tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
1====================================
2Coherent Accelerator Interface (CXL)
3====================================
4
5Introduction
6============
7
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).
12
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.
16
17 Coherent in this context means that the accelerator and CPUs can
18 both access system memory directly and with the same effective
19 addresses.
20
21 **This driver is deprecated and will be removed in a future release.**
22
23Hardware overview
24=================
25
26 ::
27
28 POWER8/9 FPGA
29 +----------+ +---------+
30 | | | |
31 | CPU | | AFU |
32 | | | |
33 | | | |
34 | | | |
35 +----------+ +---------+
36 | PHB | | |
37 | +------+ | PSL |
38 | | CAPP |<------>| |
39 +---+------+ PCIE +---------+
40
41 The POWER8/9 chip has a Coherently Attached Processor Proxy (CAPP)
42 unit which is part of the PCIe Host Bridge (PHB). This is managed
43 by Linux by calls into OPAL. Linux doesn't directly program the
44 CAPP.
45
46 The FPGA (or coherently attached device) consists of two parts.
47 The POWER Service Layer (PSL) and the Accelerator Function Unit
48 (AFU). The AFU is used to implement specific functionality behind
49 the PSL. The PSL, among other things, provides memory address
50 translation services to allow each AFU direct access to userspace
51 memory.
52
53 The AFU is the core part of the accelerator (eg. the compression,
54 crypto etc function). The kernel has no knowledge of the function
55 of the AFU. Only userspace interacts directly with the AFU.
56
57 The PSL provides the translation and interrupt services that the
58 AFU needs. This is what the kernel interacts with. For example, if
59 the AFU needs to read a particular effective address, it sends
60 that address to the PSL, the PSL then translates it, fetches the
61 data from memory and returns it to the AFU. If the PSL has a
62 translation miss, it interrupts the kernel and the kernel services
63 the fault. The context to which this fault is serviced is based on
64 who owns that acceleration function.
65
66 - POWER8 and PSL Version 8 are compliant to the CAIA Version 1.0.
67 - POWER9 and PSL Version 9 are compliant to the CAIA Version 2.0.
68
69 This PSL Version 9 provides new features such as:
70
71 * Interaction with the nest MMU on the P9 chip.
72 * Native DMA support.
73 * Supports sending ASB_Notify messages for host thread wakeup.
74 * Supports Atomic operations.
75 * etc.
76
77 Cards with a PSL9 won't work on a POWER8 system and cards with a
78 PSL8 won't work on a POWER9 system.
79
80AFU Modes
81=========
82
83 There are two programming modes supported by the AFU. Dedicated
84 and AFU directed. AFU may support one or both modes.
85
86 When using dedicated mode only one MMU context is supported. In
87 this mode, only one userspace process can use the accelerator at
88 time.
89
90 When using AFU directed mode, up to 16K simultaneous contexts can
91 be supported. This means up to 16K simultaneous userspace
92 applications may use the accelerator (although specific AFUs may
93 support fewer). In this mode, the AFU sends a 16 bit context ID
94 with each of its requests. This tells the PSL which context is
95 associated with each operation. If the PSL can't translate an
96 operation, the ID can also be accessed by the kernel so it can
97 determine the userspace context associated with an operation.
98
99
100MMIO space
101==========
102
103 A portion of the accelerator MMIO space can be directly mapped
104 from the AFU to userspace. Either the whole space can be mapped or
105 just a per context portion. The hardware is self describing, hence
106 the kernel can determine the offset and size of the per context
107 portion.
108
109
110Interrupts
111==========
112
113 AFUs may generate interrupts that are destined for userspace. These
114 are received by the kernel as hardware interrupts and passed onto
115 userspace by a read syscall documented below.
116
117 Data storage faults and error interrupts are handled by the kernel
118 driver.
119
120
121Work Element Descriptor (WED)
122=============================
123
124 The WED is a 64-bit parameter passed to the AFU when a context is
125 started. Its format is up to the AFU hence the kernel has no
126 knowledge of what it represents. Typically it will be the
127 effective address of a work queue or status block where the AFU
128 and userspace can share control and status information.
129
130
131
132
133User API
134========
135
1361. AFU character devices
137^^^^^^^^^^^^^^^^^^^^^^^^
138
139 For AFUs operating in AFU directed mode, two character device
140 files will be created. /dev/cxl/afu0.0m will correspond to a
141 master context and /dev/cxl/afu0.0s will correspond to a slave
142 context. Master contexts have access to the full MMIO space an
143 AFU provides. Slave contexts have access to only the per process
144 MMIO space an AFU provides.
145
146 For AFUs operating in dedicated process mode, the driver will
147 only create a single character device per AFU called
148 /dev/cxl/afu0.0d. This will have access to the entire MMIO space
149 that the AFU provides (like master contexts in AFU directed).
150
151 The types described below are defined in include/uapi/misc/cxl.h
152
153 The following file operations are supported on both slave and
154 master devices.
155
156 A userspace library libcxl is available here:
157
158 https://github.com/ibm-capi/libcxl
159
160 This provides a C interface to this kernel API.
161
162open
163----
164
165 Opens the device and allocates a file descriptor to be used with
166 the rest of the API.
167
168 A dedicated mode AFU only has one context and only allows the
169 device to be opened once.
170
171 An AFU directed mode AFU can have many contexts, the device can be
172 opened once for each context that is available.
173
174 When all available contexts are allocated the open call will fail
175 and return -ENOSPC.
176
177 Note:
178 IRQs need to be allocated for each context, which may limit
179 the number of contexts that can be created, and therefore
180 how many times the device can be opened. The POWER8 CAPP
181 supports 2040 IRQs and 3 are used by the kernel, so 2037 are
182 left. If 1 IRQ is needed per context, then only 2037
183 contexts can be allocated. If 4 IRQs are needed per context,
184 then only 2037/4 = 509 contexts can be allocated.
185
186
187ioctl
188-----
189
190 CXL_IOCTL_START_WORK:
191 Starts the AFU context and associates it with the current
192 process. Once this ioctl is successfully executed, all memory
193 mapped into this process is accessible to this AFU context
194 using the same effective addresses. No additional calls are
195 required to map/unmap memory. The AFU memory context will be
196 updated as userspace allocates and frees memory. This ioctl
197 returns once the AFU context is started.
198
199 Takes a pointer to a struct cxl_ioctl_start_work
200
201 ::
202
203 struct cxl_ioctl_start_work {
204 __u64 flags;
205 __u64 work_element_descriptor;
206 __u64 amr;
207 __s16 num_interrupts;
208 __s16 reserved1;
209 __s32 reserved2;
210 __u64 reserved3;
211 __u64 reserved4;
212 __u64 reserved5;
213 __u64 reserved6;
214 };
215
216 flags:
217 Indicates which optional fields in the structure are
218 valid.
219
220 work_element_descriptor:
221 The Work Element Descriptor (WED) is a 64-bit argument
222 defined by the AFU. Typically this is an effective
223 address pointing to an AFU specific structure
224 describing what work to perform.
225
226 amr:
227 Authority Mask Register (AMR), same as the powerpc
228 AMR. This field is only used by the kernel when the
229 corresponding CXL_START_WORK_AMR value is specified in
230 flags. If not specified the kernel will use a default
231 value of 0.
232
233 num_interrupts:
234 Number of userspace interrupts to request. This field
235 is only used by the kernel when the corresponding
236 CXL_START_WORK_NUM_IRQS value is specified in flags.
237 If not specified the minimum number required by the
238 AFU will be allocated. The min and max number can be
239 obtained from sysfs.
240
241 reserved fields:
242 For ABI padding and future extensions
243
244 CXL_IOCTL_GET_PROCESS_ELEMENT:
245 Get the current context id, also known as the process element.
246 The value is returned from the kernel as a __u32.
247
248
249mmap
250----
251
252 An AFU may have an MMIO space to facilitate communication with the
253 AFU. If it does, the MMIO space can be accessed via mmap. The size
254 and contents of this area are specific to the particular AFU. The
255 size can be discovered via sysfs.
256
257 In AFU directed mode, master contexts are allowed to map all of
258 the MMIO space and slave contexts are allowed to only map the per
259 process MMIO space associated with the context. In dedicated
260 process mode the entire MMIO space can always be mapped.
261
262 This mmap call must be done after the START_WORK ioctl.
263
264 Care should be taken when accessing MMIO space. Only 32 and 64-bit
265 accesses are supported by POWER8. Also, the AFU will be designed
266 with a specific endianness, so all MMIO accesses should consider
267 endianness (recommend endian(3) variants like: le64toh(),
268 be64toh() etc). These endian issues equally apply to shared memory
269 queues the WED may describe.
270
271
272read
273----
274
275 Reads events from the AFU. Blocks if no events are pending
276 (unless O_NONBLOCK is supplied). Returns -EIO in the case of an
277 unrecoverable error or if the card is removed.
278
279 read() will always return an integral number of events.
280
281 The buffer passed to read() must be at least 4K bytes.
282
283 The result of the read will be a buffer of one or more events,
284 each event is of type struct cxl_event, of varying size::
285
286 struct cxl_event {
287 struct cxl_event_header header;
288 union {
289 struct cxl_event_afu_interrupt irq;
290 struct cxl_event_data_storage fault;
291 struct cxl_event_afu_error afu_error;
292 };
293 };
294
295 The struct cxl_event_header is defined as
296
297 ::
298
299 struct cxl_event_header {
300 __u16 type;
301 __u16 size;
302 __u16 process_element;
303 __u16 reserved1;
304 };
305
306 type:
307 This defines the type of event. The type determines how
308 the rest of the event is structured. These types are
309 described below and defined by enum cxl_event_type.
310
311 size:
312 This is the size of the event in bytes including the
313 struct cxl_event_header. The start of the next event can
314 be found at this offset from the start of the current
315 event.
316
317 process_element:
318 Context ID of the event.
319
320 reserved field:
321 For future extensions and padding.
322
323 If the event type is CXL_EVENT_AFU_INTERRUPT then the event
324 structure is defined as
325
326 ::
327
328 struct cxl_event_afu_interrupt {
329 __u16 flags;
330 __u16 irq; /* Raised AFU interrupt number */
331 __u32 reserved1;
332 };
333
334 flags:
335 These flags indicate which optional fields are present
336 in this struct. Currently all fields are mandatory.
337
338 irq:
339 The IRQ number sent by the AFU.
340
341 reserved field:
342 For future extensions and padding.
343
344 If the event type is CXL_EVENT_DATA_STORAGE then the event
345 structure is defined as
346
347 ::
348
349 struct cxl_event_data_storage {
350 __u16 flags;
351 __u16 reserved1;
352 __u32 reserved2;
353 __u64 addr;
354 __u64 dsisr;
355 __u64 reserved3;
356 };
357
358 flags:
359 These flags indicate which optional fields are present in
360 this struct. Currently all fields are mandatory.
361
362 address:
363 The address that the AFU unsuccessfully attempted to
364 access. Valid accesses will be handled transparently by the
365 kernel but invalid accesses will generate this event.
366
367 dsisr:
368 This field gives information on the type of fault. It is a
369 copy of the DSISR from the PSL hardware when the address
370 fault occurred. The form of the DSISR is as defined in the
371 CAIA.
372
373 reserved fields:
374 For future extensions
375
376 If the event type is CXL_EVENT_AFU_ERROR then the event structure
377 is defined as
378
379 ::
380
381 struct cxl_event_afu_error {
382 __u16 flags;
383 __u16 reserved1;
384 __u32 reserved2;
385 __u64 error;
386 };
387
388 flags:
389 These flags indicate which optional fields are present in
390 this struct. Currently all fields are Mandatory.
391
392 error:
393 Error status from the AFU. Defined by the AFU.
394
395 reserved fields:
396 For future extensions and padding
397
398
3992. Card character device (powerVM guest only)
400^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
401
402 In a powerVM guest, an extra character device is created for the
403 card. The device is only used to write (flash) a new image on the
404 FPGA accelerator. Once the image is written and verified, the
405 device tree is updated and the card is reset to reload the updated
406 image.
407
408open
409----
410
411 Opens the device and allocates a file descriptor to be used with
412 the rest of the API. The device can only be opened once.
413
414ioctl
415-----
416
417CXL_IOCTL_DOWNLOAD_IMAGE / CXL_IOCTL_VALIDATE_IMAGE:
418 Starts and controls flashing a new FPGA image. Partial
419 reconfiguration is not supported (yet), so the image must contain
420 a copy of the PSL and AFU(s). Since an image can be quite large,
421 the caller may have to iterate, splitting the image in smaller
422 chunks.
423
424 Takes a pointer to a struct cxl_adapter_image::
425
426 struct cxl_adapter_image {
427 __u64 flags;
428 __u64 data;
429 __u64 len_data;
430 __u64 len_image;
431 __u64 reserved1;
432 __u64 reserved2;
433 __u64 reserved3;
434 __u64 reserved4;
435 };
436
437 flags:
438 These flags indicate which optional fields are present in
439 this struct. Currently all fields are mandatory.
440
441 data:
442 Pointer to a buffer with part of the image to write to the
443 card.
444
445 len_data:
446 Size of the buffer pointed to by data.
447
448 len_image:
449 Full size of the image.
450
451
452Sysfs Class
453===========
454
455 A cxl sysfs class is added under /sys/class/cxl to facilitate
456 enumeration and tuning of the accelerators. Its layout is
457 described in Documentation/ABI/obsolete/sysfs-class-cxl
458
459
460Udev rules
461==========
462
463 The following udev rules could be used to create a symlink to the
464 most logical chardev to use in any programming mode (afuX.Yd for
465 dedicated, afuX.Ys for afu directed), since the API is virtually
466 identical for each::
467
468 SUBSYSTEM=="cxl", ATTRS{mode}=="dedicated_process", SYMLINK="cxl/%b"
469 SUBSYSTEM=="cxl", ATTRS{mode}=="afu_directed", \
470 KERNEL=="afu[0-9]*.[0-9]*s", SYMLINK="cxl/%b"