|SPU_CREATE(2)||Linux Programmer's Manual||SPU_CREATE(2)|
#include <sys/spu.h> /* Definition of SPU_* constants */ #include <sys/syscall.h> /* Definition of SYS_* constants */ #include <unistd.h>
int syscall(SYS_spu_create, const char *pathname, unsigned int flags, mode_t mode, int neighbor_fd);
Note: glibc provides no wrapper for spu_create(), necessitating the use of syscall(2).
When a context is created, the returned file descriptor can only be passed to spu_run(2), used as the dirfd argument to the *at family of system calls (e.g., openat(2)), or closed; other operations are not defined. A logical SPU context is destroyed (along with all files created within the context's pathname directory) once the last reference to the context has gone; this usually occurs when the file descriptor returned by spu_create() is closed.
The neighbor_fd is used only when the SPU_CREATE_AFFINITY_SPU flag is specified; see below.
The flags argument can be zero or any bitwise OR-ed combination of the following constants:
- Rather than using signals for reporting DMA errors, use the event argument to spu_run(2).
- Create an SPU gang instead of a context. (A gang is a group of SPU contexts that are functionally related to each other and which share common scheduling parameters—priority and policy. In the future, gang scheduling may be implemented causing the group to be switched in and out as a single unit.)
- A new directory will be created at the location specified by the pathname argument. This gang may be used to hold other SPU contexts, by providing a pathname that is within the gang directory to further calls to spu_create().
- Create a context that is not affected by the SPU scheduler. Once the context is run, it will not be scheduled out until it is destroyed by the creating process.
- Because the context cannot be removed from the SPU, some functionality is disabled for SPU_CREATE_NOSCHED contexts. Only a subset of the files will be available in this context directory in spufs. Additionally, SPU_CREATE_NOSCHED contexts cannot dump a core file when crashing.
- Creating SPU_CREATE_NOSCHED contexts requires the CAP_SYS_NICE capability.
- Create an isolated SPU context. Isolated contexts are protected from some PPE (PowerPC Processing Element) operations, such as access to the SPU local store and the NPC register.
- Creating SPU_CREATE_ISOLATE contexts also requires the SPU_CREATE_NOSCHED flag.
- SPU_CREATE_AFFINITY_SPU (since Linux 2.6.23)
- Create a context with affinity to another SPU context. This affinity information is used within the SPU scheduling algorithm. Using this flag requires that a file descriptor referring to the other SPU context be passed in the neighbor_fd argument.
- SPU_CREATE_AFFINITY_MEM (since Linux 2.6.23)
- Create a context with affinity to system memory. This affinity information is used within the SPU scheduling algorithm.
- The current user does not have write access to the spufs(7) mount point.
- An SPU context already exists at the given pathname.
- pathname is not a valid string pointer in the calling process's address space.
- pathname is not a directory in the spufs(7) mount point, or invalid flags have been provided.
- Too many symbolic links were found while resolving pathname.
- The per-process limit on the number of open file descriptors has been reached.
- pathname is too long.
- The system-wide limit on the total number of open files has been reached.
- An isolated context was requested, but the hardware does not support SPU isolation.
- Part of pathname could not be resolved.
- The kernel could not allocate all resources required.
- There are not enough SPU resources available to create a new context or the user-specific limit for the number of SPU contexts has been reached.
- The functionality is not provided by the current system, because either the hardware does not provide SPUs or the spufs module is not loaded.
- A part of pathname is not a directory.
- The SPU_CREATE_NOSCHED flag has been given, but the user does not have the CAP_SYS_NICE capability.
Prior to the addition of the SPU_CREATE_AFFINITY_SPU flag in Linux 2.6.23, the spu_create() system call took only three arguments (i.e., there was no neighbor_fd argument).