diff options
Diffstat (limited to 'include/linux/scif.h')
| -rw-r--r-- | include/linux/scif.h | 366 |
1 files changed, 356 insertions, 10 deletions
diff --git a/include/linux/scif.h b/include/linux/scif.h index 44f4f3898..49a35d6ed 100644 --- a/include/linux/scif.h +++ b/include/linux/scif.h @@ -55,6 +55,7 @@ #include <linux/types.h> #include <linux/poll.h> +#include <linux/device.h> #include <linux/scif_ioctl.h> #define SCIF_ACCEPT_SYNC 1 @@ -92,6 +93,70 @@ enum { #define SCIF_PORT_RSVD 1088 typedef struct scif_endpt *scif_epd_t; +typedef struct scif_pinned_pages *scif_pinned_pages_t; + +/** + * struct scif_range - SCIF registered range used in kernel mode + * @cookie: cookie used internally by SCIF + * @nr_pages: number of pages of PAGE_SIZE + * @prot_flags: R/W protection + * @phys_addr: Array of bus addresses + * @va: Array of kernel virtual addresses backed by the pages in the phys_addr + * array. The va is populated only when called on the host for a remote + * SCIF connection on MIC. This is required to support the use case of DMA + * between MIC and another device which is not a SCIF node e.g., an IB or + * ethernet NIC. + */ +struct scif_range { + void *cookie; + int nr_pages; + int prot_flags; + dma_addr_t *phys_addr; + void __iomem **va; +}; + +/** + * struct scif_pollepd - SCIF endpoint to be monitored via scif_poll + * @epd: SCIF endpoint + * @events: requested events + * @revents: returned events + */ +struct scif_pollepd { + scif_epd_t epd; + short events; + short revents; +}; + +/** + * scif_peer_dev - representation of a peer SCIF device + * + * Peer devices show up as PCIe devices for the mgmt node but not the cards. + * The mgmt node discovers all the cards on the PCIe bus and informs the other + * cards about their peers. Upon notification of a peer a node adds a peer + * device to the peer bus to maintain symmetry in the way devices are + * discovered across all nodes in the SCIF network. + * + * @dev: underlying device + * @dnode - The destination node which this device will communicate with. + */ +struct scif_peer_dev { + struct device dev; + u8 dnode; +}; + +/** + * scif_client - representation of a SCIF client + * @name: client name + * @probe - client method called when a peer device is registered + * @remove - client method called when a peer device is unregistered + * @si - subsys_interface used internally for implementing SCIF clients + */ +struct scif_client { + const char *name; + void (*probe)(struct scif_peer_dev *spdev); + void (*remove)(struct scif_peer_dev *spdev); + struct subsys_interface si; +}; #define SCIF_OPEN_FAILED ((scif_epd_t)-1) #define SCIF_REGISTER_FAILED ((off_t)-1) @@ -345,7 +410,6 @@ int scif_close(scif_epd_t epd); * Errors: * EBADF, ENOTTY - epd is not a valid endpoint descriptor * ECONNRESET - Connection reset by peer - * EFAULT - An invalid address was specified for a parameter * EINVAL - flags is invalid, or len is negative * ENODEV - The remote node is lost or existed, but is not currently in the * network since it may have crashed @@ -398,7 +462,6 @@ int scif_send(scif_epd_t epd, void *msg, int len, int flags); * EAGAIN - The destination node is returning from a low power state * EBADF, ENOTTY - epd is not a valid endpoint descriptor * ECONNRESET - Connection reset by peer - * EFAULT - An invalid address was specified for a parameter * EINVAL - flags is invalid, or len is negative * ENODEV - The remote node is lost or existed, but is not currently in the * network since it may have crashed @@ -461,9 +524,6 @@ int scif_recv(scif_epd_t epd, void *msg, int len, int flags); * SCIF_PROT_READ - allow read operations from the window * SCIF_PROT_WRITE - allow write operations to the window * - * The map_flags argument can be set to SCIF_MAP_FIXED which interprets a - * fixed offset. - * * Return: * Upon successful completion, scif_register() returns the offset at which the * mapping was placed (po); otherwise in user mode SCIF_REGISTER_FAILED (that @@ -476,7 +536,6 @@ int scif_recv(scif_epd_t epd, void *msg, int len, int flags); * EAGAIN - The mapping could not be performed due to lack of resources * EBADF, ENOTTY - epd is not a valid endpoint descriptor * ECONNRESET - Connection reset by peer - * EFAULT - Addresses in the range [addr, addr + len - 1] are invalid * EINVAL - map_flags is invalid, or prot_flags is invalid, or SCIF_MAP_FIXED is * set in flags, and offset is not a multiple of the page size, or addr is not a * multiple of the page size, or len is not a multiple of the page size, or is @@ -759,7 +818,6 @@ int scif_writeto(scif_epd_t epd, off_t loffset, size_t len, off_t * EACCESS - Attempt to write to a read-only range * EBADF, ENOTTY - epd is not a valid endpoint descriptor * ECONNRESET - Connection reset by peer - * EFAULT - Addresses in the range [addr, addr + len - 1] are invalid * EINVAL - rma_flags is invalid * ENODEV - The remote node is lost or existed, but is not currently in the * network since it may have crashed @@ -840,7 +898,6 @@ int scif_vreadfrom(scif_epd_t epd, void *addr, size_t len, off_t roffset, * EACCESS - Attempt to write to a read-only range * EBADF, ENOTTY - epd is not a valid endpoint descriptor * ECONNRESET - Connection reset by peer - * EFAULT - Addresses in the range [addr, addr + len - 1] are invalid * EINVAL - rma_flags is invalid * ENODEV - The remote node is lost or existed, but is not currently in the * network since it may have crashed @@ -984,10 +1041,299 @@ int scif_fence_signal(scif_epd_t epd, off_t loff, u64 lval, off_t roff, * online nodes in the SCIF network including 'self'; otherwise in user mode * -1 is returned and errno is set to indicate the error; in kernel mode no * errors are returned. + */ +int scif_get_node_ids(u16 *nodes, int len, u16 *self); + +/** + * scif_pin_pages() - Pin a set of pages + * @addr: Virtual address of range to pin + * @len: Length of range to pin + * @prot_flags: Page protection flags + * @map_flags: Page classification flags + * @pinned_pages: Handle to pinned pages + * + * scif_pin_pages() pins (locks in physical memory) the physical pages which + * back the range of virtual address pages starting at addr and continuing for + * len bytes. addr and len are constrained to be multiples of the page size. A + * successful scif_pin_pages() call returns a handle to pinned_pages which may + * be used in subsequent calls to scif_register_pinned_pages(). + * + * The pages will remain pinned as long as there is a reference against the + * scif_pinned_pages_t value returned by scif_pin_pages() and until + * scif_unpin_pages() is called, passing the scif_pinned_pages_t value. A + * reference is added to a scif_pinned_pages_t value each time a window is + * created by calling scif_register_pinned_pages() and passing the + * scif_pinned_pages_t value. A reference is removed from a + * scif_pinned_pages_t value each time such a window is deleted. + * + * Subsequent operations which change the memory pages to which virtual + * addresses are mapped (such as mmap(), munmap()) have no effect on the + * scif_pinned_pages_t value or windows created against it. + * + * If the process will fork(), it is recommended that the registered + * virtual address range be marked with MADV_DONTFORK. Doing so will prevent + * problems due to copy-on-write semantics. + * + * The prot_flags argument is formed by OR'ing together one or more of the + * following values. + * SCIF_PROT_READ - allow read operations against the pages + * SCIF_PROT_WRITE - allow write operations against the pages + * The map_flags argument can be set as SCIF_MAP_KERNEL to interpret addr as a + * kernel space address. By default, addr is interpreted as a user space + * address. + * + * Return: + * Upon successful completion, scif_pin_pages() returns 0; otherwise the + * negative of one of the following errors is returned. * * Errors: - * EFAULT - Bad address + * EINVAL - prot_flags is invalid, map_flags is invalid, or offset is negative + * ENOMEM - Not enough space */ -int scif_get_node_ids(u16 *nodes, int len, u16 *self); +int scif_pin_pages(void *addr, size_t len, int prot_flags, int map_flags, + scif_pinned_pages_t *pinned_pages); + +/** + * scif_unpin_pages() - Unpin a set of pages + * @pinned_pages: Handle to pinned pages to be unpinned + * + * scif_unpin_pages() prevents scif_register_pinned_pages() from registering new + * windows against pinned_pages. The physical pages represented by pinned_pages + * will remain pinned until all windows previously registered against + * pinned_pages are deleted (the window is scif_unregister()'d and all + * references to the window are removed (see scif_unregister()). + * + * pinned_pages must have been obtain from a previous call to scif_pin_pages(). + * After calling scif_unpin_pages(), it is an error to pass pinned_pages to + * scif_register_pinned_pages(). + * + * Return: + * Upon successful completion, scif_unpin_pages() returns 0; otherwise the + * negative of one of the following errors is returned. + * + * Errors: + * EINVAL - pinned_pages is not valid + */ +int scif_unpin_pages(scif_pinned_pages_t pinned_pages); + +/** + * scif_register_pinned_pages() - Mark a memory region for remote access. + * @epd: endpoint descriptor + * @pinned_pages: Handle to pinned pages + * @offset: Registered address space offset + * @map_flags: Flags which control where pages are mapped + * + * The scif_register_pinned_pages() function opens a window, a range of whole + * pages of the registered address space of the endpoint epd, starting at + * offset po. The value of po, further described below, is a function of the + * parameters offset and pinned_pages, and the value of map_flags. Each page of + * the window represents a corresponding physical memory page of the range + * represented by pinned_pages; the length of the window is the same as the + * length of range represented by pinned_pages. A successful + * scif_register_pinned_pages() call returns po as the return value. + * + * When SCIF_MAP_FIXED is set in the map_flags argument, po will be offset + * exactly, and offset is constrained to be a multiple of the page size. The + * mapping established by scif_register_pinned_pages() will not replace any + * existing registration; an error is returned if any page of the new window + * would intersect an existing window. + * + * When SCIF_MAP_FIXED is not set, the implementation uses offset in an + * implementation-defined manner to arrive at po. The po so chosen will be an + * area of the registered address space that the implementation deems suitable + * for a mapping of the required size. An offset value of 0 is interpreted as + * granting the implementation complete freedom in selecting po, subject to + * constraints described below. A non-zero value of offset is taken to be a + * suggestion of an offset near which the mapping should be placed. When the + * implementation selects a value for po, it does not replace any extant + * window. In all cases, po will be a multiple of the page size. + * + * The physical pages which are so represented by a window are available for + * access in calls to scif_get_pages(), scif_readfrom(), scif_writeto(), + * scif_vreadfrom(), and scif_vwriteto(). While a window is registered, the + * physical pages represented by the window will not be reused by the memory + * subsystem for any other purpose. Note that the same physical page may be + * represented by multiple windows. + * + * Windows created by scif_register_pinned_pages() are unregistered by + * scif_unregister(). + * + * The map_flags argument can be set to SCIF_MAP_FIXED which interprets a + * fixed offset. + * + * Return: + * Upon successful completion, scif_register_pinned_pages() returns the offset + * at which the mapping was placed (po); otherwise the negative of one of the + * following errors is returned. + * + * Errors: + * EADDRINUSE - SCIF_MAP_FIXED is set in map_flags and pages in the new window + * would intersect an existing window + * EAGAIN - The mapping could not be performed due to lack of resources + * ECONNRESET - Connection reset by peer + * EINVAL - map_flags is invalid, or SCIF_MAP_FIXED is set in map_flags, and + * offset is not a multiple of the page size, or offset is negative + * ENODEV - The remote node is lost or existed, but is not currently in the + * network since it may have crashed + * ENOMEM - Not enough space + * ENOTCONN - The endpoint is not connected + */ +off_t scif_register_pinned_pages(scif_epd_t epd, + scif_pinned_pages_t pinned_pages, + off_t offset, int map_flags); + +/** + * scif_get_pages() - Add references to remote registered pages + * @epd: endpoint descriptor + * @offset: remote registered offset + * @len: length of range of pages + * @pages: returned scif_range structure + * + * scif_get_pages() returns the addresses of the physical pages represented by + * those pages of the registered address space of the peer of epd, starting at + * offset and continuing for len bytes. offset and len are constrained to be + * multiples of the page size. + * + * All of the pages in the specified range [offset, offset + len - 1] must be + * within a single window of the registered address space of the peer of epd. + * + * The addresses are returned as a virtually contiguous array pointed to by the + * phys_addr component of the scif_range structure whose address is returned in + * pages. The nr_pages component of scif_range is the length of the array. The + * prot_flags component of scif_range holds the protection flag value passed + * when the pages were registered. + * + * Each physical page whose address is returned by scif_get_pages() remains + * available and will not be released for reuse until the scif_range structure + * is returned in a call to scif_put_pages(). The scif_range structure returned + * by scif_get_pages() must be unmodified. + * + * It is an error to call scif_close() on an endpoint on which a scif_range + * structure of that endpoint has not been returned to scif_put_pages(). + * + * Return: + * Upon successful completion, scif_get_pages() returns 0; otherwise the + * negative of one of the following errors is returned. + * Errors: + * ECONNRESET - Connection reset by peer. + * EINVAL - offset is not a multiple of the page size, or offset is negative, or + * len is not a multiple of the page size + * ENODEV - The remote node is lost or existed, but is not currently in the + * network since it may have crashed + * ENOTCONN - The endpoint is not connected + * ENXIO - Offsets in the range [offset, offset + len - 1] are invalid + * for the registered address space of the peer epd + */ +int scif_get_pages(scif_epd_t epd, off_t offset, size_t len, + struct scif_range **pages); + +/** + * scif_put_pages() - Remove references from remote registered pages + * @pages: pages to be returned + * + * scif_put_pages() releases a scif_range structure previously obtained by + * calling scif_get_pages(). The physical pages represented by pages may + * be reused when the window which represented those pages is unregistered. + * Therefore, those pages must not be accessed after calling scif_put_pages(). + * + * Return: + * Upon successful completion, scif_put_pages() returns 0; otherwise the + * negative of one of the following errors is returned. + * Errors: + * EINVAL - pages does not point to a valid scif_range structure, or + * the scif_range structure pointed to by pages was already returned + * ENODEV - The remote node is lost or existed, but is not currently in the + * network since it may have crashed + * ENOTCONN - The endpoint is not connected + */ +int scif_put_pages(struct scif_range *pages); + +/** + * scif_poll() - Wait for some event on an endpoint + * @epds: Array of endpoint descriptors + * @nepds: Length of epds + * @timeout: Upper limit on time for which scif_poll() will block + * + * scif_poll() waits for one of a set of endpoints to become ready to perform + * an I/O operation. + * + * The epds argument specifies the endpoint descriptors to be examined and the + * events of interest for each endpoint descriptor. epds is a pointer to an + * array with one member for each open endpoint descriptor of interest. + * + * The number of items in the epds array is specified in nepds. The epd field + * of scif_pollepd is an endpoint descriptor of an open endpoint. The field + * events is a bitmask specifying the events which the application is + * interested in. The field revents is an output parameter, filled by the + * kernel with the events that actually occurred. The bits returned in revents + * can include any of those specified in events, or one of the values POLLERR, + * POLLHUP, or POLLNVAL. (These three bits are meaningless in the events + * field, and will be set in the revents field whenever the corresponding + * condition is true.) + * + * If none of the events requested (and no error) has occurred for any of the + * endpoint descriptors, then scif_poll() blocks until one of the events occurs. + * + * The timeout argument specifies an upper limit on the time for which + * scif_poll() will block, in milliseconds. Specifying a negative value in + * timeout means an infinite timeout. + * + * The following bits may be set in events and returned in revents. + * POLLIN - Data may be received without blocking. For a connected + * endpoint, this means that scif_recv() may be called without blocking. For a + * listening endpoint, this means that scif_accept() may be called without + * blocking. + * POLLOUT - Data may be sent without blocking. For a connected endpoint, this + * means that scif_send() may be called without blocking. POLLOUT may also be + * used to block waiting for a non-blocking connect to complete. This bit value + * has no meaning for a listening endpoint and is ignored if specified. + * + * The following bits are only returned in revents, and are ignored if set in + * events. + * POLLERR - An error occurred on the endpoint + * POLLHUP - The connection to the peer endpoint was disconnected + * POLLNVAL - The specified endpoint descriptor is invalid. + * + * Return: + * Upon successful completion, scif_poll() returns a non-negative value. A + * positive value indicates the total number of endpoint descriptors that have + * been selected (that is, endpoint descriptors for which the revents member is + * non-zero). A value of 0 indicates that the call timed out and no endpoint + * descriptors have been selected. Otherwise in user mode -1 is returned and + * errno is set to indicate the error; in kernel mode the negative of one of + * the following errors is returned. + * + * Errors: + * EINTR - A signal occurred before any requested event + * EINVAL - The nepds argument is greater than {OPEN_MAX} + * ENOMEM - There was no space to allocate file descriptor tables + */ +int scif_poll(struct scif_pollepd *epds, unsigned int nepds, long timeout); + +/** + * scif_client_register() - Register a SCIF client + * @client: client to be registered + * + * scif_client_register() registers a SCIF client. The probe() method + * of the client is called when SCIF peer devices come online and the + * remove() method is called when the peer devices disappear. + * + * Return: + * Upon successful completion, scif_client_register() returns a non-negative + * value. Otherwise the return value is the same as subsys_interface_register() + * in the kernel. + */ +int scif_client_register(struct scif_client *client); + +/** + * scif_client_unregister() - Unregister a SCIF client + * @client: client to be unregistered + * + * scif_client_unregister() unregisters a SCIF client. + * + * Return: + * None + */ +void scif_client_unregister(struct scif_client *client); #endif /* __SCIF_H__ */ |