M: LU-16587 utils: give lfs migrate a larger buffer (https://review.whamcloud.com/c/fs/lustre-release/+/50118)

[Gerrit Updater]

"Etienne AUJAMES <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/51593
Subject: LU-16948 utils: fix lfs migrate for sparse files
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 5f0826104c694db56ec22d40b094f8a74e9d9ffe

[Etienne Aujames]

Sorry, my mistake, the stripe_count is not the culprit (I can reproduce this with stripe_count=1).

[Patrick Farrell]

"On 2.15.3 sparse file migration with a stripe count > 2 does not keep the same number of blocks if the extents are non-aligned on the stripe_size (on master the extent must be > 64M):"

Etienne, I don't follow the second part here - What is the bug on master exactly?

Also, did you mean stripe count > 1 ?