[无双]

英文的
原文在http://www.lisoleg.net/lisoleg/applications/ioctl.htm
哪个兄弟有空的话可以把它翻译


In general an ioctl call in a user program looks like ioctl(int fd, int command, (char *) argstruct). For ioctl calls related to the networking code (these are the only ones we will deal with in this note), the file descriptor fd is actually a socket descriptor returned by the socket() system call. The command could be any one of those listed in /usr/include/linux/sockios.h. These commands are subdivided into a number of categories depending on what aspect of networking they deal with:

changing the routing table (e.g. SIOCADDRT, SIOCDELRT),
reading/updating the ARP/RARP caches (e.g. SIOCDARP, SIOCSRARP),
generic functions related to network interfaces (e.g. SIOCGIFNAME, SIOCSIFADDR etc)
The Goodies directory contains a number of sample programs illustrating the use of networking ioctl calls. As you look at these programs, notice how the structure used for argstruct depends on the ioctl command type. For example, routing table related ioctls use the rtentry structure defined in /usr/include/linux/route.h (see adddefault.c for an example) and ARP related ioctls use the arpreq structure defined in /usr/include/linux/if_arp.h (see arpread.c).

[无双]

Network interface related ioctl commands typically look like SIOCxIFyyyy where x is either S (set, write) or G (get, read). The getifinfo.c program uses such commands to read the IP address, hardware address, broadcast address and flags associated with a network interface. For these ioctls, the third argument is an ifreq structure which is defined in /usr/include/linux/if.h. In some cases, new ioctl commands may be needed in addition to those defined in sockios.h, e.g. the WaveLAN wireless networking card maintains information about wireless signal strength which may be of use to a user program. How should user programs be allowed access to this information? Our first instinct may be to define a new ioctl command in sockios.h, e.g. SIOCGIFWVLNSS (Get WaVeLaN Signal Strength). Unfortunately, this command makes no sense at all for other interfaces (e.g. loopback) and attempts to use this ioctl command on interfaces other than WaveLAN cards should be disallowed. What we need, then, is a mechanism to define interface specific ioctl commands. Luckily, the Linux OS already has built-in hooks for this purpose. If you look at sockios.h again, you will notice that each device has a predefined SIOCDEVPRIVATE ioctl command. The implementation of this command is left totally upto the person writing the corresponding device driver.

[无双]

By convention, a user program invokes a device specific ioctl command as ioctl(sockid, SIOCDEVPRIVATE, (char *) &ifr) where ifr is defined as struct ifreq ifr. It fills ifr.ifr_name with the interface name associated with the device, e.g. on tether, the WaveLAN card is named eth1. Typically, a user program will also need to exchange command arguments and results with the kernel and that is done through the ifr.ifr_data field, e.g. the signal strength infromation for the WaveLAN card could be returned in this field. The Linux source code already includes two devices de4x5 and ewrk3 that define and implement device specific ioctl commands. The source code for these drivers is in de4x5.h, de4x5.c, ewrk3.h, ewrk3.c, (in /usr/src/linux/drivers/net/). Both drivers define their own private structures (struct ewrk3_ioctl and struct de4x5_ioctl) for exchanging information between user programs and the drivers. Before the ioctl call, the user program fills out the necessary fields in this structure and points ifr.ifr_data to it.

[无双]

Before we go any further into the driver code for ewrk3 and de4x5, let us trace through various steps in the processing of an ioctl call. All interface-type ioctl requests (SIOCxIFyyyy and SIOCDEVPRIVATE) result in dev_ioctl() (in /usr/src/linux/net/core/dev.c) being called. This is just a wrapper and most of the real action is left for dev_ifsioc() (also in dev.c). About the only thing dev_ioctl() does is check whether a calling process has the appropriate permissions to issue the command (e.g. commands to alter the routing tables require root permissions). One of the first things dev_ifsioc() does is get the device structre (struct device defined in /usr/include/linux/netdevice.h) corresponding to the device named in ifr.ifr_name. This is followed by the code to implement generic interface commands (e.g. SIOCGIFADDR) inside a giant switch statement. The SIOCDEVPRIVATE command and any others with codes from 0x89F0 through 0x89FF end up in the default: branch for this switch. Here, the kernel checks to see if a device specific ioctl handler has been set up in the device structure. The handler is maintained as a function pointer in the do_ioctl field of the device structure. If the handler has been set, the kernel invokes it.

[无双]

So, to implement device specific ioctls, all one needs to do is write a device specific ioctl handler and have the do_ioctl field in the corresponding device structure point to it. For the ewrk3 device, this function is called ewrk3_ioctl() (in ewrk3.c) and the corresponding device structure is initialized in ewrk3_init(). The ewrk3_ioctl() code clearly indicates the use of ifr.ifr_data for exchanging information between the device driver and user program. Note that this area of memory can be used for bidirectional information exchange. For example, in the ewrk3 driver code, the first two bytes of ifr.ifr_data are used to convey the specific action (e.g. EWRK3_SET_PROM, EWRK3_CLR_PROM) desired by the user (the driver implements multiple device specific commands that are all invoked through SIOCDEVPRIVATE). Also, and the buffer pointed to by the fifth byte in ifr.ifr_data is used to exchange other information (hardware address when using EWRK3_SET_HWADDR or EWRK3_GET_HWADDR).

[无双]

As you go through ewrk3_ioctl(), keep in mind that normally a user process cannot directly access kernel memory. For this reason, two special procedures memcpy_tofs() and memcpy_fromfs() are provided to driver writers. The kernel procedure memcpy_tofs(arg1, arg2, arg3) copies arg3 bytes from starting address arg2 (in kernel space) to address arg1 (in user space). Similarly, memcpy_fromfs(arg1, arg2, arg3) copies arg3 bytes from starting address arg2 (in user space) to address arg1 (in kernel space). These procedures are preceded by calls to verify_area() to verify that the process has appropriate access permissions. Also notice the use of the printk() function to print debugging information. This function is similar to printf() but cannot handle floating point. The printf() function is not available to kernel code. The output generated by printk() is logged into /usr/adm/messages. For more information on these and related procedures, look at the section titled Supporting Functions in Michael K. Johnson's "Linux Kernel Hacker's Guide" which is accessible from the Linux Documentation Home Page.

[空心菜]

不懂啊

[kj501]

这个属于比较高级的内容了。要有基础才能理解的。

[moonlight]

ioctl一般是用于操纵设备的，需要操纵硬件的地方可能会用到。

[zhanxw]

为感谢kj501版主对我的帮助以及表示对LinuxSir网站的喜爱，
我将这篇文章粗粗翻译了一下。
（有些名次把握不好，限于自己水平有限，请见谅）

[zhanxw]

一般来讲ioctl在用户程序中的调用是：ioctl(int fd,int command, (char*)argstruct)。ioctl调用与网络编程有关（本文只讨论这一点），文件描述符fd实际上是由socket()系统调用返回的。参数command的取值由/usr/include/linux/sockios.h所规定。这些command的由于功能的不同，可分为以下几个小类：
• 改变路由表 (例如 SIOCADDRT, SIOCDELRT),
• 读/更新 ARP/RARP 缓存(如：SIOCDARP, SIOCSRARP),
• 一般的与网络接口有关的(例如 SIOCGIFNAME, SIOCSIFADDR 等等)
在Gooodies目录下有很多样例程序展示了如何使用ioctl。当你看这些程序时，注意参数argstruct是与参数command相关的。例如，与路由表相关的ioctl使用rtentry这种结构，rtentry定义在/usr/include/linux/route.h（参见例子adddefault.c）。与ARP有关的ioctl调用使用arpreq结构，arpreq定义在/usr/include/linux/if_arp.h（参见例子arpread.c）

[zhanxw]

与网络接口有关的ioctl调用使用的command参数通常看起来像SIOCxIFyyyy的形式，这里x要么是S（设定set，写write），要么是G（得到get，读read）。在getifinfo.c程序中就使用了这种形式的command参数来读IP地址，硬件地址，广播地址和得到与网络接口有关的一些标志（flag）。在这些ioctl调用中，第三个参数是ifreq结构，它在/usr/include/linux/if.h中定义。在某些情况下， ioctrl调用可能会使用到在sockios.h之外的新的定义，例如，WaveLAN无线网络卡会保存有关无线网络信号强度的信息，这对用户的程序可能有用。但用户怎么得到这种信息呢？我们的第一个本能是在sockios.h中定义新的ioctl命令，例如SIOCGIFWVLNSS（它的英文缩写表示WaveLAN的信号强度）。但不幸的是，这种命令不是对所有其他的网络接口（例如：loopback环回接口）有意义，而且不应当允许对于WAVLAN卡以外的网络接口使用ioctl命令。那么，我们需要的是这样一种机制：它能够定义一种与网络接口相关的ioctl命令。幸运的是，在Linux操作系统中已经为实现这个目的内建了一种挂钩（hook）机制。当你再次看sockios.h文件时，你将发现每一种设备已经预先定义了SIOCDEVPRIVATE的ioctl命令。而它的实现将留给开发相应驱动程序的人去完成。
通常，一个用户程序使用ioctl(sockid,SIOCDEVPRIVATE,(char*)&ifr)来调用与某种设备（指像WaveLAN那样的特殊设备）相关的ioctl命令，这里ifr是struct ifreq ifr形式的变量。用户程序应当在ifr.ifr_name中填充与这个设备相关的名字，例如，假设WaveLAN使用的接口号为eth1。一般的，一个用户程序还需要与内核互相交换ioctl的command参数和结果，这可以通过ifr.ifr_data这个变量来实现，例如，想得到WaveLAN中表示信号强度的信息时，可以通过返回这个变量来实现。Linux的源代码已经包括了两种设备de4x5和ewrk3，它们定义并且实现了特定的ioctl调用。这两个设备的源代码在de4x5.h,de4x5.c,ewrk3.h,ewrk3.c中（在/usr/src/linux/drivers/net/目录中）。这两种设备都定义了它们特有的结构（struct ewrk3_ioctl 和 struct de4x5_ioctl）来方便用户程序和设备驱动之间交换信息。每次调用ioctl前，用户程序应当在相应的结构变量中设定合适的初值，并且将ifr.ifr_data指向该值。

[zhanxw]

在我们进一步讨论ewrk3和de4x5的代码前，让我们仔细看看ioctl调用是如何一步步地实现的。所有的和接口相关的ioctl请求(SIOCxIFyyyy 和 SIOCDEVPRIVATE)将会调用dev_ioctl()（在/usr/src/linux/net/core/dev.c中）。但这只是一个包装器（wrapper），实际的动作将由dev_ifsioc()（也在dev.c中）来实现。差不多dev_ioctl()这个函数所做的所有工作只是检查这个调用是否已经有了正当的权限（例如，改变路由表需要有root的权限）。而dev_ifsioc()这个函数首先要做的一些事情包括得到与ifr.ifr_name相匹配的设备的结构（在/usr/include/linux/netdevice.h中定义）。但这是在实现特定的接口命令（例如：SIOCGIFADDR）之后。这些特定的接口命令被放置到一个巨大的switch语句之中。其中SIOCDEVPRIVATE命令和其他的在0x89F0到0x89FF之间的代码将出现在switch语句中的一个分支——default语句中。内核会检查表示设备的结构变量中，是否已经定义了一个与设备相关的ioctl句柄（handler）。这里的句柄是一个函数指针，它在表示设备的结构变量中do_ioctl部分。如果已经设置了这个句柄，那么内核将会执行它。
所以，如果要实现一个与设备相关的ioctl命令，所要做的只是编写一个与这个设备相关的ioctl句柄，并且将表示这个设备的结构变量中do_ioctl部分指向这个句柄。对于ewrk3这个设备，它的句柄是ewrk3_ioctl()（在ewrk3.c里面）并且相应的表示该设备的结构变量由ewrk3_init()来初始化。在ewrk3_ioctl()的代码中清晰的指出ifr.ifr_data是用作设备驱动程序和用户程序之间交换信息的。注意，这部分的内存可以双向的交流信息。例如，在ewrk3的驱动程序代码中，if.ifr_data的头两个字节是用来表示特殊的动作（例如，EWRK3_SET_PROM,EWRK3_CLR_PROM），而这个动作是符合使用者（驱动程序实现了多个与设备相关的、由SIOCDEVPRIVATE调用的命令）的要求的。另外，ifr.ifr_data中第5个字节指向的缓冲区(buffer)被用来交换其他的信息（如：当使用EWRK3_SET_HWADDR和EWRK3_GET_HWADDR时为硬件地址）

[zhanxw]

在你深入ewrk3_ioctl()时，请注意一般情况下一个用户进程不能直接访问内核所在的内存。为此，驱动开发者可以使用两个特殊的函数memcpy_tofs()和memcpy_fromfs()。内核函数memcpy_tofs(arg1, arg2, arg3) 从地址arg2（用户空间）向地址arg1(内核空间)拷贝arg3个字节。类似的，memcpy_fromfs(arg1,arg2,arg3)从地址arg2（用户空间）向地址arg1（内核空间）拷贝arg3个字节。在这些调用之前，verify_area()将会检查这个进程是否拥有合适的访问权限。另外，注意使用printk()函数可以输出debug信息。这个函数与printf()函数类似，但不能处理浮点类型的数。内核代码不能够使用printf()函数。printk()函数产生的结果将记录在/usr/adm/messages里。如果想知道更多的关于这些函数的或者与它们相关的信息，可以参考《Linux Kernel Hacker’s Guide》（在Linux文档网站的首页） 这本书中Supporting Functions部分。
翻译：Jun 4,2004 by zhanxw
联系：zhanxw00@hotmail.com

[听峰]

谢谢