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 POWER8/9 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.
62 POWER8 <-----> PSL Version 8 is compliant to the CAIA Version 1.0.
63 POWER9 <-----> PSL Version 9 is compliant to the CAIA Version 2.0.
64 This PSL Version 9 provides new features such as:
65 * Interaction with the nest MMU on the P9 chip.
67 * Supports sending ASB_Notify messages for host thread wakeup.
68 * Supports Atomic operations.
71 Cards with a PSL9 won't work on a POWER8 system and cards with a
72 PSL8 won't work on a POWER9 system.
77 There are two programming modes supported by the AFU. Dedicated
78 and AFU directed. AFU may support one or both modes.
80 When using dedicated mode only one MMU context is supported. In
81 this mode, only one userspace process can use the accelerator at
84 When using AFU directed mode, up to 16K simultaneous contexts can
85 be supported. This means up to 16K simultaneous userspace
86 applications may use the accelerator (although specific AFUs may
87 support fewer). In this mode, the AFU sends a 16 bit context ID
88 with each of its requests. This tells the PSL which context is
89 associated with each operation. If the PSL can't translate an
90 operation, the ID can also be accessed by the kernel so it can
91 determine the userspace context associated with an operation.
97 A portion of the accelerator MMIO space can be directly mapped
98 from the AFU to userspace. Either the whole space can be mapped or
99 just a per context portion. The hardware is self describing, hence
100 the kernel can determine the offset and size of the per context
107 AFUs may generate interrupts that are destined for userspace. These
108 are received by the kernel as hardware interrupts and passed onto
109 userspace by a read syscall documented below.
111 Data storage faults and error interrupts are handled by the kernel
115 Work Element Descriptor (WED)
116 =============================
118 The WED is a 64-bit parameter passed to the AFU when a context is
119 started. Its format is up to the AFU hence the kernel has no
120 knowledge of what it represents. Typically it will be the
121 effective address of a work queue or status block where the AFU
122 and userspace can share control and status information.
130 1. AFU character devices
132 For AFUs operating in AFU directed mode, two character device
133 files will be created. /dev/cxl/afu0.0m will correspond to a
134 master context and /dev/cxl/afu0.0s will correspond to a slave
135 context. Master contexts have access to the full MMIO space an
136 AFU provides. Slave contexts have access to only the per process
137 MMIO space an AFU provides.
139 For AFUs operating in dedicated process mode, the driver will
140 only create a single character device per AFU called
141 /dev/cxl/afu0.0d. This will have access to the entire MMIO space
142 that the AFU provides (like master contexts in AFU directed).
144 The types described below are defined in include/uapi/misc/cxl.h
146 The following file operations are supported on both slave and
149 A userspace library libcxl is available here:
150 https://github.com/ibm-capi/libcxl
151 This provides a C interface to this kernel API.
156 Opens the device and allocates a file descriptor to be used with
159 A dedicated mode AFU only has one context and only allows the
160 device to be opened once.
162 An AFU directed mode AFU can have many contexts, the device can be
163 opened once for each context that is available.
165 When all available contexts are allocated the open call will fail
168 Note: IRQs need to be allocated for each context, which may limit
169 the number of contexts that can be created, and therefore
170 how many times the device can be opened. The POWER8 CAPP
171 supports 2040 IRQs and 3 are used by the kernel, so 2037 are
172 left. If 1 IRQ is needed per context, then only 2037
173 contexts can be allocated. If 4 IRQs are needed per context,
174 then only 2037/4 = 509 contexts can be allocated.
180 CXL_IOCTL_START_WORK:
181 Starts the AFU context and associates it with the current
182 process. Once this ioctl is successfully executed, all memory
183 mapped into this process is accessible to this AFU context
184 using the same effective addresses. No additional calls are
185 required to map/unmap memory. The AFU memory context will be
186 updated as userspace allocates and frees memory. This ioctl
187 returns once the AFU context is started.
189 Takes a pointer to a struct cxl_ioctl_start_work:
191 struct cxl_ioctl_start_work {
193 __u64 work_element_descriptor;
195 __s16 num_interrupts;
205 Indicates which optional fields in the structure are
208 work_element_descriptor:
209 The Work Element Descriptor (WED) is a 64-bit argument
210 defined by the AFU. Typically this is an effective
211 address pointing to an AFU specific structure
212 describing what work to perform.
215 Authority Mask Register (AMR), same as the powerpc
216 AMR. This field is only used by the kernel when the
217 corresponding CXL_START_WORK_AMR value is specified in
218 flags. If not specified the kernel will use a default
222 Number of userspace interrupts to request. This field
223 is only used by the kernel when the corresponding
224 CXL_START_WORK_NUM_IRQS value is specified in flags.
225 If not specified the minimum number required by the
226 AFU will be allocated. The min and max number can be
230 For ABI padding and future extensions
232 CXL_IOCTL_GET_PROCESS_ELEMENT:
233 Get the current context id, also known as the process element.
234 The value is returned from the kernel as a __u32.
240 An AFU may have an MMIO space to facilitate communication with the
241 AFU. If it does, the MMIO space can be accessed via mmap. The size
242 and contents of this area are specific to the particular AFU. The
243 size can be discovered via sysfs.
245 In AFU directed mode, master contexts are allowed to map all of
246 the MMIO space and slave contexts are allowed to only map the per
247 process MMIO space associated with the context. In dedicated
248 process mode the entire MMIO space can always be mapped.
250 This mmap call must be done after the START_WORK ioctl.
252 Care should be taken when accessing MMIO space. Only 32 and 64-bit
253 accesses are supported by POWER8. Also, the AFU will be designed
254 with a specific endianness, so all MMIO accesses should consider
255 endianness (recommend endian(3) variants like: le64toh(),
256 be64toh() etc). These endian issues equally apply to shared memory
257 queues the WED may describe.
263 Reads events from the AFU. Blocks if no events are pending
264 (unless O_NONBLOCK is supplied). Returns -EIO in the case of an
265 unrecoverable error or if the card is removed.
267 read() will always return an integral number of events.
269 The buffer passed to read() must be at least 4K bytes.
271 The result of the read will be a buffer of one or more events,
272 each event is of type struct cxl_event, of varying size.
275 struct cxl_event_header header;
277 struct cxl_event_afu_interrupt irq;
278 struct cxl_event_data_storage fault;
279 struct cxl_event_afu_error afu_error;
283 The struct cxl_event_header is defined as:
285 struct cxl_event_header {
288 __u16 process_element;
293 This defines the type of event. The type determines how
294 the rest of the event is structured. These types are
295 described below and defined by enum cxl_event_type.
298 This is the size of the event in bytes including the
299 struct cxl_event_header. The start of the next event can
300 be found at this offset from the start of the current
304 Context ID of the event.
307 For future extensions and padding.
309 If the event type is CXL_EVENT_AFU_INTERRUPT then the event
310 structure is defined as:
312 struct cxl_event_afu_interrupt {
314 __u16 irq; /* Raised AFU interrupt number */
319 These flags indicate which optional fields are present
320 in this struct. Currently all fields are mandatory.
323 The IRQ number sent by the AFU.
326 For future extensions and padding.
328 If the event type is CXL_EVENT_DATA_STORAGE then the event
329 structure is defined as:
331 struct cxl_event_data_storage {
341 These flags indicate which optional fields are present in
342 this struct. Currently all fields are mandatory.
345 The address that the AFU unsuccessfully attempted to
346 access. Valid accesses will be handled transparently by the
347 kernel but invalid accesses will generate this event.
350 This field gives information on the type of fault. It is a
351 copy of the DSISR from the PSL hardware when the address
352 fault occurred. The form of the DSISR is as defined in the
356 For future extensions
358 If the event type is CXL_EVENT_AFU_ERROR then the event structure
361 struct cxl_event_afu_error {
369 These flags indicate which optional fields are present in
370 this struct. Currently all fields are Mandatory.
373 Error status from the AFU. Defined by the AFU.
376 For future extensions and padding
379 2. Card character device (powerVM guest only)
381 In a powerVM guest, an extra character device is created for the
382 card. The device is only used to write (flash) a new image on the
383 FPGA accelerator. Once the image is written and verified, the
384 device tree is updated and the card is reset to reload the updated
390 Opens the device and allocates a file descriptor to be used with
391 the rest of the API. The device can only be opened once.
396 CXL_IOCTL_DOWNLOAD_IMAGE:
397 CXL_IOCTL_VALIDATE_IMAGE:
398 Starts and controls flashing a new FPGA image. Partial
399 reconfiguration is not supported (yet), so the image must contain
400 a copy of the PSL and AFU(s). Since an image can be quite large,
401 the caller may have to iterate, splitting the image in smaller
404 Takes a pointer to a struct cxl_adapter_image:
405 struct cxl_adapter_image {
417 These flags indicate which optional fields are present in
418 this struct. Currently all fields are mandatory.
421 Pointer to a buffer with part of the image to write to the
425 Size of the buffer pointed to by data.
428 Full size of the image.
434 A cxl sysfs class is added under /sys/class/cxl to facilitate
435 enumeration and tuning of the accelerators. Its layout is
436 described in Documentation/ABI/testing/sysfs-class-cxl
442 The following udev rules could be used to create a symlink to the
443 most logical chardev to use in any programming mode (afuX.Yd for
444 dedicated, afuX.Ys for afu directed), since the API is virtually
447 SUBSYSTEM=="cxl", ATTRS{mode}=="dedicated_process", SYMLINK="cxl/%b"
448 SUBSYSTEM=="cxl", ATTRS{mode}=="afu_directed", \
449 KERNEL=="afu[0-9]*.[0-9]*s", SYMLINK="cxl/%b"