1 Coherent Accelerator Interface (CXL)
2 ====================================
7 The coherent accelerator interface is designed to allow the
8 coherent connection of accelerators (FPGAs and other devices) to a
9 POWER system. These devices need to adhere to the Coherent
10 Accelerator Interface Architecture (CAIA).
12 IBM refers to this as the Coherent Accelerator Processor Interface
13 or CAPI. In the kernel it's referred to by the name CXL to avoid
14 confusion with the ISDN CAPI subsystem.
16 Coherent in this context means that the accelerator and CPUs can
17 both access system memory directly and with the same effective
25 +----------+ +---------+
31 +----------+ +---------+
35 +---+------+ PCIE +---------+
37 The POWER8 chip has a Coherently Attached Processor Proxy (CAPP)
38 unit which is part of the PCIe Host Bridge (PHB). This is managed
39 by Linux by calls into OPAL. Linux doesn't directly program the
42 The FPGA (or coherently attached device) consists of two parts.
43 The POWER Service Layer (PSL) and the Accelerator Function Unit
44 (AFU). The AFU is used to implement specific functionality behind
45 the PSL. The PSL, among other things, provides memory address
46 translation services to allow each AFU direct access to userspace
49 The AFU is the core part of the accelerator (eg. the compression,
50 crypto etc function). The kernel has no knowledge of the function
51 of the AFU. Only userspace interacts directly with the AFU.
53 The PSL provides the translation and interrupt services that the
54 AFU needs. This is what the kernel interacts with. For example, if
55 the AFU needs to read a particular effective address, it sends
56 that address to the PSL, the PSL then translates it, fetches the
57 data from memory and returns it to the AFU. If the PSL has a
58 translation miss, it interrupts the kernel and the kernel services
59 the fault. The context to which this fault is serviced is based on
60 who owns that acceleration function.
66 There are two programming modes supported by the AFU. Dedicated
67 and AFU directed. AFU may support one or both modes.
69 When using dedicated mode only one MMU context is supported. In
70 this mode, only one userspace process can use the accelerator at
73 When using AFU directed mode, up to 16K simultaneous contexts can
74 be supported. This means up to 16K simultaneous userspace
75 applications may use the accelerator (although specific AFUs may
76 support fewer). In this mode, the AFU sends a 16 bit context ID
77 with each of its requests. This tells the PSL which context is
78 associated with each operation. If the PSL can't translate an
79 operation, the ID can also be accessed by the kernel so it can
80 determine the userspace context associated with an operation.
86 A portion of the accelerator MMIO space can be directly mapped
87 from the AFU to userspace. Either the whole space can be mapped or
88 just a per context portion. The hardware is self describing, hence
89 the kernel can determine the offset and size of the per context
96 AFUs may generate interrupts that are destined for userspace. These
97 are received by the kernel as hardware interrupts and passed onto
98 userspace by a read syscall documented below.
100 Data storage faults and error interrupts are handled by the kernel
104 Work Element Descriptor (WED)
105 =============================
107 The WED is a 64-bit parameter passed to the AFU when a context is
108 started. Its format is up to the AFU hence the kernel has no
109 knowledge of what it represents. Typically it will be the
110 effective address of a work queue or status block where the AFU
111 and userspace can share control and status information.
119 1. AFU character devices
121 For AFUs operating in AFU directed mode, two character device
122 files will be created. /dev/cxl/afu0.0m will correspond to a
123 master context and /dev/cxl/afu0.0s will correspond to a slave
124 context. Master contexts have access to the full MMIO space an
125 AFU provides. Slave contexts have access to only the per process
126 MMIO space an AFU provides.
128 For AFUs operating in dedicated process mode, the driver will
129 only create a single character device per AFU called
130 /dev/cxl/afu0.0d. This will have access to the entire MMIO space
131 that the AFU provides (like master contexts in AFU directed).
133 The types described below are defined in include/uapi/misc/cxl.h
135 The following file operations are supported on both slave and
138 A userspace library libcxl is available here:
139 https://github.com/ibm-capi/libcxl
140 This provides a C interface to this kernel API.
145 Opens the device and allocates a file descriptor to be used with
148 A dedicated mode AFU only has one context and only allows the
149 device to be opened once.
151 An AFU directed mode AFU can have many contexts, the device can be
152 opened once for each context that is available.
154 When all available contexts are allocated the open call will fail
157 Note: IRQs need to be allocated for each context, which may limit
158 the number of contexts that can be created, and therefore
159 how many times the device can be opened. The POWER8 CAPP
160 supports 2040 IRQs and 3 are used by the kernel, so 2037 are
161 left. If 1 IRQ is needed per context, then only 2037
162 contexts can be allocated. If 4 IRQs are needed per context,
163 then only 2037/4 = 509 contexts can be allocated.
169 CXL_IOCTL_START_WORK:
170 Starts the AFU context and associates it with the current
171 process. Once this ioctl is successfully executed, all memory
172 mapped into this process is accessible to this AFU context
173 using the same effective addresses. No additional calls are
174 required to map/unmap memory. The AFU memory context will be
175 updated as userspace allocates and frees memory. This ioctl
176 returns once the AFU context is started.
178 Takes a pointer to a struct cxl_ioctl_start_work:
180 struct cxl_ioctl_start_work {
182 __u64 work_element_descriptor;
184 __s16 num_interrupts;
194 Indicates which optional fields in the structure are
197 work_element_descriptor:
198 The Work Element Descriptor (WED) is a 64-bit argument
199 defined by the AFU. Typically this is an effective
200 address pointing to an AFU specific structure
201 describing what work to perform.
204 Authority Mask Register (AMR), same as the powerpc
205 AMR. This field is only used by the kernel when the
206 corresponding CXL_START_WORK_AMR value is specified in
207 flags. If not specified the kernel will use a default
211 Number of userspace interrupts to request. This field
212 is only used by the kernel when the corresponding
213 CXL_START_WORK_NUM_IRQS value is specified in flags.
214 If not specified the minimum number required by the
215 AFU will be allocated. The min and max number can be
219 For ABI padding and future extensions
221 CXL_IOCTL_GET_PROCESS_ELEMENT:
222 Get the current context id, also known as the process element.
223 The value is returned from the kernel as a __u32.
229 An AFU may have an MMIO space to facilitate communication with the
230 AFU. If it does, the MMIO space can be accessed via mmap. The size
231 and contents of this area are specific to the particular AFU. The
232 size can be discovered via sysfs.
234 In AFU directed mode, master contexts are allowed to map all of
235 the MMIO space and slave contexts are allowed to only map the per
236 process MMIO space associated with the context. In dedicated
237 process mode the entire MMIO space can always be mapped.
239 This mmap call must be done after the START_WORK ioctl.
241 Care should be taken when accessing MMIO space. Only 32 and 64-bit
242 accesses are supported by POWER8. Also, the AFU will be designed
243 with a specific endianness, so all MMIO accesses should consider
244 endianness (recommend endian(3) variants like: le64toh(),
245 be64toh() etc). These endian issues equally apply to shared memory
246 queues the WED may describe.
252 Reads events from the AFU. Blocks if no events are pending
253 (unless O_NONBLOCK is supplied). Returns -EIO in the case of an
254 unrecoverable error or if the card is removed.
256 read() will always return an integral number of events.
258 The buffer passed to read() must be at least 4K bytes.
260 The result of the read will be a buffer of one or more events,
261 each event is of type struct cxl_event, of varying size.
264 struct cxl_event_header header;
266 struct cxl_event_afu_interrupt irq;
267 struct cxl_event_data_storage fault;
268 struct cxl_event_afu_error afu_error;
272 The struct cxl_event_header is defined as:
274 struct cxl_event_header {
277 __u16 process_element;
282 This defines the type of event. The type determines how
283 the rest of the event is structured. These types are
284 described below and defined by enum cxl_event_type.
287 This is the size of the event in bytes including the
288 struct cxl_event_header. The start of the next event can
289 be found at this offset from the start of the current
293 Context ID of the event.
296 For future extensions and padding.
298 If the event type is CXL_EVENT_AFU_INTERRUPT then the event
299 structure is defined as:
301 struct cxl_event_afu_interrupt {
303 __u16 irq; /* Raised AFU interrupt number */
308 These flags indicate which optional fields are present
309 in this struct. Currently all fields are mandatory.
312 The IRQ number sent by the AFU.
315 For future extensions and padding.
317 If the event type is CXL_EVENT_DATA_STORAGE then the event
318 structure is defined as:
320 struct cxl_event_data_storage {
330 These flags indicate which optional fields are present in
331 this struct. Currently all fields are mandatory.
334 The address that the AFU unsuccessfully attempted to
335 access. Valid accesses will be handled transparently by the
336 kernel but invalid accesses will generate this event.
339 This field gives information on the type of fault. It is a
340 copy of the DSISR from the PSL hardware when the address
341 fault occurred. The form of the DSISR is as defined in the
345 For future extensions
347 If the event type is CXL_EVENT_AFU_ERROR then the event structure
350 struct cxl_event_afu_error {
358 These flags indicate which optional fields are present in
359 this struct. Currently all fields are Mandatory.
362 Error status from the AFU. Defined by the AFU.
365 For future extensions and padding
368 2. Card character device (powerVM guest only)
370 In a powerVM guest, an extra character device is created for the
371 card. The device is only used to write (flash) a new image on the
372 FPGA accelerator. Once the image is written and verified, the
373 device tree is updated and the card is reset to reload the updated
379 Opens the device and allocates a file descriptor to be used with
380 the rest of the API. The device can only be opened once.
385 CXL_IOCTL_DOWNLOAD_IMAGE:
386 CXL_IOCTL_VALIDATE_IMAGE:
387 Starts and controls flashing a new FPGA image. Partial
388 reconfiguration is not supported (yet), so the image must contain
389 a copy of the PSL and AFU(s). Since an image can be quite large,
390 the caller may have to iterate, splitting the image in smaller
393 Takes a pointer to a struct cxl_adapter_image:
394 struct cxl_adapter_image {
406 These flags indicate which optional fields are present in
407 this struct. Currently all fields are mandatory.
410 Pointer to a buffer with part of the image to write to the
414 Size of the buffer pointed to by data.
417 Full size of the image.
423 A cxl sysfs class is added under /sys/class/cxl to facilitate
424 enumeration and tuning of the accelerators. Its layout is
425 described in Documentation/ABI/testing/sysfs-class-cxl
431 The following udev rules could be used to create a symlink to the
432 most logical chardev to use in any programming mode (afuX.Yd for
433 dedicated, afuX.Ys for afu directed), since the API is virtually
436 SUBSYSTEM=="cxl", ATTRS{mode}=="dedicated_process", SYMLINK="cxl/%b"
437 SUBSYSTEM=="cxl", ATTRS{mode}=="afu_directed", \
438 KERNEL=="afu[0-9]*.[0-9]*s", SYMLINK="cxl/%b"