Buffers and Buffering

Data buffering and management is an essential service provided by the DTrace framework for its clients, such as dtrace(1M). This chapter explores data buffering in detail and describes options you can use to change DTrace's buffer management policies.

11.1. Principal Buffers

The principal buffer is present in every DTrace invocation and is the buffer to which tracing actions record their data by default. These actions include:

`exit`	`printf`	`trace`	`ustack`
`printa`	`stack`	`tracemem`

The principal buffers are always allocated on a per-CPU basis. This policy is not tunable, but tracing and buffer allocation can be restricted to a single CPU by using the cpu option.

11.2. Principal Buffer Policies

DTrace permits tracing in highly constrained contexts in the kernel. In particular, DTrace permits tracing in contexts in which kernel software may not reliably allocate memory. The consequence of this flexibility of context is that there always exists a possibility that DTrace will attempt to trace data when there isn't space available. DTrace must have a policy to deal with such situations when they arise, but you might wish to tune the policy based on the needs of a given experiment. Sometimes the appropriate policy might be to discard the new data. Other times it might be desirable to reuse the space containing the oldest recorded data to trace new data. Most often, the desired policy is to minimize the likelihood of running out of available space in the first place. To accommodate these varying demands, DTrace supports several different buffer policies. This support is implemented with the bufpolicy option, and can be set on a per-consumer basis. See Options and Tunables for more details on setting options.

11.2.1. switch Policy

By default, the principal buffer has a switch buffer policy. Under this policy, per-CPU buffers are allocated in pairs: one buffer is active and the other buffer is inactive. When a DTrace consumer attempts to read a buffer, the kernel firsts switches the inactive and active buffers. Buffer switching is done in such a manner that there is no window in which tracing data may be lost. Once the buffers are switched, the newly inactive buffer is copied out to the DTrace consumer. This policy assures that the consumer always sees a self-consistent buffer: a buffer is never simultaneously traced to and copied out. This technique also avoids introducing a window in which tracing is paused or otherwise prevented. The rate at which the buffer is switched and read out is controlled by the consumer with the switchrate option. As with any rate option, switchrate may be specified with any time suffix, but defaults to rate-per-second. For more details on switchrate and other options, see Options and Tunables.

To process the principal buffer at user-level at a rate faster than the default of once per second, tune the value of switchrate. The system processes actions that induce user-level activity (such as printa and system) when the corresponding record in the principal buffer is processed. The value of switchrate dictates the rate at which the system processes such actions.

Under the switch policy, if a given enabled probe would trace more data than there is space available in the active principal buffer, the data is dropped and a per-CPU drop count is incremented. In the event of one or more drops, dtrace(1M) displays a message similar to the following example:

dtrace: 11 drops on CPU 0

If a given record is larger than the total buffer size, the record will be dropped regardless of buffer policy. You can reduce or eliminate drops by either increasing the size of the principal buffer with the bufsize option or by increasing the switching rate with the switchrate option.

Under the switch policy, scratch space for copyin, copyinstr, and alloca is allocated out of the active buffer.

11.2.2. fill Policy

For some problems, you might wish to use a single in-kernel buffer. While this approach can be implemented with the switch policy and appropriate D constructs by incrementing a variable in D and predicating an exit action appropriately, such an implementation does not eliminate the possibility of drops. To request a single, large in-kernel buffer, and continue tracing until one or more of the per-CPU buffers has filled, use the fill buffer policy. Under this policy, tracing continues until an enabled probe attempts to trace more data than can fit in the remaining principal buffer space. When insufficient space remains, the buffer is marked as filled and the consumer is notified that at least one of its per-CPU buffers has filled. Once dtrace(1M) detects a single filled buffer, tracing is stopped, all buffers are processed and dtrace exits. No further data will be traced to a filled buffer even if the data would fit in the buffer.

To use the fill policy, set the bufpolicy option to fill. For example, the following command traces every system call entry into a per-CPU 2K buffer with the buffer policy set to fill:

# dtrace -n syscall:::entry -b 2k -x bufpolicy=fill

fill Policy and END Probes

END probes normally do not fire until tracing has been explicitly stopped by the DTrace consumer. END probes are guaranteed to only fire on one CPU, but the CPU on which the probe fires is undefined. With fill buffers, tracing is explicitly stopped when at least one of the per-CPU principal buffers has been marked as filled. If the fill policy is selected, the END probe may fire on a CPU that has a filled buffer. To accommodate END tracing in fill buffers, DTrace calculates the amount of space potentially consumed by END probes and subtracts this space from the size of the principal buffer. If the net size is negative, DTrace will refuse to start, and dtrace(1M) will output a corresponding error message:

dtrace: END enablings exceed size of principal buffer

The reservation mechanism ensures that a full buffer always has sufficient space for any END probes.

11.2.3. ring Policy

The DTrace ring buffer policy helps you trace the events leading up to a failure. If reproducing the failure takes hours or days, you might wish to keep only the most recent data. Once a principal buffer has filled, tracing wraps around to the first entry, thereby overwriting older tracing data. You establish the ring buffer by setting the bufpolicy option to the string ring:

# dtrace -s foo.d -x bufpolicy=ring

When used to create a ring buffer, dtrace(1M) will not display any output until the process is terminated. At that time, the ring buffer is consumed and processed. dtrace processes each ring buffer in CPU order. Within a CPU's buffer, trace records will be displayed in order from oldest to youngest. Just as with the switch buffering policy, no ordering exists between records from different CPUs are made. If such an ordering is required, you should trace the timestamp variable as part of your tracing request.

The following example demonstrates the use of a #pragma option directive to enable ring buffering:

#pragma D option bufpolicy=ring
#pragma D option bufsize=16k

syscall:::entry
/execname == $1/
{
	trace(timestamp);
}

syscall::rexit:entry
{
	exit(0);
}

11.3. Other Buffers

Principal buffers exist in every DTrace enabling. Beyond principal buffers, some DTrace consumers may have additional in-kernel data buffers: an aggregation buffer, discussed in Aggregations, and one or more speculative buffers, discussed in Speculative Tracing.

11.4. Buffer Sizes

The size of each buffer can be tuned on a per-consumer basis. Separate options are provided to tune each buffer size, as shown in the following table:

Buffer	Size Option
Principal	`bufsize`
Speculative	`specsize`
Aggregation	`aggsize`

Each of these options is set with a value that denotes the size. As with any size option, the value may have an optional size suffix. See Options and Tunables for more details. For example, to set the buffer size to one megabyte on the command line to dtrace, you can use -x to set the option:

# dtrace -P syscall -x bufsize=1m

Alternatively, you can use the -b option to dtrace:

# dtrace -P syscall -b 1m

Finally, you could set bufsize using #pragma D option:

#pragma D option bufsize=1m

The buffer size you select denotes the size of the buffer on each CPU. Moreover, for the switch buffer policy, bufsize denotes the size of each buffer on each CPU. The buffer size defaults to four megabytes.

11.5. Buffer Resizing Policy

Occasionally, the system might not have adequate free kernel memory to allocate a buffer of desired size either because not enough memory is available or because the DTrace consumer has exceeded one of the tunable limits described in Options and Tunables. You can configure the policy for buffer allocation failure using bufresize option, which defaults to auto. Under the auto buffer resize policy, the size of a buffer is halved until a successful allocation occurs. dtrace(1M) generates a message if a buffer as allocated is smaller than the requested size:

# dtrace -P syscall -b 4g
dtrace: description 'syscall' matched 430 probes
dtrace: buffer size lowered to 128m
...

or:

# dtrace -P syscall'{@a[probefunc] = count()}' -x aggsize=1g
dtrace: description 'syscall' matched 430 probes
dtrace: aggregation size lowered to 128m
...

Alternatively, you can require manual intervention after buffer allocation failure by setting bufresize to manual. Under this policy, a failure to allocate will cause DTrace to fail to start:

# dtrace -P syscall -x bufsize=1g -x bufresize=manual
dtrace: description 'syscall' matched 430 probes
dtrace: could not enable tracing: Not enough space
#

The buffer resizing policy of all buffers, principal, speculative and aggregation, is dictated by the bufresize option.

11.6. Buffer Ordering Policy

DTrace consumes its principal buffers on a per-CPU basis. This causes output to be ordered first by the order that it retrieved buffers from the CPUs and secondly by the ordering within each principal buffer. Look at the output of the following script:

syscall:::entry
{
	trace(timestamp);
}


CPU     ID                    FUNCTION:NAME
 23     24                      close:entry  3302220933052713
 23     24                      close:entry  3302220933064286
 23     24                      close:entry  3302220933066326
 23     16                      rexit:entry  3302220933111500
  1     20                      write:entry  3302220705802875
  1     20                      write:entry  3302220705807694
  1     20                      write:entry  3302220705812112
  1    106                      ioctl:entry  3302220705815463

Notice how the timestamps are not in the order that you might expect. All of the events on CPU23 are ordered and all the events on CPU 1 are ordered, however there is no total ordering based on time.

To instead order this based on time, one would use the temporal option. This can be controlled on a per-consumer basis.