4. IOAT Rawdev Driver

The ioat rawdev driver provides a poll-mode driver (PMD) for Intel® Data Streaming Accelerator (Intel DSA) and for Intel® QuickData Technology, part of Intel® I/O Acceleration Technology (Intel I/OAT). This PMD, when used on supported hardware, allows data copies, for example, cloning packet data, to be accelerated by that hardware rather than having to be done by software, freeing up CPU cycles for other tasks.

4.1. Hardware Requirements

The dpdk-devbind.py script, included with DPDK, can be used to show the presence of supported hardware. Running dpdk-devbind.py --status-dev misc will show all the miscellaneous, or rawdev-based devices on the system. For Intel® QuickData Technology devices, the hardware will be often listed as “Crystal Beach DMA”, or “CBDMA”. For Intel® DSA devices, they are currently (at time of writing) appearing as devices with type “0b25”, due to the absence of pci-id database entries for them at this point.

4.2. Compilation

For builds using meson and ninja, the driver will be built when the target platform is x86-based. No additional compilation steps are necessary.

4.3. Device Setup

Depending on support provided by the PMD, HW devices can either use the kernel configured driver or be bound to a user-space IO driver for use. For example, Intel® DSA devices can use the IDXD kernel driver or DPDK-supported drivers, such as vfio-pci.

4.3.1. Intel® DSA devices using idxd kernel driver

To use a Intel® DSA device bound to the IDXD kernel driver, the device must first be configured. The accel-config utility library can be used for configuration.

Note

The device configuration can also be done by directly interacting with the sysfs nodes. An example of how this may be done can be seen in the script dpdk_idxd_cfg.py included in the driver source directory.

There are some mandatory configuration steps before being able to use a device with an application. The internal engines, which do the copies or other operations, and the work-queues, which are used by applications to assign work to the device, need to be assigned to groups, and the various other configuration options, such as priority or queue depth, need to be set for each queue.

To assign an engine to a group:

$ accel-config config-engine dsa0/engine0.0 --group-id=0
$ accel-config config-engine dsa0/engine0.1 --group-id=1

To assign work queues to groups for passing descriptors to the engines a similar accel-config command can be used. However, the work queues also need to be configured depending on the use case. Some configuration options include:

  • mode (Dedicated/Shared): Indicates whether a WQ may accept jobs from multiple queues simultaneously.
  • priority: WQ priority between 1 and 15. Larger value means higher priority.
  • wq-size: the size of the WQ. Sum of all WQ sizes must be less that the total-size defined by the device.
  • type: WQ type (kernel/mdev/user). Determines how the device is presented.
  • name: identifier given to the WQ.

Example configuration for a work queue:

$ accel-config config-wq dsa0/wq0.0 --group-id=0 \
   --mode=dedicated --priority=10 --wq-size=8 \
   --type=user --name=app1

Once the devices have been configured, they need to be enabled:

$ accel-config enable-device dsa0
$ accel-config enable-wq dsa0/wq0.0

Check the device configuration:

$ accel-config list

4.3.2. Devices using VFIO/UIO drivers

The HW devices to be used will need to be bound to a user-space IO driver for use. The dpdk-devbind.py script can be used to view the state of the devices and to bind them to a suitable DPDK-supported driver, such as vfio-pci. For example:

$ dpdk-devbind.py -b vfio-pci 00:04.0 00:04.1

4.3.3. Device Probing and Initialization

For devices bound to a suitable DPDK-supported VFIO/UIO driver, the HW devices will be found as part of the device scan done at application initialization time without the need to pass parameters to the application.

If the device is bound to the IDXD kernel driver (and previously configured with sysfs), then a specific work queue needs to be passed to the application via a vdev parameter. This vdev parameter take the driver name and work queue name as parameters. For example, to use work queue 0 on Intel® DSA instance 0:

$ dpdk-test --no-pci --vdev=rawdev_idxd,wq=0.0

Once probed successfully, the device will appear as a rawdev, that is a “raw device type” inside DPDK, and can be accessed using APIs from the rte_rawdev library.

4.4. Using IOAT Rawdev Devices

To use the devices from an application, the rawdev API can be used, along with definitions taken from the device-specific header file rte_ioat_rawdev.h. This header is needed to get the definition of structure parameters used by some of the rawdev APIs for IOAT rawdev devices, as well as providing key functions for using the device for memory copies.

4.4.1. Getting Device Information

Basic information about each rawdev device can be queried using the rte_rawdev_info_get() API. For most applications, this API will be needed to verify that the rawdev in question is of the expected type. For example, the following code snippet can be used to identify an IOAT rawdev device for use by an application:

for (i = 0; i < count && !found; i++) {
        struct rte_rawdev_info info = { .dev_private = NULL };
        found = (rte_rawdev_info_get(i, &info, 0) == 0 &&
                        strcmp(info.driver_name,
                                        IOAT_PMD_RAWDEV_NAME_STR) == 0);
}

When calling the rte_rawdev_info_get() API for an IOAT rawdev device, the dev_private field in the rte_rawdev_info struct should either be NULL, or else be set to point to a structure of type rte_ioat_rawdev_config, in which case the size of the configured device input ring will be returned in that structure.

4.4.2. Device Configuration

Configuring an IOAT rawdev device is done using the rte_rawdev_configure() API, which takes the same structure parameters as the, previously referenced, rte_rawdev_info_get() API. The main difference is that, because the parameter is used as input rather than output, the dev_private structure element cannot be NULL, and must point to a valid rte_ioat_rawdev_config structure, containing the ring size to be used by the device. The ring size must be a power of two, between 64 and 4096. If it is not needed, the tracking by the driver of user-provided completion handles may be disabled by setting the hdls_disable flag in the configuration structure also.

The following code shows how the device is configured in test_ioat_rawdev.c:

#define IOAT_TEST_RINGSIZE 512
     struct rte_ioat_rawdev_config p = { .ring_size = -1 };
     struct rte_rawdev_info info = { .dev_private = &p };

     /* ... */

     p.ring_size = IOAT_TEST_RINGSIZE;
     if (rte_rawdev_configure(dev_id, &info, sizeof(p)) != 0) {
             printf("Error with rte_rawdev_configure()\n");
             return -1;
     }

Once configured, the device can then be made ready for use by calling the rte_rawdev_start() API.

4.4.3. Performing Data Copies

To perform data copies using IOAT rawdev devices, the functions rte_ioat_enqueue_copy() and rte_ioat_perform_ops() should be used. Once copies have been completed, the completion will be reported back when the application calls rte_ioat_completed_ops().

The rte_ioat_enqueue_copy() function enqueues a single copy to the device ring for copying at a later point. The parameters to that function include the IOVA addresses of both the source and destination buffers, as well as two “handles” to be returned to the user when the copy is completed. These handles can be arbitrary values, but two are provided so that the library can track handles for both source and destination on behalf of the user, e.g. virtual addresses for the buffers, or mbuf pointers if packet data is being copied.

While the rte_ioat_enqueue_copy() function enqueues a copy operation on the device ring, the copy will not actually be performed until after the application calls the rte_ioat_perform_ops() function. This function informs the device hardware of the elements enqueued on the ring, and the device will begin to process them. It is expected that, for efficiency reasons, a burst of operations will be enqueued to the device via multiple enqueue calls between calls to the rte_ioat_perform_ops() function.

The following code from test_ioat_rawdev.c demonstrates how to enqueue a burst of copies to the device and start the hardware processing of them:

struct rte_mbuf *srcs[32], *dsts[32];
unsigned int j;

for (i = 0; i < RTE_DIM(srcs); i++) {
        char *src_data;

        srcs[i] = rte_pktmbuf_alloc(pool);
        dsts[i] = rte_pktmbuf_alloc(pool);
        srcs[i]->data_len = srcs[i]->pkt_len = length;
        dsts[i]->data_len = dsts[i]->pkt_len = length;
        src_data = rte_pktmbuf_mtod(srcs[i], char *);

        for (j = 0; j < length; j++)
                src_data[j] = rand() & 0xFF;

        if (rte_ioat_enqueue_copy(dev_id,
                        srcs[i]->buf_iova + srcs[i]->data_off,
                        dsts[i]->buf_iova + dsts[i]->data_off,
                        length,
                        (uintptr_t)srcs[i],
                        (uintptr_t)dsts[i]) != 1) {
                printf("Error with rte_ioat_enqueue_copy for buffer %u\n",
                                i);
                return -1;
        }
}
rte_ioat_perform_ops(dev_id);

To retrieve information about completed copies, the API rte_ioat_completed_ops() should be used. This API will return to the application a set of completion handles passed in when the relevant copies were enqueued.

The following code from test_ioat_rawdev.c shows the test code retrieving information about the completed copies and validating the data is correct before freeing the data buffers using the returned handles:

if (rte_ioat_completed_ops(dev_id, 64, (void *)completed_src,
                (void *)completed_dst) != RTE_DIM(srcs)) {
        printf("Error with rte_ioat_completed_ops\n");
        return -1;
}
for (i = 0; i < RTE_DIM(srcs); i++) {
        char *src_data, *dst_data;

        if (completed_src[i] != srcs[i]) {
                printf("Error with source pointer %u\n", i);
                return -1;
        }
        if (completed_dst[i] != dsts[i]) {
                printf("Error with dest pointer %u\n", i);
                return -1;
        }

        src_data = rte_pktmbuf_mtod(srcs[i], char *);
        dst_data = rte_pktmbuf_mtod(dsts[i], char *);
        for (j = 0; j < length; j++)
                if (src_data[j] != dst_data[j]) {
                        printf("Error with copy of packet %u, byte %u\n",
                                        i, j);
                        return -1;
                }
        rte_pktmbuf_free(srcs[i]);
        rte_pktmbuf_free(dsts[i]);
}

4.4.4. Filling an Area of Memory

The IOAT driver also has support for the fill operation, where an area of memory is overwritten, or filled, with a short pattern of data. Fill operations can be performed in much the same was as copy operations described above, just using the rte_ioat_enqueue_fill() function rather than the rte_ioat_enqueue_copy() function.

4.4.5. Querying Device Statistics

The statistics from the IOAT rawdev device can be got via the xstats functions in the rte_rawdev library, i.e. rte_rawdev_xstats_names_get(), rte_rawdev_xstats_get() and rte_rawdev_xstats_by_name_get. The statistics returned for each device instance are:

  • failed_enqueues
  • successful_enqueues
  • copies_started
  • copies_completed