Linux设备驱动之pci设备的枚举(linux初始化时PCI设备识别)

一:前言
Pci,是Peripheral Component  Interconnect的缩写,翻译成中文即为外部设备互联.与传统的总线相比.它的传输速率较高.能为用户提供动态查询pci  deivce.和局部总线信息的方法,此外,它还能自动为总线提供仲裁.在近几年的发展过程中,被广泛应用于多种平台.
pci协议比较复杂,关于它的详细说明,请查阅有关pci规范的资料,本文不会重复这些部份.
对于驱动工程师来说,Pci设备的枚举是pci设备驱动编写最复杂的操作.分析和理解这部份,是进行深入分析pci设备驱动架构的基础.
我们也顺便来研究一下,linux是怎么对这个庞然大物进行封装的.
二:pci架构概貌
上图展现了pci驱动架构中,pci_bus.pci_dev之间的关系.
如上图所示:所有的根总线都链接在pci_root_buses链表中. Pci_bus ->device链表链接着该总线下的所有设备.而pci_bus->children链表链接着它的下层总线.
对于pci_dev来说,pci_dev->bus指向它所属的pci_bus. Pci_dev->bus_list链接在它所属bus的device链表上.此外,所有pci设备都链接在pci_device链表中.
三:pci设备的配置空间
每个pci设备都有最多256个连续的配置空间.配置空间中包含了设备的厂商ID,设备ID,IRQ,设备存储区信息等.摘下LDD3中的一副说明图,如下:
要注意了,上图是以字节为单位的,而不是以位为单位.
那怎么去读取每个设备的配置空间呢?我们在开篇的时候提到过,pci总线为用户提供了动态查询pci设备信息的方法.
在x86上,保留了0xCF8~0xCFF的8个寄存器.实际上就是对应地址为0xCF8的32位寄存器和地址为0xCFC的32位寄存器.
在0xCF8寄存中写入要访问设备对应的总线号, 设备号、功能号和寄存器号组成的一个32位数写入0xCF8.然后从0xCFC上就可以取出对应pci设备的信息.
写入到0xCF8寄存器的格式如下:
低八位(0~7):            (寄存器地址)&0xFC.低二位为零
8~10:功能位.            有时候,一个pci设备对应多个功能.将每个功能单元分离出来,对应一个独立的pci device
11~15位:设备号        对应该pci总线上的设备序号
16~23位:总线号        根总线的总线号为0.每遍历到下层总线,总线号+1
31:有效位                    如果该位为1.则说明写入的数据有效,否则无效
例如:要读取n总线号m设备号f功能号对应设备的vendor id和Device id.过程如下:
要写入到0xCF8中的数为: l = 0x80<<23 | n<<16 | m<<11 | f<<8 | 0x00
即:outl(l,0xCF8)
从0xCFC中读相关信息:
L = Inw(0xCFC)  (从上图中看到,vendor id和device id总共占四个字节.因此用inw)
所以:device id = L&0xFF
         Vendor id = (L>>8)&0xFF
 
四:总线枚举入口分析
Pci的代码分为两个部份.一个部份是与平台相关的部份.存放在linux-2.6.25\arch\XXX\pci.在x86,对应为 linux-2.6.25\arch\x86\pci\  另一个部份是平台无关的代码,存放在linux-2.6.25\driver\pci\下面.
大致浏览一下这两个地方的init函数.发现可能枚举pci设备是由函数pcibios_scan_root()完成的.不过搜索源代码后,发 现有两个地方会调用这个调数.一个是在linux-2.6.25\arch\x86\pci\numa.c  另一个是linux-2.6.25\arch\x86\pci\Legacy.c
这两个地方都是封装在一个subsys_initcall()所引用的初始化函数呢? 到底哪一个文件才是我们要分析的呢?
分析一下linux-2.6.25\arch\x86\pci\下的Makefile_32.内容如下:
obj-y                             := i386.o init.o
 
obj-$(CONFIG_PCI_BIOS)                 += pcbios.o
obj-$(CONFIG_PCI_MMCONFIG)     += mmconfig_32.o direct.o mmconfig-shared.o
obj-$(CONFIG_PCI_DIRECT)  += direct.o
 
pci-y                             := fixup.o
pci-$(CONFIG_ACPI)                  += acpi.o
pci-y                             += legacy.o irq.o
 
pci-$(CONFIG_X86_VISWS)              := visws.o fixup.o
pci-$(CONFIG_X86_NUMAQ)            := numa.o irq.o
 
obj-y                             += $(pci-y) common.o early.o
从这个makefile中可以看出:legacy.c是一定会编译到了.而numa.c只有在编译选择了CONFIG_X86_NUMAQ的时候才起效.所以,我们可以毫不犹豫的将眼光放到了legacy.c中.
该文件中的初始化函数如下:
static int __init pci_legacy_init(void)
{
         if (!raw_pci_ops) {
                   printk("PCI: System does not support PCI\n");
                   return 0;
         }
 
         if (pcibios_scanned++)
                   return 0;
 
         printk("PCI: Probing PCI hardware\n");
         pci_root_bus = pcibios_scan_root(0);
         if (pci_root_bus)
                   pci_bus_add_devices(pci_root_bus);
 
         pcibios_fixup_peer_bridges();
 
         return 0;
}
 
subsys_initcall(pci_legacy_init);
由subsys_initcall()引用的函数都会放在init区域,这里面的函数是kernel启动的时候会自己执行的函数.首先我们碰到 的问题是raw_pci_ops是在什么地方被赋值的.搜索整个代码树,发现是在pci_access_init()中初始化的.如下:
static __init int pci_access_init(void)
{
         int type __maybe_unused = 0;
 
#ifdef CONFIG_PCI_DIRECT
         type = pci_direct_probe();
#endif
#ifdef CONFIG_PCI_MMCONFIG
         pci_mmcfg_init(type);
#endif
         if (raw_pci_ops)
                   return 0;
#ifdef CONFIG_PCI_BIOS
         pci_pcbios_init();
#endif
         /*
          * don't check for raw_pci_ops here because we want pcbios as last
          * fallback, yet it's needed to run first to set pcibios_last_bus
          * in case legacy PCI probing is used. otherwise detecting peer busses
          * fails.
          */
#ifdef CONFIG_PCI_DIRECT
         pci_direct_init(type);
#endif
         if (!raw_pci_ops)
                   printk(KERN_ERR
                   "PCI: Fatal: No config space access function found\n");
 
         return 0;
}
arch_initcall(pci_access_init);
由于arch_initcall()的优先级比subsys_initcall要高.因此,会先运行完pci_access_init之后,才会执行pci_legacy_init.
 
上面的代码看起来很复杂,没关系,去掉几个我们没有用到的编译代码就简单了.  在x86中,bios其实提供了pci设备的枚举功能.这也是CONFIG_PCI_BIOS的作用,如果对它进行了定义,那么就用bios的pci枚举 功能.如果没有定义,说明不采用bios的功能,而是自己手动去枚举,这就是CONFIG_PCI_DIRECT的作用.为了一般性,我们分析 CONFIG_PCI_DIRECT的过程.把其它不相关的代码略掉.剩余的就简单了.
 
在pci规范中,定义了两种操作配置空间的方法,即type1 和type2.在新的设计中,type2的配置机制不会被采用,通常会使用type1.因此,在代码中pci_direct_probe()一般会返回1,即使用type1.
pci_direct_init()的代码如下:
void __init pci_direct_init(int type)
{
         if (type == 0)
                   return;
         printk(KERN_INFO "PCI: Using configuration type %d\n", type);
         if (type == 1)
                   raw_pci_ops = &pci_direct_conf1;
         else
                   raw_pci_ops = &pci_direct_conf2;
}
在这里看到,ram_pci_ops最终会指向pci_direct_conf1.顺便看下这个结构:
struct pci_raw_ops pci_direct_conf1 = {
         .read =               pci_conf1_read,
         .write =      pci_conf1_write,
};
这个结构其实就是pci设备配置空间操作的接口.
 
五:pci设备的枚举过程
返回到pci_legacy_init()中
static int __init pci_legacy_init(void)
{
         ……
         printk("PCI: Probing PCI hardware\n");
         pci_root_bus = pcibios_scan_root(0);
         if (pci_root_bus)
                   pci_bus_add_devices(pci_root_bus);
         ……
}
Pci设备的枚举过程是由pcibios_scan_root()完成的.在这里调用是以0为参数.说明是从根总线起开始枚举.
 
pcibios_scan_root()代码如下:
struct pci_bus * __devinit pcibios_scan_root(int busnum)
{
         struct pci_bus *bus = NULL;
         struct pci_sysdata *sd;
 
         dmi_check_system(pciprobe_dmi_table);
         while ((bus = pci_find_next_bus(bus)) != NULL) {
                   if (bus->number == busnum) {
                            /* Already scanned */
                            return bus;
                   }
         }
 
         /* Allocate per-root-bus (not per bus) arch-specific data.
          * TODO: leak; this memory is never freed.
          * It's arguable whether it's worth the trouble to care.
          */
         sd = kzalloc(sizeof(*sd), GFP_KERNEL);
         if (!sd) {
                   printk(KERN_ERR "PCI: OOM, not probing PCI bus %02x\n", busnum);
                   return NULL;
         }
 
         printk(KERN_DEBUG "PCI: Probing PCI hardware (bus %02x)\n", busnum);
 
         return pci_scan_bus_parented(NULL, busnum, &pci_root_ops, sd);
}
 
先在pci_root_buses中判断是否存在这个根总线对应的总线号.如果存在,说明这条总线已经遍历过了,直接退出.
Pci_root_ops这是定义的pci设备配置空间的操作.在没有选择CONFIG_PCI_MMCONFIG的情况下,它的操作都会转入我们在上面的分析的,ram_pci_ops中.这个过程非常简单,可以自行分析.
然后,流程转入pci_scan_bus_parented().代码如下:
struct pci_bus * __devinit pci_scan_bus_parented(struct device *parent,
                   int bus, struct pci_ops *ops, void *sysdata)
{
         struct pci_bus *b;
 
         b = pci_create_bus(parent, bus, ops, sysdata);
         if (b)
                   b->subordinate = pci_scan_child_bus(b);
         return b;
}
在pci_create_bus()中,为对应总线号构建pci_bus,然后将其挂入到pci_root_buses链表.该函数代码比较简 单,请自行分析.然后,调用然后pci_scan_child_bus枚举该总线下的所有设备.pci_bus->subordinate表示下流 总线的最大总线号.pci_sacn_child_bus()代码如下:
 
 
unsigned int __devinit pci_scan_child_bus(struct pci_bus *bus)
{
         unsigned int devfn, pass, max = bus->secondary;
         struct pci_dev *dev;
 
         pr_debug("PCI: Scanning bus %04x:%02x\n", pci_domain_nr(bus), bus->number);
 
         /* Go find them, Rover! */
         //按功能号扫描设备号对应的pci 设备
         for (devfn = 0; devfn < 0x100; devfn += 8)
                   pci_scan_slot(bus, devfn);
 
         /*
          * After performing arch-dependent fixup of the bus, look behind
          * all PCI-to-PCI bridges on this bus.
          */
         pr_debug("PCI: Fixups for bus %04x:%02x\n", pci_domain_nr(bus), bus->number);
         pcibios_fixup_bus(bus);
         for (pass=0; pass < 2; pass++)
                   list_for_each_entry(dev, &bus->devices, bus_list) {
                            if (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE ||
                                dev->hdr_type == PCI_HEADER_TYPE_CARDBUS)
                                     max = pci_scan_bridge(bus, dev, max, pass);
                   }
 
         /*
          * We've scanned the bus and so we know all about what's on
          * the other side of any bridges that may be on this bus plus
          * any devices.
          *
          * Return how far we've got finding sub-buses.
          */
         pr_debug("PCI: Bus scan for %04x:%02x returning with max=%02x\n",
                   pci_domain_nr(bus), bus->number, max);
         return max;
}
这节的难点就是在这个地方了,从我们之前分析的pci设备配置空间的读写方式可得知.对特定总线.下面最多个32个设备号.每个设备号又对应8 个功能号.我们可以将设备号和功能号放到一起,即占8~15位.在这面的代码中.对每个设备号调用pci_scan_slot()去扫描它下面的8个功能 号对应的设备.总而言之,把该总线下面的所有设备都要枚举完.
 
pci_scan_slot()代码如下:
nt pci_scan_slot(struct pci_bus *bus, int devfn)
{
         int func, nr = 0;
         int scan_all_fns;
 
         scan_all_fns = pcibios_scan_all_fns(bus, devfn);
 
         for (func = 0; func < 8; func++, devfn++) {
                   struct pci_dev *dev;
 
                   dev = pci_scan_single_device(bus, devfn);
                   if (dev) {
                            nr++;
 
                            /*
                            * If this is a single function device,
                            * don't scan past the first function.
                            */
                            if (!dev->multifunction) {
                                     if (func > 0) {
                                               dev->multifunction = 1;
                                     } else {
                                              break;
                                     }
                            }
                   } else {
                            if (func == 0 && !scan_all_fns)
                                     break;
                   }
         }
         return nr;
}
对其它的每个设备都会调用pci_scan_single_device().如果是单功能设备(dev->multifunction == 0).则只要判断它的第一个功能号可以了,不需要判断之后功能号对应的设备.
Pci_scan_single_device()代码如下:
struct pci_dev *__ref pci_scan_single_device(struct pci_bus *bus, int devfn)
{
         struct pci_dev *dev;
 
         dev = pci_scan_device(bus, devfn);
         if (!dev)
                   return NULL;
 
         //将pci_dev加至pci_bus->devices
         pci_device_add(dev, bus);
 
         return dev;
}
对每个设备,都会调用pci_scan_device()执行扫描的过程,如果该设备存在,就会将该设备加入到所属总线的devices链表 上.这是在pci_device_add()函数中完成的,这个函数比较简单.这里不做详细分析.我们把注意力集中到 pci_scan_device(),这函数有点长,分段分析如下:
 
static struct pci_dev * __devinit
pci_scan_device(struct pci_bus *bus, int devfn)
{
         struct pci_dev *dev;
         u32 l;
         u8 hdr_type;
         int delay = 1;
 
         if (pci_bus_read_config_dword(bus, devfn, PCI_VENDOR_ID, &l))
                   return NULL;
 
         /* some broken boards return 0 or ~0 if a slot is empty: */
         if (l == 0xffffffff || l == 0x00000000 ||
             l == 0x0000ffff || l == 0xffff0000)
                   return NULL;
 
         /* Configuration request Retry Status */
         while (l == 0xffff0001) {
                   msleep(delay);
                   delay *= 2;
                   if (pci_bus_read_config_dword(bus, devfn, PCI_VENDOR_ID, &l))
                            return NULL;
                   /* Card hasn't responded in 60 seconds?  Must be stuck. */
                   if (delay > 60 * 1000) {
                            printk(KERN_WARNING "Device %04x:%02x:%02x.%d not "
                                               "responding\n", pci_domain_nr(bus),
                                               bus->number, PCI_SLOT(devfn),
                                               PCI_FUNC(devfn));
                            return NULL;
                   }
         }
从配置空间中读取该设备对应的vendor id和device id.如果读出来的值,有一个是空的,则说明该功能号对应的设备不存在,或者是配置非法.
如果读出来的是0xffff0001.则需要重新读一次,如果重读次数过多,也会退出
 
         if (pci_bus_read_config_byte(bus, devfn, PCI_HEADER_TYPE, &hdr_type))
                   return NULL;
 
         dev = alloc_pci_dev();
         if (!dev)
                   return NULL;
 
         dev->bus = bus;
         dev->sysdata = bus->sysdata;
         dev->dev.parent = bus->bridge;
         dev->dev.bus = &pci_bus_type;
         dev->devfn = devfn;
         dev->hdr_type = hdr_type & 0x7f;
         dev->multifunction = !!(hdr_type & 0x80);
         dev->vendor = l & 0xffff;
         dev->device = (l >> 16) & 0xffff;
         dev->cfg_size = pci_cfg_space_size(dev);
         dev->error_state = pci_channel_io_normal;
         set_pcie_port_type(dev);
 
         /* Assume 32-bit PCI; let 64-bit PCI cards (which are far rarer)
            set this higher, assuming the system even supports it.  */
         dev->dma_mask = 0xffffffff;
接着,将不同类型设备的共同头部配置读出来,然后赋值给pci_dev的相应成员.这里有个特别要值得注意的地方:  dev->dev.bus = & pci_bus_type.即将pci_dev里面封装的device结构的bus设置为了pci_bus_type.这个是很核心的一个步 骤.我们先将它放到这里,之后的再来详细分析
特别的, HEADER_TYPE的最高位为0,表示该设备是一个单功能设备
 
         if (pci_setup_device(dev) < 0) {
                   kfree(dev);
                   return NULL;
         }
 
         return dev;
}
最后,流程就会转入到pci_setup_deivce()对特定类型的设备配置都行读取操作了.代码如下:
static int pci_setup_device(struct pci_dev * dev)
{
         u32 class;
 
         sprintf(pci_name(dev), "%04x:%02x:%02x.%d", pci_domain_nr(dev->bus),
                   dev->bus->number, PCI_SLOT(dev->devfn), PCI_FUNC(dev->devfn));
 
         pci_read_config_dword(dev, PCI_CLASS_REVISION, &class);
         dev->revision = class & 0xff;
         class >>= 8;                                      /* upper 3 bytes */
         dev->class = class;
         class >>= 8;
 
         pr_debug("PCI: Found %s [%04x/%04x] %06x %02x\n", pci_name(dev),
                    dev->vendor, dev->device, class, dev->hdr_type);
 
         /* "Unknown power state" */
         dev->current_state = PCI_UNKNOWN;
 
         /* Early fixups, before probing the BARs */
         pci_fixup_device(pci_fixup_early, dev);
         class = dev->class >> 8;
 
         switch (dev->hdr_type) {                 /* header type */
         case PCI_HEADER_TYPE_NORMAL:                    /* standard header */
                   if (class == PCI_CLASS_BRIDGE_PCI)
                            goto bad;
                   pci_read_irq(dev);
                   pci_read_bases(dev, 6, PCI_ROM_ADDRESS);
                   pci_read_config_word(dev, PCI_SUBSYSTEM_VENDOR_ID, &dev->subsystem_vendor);
                   pci_read_config_word(dev, PCI_SUBSYSTEM_ID, &dev->subsystem_device);
 
                   /*
                    *      Do the ugly legacy mode stuff here rather than broken chip
                    *      quirk code. Legacy mode ATA controllers have fixed
                    *      addresses. These are not always echoed in BAR0-3, and
                    *      BAR0-3 in a few cases contain junk!
                    */
                   if (class == PCI_CLASS_STORAGE_IDE) {
                            u8 progif;
                            pci_read_config_byte(dev, PCI_CLASS_PROG, &progif);
                            if ((progif & 1) == 0) {
                                     dev->resource[0].start = 0x1F0;
                                     dev->resource[0].end = 0x1F7;
                                     dev->resource[0].flags = LEGACY_IO_RESOURCE;
                                     dev->resource[1].start = 0x3F6;
                                     dev->resource[1].end = 0x3F6;
                                     dev->resource[1].flags = LEGACY_IO_RESOURCE;
                            }
                            if ((progif & 4) == 0) {
                                     dev->resource[2].start = 0x170;
                                     dev->resource[2].end = 0x177;
                                     dev->resource[2].flags = LEGACY_IO_RESOURCE;
                                     dev->resource[3].start = 0x376;
                                     dev->resource[3].end = 0x376;
                                     dev->resource[3].flags = LEGACY_IO_RESOURCE;
                            }
                   }
                   break;
 
         case PCI_HEADER_TYPE_BRIDGE:                     /* bridge header */
                   if (class != PCI_CLASS_BRIDGE_PCI)
                            goto bad;
                   /* The PCI-to-PCI bridge spec requires that subtractive
                      decoding (i.e. transparent) bridge must have programming
                      interface code of 0x01. */
                   pci_read_irq(dev);
                   dev->transparent = ((dev->class & 0xff) == 1);
                   pci_read_bases(dev, 2, PCI_ROM_ADDRESS1);
                   break;
 
         case PCI_HEADER_TYPE_CARDBUS:                 /* CardBus bridge header */
                   if (class != PCI_CLASS_BRIDGE_CARDBUS)
                            goto bad;
                   pci_read_irq(dev);
                   pci_read_bases(dev, 1, 0);
                   pci_read_config_word(dev, PCI_CB_SUBSYSTEM_VENDOR_ID, &dev->subsystem_vendor);
                   pci_read_config_word(dev, PCI_CB_SUBSYSTEM_ID, &dev->subsystem_device);
                   break;
 
         default:                                     /* unknown header */
                   printk(KERN_ERR "PCI: device %s has unknown header type %02x, ignoring.\n",
                            pci_name(dev), dev->hdr_type);
                   return -1;
 
         bad:
                   printk(KERN_ERR "PCI: %s: class %x doesn't match header type %02x. Ignoring class.\n",
                          pci_name(dev), class, dev->hdr_type);
                   dev->class = PCI_CLASS_NOT_DEFINED;
         }
 
         /* We found a fine healthy device, go go go... */
         return 0;
}
总共有三种类型的设备,分别为常规设备(PCI_HEADER_TYPE_NORMAL)  ,pci-pci桥设备(PCI_HEADER_TYPE_BRIDGE),笔记本电脑上使用的 cardbus(PCI_HEADER_TYPE_CARDBUS).这里的操作不外乎是IRQ的确定,设备存储区间映射等.先将这几个操作分析如下:
1: IRQ号的确定
该操作接口为pci_read_irq():
static void pci_read_irq(struct pci_dev *dev)
{
         unsigned char irq;
 
         pci_read_config_byte(dev, PCI_INTERRUPT_PIN, &irq);
         dev->pin = irq;
         if (irq)
                   pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &irq);
         dev->irq = irq;
}
在PCI_INTERRUPT_PIN中存放的是将INTA~INTD的哪一个引脚连接到了中断控制器,如果该值为零.说明并末将引脚连接至中断控制器.自然也就不能产生中断信号.
其实,在PCI_INTERRUPT_LINE存放的是该设备的中断线连接在中断控制器的哪一个IRQ线上.也就是对应设备的IRQ.
注意这里的寄存器只读有意义,并不是更改寄存器的值就更改该设备的IRQ
 
2:内部存储区间的确定
从之前的pci设备配置寄存器图中可以看到.有从0x10~0x27的6个base address寄存器.里面存放的就是内部存储器的地起地址和长度,及其类型.
首先将对应寄存器的值取出.如果最低位为1.则说明该区域是I/O端口,高29位是端口地址的高29位,低3位为零.否则是存储映射区间.前28位是存储区的高28位,低四位为零.
然后,将该寄存器全部置1.再读,取得的是长度信息. 如果是I/O端口,屏弊其低三位,如果是存储区间,屏弊其低四位.最后取第1个位为1对应的大小,即为相应区间的长度.
例如,取出来的值是0xC107.假设是I/O端口
屏蔽掉低三位,为0xC100.第一个为1的值对应的值为0x0100.即0x100
另上,ROM的操作也跟此类似.
在上面的代码中,内部存储区间的确定是由pci_read_bases()完成的.这个函数代码比较长.涉及到的东西又不多,因此不做详细分析.结构上面的分析,应该很容易看懂代码了.
 
从上面的代码可以看出,对于常规设备,有6个存储区间和一个ROM。Pci briage只有2个存储区间和一个ROM。Cardbus只有一个存储区间没有ROM。
好了,再这里,每一类设备的信息都已经完全读取出来了,并存放在pci_dev的相关字段。此后在驱动中就可以直接找到pci_dev.取得相应的信息,而不需要再次去枚举了.
 
再这里,万里长征只是迈出了一小步。我们知道,pci总线可以通过pci bridge再连一层pci总线。这个问题显然是一个递归过程。我们接下来看pci桥的处理。
返回到pci_scan_child_bus()中。我们将下面要分析的代码列出来:
unsigned int __devinit pci_scan_child_bus(struct pci_bus *bus)
{
         ……
         ……
         for (devfn = 0; devfn < 0x100; devfn += 8)
                   pci_scan_slot(bus, devfn);
         pcibios_fixup_bus(bus);
         for (pass=0; pass < 2; pass++)
                   list_for_each_entry(dev, &bus->devices, bus_list) {
                            if (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE ||
                                dev->hdr_type == PCI_HEADER_TYPE_CARDBUS)
                                     max = pci_scan_bridge(bus, dev, max, pass);
                   }
         ……
}
Pcibios_fixup_bus()这个函数看名字是用来修正总线的。芯片厂商在发布产品后,又检测到上次发布的产品有问题。回厂升级是不 可能的了。只能提供软件修改的手段,发布一些修正包。Linux将很多厂商的修改正集合在一起。这也就是pcibios_fixup_bus()要进行的 操作。具体设备的修正功能,我们就不再研究了。这个函数里还有一个重要的操作。列出代码如下:
void __devinit  pcibios_fixup_bus(struct pci_bus *b)
{
         struct pci_dev *dev;
 
         pcibios_fixup_ghosts(b);
         pci_read_bridge_bases(b);
         list_for_each_entry(dev, &b->devices, bus_list)
                   pcibios_fixup_device_resources(dev);
}
我们所在讨论的重要的操作就是在pci_read_bridge_bases()中完成的。除了之上分析的配置字段外,其实pci桥还有一个很重要的配置项。即:过滤窗口。
过滤窗口决定了访问的方向。例如:如果cpu一侧要经过pci bridge访问pci总线,则它的地址必须要落在这个pci桥的过滤窗口内才可以通过。另外,pci bridge下游的pci bus要访问cpu侧。则地址必须要落在过滤窗口外才可以。
此外,pci  bridge还提供了一个命令寄存器来控制“memory access enable“和“I/O access  enable”两个位来控制两个功能。如果全为0.则两个方向都会关闭。在pci初始化前,为了防止对cpu侧造成干扰, 这两个功能都关闭的,
Pci bridge有三个这样的窗口,分别如下:
1:起始地址在PCI_IO_BASE中,长度在PCI_IO_LIMIT中。如果是32位,还要通过PCI_IO_BASE_UPPER16和PCI_IO_LIMIT_UPPER16提供高16位。
2:起始地址在PCI_MEMORY_BASE,长度在PCI_MEMORY_LIMIT中。这个是一个16位的窗口。
3:起始地址在PCI_PREF_MEMORY_BASE,长度在PCI_PREF_MEMORY_LIMIT.默认是32位。如果是64,则需要PCI_PREF_BASE_UPPER32和PCI_PREF_LIMIT_UPPER32提供高32位.
存储区间在这里看起来有点繁杂。以图的形式总结如下:
 
结合上面说的,理解pci_read_bridge_bases()的代码就不难了。这里不再做详细分析。
 
现在终于把应该读的配置读完了,可以进行下层pci总线的遍历了。
列出这段代码:
unsigned int __devinit pci_scan_child_bus(struct pci_bus *bus)
{
……
         ……
for (pass=0; pass < 2; pass++)
                   list_for_each_entry(dev, &bus->devices, bus_list) {
                            if (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE ||
                                dev->hdr_type == PCI_HEADER_TYPE_CARDBUS)
                                     max = pci_scan_bridge(bus, dev, max, pass);
                   }
         ……
         ……
}
上面的操作基本上就是遍历挂在pci_bus->devices上面的设备(是否还记得上面在分析的时候,每枚举到一个设备都会加入到 pci_bus->device呢*^_^*)。如果是pci桥或者是cardbus。就会调用pci_scan_bridge()来遍历桥下面的 设备.
这里让人疑惑的是,为什么要遍历二次呢?
这是因为,在x86上,系统启动的时候,bios会枚举一次pci设备。所以有些pci  bridge是经过bios处理过的。而有些是bios可能没有枚举到的。这就需要分两次处理。一次来处理那里已经由bios处理过的pci  bridge.一次是处理全新的pci bridge.这样做,这样做是因为,每次枚举总线后,要为其分配一个总线号,而bios处理后的pci  bridge的总线号全部都由bios分配好了,要为新的pci bridge分配总线号。而必须要处理完旧的pci  bridge才会知道可用的总线号是多少。
跟进pci_sacn_bridge()的代码,这段代码较长,分段分析如下:
 
int __devinit pci_scan_bridge(struct pci_bus *bus, struct pci_dev *dev, int max, int pass)
{
         struct pci_bus *child;
         int is_cardbus = (dev->hdr_type == PCI_HEADER_TYPE_CARDBUS);
         u32 buses, i, j = 0;
         u16 bctl;
 
         pci_read_config_dword(dev, PCI_PRIMARY_BUS, &buses);
 
         pr_debug("PCI: Scanning behind PCI bridge %s, config %06x, pass %d\n",
                    pci_name(dev), buses & 0xffffff, pass);
 
         /* Disable MasterAbortMode during probing to avoid reporting
            of bus errors (in some architectures) */
         pci_read_config_word(dev, PCI_BRIDGE_CONTROL, &bctl);
         pci_write_config_word(dev, PCI_BRIDGE_CONTROL,
                                  bctl & ~PCI_BRIDGE_CTL_MASTER_ABORT);
         if ((buses & 0xffff00) && !pcibios_assign_all_busses() && !is_cardbus) {
                   unsigned int cmax, busnr;
                   /*
                    * Bus already configured by firmware, process it in the first
                    * pass and just note the configuration.
                    */
                   if (pass)
                            goto out;
                   //次总线号,也就是下游pci的总线号
                   busnr = (buses >> 8) & 0xFF;
 
                   /*
                    * If we already got to this bus through a different bridge,
                    * ignore it.  This can happen with the i450NX chipset.
                    */
                    //如果该总线号已经在我们的遍历树中了,说明这条总线已经是处理过的
                   if (pci_find_bus(pci_domain_nr(bus), busnr)) {
                            printk(KERN_INFO "PCI: Bus %04x:%02x already known\n",
                                               pci_domain_nr(bus), busnr);
                            goto out;
                   }
 
                   //构造一个pci_bus.并将其链入到父总线的children链表上
                   child = pci_add_new_bus(bus, dev, busnr);
                   if (!child)
                            goto out;
                   child->primary = buses & 0xFF;
                   child->subordinate = (buses >> 16) & 0xFF;
                   child->bridge_ctl = bctl;
 
                   //递归遍历子总线.返回的是下层最大的总线号
                   cmax = pci_scan_child_bus(child);
                   if (cmax > max)
                            max = cmax;
                   if (child->subordinate > max)
                            max = child->subordinate;
         } else {
                   /*
                    * We need to assign a number to this bus which we always
                    * do in the second pass.
                    */
                    //在第一次遍历的时候。是不会处理新的pci bus的
                   if (!pass) {
                            if (pcibios_assign_all_busses())
                                     /* Temporarily disable forwarding of the
                                        configuration cycles on all bridges in
                                        this bus segment to avoid possible
                                        conflicts in the second pass between two
                                        bridges programmed with overlapping
                                        bus ranges. */
                                     pci_write_config_dword(dev, PCI_PRIMARY_BUS,
                                                               buses & ~0xffffff);
                            goto out;
                   }
 
                   /* Clear errors */
                  //往状态寄存器全写1
                   pci_write_config_word(dev, PCI_STATUS, 0xffff);
 
                   /* Prevent assigning a bus number that already exists.
                    * This can happen when a bridge is hot-plugged */
                    //要处理的总线编号是否已经存在了
                   if (pci_find_bus(pci_domain_nr(bus), max+1))
                            goto out;
                   child = pci_add_new_bus(bus, dev, ++max);
                   buses = (buses & 0xff000000)
                         | ((unsigned int)(child->primary)     <<  0)
                         | ((unsigned int)(child->secondary)   <<  8)
                         | ((unsigned int)(child->subordinate) << 16);
 
                   /*
                    * yenta.c forces a secondary latency timer of 176.
                    * Copy that behaviour here.
                    */
                   if (is_cardbus) {
                            buses &= ~0xff000000;
                            buses |= CARDBUS_LATENCY_TIMER << 24;
                   }
                           
                  
                  pci_write_config_dword(dev, PCI_PRIMARY_BUS, buses);
 
                   if (!is_cardbus) {
                            child->bridge_ctl = bctl;
                           
                            pci_fixup_parent_subordinate_busnr(child, max);
                            /* Now we can scan all subordinate buses... */
                            max = pci_scan_child_bus(child);
                            pci_fixup_parent_subordinate_busnr(child, max);
                   } else {
                                     for (i=0; i
                                     struct pci_bus *parent = bus;
                                     if (pci_find_bus(pci_domain_nr(bus),
                                                                 max+i+1))
                                               break;
                                     while (parent->parent) {
                                               if ((!pcibios_assign_all_busses()) &&
                                                   (parent->subordinate > max) &&
                                                   (parent->subordinate <= max+i)) {
                                                        j = 1;
                                               }
                                               parent = parent->parent;
                                     }
                                     if (j) {
                                               i /= 2;
                                               break;
                                     }
                            }
                            max += i;
                            pci_fixup_parent_subordinate_busnr(child, max);
                   }
                   child->subordinate = max;
                   pci_write_config_byte(dev, PCI_SUBORDINATE_BUS, max);
         }
 
         sprintf(child->name, (is_cardbus ? "PCI CardBus #%02x" : "PCI Bus #%02x"), child->number);
 
         /* Has only triggered on CardBus, fixup is in yenta_socket */
         while (bus->parent) {
                   if ((child->subordinate > bus->subordinate) ||
                       (child->number > bus->subordinate) ||
                       (child->number number) ||
                       (child->subordinate number)) {
                            pr_debug("PCI: Bus #%02x (-#%02x) is %s "
                                     "hidden behind%s bridge #%02x (-#%02x)\n",
                                     child->number, child->subordinate,
                                     (bus->number > child->subordinate &&
                                      bus->subordinate number) ?
                                               "wholly" : "partially",
                                     bus->self->transparent ? " transparent" : "",
                                     bus->number, bus->subordinate);
                   }
                   bus = bus->parent;
         }
 
out:
         pci_write_config_word(dev, PCI_BRIDGE_CONTROL, bctl);
 
         return max;
}
忽略有关cardbus总线相关的部份。PCI_PRIMARY_BUS寄存器中的值的含义为:从低位到高位分别为:主总线号,次总线号,子层 最大线号。各占两位。如果从PCI_PRIMARY_BUS取出来的值,次总线号和子层最大线号有意义。说明该pci-bridge是被bios处理过 的。 这里为什么没有判断主总线号的值呢?这是因为总主线号可能为零,如根总线。而次总线是每枚举到一个pci-bridge就会增1。
如果该pci bridge是被bios处理过的,那直接构造一个pci_bus(pci_add_new_bus()).再递归枚举这个pci_bus下的设备就可以了。
相反,如果该pci-bridge没有被bios处理过,那就需要我们手动去处理了。这时,为它分配一个可能的总线号。然后将总线号写入PCI_PRIMARY_BUS寄存器。再构造一个pci_bus.递归枚举其下的设备。
最后的一个while()循环是打印出一些DEBUG信息。不需要理会
 
TODO:在上述代码红色标识部份。如果是没有由bios处理过的pci-bridge.有效设置只是次总线号和下层最大总线号。难道不需要设置主总线号么?
特别注意.在遍其次级总线下层pci时.此时还不知道下层最大总号是多少.所以将pci_bus->subordinate赋值为了 0xFF.即其下的所有设备都可以透过这个pci-bridge(参考pci_alloc_child_bus()中的处理).然后等下层的子总线遍历完 了之后,再来确定子总线的最大总线号,将其更新至pci_bus->subordinate.
对于上次主总线号的疑问:其实这个总主线号在访问下层总线是不会被使用到的,因为在配置的时候,只会比较pci_bridge的次总线号和下层最大总线号.如果总线号落在这个区间中间.将其透传到下一层总线.否则,忽略这个请求.
递归完成之后,pci总线上的所有信息都被找到了。所有pci_bus被存放在pci_root_buses为根的倒立树中.而总线上对应的pci 设备存放在pci_bus->device链表中.
返回到我们开篇时的初始化函数pci_legacy_init().这个函数还剩下一部份.代码如下:
static int __init pci_legacy_init(void)
{
         if (!raw_pci_ops) {
                   printk("PCI: System does not support PCI\n");
                   return 0;
         }
 
         if (pcibios_scanned++)
                   return 0;
 
         printk("PCI: Probing PCI hardware\n");
         pci_root_bus = pcibios_scan_root(0);
         if (pci_root_bus)
                   pci_bus_add_devices(pci_root_bus);
 
         pcibios_fixup_peer_bridges();
 
         return 0;
}
如果总线遍历成功,就会转入pci_bus_add_devices().代码片段如下:
void pci_bus_add_devices(struct pci_bus *bus)
{
         ……
         ……
pci_bus_add_device(dev)
……
……
}
该函数会遍历总线上的device链表.然后对每个设备调用pci_bus_add_device().代码如下:
int pci_bus_add_device(struct pci_dev *dev)
{
         int retval;
         retval = device_add(&dev->dev);
         if (retval)
                   return retval;
 
         down_write(&pci_bus_sem);
         list_add_tail(&dev->global_list, &pci_devices);
         up_write(&pci_bus_sem);
 
         pci_proc_attach_device(dev);
         pci_create_sysfs_dev_files(dev);
         return 0;
}
从上面的代码可以看出.它先将设备添加,然后再将设备挂载到全局的pci_devices链表上.
这样顺着pci_devices就可以找到所有的设备信息了.
另外,对于pci_dev的初始化,我们之前曾强调过.初始化代码片段如下:
static struct pci_dev * __devinit
pci_scan_device(struct pci_bus *bus, int devfn)
{
         ……
         ……
         dev->dev.parent = bus->bridge;
         dev->dev.bus = &pci_bus_type;
         ……
         ……
}
即该设备是属于pci_bus_type总线的.事实上,我们编写的pci驱动程序,也是基于pci_bus_type.这样,就可以在添加驱动的时候,就可以匹配想要的设备了.
在这里,特别注意一下,经过这里的枚举,只是枚举完了第一条根总线.那其它的根总线是在什么地方被枚举的呢? 在下一节里,再来分析这个问题.
 
六:小结
在linux的pci架构中,大量运用了深度优先遍历算法.这是由pci总线结构所决定的.经过这一章的分析过后,我们应该对pci架构有了一定的了解.同时在这一章还留下了一个问题.到下一节再来进行分析.

你可能感兴趣的:(linux)