1 SPUFS(2) Linux Programmer's Manual SPUFS(2)
6 spufs - the SPU file system
10 The SPU file system is used on PowerPC machines that implement the Cell
11 Broadband Engine Architecture in order to access Synergistic Processor
14 The file system provides a name space similar to posix shared memory or
15 message queues. Users that have write permissions on the file system
16 can use spu_create(2) to establish SPU contexts in the spufs root.
18 Every SPU context is represented by a directory containing a predefined
19 set of files. These files can be used for manipulating the state of the
20 logical SPU. Users can change permissions on those files, but not actu-
21 ally add or remove files.
26 set the user owning the mount point, the default is 0 (root).
29 set the group owning the mount point, the default is 0 (root).
33 The files in spufs mostly follow the standard behavior for regular sys-
34 tem calls like read(2) or write(2), but often support only a subset of
35 the operations supported on regular file systems. This list details the
36 supported operations and the deviations from the behaviour in the
39 All files that support the read(2) operation also support readv(2) and
40 all files that support the write(2) operation also support writev(2).
41 All files support the access(2) and stat(2) family of operations, but
42 only the st_mode, st_nlink, st_uid and st_gid fields of struct stat
43 contain reliable information.
45 All files support the chmod(2)/fchmod(2) and chown(2)/fchown(2) opera-
46 tions, but will not be able to grant permissions that contradict the
47 possible operations, e.g. read access on the wbox file.
49 The current set of files is:
53 the contents of the local storage memory of the SPU. This can be
54 accessed like a regular shared memory file and contains both code and
55 data in the address space of the SPU. The possible operations on an
58 read(2), pread(2), write(2), pwrite(2), lseek(2)
59 These operate as documented, with the exception that seek(2),
60 write(2) and pwrite(2) are not supported beyond the end of the
61 file. The file size is the size of the local storage of the SPU,
62 which normally is 256 kilobytes.
65 Mapping mem into the process address space gives access to the
66 SPU local storage within the process address space. Only
67 MAP_SHARED mappings are allowed.
71 The first SPU to CPU communication mailbox. This file is read-only and
72 can be read in units of 32 bits. The file can only be used in non-
73 blocking mode and it even poll() will not block on it. The possible
74 operations on an open mbox file are:
77 If a count smaller than four is requested, read returns -1 and
78 sets errno to EINVAL. If there is no data available in the mail
79 box, the return value is set to -1 and errno becomes EAGAIN.
80 When data has been read successfully, four bytes are placed in
81 the data buffer and the value four is returned.
85 The second SPU to CPU communication mailbox. This file is similar to
86 the first mailbox file, but can be read in blocking I/O mode, and the
87 poll family of system calls can be used to wait for it. The possible
88 operations on an open ibox file are:
91 If a count smaller than four is requested, read returns -1 and
92 sets errno to EINVAL. If there is no data available in the mail
93 box and the file descriptor has been opened with O_NONBLOCK, the
94 return value is set to -1 and errno becomes EAGAIN.
96 If there is no data available in the mail box and the file
97 descriptor has been opened without O_NONBLOCK, the call will
98 block until the SPU writes to its interrupt mailbox channel.
99 When data has been read successfully, four bytes are placed in
100 the data buffer and the value four is returned.
103 Poll on the ibox file returns (POLLIN | POLLRDNORM) whenever
104 data is available for reading.
108 The CPU to SPU communation mailbox. It is write-only and can be written
109 in units of 32 bits. If the mailbox is full, write() will block and
110 poll can be used to wait for it becoming empty again. The possible
111 operations on an open wbox file are: write(2) If a count smaller than
112 four is requested, write returns -1 and sets errno to EINVAL. If there
113 is no space available in the mail box and the file descriptor has been
114 opened with O_NONBLOCK, the return value is set to -1 and errno becomes
117 If there is no space available in the mail box and the file descriptor
118 has been opened without O_NONBLOCK, the call will block until the SPU
119 reads from its PPE mailbox channel. When data has been read success-
120 fully, four bytes are placed in the data buffer and the value four is
124 Poll on the ibox file returns (POLLOUT | POLLWRNORM) whenever
125 space is available for writing.
131 Read-only files that contain the length of the current queue, i.e. how
132 many words can be read from mbox or ibox or how many words can be
133 written to wbox without blocking. The files can be read only in 4-byte
134 units and return a big-endian binary integer number. The possible
135 operations on an open *box_stat file are:
138 If a count smaller than four is requested, read returns -1 and
139 sets errno to EINVAL. Otherwise, a four byte value is placed in
140 the data buffer, containing the number of elements that can be
141 read from (for mbox_stat and ibox_stat) or written to (for
142 wbox_stat) the respective mail box without blocking or resulting
152 Internal registers of the SPU. The representation is an ASCII string
153 with the numeric value of the next instruction to be executed. These
154 can be used in read/write mode for debugging, but normal operation of
155 programs should not rely on them because access to any of them except
156 npc requires an SPU context save and is therefore very inefficient.
158 The contents of these files are:
160 npc Next Program Counter
164 decr_status Decrementer Status
166 spu_tag_mask MFC tag mask for SPU DMA
168 event_mask Event mask for SPU interrupts
170 srr0 Interrupt Return address register
173 The possible operations on an open npc, decr, decr_status,
174 spu_tag_mask, event_mask or srr0 file are:
177 When the count supplied to the read call is shorter than the
178 required length for the pointer value plus a newline character,
179 subsequent reads from the same file descriptor will result in
180 completing the string, regardless of changes to the register by
181 a running SPU task. When a complete string has been read, all
182 subsequent read operations will return zero bytes and a new file
183 descriptor needs to be opened to read the value again.
186 A write operation on the file results in setting the register to
187 the value given in the string. The string is parsed from the
188 beginning to the first non-numeric character or the end of the
189 buffer. Subsequent writes to the same file descriptor overwrite
190 the previous setting.
194 This file gives access to the Floating Point Status and Control Regis-
195 ter as a four byte long file. The operations on the fpcr file are:
198 If a count smaller than four is requested, read returns -1 and
199 sets errno to EINVAL. Otherwise, a four byte value is placed in
200 the data buffer, containing the current value of the fpcr regis-
204 If a count smaller than four is requested, write returns -1 and
205 sets errno to EINVAL. Otherwise, a four byte value is copied
206 from the data buffer, updating the value of the fpcr register.
211 The two signal notification channels of an SPU. These are read-write
212 files that operate on a 32 bit word. Writing to one of these files
213 triggers an interrupt on the SPU. The value writting to the signal
214 files can be read from the SPU through a channel read or from host user
215 space through the file. After the value has been read by the SPU, it
216 is reset to zero. The possible operations on an open signal1 or sig-
220 If a count smaller than four is requested, read returns -1 and
221 sets errno to EINVAL. Otherwise, a four byte value is placed in
222 the data buffer, containing the current value of the specified
223 signal notification register.
226 If a count smaller than four is requested, write returns -1 and
227 sets errno to EINVAL. Otherwise, a four byte value is copied
228 from the data buffer, updating the value of the specified signal
229 notification register. The signal notification register will
230 either be replaced with the input data or will be updated to the
231 bitwise OR or the old value and the input data, depending on the
232 contents of the signal1_type, or signal2_type respectively,
238 These two files change the behavior of the signal1 and signal2 notifi-
239 cation files. The contain a numerical ASCII string which is read as
240 either "1" or "0". In mode 0 (overwrite), the hardware replaces the
241 contents of the signal channel with the data that is written to it. in
242 mode 1 (logical OR), the hardware accumulates the bits that are subse-
243 quently written to it. The possible operations on an open signal1_type
244 or signal2_type file are:
247 When the count supplied to the read call is shorter than the
248 required length for the digit plus a newline character, subse-
249 quent reads from the same file descriptor will result in com-
250 pleting the string. When a complete string has been read, all
251 subsequent read operations will return zero bytes and a new file
252 descriptor needs to be opened to read the value again.
255 A write operation on the file results in setting the register to
256 the value given in the string. The string is parsed from the
257 beginning to the first non-numeric character or the end of the
258 buffer. Subsequent writes to the same file descriptor overwrite
259 the previous setting.
264 none /spu spufs gid=spu 0 0
268 Arnd Bergmann <arndb@de.ibm.com>, Mark Nutter <mnutter@us.ibm.com>,
269 Ulrich Weigand <Ulrich.Weigand@de.ibm.com>
272 capabilities(7), close(2), spu_create(2), spu_run(2), spufs(7)
276 Linux 2005-09-28 SPUFS(2)
278 ------------------------------------------------------------------------------
280 SPU_RUN(2) Linux Programmer's Manual SPU_RUN(2)
285 spu_run - execute an spu context
291 int spu_run(int fd, unsigned int *npc, unsigned int *event);
294 The spu_run system call is used on PowerPC machines that implement the
295 Cell Broadband Engine Architecture in order to access Synergistic Pro-
296 cessor Units (SPUs). It uses the fd that was returned from spu_cre-
297 ate(2) to address a specific SPU context. When the context gets sched-
298 uled to a physical SPU, it starts execution at the instruction pointer
301 Execution of SPU code happens synchronously, meaning that spu_run does
302 not return while the SPU is still running. If there is a need to exe-
303 cute SPU code in parallel with other code on either the main CPU or
304 other SPUs, you need to create a new thread of execution first, e.g.
305 using the pthread_create(3) call.
307 When spu_run returns, the current value of the SPU instruction pointer
308 is written back to npc, so you can call spu_run again without updating
311 event can be a NULL pointer or point to an extended status code that
312 gets filled when spu_run returns. It can be one of the following con-
315 SPE_EVENT_DMA_ALIGNMENT
316 A DMA alignment error
318 SPE_EVENT_SPE_DATA_SEGMENT
319 A DMA segmentation error
321 SPE_EVENT_SPE_DATA_STORAGE
324 If NULL is passed as the event argument, these errors will result in a
325 signal delivered to the calling process.
328 spu_run returns the value of the spu_status register or -1 to indicate
329 an error and set errno to one of the error codes listed below. The
330 spu_status register value contains a bit mask of status codes and
331 optionally a 14 bit code returned from the stop-and-signal instruction
332 on the SPU. The bit masks for the status codes are:
334 0x02 SPU was stopped by stop-and-signal.
336 0x04 SPU was stopped by halt.
338 0x08 SPU is waiting for a channel.
340 0x10 SPU is in single-step mode.
342 0x20 SPU has tried to execute an invalid instruction.
344 0x40 SPU has tried to access an invalid channel.
347 The bits masked with this value contain the code returned from
350 There are always one or more of the lower eight bits set or an error
351 code is returned from spu_run.
354 EAGAIN or EWOULDBLOCK
355 fd is in non-blocking mode and spu_run would block.
357 EBADF fd is not a valid file descriptor.
359 EFAULT npc is not a valid pointer or status is neither NULL nor a valid
362 EINTR A signal occurred while spu_run was in progress. The npc value
363 has been updated to the new program counter value if necessary.
365 EINVAL fd is not a file descriptor returned from spu_create(2).
367 ENOMEM Insufficient memory was available to handle a page fault result-
368 ing from an MFC direct memory access.
370 ENOSYS the functionality is not provided by the current system, because
371 either the hardware does not provide SPUs or the spufs module is
376 spu_run is meant to be used from libraries that implement a more
377 abstract interface to SPUs, not to be used from regular applications.
378 See http://www.bsc.es/projects/deepcomputing/linuxoncell/ for the rec-
383 This call is Linux specific and only implemented by the ppc64 architec-
384 ture. Programs using this system call are not portable.
388 The code does not yet fully implement all features lined out here.
392 Arnd Bergmann <arndb@de.ibm.com>
395 capabilities(7), close(2), spu_create(2), spufs(7)
399 Linux 2005-09-28 SPU_RUN(2)
401 ------------------------------------------------------------------------------
403 SPU_CREATE(2) Linux Programmer's Manual SPU_CREATE(2)
408 spu_create - create a new spu context
412 #include <sys/types.h>
415 int spu_create(const char *pathname, int flags, mode_t mode);
418 The spu_create system call is used on PowerPC machines that implement
419 the Cell Broadband Engine Architecture in order to access Synergistic
420 Processor Units (SPUs). It creates a new logical context for an SPU in
421 pathname and returns a handle to associated with it. pathname must
422 point to a non-existing directory in the mount point of the SPU file
423 system (spufs). When spu_create is successful, a directory gets cre-
424 ated on pathname and it is populated with files.
426 The returned file handle can only be passed to spu_run(2) or closed,
427 other operations are not defined on it. When it is closed, all associ-
428 ated directory entries in spufs are removed. When the last file handle
429 pointing either inside of the context directory or to this file
430 descriptor is closed, the logical SPU context is destroyed.
432 The parameter flags can be zero or any bitwise or'd combination of the
436 Allow mapping of some of the hardware registers of the SPU into
437 user space. This flag requires the CAP_SYS_RAWIO capability, see
440 The mode parameter specifies the permissions used for creating the new
441 directory in spufs. mode is modified with the user's umask(2) value
442 and then used for both the directory and the files contained in it. The
443 file permissions mask out some more bits of mode because they typically
444 support only read or write access. See stat(2) for a full list of the
445 possible mode values.
449 spu_create returns a new file descriptor. It may return -1 to indicate
450 an error condition and set errno to one of the error codes listed
456 The current user does not have write access on the spufs mount
459 EEXIST An SPU context already exists at the given path name.
461 EFAULT pathname is not a valid string pointer in the current address
464 EINVAL pathname is not a directory in the spufs mount point.
466 ELOOP Too many symlinks were found while resolving pathname.
468 EMFILE The process has reached its maximum open file limit.
471 pathname was too long.
473 ENFILE The system has reached the global open file limit.
475 ENOENT Part of pathname could not be resolved.
477 ENOMEM The kernel could not allocate all resources required.
479 ENOSPC There are not enough SPU resources available to create a new
480 context or the user specific limit for the number of SPU con-
481 texts has been reached.
483 ENOSYS the functionality is not provided by the current system, because
484 either the hardware does not provide SPUs or the spufs module is
488 A part of pathname is not a directory.
493 spu_create is meant to be used from libraries that implement a more
494 abstract interface to SPUs, not to be used from regular applications.
495 See http://www.bsc.es/projects/deepcomputing/linuxoncell/ for the rec-
500 pathname must point to a location beneath the mount point of spufs. By
501 convention, it gets mounted in /spu.
505 This call is Linux specific and only implemented by the ppc64 architec-
506 ture. Programs using this system call are not portable.
510 The code does not yet fully implement all features lined out here.
514 Arnd Bergmann <arndb@de.ibm.com>
517 capabilities(7), close(2), spu_run(2), spufs(7)
521 Linux 2005-09-28 SPU_CREATE(2)