DESCRIPTION
The
th_define utility provides an interface to the
bus_ops fault injection
bofi device driver for defining error injection specifications (referred to as errdefs). An errdef corresponds to a specification of how to corrupt a device driver's accesses to its hardware. The command line arguments determine the precise nature of the fault to be injected. If the supplied arguments define a consistent errdef, the
th_define process will store the errdef with the
bofi driver and suspend itself until the criteria given by the errdef become satisfied (in practice, this will occur when the access counts go to zero).
You use the
th_manage(1M) command with the
start option to activate the resulting errdef. The effect of
th_manage with the
start option is that the
bofi driver acts upon the errdef by matching the number of hardware accesses—specified in
count, that are of the type specified in
acc_types, made by instance number
instance—of the driver whose name is
name, (or by the driver instance specified by
path) to the register set (or DMA handle) specified by
reg_number, that lie within the range
offset to
offset +
length from the beginning of the register set or DMA handle. It then applies
operator and
operand to the next
failcount matching accesses.
If
acc_types includes
log,
th_define runs in automatic test script generation mode, and a set of test scripts (written in the Korn shell) is created and placed in a sub-directory of the current directory with the name
<driver>.test.
<id> (for example,
glm.test.978177106). A separate, executable script is generated for each access handle that matches the logging criteria. The log of accesses is placed at the top of each script as a record of the session. If the current directory is not writable, file output is written to standard output. The base name of each test file is the driver name, and the extension is a number that discriminates between different access handles. A control script (with the same name as the created test directory) is generated that will run all the test scripts sequentially.
Executing the scripts will install, and then activate, the resulting error definitions. Error definitions are activated sequentially and the driver instance under test is taken offline and brought back online before each test (refer to the
-e option for more information). By default, logging applies to all
PIO accesses, all interrupts, and all DMA accesses to and from areas mapped for both reading and writing. You can constrain logging by specifying additional
acc_types,
reg_number,
offset and
length. Logging will continue for
count matching accesses, with an optional time limit of
collect_time seconds.
Either the
-n or
-P option must be provided. The other options are optional. If an option (other than
-a) is specified multiple times, only the final value for the option is used. If an option is not specified, its associated value is set to an appropriate default, which will provide maximal error coverage as described below.
OPTIONS
The following options are available:
-n name
Specify the name of the driver to test. (String)
-i instance
Test only the specified driver instance (-1 matches all instances of driver). (Numeric)
-P path
Specify the full device path of the driver to test. (String)
-r reg_number
Test only the given register set or DMA handle (-1 matches all register sets and DMA handles). (Numeric)
-a acc_types
Only the specified access types will be matched. Valid values for the
acc_types argument are
log,
pio,
pio_r,
pio_w,
dma,
dma_r,
dma_w and
intr. Multiple access types, separated by spaces, can be specified. The default is to match all hardware accesses.
If
acc_types is set to
log, logging will match all
PIO accesses, interrupts and DMA accesses to and from areas mapped for both reading and writing.
log can be combined with other
acc_types, in which case the matching condition for logging will be restricted to the specified additional
acc_types. Note that
dma_r will match only DMA handles mapped for reading only;
dma_w will match only DMA handles mapped for writing only;
dma will match only DMA handles mapped for both reading and writing.
-l offset [length]
Constrain the range of qualifying accesses. The offset and length arguments indicate that any access of the type specified with the -a option, to the register set or DMA handle specified with the -r option, lie at least offset bytes into the register set or DMA handle and at most offset + length bytes into it. The default for offset is 0. The default for length is the maximum value that can be placed in an offset_t C data type (see types.h). Negative values are converted into unsigned quantities. Thus, th_define -l 0 -1 is maximal.
-c count[failcount]
Wait for
count number of matching accesses, then apply an operator and operand (see the
-o option) to the next
failcount number of matching accesses. If the access type (see the
-a option) includes logging, the number of logged accesses is given by
count +
failcount - 1. The -1 is required because the last access coincides with the first faulting access.
Note that access logging may be combined with error injection if
failcount and
operator are nonzero and if the access type includes logging and any of the other access types (
pio,
dma and
intr) See the description of access types in the definition of the
-a option, above.
When the
count and
failcount fields reach zero, the status of the errdef is reported to standard output. When all active errdefs created by the
th_define process complete, the process exits. If
acc_types includes
log,
count determines how many accesses to log. If
count is not specified, a default value is used. If
failcount is set in this mode, it will simply increase the number of accesses logged by a further
failcount - 1.
-o operator [operand]
For qualifying PIO read and write accesses, the value read from or written to the hardware is corrupted according to the value of
operator:
EQ
operand is returned to the driver.
OR
operand is bitwise ORed with the real value.
AND
operand is bitwise ANDed with the real value.
XOR
operand is bitwise XORed with the real value.
For PIO write accesses, the following operator is allowed:
NO
Simply ignore the driver's attempt to write to the hardware.
Note that a driver performs PIO via the
ddi_getX(),
ddi_putX(),
ddi_rep_getX() and
ddi_rep_putX() routines (where
X is 8, 16, 32 or 64). Accesses made using
ddi_getX() and
ddi_putX() are treated as a single access, whereas an access made using the
ddi_rep_*(9F) routines are broken down into their respective number of accesses, as given by the
repcount parameter to these DDI calls. If the access is performed via a DMA handle,
operator and
value are applied to every access that comprises the DMA request. If interference with interrupts has been requested then the operator may take any of the following values:
DELAY
After count accesses (see the -c option), delay delivery of the next failcount number of interrupts for operand number of microseconds.
LOSE
After count number of interrupts, fail to deliver the next failcount number of real interrupts to the driver.
EXTRA
After count number of interrupts, start delivering operand number of extra interrupts for the next failcount number of real interrupts.
The default value for
operand and
operator is to corrupt the data access by flipping each bit (XOR with -1).
-f acc_chk
If the acc_chk parameter is set to 1 or pio, then the driver's calls to ddi_check_acc_handle(9F) return DDI_FAILURE when the access count goes to 1. If the acc_chk parameter is set to 2 or dma, then the driver's calls to ddi_check_dma_handle(9F) return DDI_FAILURE when the access count goes to 1.
-w max_wait_period [report_interval]
Constrain the period for which an error definition will remain active. The option applies only to non-logging errdefs. If an error definition remains active for
max_wait_period seconds, the test will be aborted. If
report_interval is set to a nonzero value, the current status of the error definition is reported to standard output every
report_interval seconds. The default value is zero. The status of the errdef is reported in parsable format (eight fields, each separated by a colon (
:) character, the last of which is a string enclosed by double quotes and the remaining seven fields are integers):
ft:
mt:
ac:
fc:
chk:
ec:
s:
"message" which are defined as follows:
ft
The UTC time when the fault was injected.
mt
The UTC time when the driver reported the fault.
ac
The number of remaining non-faulting accesses.
fc
The number of remaining faulting accesses.
chk
The value of the acc_chk field of the errdef.
ec
The number of fault reports issued by the driver against this errdef (mt holds the time of the initial report).
s
The severity level reported by the driver.
"message"
Textual reason why the driver has reported a fault.
-h
Display the command usage string.
-s collect_time
If acc_types is given with the -a option and includes log, the errdef will log accesses for collect_time seconds (the default is to log until the log becomes full). Note that, if the errdef specification matches multiple driver handles, multiple logging errdefs are registered with the bofi driver and logging terminates when all logs become full or when collect_time expires or when the associated errdefs are cleared. The current state of the log can be checked with the th_manage(1M) command, using the broadcast parameter. A log can be terminated by running th_manage(1M) with the clear_errdefs option or by sending a SIGALRM signal to the th_define process. See alarm(2) for the semantics of SIGALRM.
-p policy
Applicable when the
acc_types option includes
log. The parameter modifies the policy used for converting from logged accesses to errdefs. All policies are inclusive:
-
o
-
Use rare to bias error definitions toward rare accesses (default).
-
o
-
Use operator to produce a separate error definition for each operator type (default).
-
o
-
Use common to bias error definitions toward common accesses.
-
o
-
Use median to bias error definitions toward median accesses.
-
o
-
Use maximal to produce multiple error definitions for duplicate accesses.
-
o
-
Use unbiased to create unbiased error definitions.
-
o
-
Use onebyte, twobyte, fourbyte, or eightbyte to select errdefs corresponding to 1, 2, 4 or 8 byte accesses (if chosen, the -xr option is enforced in order to ensure that ddi_rep_*() calls are decomposed into multiple single accesses).
-
o
-
Use multibyte to create error definitions for multibyte accesses performed using ddi_rep_get*() and ddi_rep_put*().
Policies can be combined by adding together these options. See the NOTES section for further information.
-x flags
Applicable when the
acc_types option includes
log. The
flags parameter modifies the way in which the
bofi driver logs accesses. It is specified as a string containing any combination of the following letters:
w
Continuous logging (that is, the log will wrap when full).
t
Timestamp each log entry (access times are in seconds).
r
Log repeated I/O as individual accesses (for example, a ddi_rep_get16(9F) call which has a repcount of N is logged N times with each transaction logged as size 2 bytes. Without this option, the default logging behavior is to log this access once only, with a transaction size of twice the repcount).
-C comment_string
Applicable when the acc_types option includes log. It provides a comment string to be placed in any generated test scripts. The string must be enclosed in double quotes.
-e fixup_script [args]
Applicable when the
acc_types option includes
log. The output of a logging errdefs is to generate a test script for each driver access handle. Use this option to embed a command in the resulting script before the errors are injected. The generated test scripts will take an instance offline and bring it back online before injecting errors in order to bring the instance into a known fault-free state. The executable
fixup_script will be called twice with the set of optional
args— once just before the instance is taken offline and again after the instance has been brought online. The following variables are passed into the environment of the called executable:
DRIVER_PATH
Identifies the device path of the instance.
DRIVER_INSTANCE
Identifies the instance number of the device.
DRIVER_UNCONFIGURE
Has the value 1 when the instance is about to be taken offline.
DRIVER_CONFIGURE
Has the value 1 when the instance has just been brought online.
Typically, the executable ensures that the device under test is in a suitable state to be taken offline (unconfigured) or in a suitable state for error injection (for example configured, error free and servicing a workload). A minimal script for a network driver could be:
#!/bin/ksh
driver=xyznetdriver
ifnum=$driver$DRIVER_INSTANCE
if [[ $DRIVER_CONFIGURE = 1 ]]; then
ifconfig $ifnum plumb
ifconfig $ifnum ...
ifworkload start $ifnum
elif [[ $DRIVER_UNCONFIGURE = 1 ]]; then
ifworkload stop $ifnum
ifconfig $ifnum down
ifconfig $ifnum unplumb
fi
exit $?
The
-e option must be the last option on the command line.
If the
-a log option is selected but the
-e option is not given, a default script is used. This script repeatedly attempts to detach and then re-attach the device instance under test.
EXAMPLES
Examples of Error Definitions
th_define -n foo -i 1 -a log
Logs all accesses to all handles used by instance 1 of the
foo driver while running the default workload (attaching and detaching the instance). Then generates a set of test scripts to inject appropriate errdefs while running that default workload.
th_define -n foo -i 1 -a log pio
Logs PIO accesses to each PIO handle used by instance 1 of the
foo driver while running the default workload (attaching and detaching the instance). Then generates a set of test scripts to inject appropriate errdefs while running that default workload.
th_define -n foo -i 1 -p onebyte median -e fixup arg -now
Logs all accesses to all handles used by instance 1 of the
foo driver while running the workload defined in the fixup script
fixup with arguments
arg and
-now. Then generates a set of test scripts to inject appropriate errdefs while running that workload. The resulting error definitions are requested to focus upon single byte accesses to locations that are accessed a
median number of times with respect to frequency of access to I/O addresses.
th_define -n se -l 0x20 1 -a pio_r -o OR 0x4 -c 10 1000
Simulates a stuck serial chip command by forcing 1000 consecutive read accesses made by any instance of the
se driver to its command status register, thereby returning status busy.
th_define -n foo -i 3 -r 1 -a pio_r -c 0 1 -f 1 -o OR 0x100
Causes 0x100 to be ORed into the next physical I/O read access from any register in register set 1 of instance 3 of the
foo driver. Subsequent calls in the driver to
ddi_check_acc_handle() return
DDI_FAILURE.
th_define -n foo -i 3 -r 1 -a pio_r -c 0 1 -o OR 0x0
Causes 0x0 to be ORed into the next physical I/O read access from any register in register set 1 of instance 3 of the
foo driver. This is of course a no-op.
th_define -n foo -i 3 -r 1 -l 0x8100 1 -a pio_r -c 0 10 -o EQ 0x70003
Causes the next ten next physical I/O reads from the register at offset 0x8100 in register set 1 of instance 3 of the
foo driver to return 0x70003.
th_define -n foo -i 3 -r 1 -l 0x8100 1 -a pio_w -c 100 3 -o AND 0xffffffffffffefff
The next 100 physical I/O writes to the register at offset 0x8100 in register set 1 of instance 3 of the
foo driver take place as normal. However, on each of the three subsequent accesses, the 0x1000 bit will be cleared.
th_define -n foo -i 3 -r 1 -l 0x8100 0x10 -a pio_r -c 0 1 -f 1 -o XOR 7
Causes the bottom three bits to have their values toggled for the next physical I/O read access to registers with offsets in the range 0x8100 to 0x8110 in register set 1 of instance 3 of the
foo driver. Subsequent calls in the driver to
ddi_check_acc_handle() return
DDI_FAILURE.
th_define -n foo -i 3 -a pio_w -c 0 1 -o NO 0
Prevents the next physical I/O write access to any register in any register set of instance 3 of the
foo driver from going out on the bus.
th_define -n foo -i 3 -l 0 8192 -a dma_r -c 0 1 -o OR 7
Causes 0x7 to be ORed into each
long long in the first 8192 bytes of the next DMA read, using any DMA handle for instance 3 of the
foo driver.
th_define -n foo -i 3 -r 2 -l 0 8 -a dma_r -c 0 1 -o OR 0x7070707070707070
Causes 0x70 to be ORed into each byte of the first
long long of the next DMA read, using the DMA handle with sequential allocation number 2 for instance 3 of the
foo driver.
th_define -n foo -i 3 -l 256 256 -a dma_w -c 0 1 -f 2 -o OR 7
Causes 0x7 to be ORed into each
long long in the range from offset 256 to offset 512 of the next DMA write, using any DMA handle for instance 3 of the
foo driver. Subsequent calls in the driver to
ddi_check_dma_handle() return
DDI_FAILURE.
th_define -n foo -i 3 -r 0 -l 0 8 -a dma_w -c 100 3 -o AND 0xffffffffffffefff
The next 100 DMA writes using the DMA handle with sequential allocation number 0 for instance 3 of the
foo driver take place as normal. However, on each of the three subsequent accesses, the 0x1000 bit will be cleared in the first
long long of the transfer.
th_define -n foo -i 3 -a intr -c 0 6 -o LOSE 0
Causes the next six interrupts for instance 3 of the
foo driver to be lost.
th_define -n foo -i 3 -a intr -c 30 1 -o EXTRA 10
When the thirty-first subsequent interrupt for instance 3 of the
foo driver occurs, a further ten interrupts are also generated.
th_define -n foo -i 3 -a intr -c 0 1 -o DELAY 1024
Causes the next interrupt for instance 3 of the
foo driver to be delayed by 1024 microseconds.
NOTES
The policy option in the
th_define -p syntax determines how a set of logged accesses will be converted into the set of error definitions. Each logged access will be matched against the chosen policies to determine whether an error definition should be created based on the access.
Any number of policy options can be combined to modify the generated error definitions.
Bytewise Policies
These select particular I/O transfer sizes. Specifying a byte policy will exclude other byte policies that have not been chosen. If none of the byte type policies is selected, all transfer sizes are treated equally. Otherwise, only those specified transfer sizes will be selected.
onebyte
Create errdefs for one byte accesses (ddi_get8())
twobyte
Create errdefs for two byte accesses (ddi_get16())
fourbyte
Create errdefs for four byte accesses (ddi_get32())
eightbyte
Create errdefs for eight byte accesses (ddi_get64())
multibyte
Create errdefs for repeated byte accesses (ddi_rep_get*())
Frequency of Access Policies
The frequency of access to a location is determined according to the access type, location and transfer size (for example, a two-byte read access to address A is considered distinct from a four-byte read access to address A). The algorithm is to count the number of accesses (of a given type and size) to a given location, and find the locations that were most and least accessed (let
maxa and
mina be the number of times these locations were accessed, and
mean the total number of accesses divided by total number of locations that were accessed). Then a rare access is a location that was accessed less than
(mean - mina) / 3 + mina
times. Similarly for the definition of common accesses:
maxa - (maxa - mean) / 3
A location whose access patterns lies within these cutoffs is regarded as a location that is accessed with median frequency.
rare
Create errdefs for locations that are rarely accessed.
common
Create errdefs for locations that are commonly accessed.
median
Create errdefs for locations that are accessed a median frequency.
Policies for Minimizing errdefs
If a transaction is duplicated, either a single or multiple errdefs will be written to the test scripts, depending upon the following two policies:
maximal
Create multiple errdefs for locations that are repeatedly accessed.
unbiased
Create a single errdef for locations that are repeatedly accessed.
operators
For each location, a default operator and operand is typically applied. For maximal test coverage, this default may be modified using the operators policy so that a separate errdef is created for each of the possible corruption operators.