1 ====================================
2 Coherent Accelerator Interface (CXL)
3 ====================================
8 The coherent accelerator interface is designed to allow the
9 coherent connection of accelerators (FPGAs and other devices) to a
10 POWER system. These devices need to adhere to the Coherent
11 Accelerator Interface Architecture (CAIA).
13 IBM refers to this as the Coherent Accelerator Processor Interface
14 or CAPI. In the kernel it's referred to by the name CXL to avoid
15 confusion with the ISDN CAPI subsystem.
17 Coherent in this context means that the accelerator and CPUs can
18 both access system memory directly and with the same effective
28 +----------+ +---------+
34 +----------+ +---------+
38 +---+------+ PCIE +---------+
40 The POWER8/9 chip has a Coherently Attached Processor Proxy (CAPP)
41 unit which is part of the PCIe Host Bridge (PHB). This is managed
42 by Linux by calls into OPAL. Linux doesn't directly program the
45 The FPGA (or coherently attached device) consists of two parts.
46 The POWER Service Layer (PSL) and the Accelerator Function Unit
47 (AFU). The AFU is used to implement specific functionality behind
48 the PSL. The PSL, among other things, provides memory address
49 translation services to allow each AFU direct access to userspace
52 The AFU is the core part of the accelerator (eg. the compression,
53 crypto etc function). The kernel has no knowledge of the function
54 of the AFU. Only userspace interacts directly with the AFU.
56 The PSL provides the translation and interrupt services that the
57 AFU needs. This is what the kernel interacts with. For example, if
58 the AFU needs to read a particular effective address, it sends
59 that address to the PSL, the PSL then translates it, fetches the
60 data from memory and returns it to the AFU. If the PSL has a
61 translation miss, it interrupts the kernel and the kernel services
62 the fault. The context to which this fault is serviced is based on
63 who owns that acceleration function.
65 - POWER8 and PSL Version 8 are compliant to the CAIA Version 1.0.
66 - POWER9 and PSL Version 9 are compliant to the CAIA Version 2.0.
68 This PSL Version 9 provides new features such as:
70 * Interaction with the nest MMU on the P9 chip.
72 * Supports sending ASB_Notify messages for host thread wakeup.
73 * Supports Atomic operations.
76 Cards with a PSL9 won't work on a POWER8 system and cards with a
77 PSL8 won't work on a POWER9 system.
82 There are two programming modes supported by the AFU. Dedicated
83 and AFU directed. AFU may support one or both modes.
85 When using dedicated mode only one MMU context is supported. In
86 this mode, only one userspace process can use the accelerator at
89 When using AFU directed mode, up to 16K simultaneous contexts can
90 be supported. This means up to 16K simultaneous userspace
91 applications may use the accelerator (although specific AFUs may
92 support fewer). In this mode, the AFU sends a 16 bit context ID
93 with each of its requests. This tells the PSL which context is
94 associated with each operation. If the PSL can't translate an
95 operation, the ID can also be accessed by the kernel so it can
96 determine the userspace context associated with an operation.
102 A portion of the accelerator MMIO space can be directly mapped
103 from the AFU to userspace. Either the whole space can be mapped or
104 just a per context portion. The hardware is self describing, hence
105 the kernel can determine the offset and size of the per context
112 AFUs may generate interrupts that are destined for userspace. These
113 are received by the kernel as hardware interrupts and passed onto
114 userspace by a read syscall documented below.
116 Data storage faults and error interrupts are handled by the kernel
120 Work Element Descriptor (WED)
121 =============================
123 The WED is a 64-bit parameter passed to the AFU when a context is
124 started. Its format is up to the AFU hence the kernel has no
125 knowledge of what it represents. Typically it will be the
126 effective address of a work queue or status block where the AFU
127 and userspace can share control and status information.
135 1. AFU character devices
137 For AFUs operating in AFU directed mode, two character device
138 files will be created. /dev/cxl/afu0.0m will correspond to a
139 master context and /dev/cxl/afu0.0s will correspond to a slave
140 context. Master contexts have access to the full MMIO space an
141 AFU provides. Slave contexts have access to only the per process
142 MMIO space an AFU provides.
144 For AFUs operating in dedicated process mode, the driver will
145 only create a single character device per AFU called
146 /dev/cxl/afu0.0d. This will have access to the entire MMIO space
147 that the AFU provides (like master contexts in AFU directed).
149 The types described below are defined in include/uapi/misc/cxl.h
151 The following file operations are supported on both slave and
154 A userspace library libcxl is available here:
156 https://github.com/ibm-capi/libcxl
158 This provides a C interface to this kernel API.
163 Opens the device and allocates a file descriptor to be used with
166 A dedicated mode AFU only has one context and only allows the
167 device to be opened once.
169 An AFU directed mode AFU can have many contexts, the device can be
170 opened once for each context that is available.
172 When all available contexts are allocated the open call will fail
176 IRQs need to be allocated for each context, which may limit
177 the number of contexts that can be created, and therefore
178 how many times the device can be opened. The POWER8 CAPP
179 supports 2040 IRQs and 3 are used by the kernel, so 2037 are
180 left. If 1 IRQ is needed per context, then only 2037
181 contexts can be allocated. If 4 IRQs are needed per context,
182 then only 2037/4 = 509 contexts can be allocated.
188 CXL_IOCTL_START_WORK:
189 Starts the AFU context and associates it with the current
190 process. Once this ioctl is successfully executed, all memory
191 mapped into this process is accessible to this AFU context
192 using the same effective addresses. No additional calls are
193 required to map/unmap memory. The AFU memory context will be
194 updated as userspace allocates and frees memory. This ioctl
195 returns once the AFU context is started.
197 Takes a pointer to a struct cxl_ioctl_start_work
201 struct cxl_ioctl_start_work {
203 __u64 work_element_descriptor;
205 __s16 num_interrupts;
215 Indicates which optional fields in the structure are
218 work_element_descriptor:
219 The Work Element Descriptor (WED) is a 64-bit argument
220 defined by the AFU. Typically this is an effective
221 address pointing to an AFU specific structure
222 describing what work to perform.
225 Authority Mask Register (AMR), same as the powerpc
226 AMR. This field is only used by the kernel when the
227 corresponding CXL_START_WORK_AMR value is specified in
228 flags. If not specified the kernel will use a default
232 Number of userspace interrupts to request. This field
233 is only used by the kernel when the corresponding
234 CXL_START_WORK_NUM_IRQS value is specified in flags.
235 If not specified the minimum number required by the
236 AFU will be allocated. The min and max number can be
240 For ABI padding and future extensions
242 CXL_IOCTL_GET_PROCESS_ELEMENT:
243 Get the current context id, also known as the process element.
244 The value is returned from the kernel as a __u32.
250 An AFU may have an MMIO space to facilitate communication with the
251 AFU. If it does, the MMIO space can be accessed via mmap. The size
252 and contents of this area are specific to the particular AFU. The
253 size can be discovered via sysfs.
255 In AFU directed mode, master contexts are allowed to map all of
256 the MMIO space and slave contexts are allowed to only map the per
257 process MMIO space associated with the context. In dedicated
258 process mode the entire MMIO space can always be mapped.
260 This mmap call must be done after the START_WORK ioctl.
262 Care should be taken when accessing MMIO space. Only 32 and 64-bit
263 accesses are supported by POWER8. Also, the AFU will be designed
264 with a specific endianness, so all MMIO accesses should consider
265 endianness (recommend endian(3) variants like: le64toh(),
266 be64toh() etc). These endian issues equally apply to shared memory
267 queues the WED may describe.
273 Reads events from the AFU. Blocks if no events are pending
274 (unless O_NONBLOCK is supplied). Returns -EIO in the case of an
275 unrecoverable error or if the card is removed.
277 read() will always return an integral number of events.
279 The buffer passed to read() must be at least 4K bytes.
281 The result of the read will be a buffer of one or more events,
282 each event is of type struct cxl_event, of varying size::
285 struct cxl_event_header header;
287 struct cxl_event_afu_interrupt irq;
288 struct cxl_event_data_storage fault;
289 struct cxl_event_afu_error afu_error;
293 The struct cxl_event_header is defined as
297 struct cxl_event_header {
300 __u16 process_element;
305 This defines the type of event. The type determines how
306 the rest of the event is structured. These types are
307 described below and defined by enum cxl_event_type.
310 This is the size of the event in bytes including the
311 struct cxl_event_header. The start of the next event can
312 be found at this offset from the start of the current
316 Context ID of the event.
319 For future extensions and padding.
321 If the event type is CXL_EVENT_AFU_INTERRUPT then the event
322 structure is defined as
326 struct cxl_event_afu_interrupt {
328 __u16 irq; /* Raised AFU interrupt number */
333 These flags indicate which optional fields are present
334 in this struct. Currently all fields are mandatory.
337 The IRQ number sent by the AFU.
340 For future extensions and padding.
342 If the event type is CXL_EVENT_DATA_STORAGE then the event
343 structure is defined as
347 struct cxl_event_data_storage {
357 These flags indicate which optional fields are present in
358 this struct. Currently all fields are mandatory.
361 The address that the AFU unsuccessfully attempted to
362 access. Valid accesses will be handled transparently by the
363 kernel but invalid accesses will generate this event.
366 This field gives information on the type of fault. It is a
367 copy of the DSISR from the PSL hardware when the address
368 fault occurred. The form of the DSISR is as defined in the
372 For future extensions
374 If the event type is CXL_EVENT_AFU_ERROR then the event structure
379 struct cxl_event_afu_error {
387 These flags indicate which optional fields are present in
388 this struct. Currently all fields are Mandatory.
391 Error status from the AFU. Defined by the AFU.
394 For future extensions and padding
397 2. Card character device (powerVM guest only)
399 In a powerVM guest, an extra character device is created for the
400 card. The device is only used to write (flash) a new image on the
401 FPGA accelerator. Once the image is written and verified, the
402 device tree is updated and the card is reset to reload the updated
408 Opens the device and allocates a file descriptor to be used with
409 the rest of the API. The device can only be opened once.
414 CXL_IOCTL_DOWNLOAD_IMAGE / CXL_IOCTL_VALIDATE_IMAGE:
415 Starts and controls flashing a new FPGA image. Partial
416 reconfiguration is not supported (yet), so the image must contain
417 a copy of the PSL and AFU(s). Since an image can be quite large,
418 the caller may have to iterate, splitting the image in smaller
421 Takes a pointer to a struct cxl_adapter_image::
423 struct cxl_adapter_image {
435 These flags indicate which optional fields are present in
436 this struct. Currently all fields are mandatory.
439 Pointer to a buffer with part of the image to write to the
443 Size of the buffer pointed to by data.
446 Full size of the image.
452 A cxl sysfs class is added under /sys/class/cxl to facilitate
453 enumeration and tuning of the accelerators. Its layout is
454 described in Documentation/ABI/testing/sysfs-class-cxl
460 The following udev rules could be used to create a symlink to the
461 most logical chardev to use in any programming mode (afuX.Yd for
462 dedicated, afuX.Ys for afu directed), since the API is virtually
465 SUBSYSTEM=="cxl", ATTRS{mode}=="dedicated_process", SYMLINK="cxl/%b"
466 SUBSYSTEM=="cxl", ATTRS{mode}=="afu_directed", \
467 KERNEL=="afu[0-9]*.[0-9]*s", SYMLINK="cxl/%b"