FMC Device ********** Within the Linux bus framework, the FMC device is created and registered by the carrier driver. For example, the PCI driver for the SPEC card fills a data structure for each SPEC that it drives, and registers an associated FMC device for each card. The SVEC driver can do exactly the same for the VME carrier (actually, it should do it twice, because the SVEC carries two FMC mezzanines). Similarly, an Etherbone driver will be able to register its own FMC devices, offering communication primitives through frame exchange. The contents of the EEPROM within the FMC are used for identification purposes, i.e. for matching the device with its own driver. For this reason the device structure includes a complete copy of the EEPROM (actually, the carrier driver may choose whether or not to return it - for example we most likely won't have the whole EEPROM available for Etherbone devices. The following listing shows the current structure defining a device. Please note that all the machinery is in place but some details may still change in the future. For this reason, there is a version field at the beginning of the structure. As usual, the minor number will change for compatible changes (like a new flag) and the major number will increase when an incompatible change happens (for example, a change in layout of some fmc data structures). Device writers should just set it to the value FMC_VERSION, and be ready to get back -EINVAL at registration time. struct fmc_device { unsigned long version; unsigned long flags; struct module *owner; /* char device must pin it */ struct fmc_fru_id id; /* for EEPROM-based match */ struct fmc_operations *op; /* carrier-provided */ int irq; /* according to host bus. 0 == none */ int eeprom_len; /* Usually 8kB, may be less */ int eeprom_addr; /* 0x50, 0x52 etc */ uint8_t *eeprom; /* Full contents or leading part */ char *carrier_name; /* "SPEC" or similar, for special use */ void *carrier_data; /* "struct spec *" or equivalent */ __iomem void *fpga_base; /* May be NULL (Etherbone) */ __iomem void *slot_base; /* Set by the driver */ struct fmc_device **devarray; /* Allocated by the bus */ int slot_id; /* Index in the slot array */ int nr_slots; /* Number of slots in this carrier */ unsigned long memlen; /* Used for the char device */ struct device dev; /* For Linux use */ struct device *hwdev; /* The underlying hardware device */ unsigned long sdbfs_entry; struct sdb_array *sdb; uint32_t device_id; /* Filled by the device */ char *mezzanine_name; /* Defaults to ``fmc'' */ void *mezzanine_data; }; The meaning of most fields is summarized in the code comment above. The following fields must be filled by the carrier driver before registration: * version: must be set to FMC_VERSION. * owner: set to MODULE_OWNER. * op: the operations to act on the device. * irq: number for the mezzanine; may be zero. * eeprom_len: length of the following array. * eeprom_addr: 0x50 for first mezzanine and so on. * eeprom: the full content of the I2C EEPROM. * carrier_name. * carrier_data: a unique pointer for the carrier. * fpga_base: the I/O memory address (may be NULL). * slot_id: the index of this slot (starting from zero). * memlen: if fpga_base is valid, the length of I/O memory. * hwdev: to be used in some dev_err() calls. * device_id: a slot-specific unique integer number. Please note that the carrier should read its own EEPROM memory before registering the device, as well as fill all other fields listed above. The following fields should not be assigned, because they are filled later by either the bus or the device driver: * flags. * fru_id: filled by the bus, parsing the eeprom. * slot_base: filled and used by the driver, if useful to it. * devarray: an array og all mezzanines driven by a singe FPGA. * nr_slots: set by the core at registration time. * dev: used by Linux. * sdb: FPGA contents, scanned according to driver's directions. * sdbfs_entry: SDB entry point in EEPROM: autodetected. * mezzanine_data: available for the driver. * mezzanine_name: filled by fmc-bus during identification. Note: mezzanine_data may be redundant, because Linux offers the drvdata approach, so the field may be removed in later versions of this bus implementation. As I write this, she SPEC carrier is already completely functional in the fmc-bus environment, and is a good reference to look at. The API Offered by Carriers =========================== The carrier provides a number of methods by means of the `fmc_operations' structure, which currently is defined like this (again, it is a moving target, please refer to the header rather than this document): struct fmc_operations { uint32_t (*readl)(struct fmc_device *fmc, int offset); void (*writel)(struct fmc_device *fmc, uint32_t value, int offset); int (*reprogram)(struct fmc_device *f, struct fmc_driver *d, char *gw); int (*validate)(struct fmc_device *fmc, struct fmc_driver *drv); int (*irq_request)(struct fmc_device *fmc, irq_handler_t h, char *name, int flags); void (*irq_ack)(struct fmc_device *fmc); int (*irq_free)(struct fmc_device *fmc); int (*gpio_config)(struct fmc_device *fmc, struct fmc_gpio *gpio, int ngpio); int (*read_ee)(struct fmc_device *fmc, int pos, void *d, int l); int (*write_ee)(struct fmc_device *fmc, int pos, const void *d, int l); }; The individual methods perform the following tasks: `readl' `writel' These functions access FPGA registers by whatever means the carrier offers. They are not expected to fail, and most of the time they will just make a memory access to the host bus. If the carrier provides a fpga_base pointer, the driver may use direct access through that pointer. For this reason the header offers the inline functions fmc_readl and fmc_writel that access fpga_base if the respective method is NULL. A driver that wants to be portable and efficient should use fmc_readl and fmc_writel. For Etherbone, or other non-local carriers, error-management is still to be defined. `validate' Module parameters are used to manage different applications for two or more boards of the same kind. Validation is based on the busid module parameter, if provided, and returns the matching index in the associated array. See *note Module Parameters:: in in doubt. If no match is found, `-ENOENT' is returned; if the user didn't pass `busid=', all devices will pass validation. The value returned by the validate method can be used as index into other parameters (for example, some drivers use the `lm32=' parameter in this way). Such "generic parameters" are documented in *note Module Parameters::, below. The validate method is used by `fmc-trivial.ko', described in *note fmc-trivial::. `reprogram' The carrier enumerates FMC devices by loading a standard (or golden) FPGA binary that allows EEPROM access. Each driver, then, will need to reprogram the FPGA by calling this function. If the name argument is NULL, the carrier should reprogram the golden binary. If the gateware name has been overridden through module parameters (in a carrier-specific way) the file loaded will match the parameters. Per-device gateware names can be specified using the `gateware=' parameter, see *note Module Parameters::. Note: Clients should call rhe new helper, fmc_reprogram, which both calls this method and parse the SDB tree of the FPGA. `irq_request' `irq_ack' `irq_free' Interrupt management is carrier-specific, so it is abstracted as operations. The interrupt number is listed in the device structure, and for the mezzanine driver the number is only informative. The handler will receive the fmc pointer as dev_id; the flags argument is passed to the Linux request_irq function, but fmc-specific flags may be added in the future. You'll most likely want to pass the `IRQF_SHARED' flag. `gpio_config' The method allows to configure a GPIO pin in the carrier, and read its current value if it is configured as input. See *note The GPIO Abstraction:: for details. `read_ee' `write_ee' Read or write the EEPROM. The functions are expected to be only called before reprogramming and the carrier should refuse them with `ENODEV' after reprogramming. The offset is expected to be within 8kB (the current size), but addresses up to 1MB are reserved to fit bigger I2C devices in the future. Carriers may offer access to other internal flash memories using these same methods: for example the SPEC driver may define that its carrier I2C memory is seen at offset 1M and the internal SPI flash is seen at offset 16M. This multiplexing of several flash memories in the same address space is carrier-specific and should only be used by a driver that has verified the `carrier_name' field. The GPIO Abstraction ==================== Support for GPIO pins in the fmc-bus environment is not very straightforward and deserves special discussion. While the general idea of a carrier-independent driver seems to fly, configuration of specific signals within the carrier needs at least some knowledge of the carrier itself. For this reason, the specific driver can request to configure carrier-specific GPIO pins, numbered from 0 to at most 4095. Configuration is performed by passing a pointer to an array of struct fmc_gpio items, as well as the length of the array. This is the data structure: struct fmc_gpio { char *carrier_name; int gpio; int _gpio; /* internal use by the carrier */ int mode; /* GPIOF_DIR_OUT etc, from <linux/gpio.h> */ int irqmode; /* IRQF_TRIGGER_LOW and so on */ }; By specifying a carrier_name for each pin, the driver may access different pins in different carriers. The gpio_config method is expected to return the number of pins successfully configured, ignoring requests for other carriers. However, if no pin is configured (because no structure at all refers to the current carrier_name), the operation returns an error so the caller will know that it is running under a yet-unsupported carrier. So, for example, a driver that has been developed and tested on both the SPEC and the SVEC may request configuration of two different GPIO pins, and expect one such configuration to succeed - if none succeeds it most likely means that the current carrier is a still-unknown one. If, however, your GPIO pin has a specific known role, you can pass a special number in the gpio field, using one of the following macros: #define FMC_GPIO_RAW(x) (x) /* 4096 of them */ #define FMC_GPIO_IRQ(x) ((x) + 0x1000) /* 256 of them */ #define FMC_GPIO_LED(x) ((x) + 0x1100) /* 256 of them */ #define FMC_GPIO_KEY(x) ((x) + 0x1200) /* 256 of them */ #define FMC_GPIO_TP(x) ((x) + 0x1300) /* 256 of them */ #define FMC_GPIO_USER(x) ((x) + 0x1400) /* 256 of them */ Use of virtual GPIO numbers (anything but FMC_GPIO_RAW) is allowed provided the carrier_name field in the data structure is left unspecified (NULL). Each carrier is responsible for providing a mapping between virtual and physical GPIO numbers. The carrier may then use the _gpio field to cache the result of this mapping. All carriers must map their I/O lines to the sets above starting from zero. The SPEC, for example, maps interrupt pins 0 and 1, and test points 0 through 3 (even if the test points on the PCB are called 5,6,7,8). If, for example, a driver requires a free LED and a test point (for a scope probe to be plugged at some point during development) it may ask for FMC_GPIO_LED(0) and FMC_GPIO_TP(0). Each carrier will provide suitable GPIO pins. Clearly, the person running the drivers will know the order used by the specific carrier driver in assigning leds and testpoints, so to make a carrier-dependent use of the diagnostic tools. In theory, some form of autodetection should be possible: a driver like the wr-nic (which uses IRQ(1) on the SPEC card) should configure IRQ(0), make a test with software-generated interrupts and configure IRQ(1) if the test fails. This probing step should be used because even if the wr-nic gateware is known to use IRQ1 on the SPEC, the driver should be carrier-independent and thus use IRQ(0) as a first bet - actually, the knowledge that IRQ0 may fail is carrier-dependent information, but using it doesn't make the driver unsuitable for other carriers. The return value of gpio_config is defined as follows: * If no pin in the array can be used by the carrier, `-ENODEV'. * If at least one virtual GPIO number cannot be mapped, `-ENOENT'. * On success, 0 or positive. The value returned is the number of high input bits (if no input is configured, the value for success is 0). While I admit the procedure is not completely straightforward, it allows configuration, input and output with a single carrier operation. Given the typical use case of FMC devices, GPIO operations are not expected to ever by in hot paths, and GPIO access so fare has only been used to configure the interrupt pin, mode and polarity. Especially reading inputs is not expected to be common. If your device has GPIO capabilities in the hot path, you should consider using the kernel's GPIO mechanisms.