Declared failure of DDLP to entirely solve the concurrency problem

This commit is contained in:
Thomas Schmitt 2007-04-21 12:37:24 +00:00
parent 571af4f255
commit c64cda0bde

View File

@ -1,3 +1,25 @@
-------------------------------------------------------------------------------
Users of modern desktop Linux installations report misburns with CD/DVD
recording due to concurrency problems.
This text describes two locking protocols which have been developed by our
best possible effort. But finally they rather serve as repelling example of
what would be needed in user space to achieve an insufficient partial solution.
Ted Ts'o was so friendly to help as critic with his own use cases. It turned
out that we cannot imagine a way in user space how to cover reliably the needs
of callers of libblkid and the needs of our burn programs.
-------------------------------------------------------------------------------
Content:
The "Delicate Device Locking Protocol" shall demonstrate our sincere
consideration of the problem.
"What are the Stumble Stones ?" lists reasons why the effort finally failed.
-----------------------------------------------------------------------------
Delicate Device Locking Protocol
@ -211,27 +233,23 @@ Prone to failure without further reason is:
DDLP-B
This protocol relies on proxy lock files in some filesystem directory. It can
be embedded into DDLP-A or it ican be used be used standalone, outside DDLP-A.
be embedded into DDLP-A or it can be used be used standalone, outside DDLP-A.
DDLP-A shall be kept by DDLP-B from trying to access any device file which
might already be in use. There is a problematic gesture in DDLP-A when SCSI
address parameters are to be retrieved. For now this gesture seems to be
harmless. But one never knows.
Vice versa DDLP-B may get from DDLP-A the service to search for SCSI device
file siblings. So they are best as a couple.
But they are not perfect. Not even as couple. fcntl() locking is flawed.
There is a proxy file locking protocol described in FHS:
http://www.pathname.com/fhs/pub/fhs-2.3.html#VARLOCKLOCKFILES
But it has shortcommings:
- Stale locks are possible.
- Much info is missing about the occupying process: host id, program, purpose
- It is necessary to create a file (using the _old_ meaning of O_EXCL flag ?).
- No way to indicate difference between exclusive and shared locks.
- Relies entirely on basename of device file path.
- /var/lock/ is not available early during system start and often has
restrictive permission settings.
The stale locks and the clear prescriptions in FHS make /var/lock/ entirely
unsuitable for our purpose.
But it has shortcommings (see below). Decisive obstacle for its usage are the
possibility for stale locks and the lack of shared locks.
DDLP-B rather defines a "path prefix" which is advised to be
/tmp/ddlpb-lock-
@ -244,23 +262,21 @@ then act as additional access restriction to the device files.
The use of fcntl(F_SETLK) will prevent any stale locks after the process ended.
It will also allow to obtain shared locks as well as exclusive locks.
There are several classes of device specific suffixes:
There are two classes of device specific suffixes:
- Device file path suffix. "/" gets replaced by "_-". Eventual "_-" in path
gets replaced by "_-_-".
E.g.: "_-dev_-sr0" , "_-mydevs_-burners_-nec"
- Device file path suffix. Absolute paths only. "/" gets replaced by "_-".
Eventual "_-" in path gets replaced by "_-_-". The leading group of "_-"
is always interpreted as a group of "/", though. E.g.:
/dev/sr0 <-> "_-dev_-sr0"
/mydevs/burner/nec <-> "_-mydevs_-burners_-nec"
/dev/rare_-name <-> "_-dev_-rare_-_-name"
///strange/dev/x <-> "_-_-_-strange_-dev_-x"
- st_rdev suffix. A hex representation of struct stat.st_rdev. Capital letters.
The number of characters is pare with at most one leading 0. I.e. bytewise
printf("%2.2X") beginning with the highest order byte that is not zero.
E.g. : "0B01", "2200", "01000000000004001"
- SCSI parameter suffix. A tuple of decimal numbers representing the SCSI
address if applicable for the device at all. On Linux this are the four
numbers Host,Channel,Id,Lun obtained by ioctl(SCSI_IOCTL_GET_IDLUN).
The separator is the minor letter "s".
E.g. "1s0s0s0", "0s0s3s0"
If a lockfile does not exist and cannot be created then this shall not keep
a program from working on a device. But if a lockfile exists and if permissions
or locking state do not allow to obtain a lock of the appropirate type, then
@ -270,25 +286,103 @@ immediate close(2) of an already opened device file.
The vulnerable programs shall not start their operation before they locked a
wide collection of drive representations.
Non-vulnerable programs shall take care to lock at least the suffix resulting
from the path they will be using and the suffix of the st_rdev from that path.
Non-vulnerable programs shall take care to lock the suffix resulting from the
path they will be using and the suffix from the st_rdev from that path.
The latter is to be obtained by call stat(2).
>>> Vulnerable program shall use SCSI parameter suffixes to ensure that the search
>>> for further paths and st_rdev representations of the same device does not
>>> disturb
Locks get upheld as long as their file descriptor is not closed or no other
incident as described in man 2 fcntl releases the lock.
So with shared locks there are no imandatory further activities after they
have been obtained.
In case of exclusive locks, the file has to have been opened for writing and
must be truncated to 0 bytes length immediately after obtaining the lock.
When releasing an exclusive lock it is a nice gesture to
already do this truncation.
Then a /var/lock/ compatible first line has to be written.
E.g. by: printf("%10u\n",(unsigned) getpid()) yielding " 1230\n".
Any further lines are optional. They shall have the form Name=Value and must
be printable cleartext. If such further lines exist, then the last one must
have the name "endmark".
Defined Names are:
hostid =hostname of the machine where the process number of line 1 is valid
start =start time of lock in seconds since 1970. E.g: 1177147634.592410
program =self chosen name of the program which obtained the lock
argv0 =argv[0] of that program
mainpath =device file path which will be used for operations by that program
path =device file path which lead to the lock
st_rdev =st_rdev suffix which is associated with path
scsi_hcil=eventual SCSI parameters Host,Channel,Id,Lun
scsi_bus =eventual SCSI parameter Bus
endmark =declares the info as complete.
Any undefined name or a line without "=" shall be handled as comment.
"=" in the value is allowed. Any line beginning with an "=" character is an
extension of the previous value.
If programs encounter an exclusive lock, they are invited to read the content
of the lockfile anyway. But they should be aware that the info might be in the
progress of emerging. There is a race condition possible in the short time
between obtaining the exclusive lock and erasing the file content.
If it is not crucial to obtain most accurate info then one may take the newline
of the first line as indicator of a valid process number and the "endmark"
name as indicator that the preceding lines are valid.
Very cautious readers should obtain the info twice with a decent waiting period
inbetween. Only if both results are identical they should be considered valid.
If it is sure that the device has valid SCSI address parameters then these
should be obtained first and the SCSI parameter suffix should be locked before
any further activity is started. If done so, then the open(2) flags shall
include O_NDELAY to avoid side effect. O_NDELAY may be revoked later by
fcntl(2) F_GETFL,F_SETFL.
This gesture is mandatory only for vulnerable
programs in order to obtain more path and st_rdev suffixes.
There is no implementation of DDLP-B yet.
Example: Device file path "/dev/sr1"
----------------------------------------------------------------------------
What are the Stumble Stones ?
----------------------------------------------------------------------------
Any of the considered locking mechanisms has decisive shortcommings
which keeps it from being the solution to all known legitimate use cases.
The attempt has failed to compose a waterproof locking mechanism from means of
POSIX, FHS and from hardly documented Linux open(O_EXCL) on device files.
The resulting mechanisms would need about 1000 lines of code and still do
not close all gaps resp. cover the well motivated use cases.
This attempt you see above: DDLP-A and DDLP-B.
Summary of the reasons why the established locking mechanisms do not suffice:
None of the mechanisms can take care of the double device driver identity
sr versus sg. To deduce the one device file from the other involves the need
to open many other (possibly unrelated) device files with the risk to disturb
them.
This hard to solve problem is aggravated by the following facts.
Shortcommings of Linux specific open(O_EXCL) :
- O_EXCL | O_RDONLY does not succeed with /dev/sg*
- O_EXCL cannot provide shared locks for programs which only want to lock
against burn programs but not against their own peers.
- O_EXCL keeps from obtaining information by harmless activities.
- O_EXCL already has a meaning with devices which are mounted as filesystems.
This priority meaning is more liberal than the one needed for CD/DV recording
protection.
Shortcommings of POSIX fcntl(F_SETLK) :
- fcntl() demands an open file descriptor. open(2) might have side effects.
- fcntl() locks can be released inadvertedly by submodules which just open and
close the same file (inode ?) without refering to fcntl locks in any way.
See man 2 fcntl "This is bad:".
Stacking of software modules is a widely used design pattern. But fcntl()
cannot cope with that.
Shortcommings of FHS /var/lock/ :
- Stale locks are possible.
- It is necessary to create a file (using the _old_ meaning of O_EXCL flag ?)
but /var/lock/ might not be available early during system start and it often
has restrictive permission settings.
- There is no way to indicate a difference between exclusive and shared locks.
- The FHS prescription relies entirely on the basename of the device file path.