Daniel Westermann's Blog

linux ( as well as most of the unixes ) provides the ability to integrate many different file systems at the same time. to name a few of them:

• ext2, ext3, ext4
• ocfs, ocfs2
• reiserfs
• vxfs
• brtfs
• dos, ntfs
• …

although each of them provides different features and was developed with different purposes in mind the tools to work with them stay the same:

• cp
• mv
• cd
• …

the layer which makes this possible is called the virtual filesystem ( vfs ). this layer provides a common interface for the filesystems which are plugged into the operating system. I already introduced one special kind of filesystem, the the proc filesystem. the proc filesystem does not handle any files on disk or on the network, but neitherless it is a filesystem. in addition to the above mentioned filesystems, which all are disk based, filesystem may also handle files on the network, such as nfs or cifs.

no matter what kind of filesystem you are working with: when interacting with the filesystem by using the commands of choice you are routed through the virtual filesystem:

the virtual file system

to make this possible there needs to be a standard all file system implementations must comply with, and this standard is called the common file model. the key components this model consist of are:

• the superblock which stores information about a mounted filesystem ( … that is stored in memory as a doube linked list )
• inodes which store information about a specific file ( … that are stored in memory as a doube linked list)
• the file object which stores information of the underlying files
• dentries, which represent the links to build the directory structure ( … that are stored in memory as a doube linked list)

to speed up operations on the file systems some of the information which is normally stored on disk are cached. if you recall the post about slabs, you can find an entry like the following in the /proc/slabinfo file if you have a mounted ext4 filesystem on your system:

cat /proc/slabinfo | grep ext4 | grep cache
ext4_inode_cache   34397  34408    920   17    4 : tunables    0    0    0 : slabdata   2024   2024      0

so what needs the kernel to do if, for example, a request for listing the contents of a directoy comes in and the directory resides on an ext4 filesystem? because the filesystem is mounted the kernel knows that the filesystem for the specific request is of type ext4. the ls command will then be translated ( pointed ) to the specific ls implementation of the ext4 filesystem. this operation is the same for all commands interacting with filesystems. there is a pointer for each operation that links to the specific implementation of the command in question:

directory listing

as the superblock is stored in memory and therefore may become dirty, that is not synchronized with the superblock on disk, there is the same issue that oracle must handle with its buffer pools: periodically check the dirty flag and write down the changes to disk. the same is true for inodes ( while in memory ), which contain all the information that make up a file. closing a loop to oracle again: to speed up searching the ionodes linux maintains a hash table for fast access ( remember how oracle uses hashes to identify sql statements in the shared_pool ).

when there are files, there are processes which want to work with files. once a file is opened a new file object will be created. as these are frequent operations file objects are allocated through a slab cache.

the file objects itself are visible to the user through the /proc filesystem per process:

ls -la /proc/*/fd/
/proc/832/fd/:
total 0
dr-x------ 2 root root  0 2012-05-18 14:03 .
dr-xr-xr-x 8 root root  0 2012-05-18 06:40 ..
lrwx------ 1 root root 64 2012-05-18 14:03 0 -> /dev/null
lrwx------ 1 root root 64 2012-05-18 14:03 1 -> /dev/null
lr-x------ 1 root root 64 2012-05-18 14:03 10 -> anon_inode:inotify
lrwx------ 1 root root 64 2012-05-18 14:03 2 -> /dev/null
lrwx------ 1 root root 64 2012-05-18 14:03 3 -> anon_inode:[eventfd]
lrwx------ 1 root root 64 2012-05-18 14:03 4 -> /dev/null
lrwx------ 1 root root 64 2012-05-18 14:03 5 -> anon_inode:[signalfd]
lrwx------ 1 root root 64 2012-05-18 14:03 6 -> socket:[7507]
lrwx------ 1 root root 64 2012-05-18 14:03 7 -> anon_inode:[eventfd]
lrwx------ 1 root root 64 2012-05-18 14:03 8 -> anon_inode:[eventfd]
lrwx------ 1 root root 64 2012-05-18 14:03 9 -> socket:[11878]
...

usually numbers 0 – 3 refer to the standard input, standard output and standard error of the corresponding process.

last but not least there are the dentries. as with the file objects, dentries are allocated from a slab cache, the dentry cache in this case:

cat /proc/slabinfo | grep dentry
dentry             60121  61299    192   21    1 : tunables    0    0    0 : slabdata   2919   2919      0

directories are files, too, but special in that kind that dictories may contain other files or directories. once a directory is read into memory it is transformed into a dentry object. as this operation is expensive there is the dentry cache mentioned above. thus the operations for building the dentry objects can be minimized.
another link to oracle wording: the unused dentry double linked list uses a least recently used ( lru ) algorithm to track the usage of the entries. when the kernel needs to shrink the cache the objects at the tail of the list will be removed. as with the ionodes there is hash table for the dentries and a lock protecting the lists ( dcache_spin_lock in this case ).

this should give you enough hints to go further if you are interesed …

until now we had an introduction to processes, how they are managed, what signals are and what they are used for, how the linux kernel ( and oracle ) uses double linked list to quickly look up memory structures and how critical regions like shared memory can be protected. this post gives an introduction to timing and process scheduling.

as the cpu can execute only one process at a time but because maybe hundreds or thousands of processes want to do their work the kernel must provide a mechanism to decide which process to run next ( process switching ). this is the task of the scheduler. for being able to do what it does, the scheduler must be able to make decisions, and the decisions are based on time and priorities.

lots and lots of work behind the scenes is driven by time measurements. consider cronjobs, for example. without being able to measure time they would not work. in short the kernel must be able to keep the current time and to provide a mechanism to notify programs when a specific interval has elapsed.

on the one hand there is the real time clock ( accessible through the /dev/rtc interface ) which is a special chip that continues to tick even if the computer is powered off ( there is a small battery for this chip ). the real time clock is used by linux to derive the date and time.

on the other hand there are several other mechanisms which can be used for timing:

• time stamp counter
• programmable interval timer
• cpu local timer
• high precision event timer
• acpi power management timer

one of the time related activities the kernel must perform is to determine how long a process has been running. each process is given a time slot in which it may run, which is called a quanta. if the quantum expires and the process did not terminate a process switch may occur ( another process is selected for execution ). these processes are called expired. active processes are those which did not yet consume their quantum.
additionally each process has a priority assigned, which is used by the scheduler to decide how appropriate it is to let the process do its work on the cpu.

in general processes can be divided in three classes:

• interactive: typical interactive processes are those which respond to keyboard and mouse inputs of an end user. as an user wants to see quick responses, for example when editing text, these processes must be woken up quickly
• batch: batch processes do not interact with the user and often run in the background.
• real-time: real-time processes have very strong scheduling requirements and should not be blocked by processes with lower priorities.

in general the scheduler will give more attention to interactive processes than to batch processes, although this must not always be true.

one way we can change the base priority of processes from the command line is by using the “nice” command:

nice -19 vi

if you check the process without the nice call:

ps -aux | grep vi
oracle 4185 0.5 0.0 5400 1504 pts/0 S+ 10:51 0:00 vi

… and compare it to when you call vi with a nice value:

ps -aux | grep vi
oracle 4194 1.6 0.0 5400 1496 pts/0 SN+ 10:52 0:00 vi

.. you will see that “S+” changes to “SN+” ( the “N” stands for “low-priority (nice to other users)”

processes in linux are preemptable, which means that higher priority processes may suspend lower priority processes when they enter the running state. another reason a process can be preempted is when its time quantum expires.

consider this example: a user is writing an email while copying music from a cd to her computer. the email client is considered an interactive program while the copy job is considered a batch program. each time the user presses a key on her keyboard an interrupt occurs and the scheduler selects the email program for execution. but because users tend to think when writing emails there is plenty of time ( regarding the cpu ) between the key presses to wake up the copy job and let it do its work.

the time a process is allowed to be on a cpu, the quantum, is derived from a so called “static priority” which can be in the range of 100 to 139 ( with 100 being the highest priority and 139 being the lowest ). the higher the priority the more time the process is granted ( which ranges from 800ms for the highest priority to 5ms for the lowest priority ). in addition to the static priority there is a “dynamic priority” for each process ( again ranging from 100 to 139 ). without going too much into detail again: the dynamic priority is the one the scheduler uses for its decisions. as the name suggest, this priority may change over time ( depending on the average sleep time of a process ). processes with longer sleep times usually get a bonus ( the priority will be increased ) while processes with lower sleep times will get a penalty ( the priority will be decreased ). the average sleep time is also used by the scheduler to decide if processes are interactive or batch.

recall the post about double linked lists. the most important data structure used by the scheduler is the runqueue, which in fact is another linked list. this list links together all the process descriptors of the processes which want to run ( there is one runqueue per cpu ). one process can be in one runqueue only, but processes may migrate to others runqueues if the load between the cpus becomes unbalanced.

what to keep in mind: as only one process can run on one cpu at a time the scheduler decides which process to run next and which processes to suspend in case higher priority processes enter the running state. in general interactive processes are favored over batch processes and real-time processes should not be blocked by lower priority processes.

as mentioned in the previous post about semaphores there are more things to consider when it comes to interprocess communication. as semaphores are used to protect critical regions, there must be some critical regions to protect and this is the shared memory oracle uses for its communication.

to give an example on how the shared memory addressing works we will take a look at what happens when the database starts up.
for this you’ll need two sessions to a test infrastructure ( one as the database owner, the other as root ).

session one ( oracle ):
connect to sqlplus as sysdba make sure you shutdown the database ( do not exit sqlplus once the database is down ):

sqlplus / as sysdba
shutdown immediate

session two ( root ): discover the PID for then sqlplus session above …

ps -ef | grep sqlp
oracle    3062  3036  0 09:49 pts/1    00:00:00 sqlplus

… check the shared memory segments and trace the sqlplus PID from above:

ipcs -m
------ Shared Memory Segments --------
key        shmid      owner      perms      bytes      nattch     status      
0x7401003e 1310720    root      600        4          0                       
0x74010014 1998849    root      600        4          0                       
0x00000000 2359298    root      644        80         2                       
0x74010013 1966083    root      600        4          0                       
0x00000000 2392068    root      644        16384      2                       
0x00000000 2424837    root      644        280        2                       
0x00000000 2490374    grid      640        4096       0                       
0x00000000 2523143    grid      640        4096       0                       
0x8e11371c 2555912    grid      640        4096       0
# start the trace
strace -o db_startup.log -fp 3062

it is important to specify the “-f” flag for the strace call. this will tell strace to follow the child processes spawned.

in session one startup the database…

startup

… and stop the tracing in the root session once the database is up and re-check the shared memory segments.

ipcs -m
------ Shared Memory Segments --------
key        shmid      owner      perms      bytes      nattch     status      
0x7401003e 1310720    root      600        4          0                       
0x74010014 1998849    root      600        4          0                       
0x00000000 2359298    root      644        80         2                       
0x74010013 1966083    root      600        4          0                       
0x00000000 2392068    root      644        16384      2                       
0x00000000 2424837    root      644        280        2                       
0x00000000 2490374    grid      640        4096       0                       
0x00000000 2523143    grid      640        4096       0                       
0x8e11371c 2555912    grid      640        4096       0                       
0x00000000 3538953    oracle    640        4096       0                       
0x00000000 3571722    oracle    640        4096       0                       
0x3393b3a4 3604491    oracle    640        4096       0

as you can see, three more segments appeared after the database started up.

you’ll probably noticed some trace output on the screen similar to this:

Process 3468 detached
Process 3470 attached (waiting for parent)
Process 3470 resumed (parent 3409 ready)
Process 3471 attached (waiting for parent)
Process 3471 resumed (parent 3470 ready)
Process 3469 detached
Process 3470 detached

this is because of the “-f” flag given to strace.
the complete trace output is now available in the db_startup.log trace file and we are ready to take a look at it.

the first thing that catches the eye are the various references to the “/proc” filesystem. in may trace file there are 1213 calls to it. you can check this with:

grep "/proc/" db_startup.log | wc -l

take a look at the previous post which introduces the “/proc” filesystem for more information. for the scope of this post just notice how much depends on it.

the actual startup of the database is triggered by the following line:

execve("/opt/oracle/product/base/11.2.0.3/bin/oracle", ["oracleDB112", "(DESCRIPTION=(LOCAL=YES)(ADDRESS"], [/* 22 vars */]) = 0

this is the call to the oracle binary ( execve executes the binary ) with 22 arguments omitted. from now on the oracle instance starts up.

the calls important to the shared memory stuff are the following:

• brk: changes a data segment’s size
• mmap, munmap: maps/unmaps files or devices into memory
• mprotect: sets protection on a region of memory
• shmget: allocates a shared memory segment
• shmat, shmdt: performs attach/detach operations on shared memory
• get_mempolicy: return NUMA memory policies for a process
• semget: get a semaphore identifier
• semctl: perform control operations on a semaphore
• semop, semtimedop: perform sempahore operations

for each of the above commands you can check the man-pages for more information.
as the trace file is rather large and a lot of things are happening i will focus on the minimum ( this is not about re-engineering oracle :) ):

let’s check the keys returned by the ipcs command above:

egrep "3538953|3571722|3604491" db_startup.log
...
5365  shmget(IPC_PRIVATE, 4096, IPC_CREAT|IPC_EXCL|0640) = 3538953
5365  shmget(IPC_PRIVATE, 4096, IPC_CREAT|IPC_EXCL|0640) = 3571722
5365  shmget(0x3393b3a4, 4096, IPC_CREAT|IPC_EXCL|0640) = 3604491
...

as you can see the identifiers returned by the shmget call ( 3604491,3571722,3538953 ) correspond to the ones reported by ipcs. you wonder about the size of 4096 bytes ? this is because memory_target/memory_max_target is in use by the instance. if the database is configured using sga_target/sga_max_target you would see the actual size. let’s check this:

su - oracle
sqlplus / as sysdba
alter system reset memory_max_target scope=spfile;
alter system reset memory_target scope=spfile;
alter system set sga_max_size=256m scope=spfile;
alter system set sga_target=256m scope=spfile;
alter system set pga_aggregate_target=24m scope=spfile;
startup force;
exit;
# re-check the shared memory segments
ipcs -m
------ Shared Memory Segments --------
key        shmid      owner      perms      bytes      nattch     status      
0x00000000 2359298    root      644        80         2                       
0x00000000 2392068    root      644        16384      2                       
0x00000000 2424837    root      644        280        2                       
0x00000000 2490374    grid      640        4096       0                       
0x00000000 2523143    grid      640        4096       0                       
0x8e11371c 2555912    grid      640        4096       0                       
0x00000000 3801097    oracle    640        8388608    25                      
0x00000000 3833866    oracle    640        260046848  25                      
0x3393b3a4 3866635    oracle    640        2097152    25

the “260046848” corresponds to the sga size of 256m and the nattch column shows that 25 processes are attached to it. you can double check the 25
attached processes if you want:

ps -ef | grep DB112 | grep -v LISTENER | grep -v grep | wc -l

let’s return to the memory_target/memory_max_target configuration. as oracle puts together all the memory junks ( pga and sga ) the management of memory changes to the virtual shared memory filesystem ( tmpfs ). unfortunately this is not visible with the ipcs command.
but you can map your memory_* sizes to the shm filesystem:

ls -la /dev/shm/ | grep -v "+ASM"
total 466100
drwxrwxrwt  2 root   root        2640 Apr 10 13:09 .
drwxr-xr-x 10 root   root        3400 Apr 10 09:44 ..
-rw-r-----  1 oracle asmadmin 4194304 Apr 10 13:27 ora_DB112_3932169_0
-rw-r-----  1 oracle asmadmin 4194304 Apr 10 13:09 ora_DB112_3932169_1
-rw-r-----  1 oracle asmadmin 4194304 Apr 10 13:09 ora_DB112_3964938_0
-rw-r-----  1 oracle asmadmin 4194304 Apr 10 13:20 ora_DB112_3964938_1
-rw-r-----  1 oracle asmadmin 4194304 Apr 10 13:09 ora_DB112_3964938_10
-rw-r-----  1 oracle asmadmin 4194304 Apr 10 13:20 ora_DB112_3964938_11
-rw-r-----  1 oracle asmadmin 4194304 Apr 10 13:10 ora_DB112_3964938_12

note that i have excluded the ASM stuff here. in my case each segment ( or granule ) is 4mb of size ( this depends on the avaible memory of the system ) and the sum of all the segments should get you near to your memory_* configuration.

as ipcs can not tell you much here there are other commands to use. if you want to know which process has a memory granule open:

fuser -v /dev/shm/ora_DB112_4358154_49
                     USER        PID ACCESS COMMAND
/dev/shm/ora_DB112_4358154_49:
                     oracle     6626 ....m oracle
                     oracle     6628 ....m oracle
                     oracle     6630 ....m oracle
                     oracle     6634 ....m oracle
                     oracle     6636 ....m oracle
                     oracle     6638 ....m oracle
                     oracle     6640 ....m oracle
                     oracle     6642 ....m oracle
                     oracle     6644 ....m oracle
                     oracle     6646 ....m oracle
                     oracle     6648 ....m oracle
                     oracle     6650 ....m oracle
                     oracle     6652 ....m oracle
                     oracle     6654 ....m oracle
                     oracle     6656 ....m oracle
                     oracle     6658 ....m oracle
                     oracle     6662 ....m oracle
                     oracle     6669 ....m oracle
                     oracle     6744 ....m oracle
                     oracle     6767 ....m oracle
                     oracle     6769 ....m oracle
                     oracle     6791 ....m oracle
                     oracle     7034 ....m oracle

or the other way around, if you want to know which files are opened by a specific process:

ps -ef | grep pmon | grep -v "ASM"
oracle    6626     1  0 13:40 ?        00:00:05 ora_pmon_DB112
root      7075  5338  0 14:33 pts/0    00:00:00 grep pmon
# use the pmap command on the PID
pmap 6626
6626:   ora_pmon_DB112
0000000000400000 183436K r-x--  /opt/oracle/product/base/11.2.0.3/bin/oracle
000000000b922000   1884K rwx--  /opt/oracle/product/base/11.2.0.3/bin/oracle
000000000baf9000    304K rwx--    [ anon ]
0000000010c81000    660K rwx--    [ anon ]
0000000060000000      4K r-xs-  /dev/shm/ora_DB112_4325385_0
0000000060001000   4092K rwxs-  /dev/shm/ora_DB112_4325385_0
0000000060400000   4096K rwxs-  /dev/shm/ora_DB112_4325385_1
0000000060800000   4096K rwxs-  /dev/shm/ora_DB112_4358154_0
0000000060c00000   4096K rwxs-  /dev/shm/ora_DB112_4358154_1
0000000061000000   4096K rwxs-  /dev/shm/ora_DB112_4358154_2
0000000061400000   4096K rwxs-  /dev/shm/ora_DB112_4358154_3
0000000061800000   4096K rwxs-  /dev/shm/ora_DB112_4358154_4
0000000061c00000   4096K rwxs-  /dev/shm/ora_DB112_4358154_5
0000000062000000   4096K rwxs-  /dev/shm/ora_DB112_4358154_6
...

if you have troubles starting up your instance with this configuration ( ORA-00845 ) check the size of the virtual filesystem:

df -h
Filesystem            Size  Used Avail Use% Mounted on
/dev/hdc1              28G   14G   12G  54% /
tmpfs                 741M  456M  286M  62% /dev/shm

depending on your configuration ( memory_* or sga_* parameters ) the way that memory is managed changes ( from System V to POSIX, to be exact ).

lots and lots of information. not all of it is important to keep in mind. but what you should remember:
there are several processes and memory segments that make up the oracle instance. as several processes are attached to the same memory regions there must be a way to protect them from concurrent access ( think of semaphores ) … and oracle heavily depends on shared memory. if you scroll through the trace file you’ll notice that there are thousands of operations going on when an oracle instance starts up. imagine what is going on if the instance is under heavy workload and lots and lots of things need protection.

ps: for those interested:

there is plenty of more interesting stuff which you can find in the db_startup.log trace, for example:

writing the audit files:

grep -i adump db_startup.log  | grep -v ASM
3404  open("/oradata/DB112/admin/adump/DB112_ora_3404_2.aud", O_RDWR|O_CREAT|O_EXCL, 0660) = 10
3404  write(10, "/oradata/DB112/admin/adump/DB112"..., 47) = 47
3444  open("/oradata/DB112/admin/adump/DB112_ora_3444_1.aud", O_RDWR|O_CREAT|O_EXCL, 0660) = -1 EEXIST (File exists)
3444  open("/oradata/DB112/admin/adump/DB112_ora_3444_2.aud", O_RDWR|O_CREAT|O_EXCL, 0660) = 8
3444  write(8, "/oradata/DB112/admin/adump/DB112"..., 47) = 47
3481  open("/oradata/DB112/admin/adump/DB112_ora_3481_1.aud", O_RDWR|O_CREAT|O_EXCL, 0660 
3481  write(8, "/oradata/DB112/admin/adump/DB112"..., 47) = 47

writing the alert.log:

grep -i "alert_DB112.log" db_startup.log
3404  lstat("/oradata/DB112/admin/diag/rdbms/db112/DB112/trace/alert_DB112.log", {st_mode=S_IFREG|0640, st_size=110201, ...}) = 0
3404  open("/oradata/DB112/admin/diag/rdbms/db112/DB112/trace/alert_DB112.log", O_WRONLY|O_CREAT|O_APPEND, 0660) = 5
3404  lstat("/oradata/DB112/admin/diag/rdbms/db112/DB112/trace/alert_DB112.log", {st_mode=S_IFREG|0640, st_size=110260, ...}) = 0
3404  open("/oradata/DB112/admin/diag/rdbms/db112/DB112/trace/alert_DB112.log", O_WRONLY|O_CREAT|O_APPEND, 0660) = 11

reading the oracle message files:

grep msb db_startup.log
db_startup.log:5438  open("/opt/oracle/product/base/11.2.0.3/oracore/mesg/lrmus.msb", O_RDONLY) = 18
db_startup.log:5438  open("/opt/oracle/product/base/11.2.0.3/oracore/mesg/lrmus.msb", O_RDONLY) = 18
db_startup.log:5430  open("/opt/oracle/product/base/11.2.0.3/rdbms/mesg/oraus.msb", O_RDONLY 
db_startup.log:5494  open("/opt/oracle/product/base/11.2.0.3/rdbms/mesg/oraus.msb", O_RDONLY

getting sempahores:

grep semget db_startup.log 
5365  semget(IPC_PRIVATE, 1, IPC_CREAT|IPC_EXCL|0600) = 1081346
5365  semget(IPC_PRIVATE, 124, IPC_CREAT|IPC_EXCL|0666) = 1114114
5365  semget(IPC_PRIVATE, 124, IPC_CREAT|0660) = 1146882
5365  semget(0x710dfe10, 0, 0)          = -1 ENOENT (No such file or directory)
5365  semget(0x46db3f80, 0, 0)          = -1 ENOENT (No such file or directory)
5365  semget(0x9ae46084, 0, 0)          = -1 ENOENT (No such file or directory)
5365  semget(0xf6dcc368, 0, 0)          = -1 ENOENT (No such file or directory)
5365  semget(0x710dfe10, 124, IPC_CREAT|IPC_EXCL|0640) = 1179650

some exadata stuff:

3404  open("/etc/oracle/cell/network-config/cellinit.ora", O_RDONLY) = -1 ENOENT (No such file or directory)

and … and …