Weblog

Solaris Devfs Solaris originated from BSD and SVR4 UNIX. Over the years, many enhancements have been made to address business needs. One area of big change is the I/O framework and device name management.
Traditional UNIX kernel configures all devices at boot time. Device access is supported via two indexed arrays, bdevsw[] and cdevsw[], for block and character devices, respectively. The array elements contain references to driver entry points compiled into the kernel. Applications access device by opening device special files, created via the mknod(2) syscall. A device special file has a type (block or char) and a device number (dev_t). The type informs the kernel whether to use bdevsw[] or cdevsw[]. The device number contains two parts, major and minor. The major number is used to index into the arrays, and minor number is used by driver only, typically to determine which device instance to access.
Solaris modified and extended the model in many ways.

• bdevsw[] and cdevsw[] are merged into a single array, devopsp[], indexed by a major number common to both block and character drivers. The elements of devopsp[] reference driver's dev_ops structure, containing driver entry points for device autoconfiguration (probe, attach, detach), bus nexus oriented operations (bus_ops), and block/char operations (cb_ops).
• All drivers are loadable kernel modules. The modules are loaded on-demand. During system startup, the kernel only loads those driver modules required to boot the system. As a result, a normal system boot may not initialize all hardware attached to the system. To initialize all hardware at boot time, a reconfiguration boot (boot -r) is required.
• Devices are represented in the kernel by a tree of device information (struct dev_info) nodes. Inner (nexus) nodes represent bus controllers and adaptors while leaf nodes represent devices. Leaf nodes bind to "leaf" drivers, which implements cb_ops to handle I/O requests. Inner nodes bound to bus nexus drivers, which implements bus_ops to satisfy leaf driver requests or pass the requests to parent bus or the hardware platform. This design allows generic leaf drivers to be written without knowing the details of the transport. For example, the scsi disk driver (sd) can be used to control many types of disks such as SCSI-2, USB storage, Fibre-channel, and atapi CD-ROM.
• A private namespace, /devices, is introduced to mirror device names in the Open Boot PROM (obp) defined in the IEEE 1275 standard. The namespace reflects the physical topology of I/O devices and bus interconnects. This namespace is controlled by a filesystem named "devfs", first introduced in Solaris 10. A key feature of devfs is that a filesystem lookup operation actually drives configuration of the specific device instance corresponding to the pathname. For example, # ls /devices/pci@0,0/pci-ide@11,1/ide@1/sd@0,0:a would cause the ATAPI cdrom drive to attach even if it is currently not configured in the kernel.
• The public names in /dev are symbolic links to a pathname in /devices. The /dev names are created at Solaris Install time by devfsadm(1M). When new devices are added, the /dev name space is updated via devfsadmd(1M), the daemon version of devfsadm.

The current Solaris I/O framework is flexible and scales well from a single CPU system to high-end servers with 100+ CPUs and 1000+ devices. In addition, I/O devices can be reconfigured dynamically without rebooting the system. This functionality is also referred to as Dynamic Reconfiguration or hotplugging. In a future blog entry, I hope to explain in more detail the inner workings of devfs and the kernel device tree. ( Jun 14 2005, 08:38:19 AM PDT ) Permalink Comments [2]