Chinaunix首页 | 论坛 | 博客
  • 博客访问: 1940348
  • 博文数量: 1000
  • 博客积分: 0
  • 博客等级: 民兵
  • 技术积分: 7921
  • 用 户 组: 普通用户
  • 注册时间: 2013-08-20 09:23
个人简介

storage R&D guy.

文章分类

全部博文(1000)

文章存档

2019年(5)

2017年(47)

2016年(38)

2015年(539)

2014年(193)

2013年(178)

分类: LINUX

2013-09-10 10:12:38

 
Pci设备的I/O和内存是一个比较复杂的问题.如下的总线结构:
在上图的总线结构中,ethernet设备和pci-pci bridge的同类型资源空间必须要是pci bus0的一个子集
例如,pci bus 0的I/O端口资源是0x00CC~0x01CC. Ethernet设备的I/O范围的是0x00CC~0x0xE0.那么pci-pci bridge的I/O端口范围就必须要在0x0xE0~0x01CC之间.
同样,SCSI和VIDEO同类型资源必须要是pci_bus1的子集.pci bus1上有一个pci桥,对应的资源也就是它所连桥上的资源.即pci_bus->self.
也就是说,下层总线的资源是它上层总线资源的子集。上层总线资源是下层总线资源的父集。
其实,每个PCI设备的资源地始地址都是由操作系统设置的.在x86上,都由bios设置好了.假若没有bios的时候,我们应该怎么去设置设备的资源起始范围呢?
可能在pci枚举完成之后:
1:从根总线开始,设置根总线的资源范围是从0开始,到0xFFFF或者0xFFFFFFFF的最大范围.
2:对其它的设备,可往其资源寄存器全部写入1,就可以求得该资源项的类型和长度.
3:设备从根总线的资源那里分得对应长度的资源.
4:如果设备是pci-pci bridge,则递归配置它.
 
可能有人会有这样迷惑,对应于上图,如果pci-pci bridge的资源大小是N.而SCSI和video资源范围超过了N怎么办呢?
我们必须要注意一点,总线的区间是可以自已设定的,而设备资源的区间是在设计的时候就已经确定好了.也就是说,我们可以更改pci device区间的起始地址,但我们不能改变它的大小.
因此,出现了上面所说的这种情况.可能是由bios在处理PCI的时候出现了BUG.我们需要调整总线的资源区间.
 
其实对于pci_bus的资源范围就是它的过滤窗口.对于过滤窗口的作用,我们在枚举的时候分析的很清楚了.
 
CPU访问PC过程是这样的(只有一个根总线和pci-pci bridge过滤窗口功能打开的情况):
1:cpu向pci发出一个I/O请求.首先经过根总线.它会判断是否在它的资源范围内.如果在它的范围,它就会丢向总线所在的每一个设备.包括pci bridge. 如果没有在根总线的资源范围,则不会处理这个请求.
2:如果pci设备判断该地址属于它的资源范围,则处理后发出应答
4:pci bridge接收到这个请求,它会判断I/O地址是否在它的资源范围内.如果在它的范围,它会把它丢到它的下层子线.
5:下层总线经过经过相同的处理后,就会找到这个PCI设备了
 
一个PCI设备访问其它PCI设备或者其它外设的过程:
1:首先这个PCI发出一个请求,这个请求会在总线上广播
2:如果要请求的设备是在同级总线,就会产生应答
3:请求的设备不是在同层总线,就会进行pci bridge.pci桥判断该请求不在它的范围内(目的地不是它下层的设备),就会将它丢向上层.
4:这样辗转之后,就能找到对应的设备了
 
经过这样的分析过来,相信对pci bridge的过滤窗口有更深的理解了.
 
Linux中使用struct resource的结构来表示I/O端口或者是设备内存。定义如下:
struct resource {
         resource_size_t start;
         resource_size_t end;
         const char *name;
         unsigned long flags;
         struct resource *parent, *sibling, *child;
};
Start: 表示它所占资源的起始地址。
End: 表示它所占资源的未尾地址
Name: 所占资源的名字
Flags: 资源的类型。目前有I/O和内存两种
Parent.sibling.child:用来表示资源的所属关系。分别表示它的父结点,兄弟结点和子结点。
 
从前面的分析可以看到,有一些总线可能bios没有遍历到或许bios的处理有错误,所以需要对整个系统的PCI总线和PCI设备的资源占用情况遍历一次。完整的建立上述的struct resource结构(在之前枚举的时候,只是处理了start和end成员).。这个过程是在pcibios_resource_survey( )完成的。如下所示:
 
subsys_initcall(pcibios_init);
static int __init pcibios_init(void)
{
         ……
         …….
         pcibios_resource_survey();
}
 
pcibios_init这个函数是被fs_initcall()所描述的。在kernel启动的时候,会调用宏所描述的函数。在pcibios_init ()又会调用pcibios_assign_resources(),它的代码如下所示:
void __init pcibios_resource_survey(void)
{
         DBG("PCI: Allocating resources\n");
         pcibios_allocate_bus_resources(&pci_root_buses);
         pcibios_allocate_resources(0);
         pcibios_allocate_resources(1);
}
它先对总线的资源进行处理。然后再对PCI设备的资源进行处理。我们先看pcibios_allocate_bus_resources()
static void __init pcibios_allocate_bus_resources(struct list_head *bus_list)
{
         struct pci_bus *bus;
         struct pci_dev *dev;
         int idx;
         struct resource *r, *pr;
 
         /* Depth-First Search on bus tree */
         list_for_each_entry(bus, bus_list, node) {
 
                   //pci-bridge
                   if ((dev = bus->self)) {
                            for (idx = PCI_BRIDGE_RESOURCES;
                                idx < PCI_NUM_RESOURCES; idx++) {
                                     r = &dev->resource[idx];
                                     if (!r->flags)
                                               continue;
                                     pr = pci_find_parent_resource(dev, r);
                                     if (!r->start || !pr ||
                                         request_resource(pr, r) < 0) {
                                               printk(KERN_ERR "PCI: Cannot allocate "
                                                        "resource region %d "
                                                        "of bridge %s\n",
                                                        idx, pci_name(dev));
                                               /*
                                                * Something is wrong with the region.
                                                * Invalidate the resource to prevent
                                                * child resource allocations in this
                                                * range.
                                                */
                                               r->flags = 0;
                                     }
                            }
                   }
                   pcibios_allocate_bus_resources(&bus->children);
         }
}
这个是一个深度优先搜索算法。类似的算法在pci结构中用得很多。
它遍历pci_root_buses中的每一个根总线下面的所有总线。如果该总线有对应的pci-pci bridge,就先处理这个pci桥的资源.
PCI桥的资源范围是PCI_BRIDGE_RESOURCES~ PCI_NUM_RESOURCES.对于它的每个资源区间。都要从它的上层总线中获得.代码中遍历PCI桥的每一个资源区间,然后找到它在上层总线的对应区间。然后为它建立结构关系。
pci_find_parent_resource()就是为PCI的资源区间找到一个合适的父结点。代码如下:
struct resource *
pci_find_parent_resource(const struct pci_dev *dev, struct resource *res)
{
         const struct pci_bus *bus = dev->bus;
         int i;
         struct resource *best = NULL;
 
         for(i = 0; i < PCI_BUS_NUM_RESOURCES; i++) {
                   struct resource *r = bus->resource[i];
                   if (!r)
                            continue;
                   if (res->start && !(res->start >= r->start && res->end <= r->end))
                            continue;  /* Not contained */
                   if ((res->flags ^ r->flags) & (IORESOURCE_IO | IORESOURCE_MEM))
                            continue;  /* Wrong type */
                   if (!((res->flags ^ r->flags) & IORESOURCE_PREFETCH))
                            return r;     /* Exact match */
                   if ((res->flags & IORESOURCE_PREFETCH) && !(r->flags & IORESOURCE_PREFETCH))
                            best = r;    /* Approximating prefetchable by non-prefetchable */
         }
         return best;
}
首先从pci_dev ->bus就找到了它的上层总线,每条总线拥有PCI_BUS_NUM_RESOURCES个资源区间. 所要寻找的父结点必须要满足以后几个条件:
1:子结点的范围必须要落在父结点的区间范围内
2:父子区间的基本类型应该一致。(基本类型即IO或者内存)
3:如果父子窗口都是可预读的,就完全匹配了
4:如果子结点可预读,而父结点不可预读。也将就着可以了.
注意:父结点可预读而子结点不可预读是不允许的。
 
找到它所属的父结点之后,会调用request_resource()从父结点中请求资源。代码如下:
int request_resource(struct resource *root, struct resource *new)
{
         struct resource *conflict;
 
         write_lock(&resource_lock);
         conflict = __request_resource(root, new);
         write_unlock(&resource_lock);
         return conflict ? -EBUSY : 0;
}
__request_resource()代码如下:
static struct resource * __request_resource(struct resource *root, struct resource *new)
{
         resource_size_t start = new->start;
         resource_size_t end = new->end;
         struct resource *tmp, **p;
 
         if (end < start)
                   return root;
         if (start < root->start)
                   return root;
         if (end > root->end)
                   return root;
         p = &root->child;
         for (;;) {
                   tmp = *p;
                   if (!tmp || tmp->start > end) {
                            new->sibling = tmp;
                            *p = new;
                            new->parent = root;
                            return NULL;
                   }
                   p = &tmp->sibling;
                   if (tmp->end < start)
                            continue;
                   return tmp;
         }
}
这个函数的逻辑比较简单。即在它父节点的子节点中找个合适的位置插下去。父结点的子结点都是按照起始资源地址从小到大的顺序排列的。
 
返回到pcibios_allocate_bus_resources()中,如果它的资源分配过程失败,它会怎么处理呢?看下面的代码片段:
static void __init pcibios_allocate_bus_resources(struct list_head *bus_list)
{
         ……
         ……
if (!r->start || !pr || request_resource(pr, r) < 0) {
                                               printk(KERN_ERR "PCI: Cannot allocate "
                                                        "resource region %d "
                                                        "of bridge %s\n",
                                                        idx, pci_name(dev));
                                               r->flags = 0;
         }
         ……
         ……
}
也就是说,如果分配失败了,它会将资源的flags标志置为0
 
回到pcibios_resource_survey()中,接着往下来,会发现它以不同的参数调用了pcibios_allocate_resources()两次。跟进这个函数的代码进行分析:
static void __init pcibios_allocate_resources(int pass)
{
         struct pci_dev *dev = NULL;
         int idx, disabled;
         u16 command;
         struct resource *r, *pr;
 
         for_each_pci_dev(dev) {
                   pci_read_config_word(dev, PCI_COMMAND, &command);
                   for (idx = 0; idx < PCI_ROM_RESOURCE; idx++) {
                            r = &dev->resource[idx];
                            if (r->parent)               /* Already allocated */
                                     continue;
                            if (!r->start)                  /* Address not assigned at all */
                                     continue;
                            if (r->flags & IORESOURCE_IO)
                                     disabled = !(command & PCI_COMMAND_IO);
                            else
                                     disabled = !(command & PCI_COMMAND_MEMORY);
                            //对于已经启用的,在第一次扫描的时候就将其配制
                            //否则。要等到第二次
                            if (pass == disabled) {
                                     DBG("PCI: Resource %08lx-%08lx "
                                         "(f=%lx, d=%d, p=%d)\n",
                                         r->start, r->end, r->flags, disabled, pass);
                                     pr = pci_find_parent_resource(dev, r);
                                     if (!pr || request_resource(pr, r) < 0) {
                                               printk(KERN_ERR "PCI: Cannot allocate "
                                                        "resource region %d "
                                                        "of device %s\n",
                                                        idx, pci_name(dev));
                                               /* We'll assign a new address later */
                                               r->end -= r->start;
                                               r->start = 0;
                                     }
                            }
                   }
                   if (!pass) {
 
                            //对于ROM。在第一次扫描时就将它关闭
                            r = &dev->resource[PCI_ROM_RESOURCE];
                            if (r->flags & IORESOURCE_ROM_ENABLE) {
                                     /* Turn the ROM off, leave the resource region,
                                      * but keep it unregistered. */
                                     u32 reg;
                                     DBG("PCI: Switching off ROM of %s\n",
                                               pci_name(dev));
                                     r->flags &= ~IORESOURCE_ROM_ENABLE;
                                     pci_read_config_dword(dev,
                                                        dev->rom_base_reg, ®);
                                     pci_write_config_dword(dev, dev->rom_base_reg,
                                                        reg & ~PCI_ROM_ADDRESS_ENABLE);
                            }
                   }
         }
}
该函数遍历整个pci设备。如果该设备的相应空间已经启用了(I/O或者内存)。那在以0为参数调用的时候就让它分配好资源。对于没有被启用的资源。要等到第二次以1为参数调用参数时才会处理。
另外:在第一次处理中就把设备的ROM区间关闭。要等到使用设备的时候再把区间打开。这个打开的过程一般在设备驱动程序里完成。
到这里,我们终于知道为什么要用不同的参数调用函数二次。这样做是为了优先让已经被启用的资源从父节点中分得资源。
如果资源分配失败了。就会将相应资源的start设为0。End成员则设为这个区间的长度。
 
有必要提一下linux2.4.12中这部份的处理,它是这样子的:
对于pci bus的资源分配,如果分配失败.不做任何事情,只是打印出一条警告语句.
而在linux2.6.25中.如果pci bus资源分配失败,会将资源的flag置为0.
为什么2.6要这样做呢?
其实2.4.12中的处理是一个BUG.如果pci bus资源分配失败没有任何的处理的话.那接下来为该总线下面的pci device分配资源的时候,是从该总线的资源中请求的.而事实上,该总线的资源又是有冲突的.这样也造成了该总线下层设备的资源无效.
而在linux 2.6中.将分配失败总线的flag置为了0.下层设备在请求资源的分配的时候,就会类型匹配失败,而避免了上面说的这个部问题.
这个BUG不知道在后面的版本中有没有被修正 :-)
 
pcibios_resource_survey()执先完了之后。Pci的所有总线和设备的资源都被验证分配了一次。对于不能正确分配资源的设备。也做好了标记。接下来,我们来看一下怎么处理资源分配失败的设备。
 
看下面的这段代码:
fs_initcall(pcibios_assign_resources);
fs_initcall()的优先级比subsys_initcall的优先级小,它在pcibios_init()之后才会得到运行。看一下它的代码:
static int __init pcibios_assign_resources(void)
{
         struct pci_dev *dev = NULL;
         struct resource *r, *pr;
 
         if (!(pci_probe & PCI_ASSIGN_ROMS)) {
                   /*
                    * Try to use BIOS settings for ROMs, otherwise let
                    * pci_assign_unassigned_resources() allocate the new
                    * addresses.
                    */
                   for_each_pci_dev(dev) {
                            r = &dev->resource[PCI_ROM_RESOURCE];
                            if (!r->flags || !r->start)
                                     continue;
                            pr = pci_find_parent_resource(dev, r);
                            if (!pr || request_resource(pr, r) < 0) {
                                     r->end -= r->start;
                                     r->start = 0;
                            }
                   }
         }
 
         pci_assign_unassigned_resources();
 
         return 0;
}
之前在pcibios_resource_survey()中没有处理ROM的区间。在这里,先遍历每个设备,处理一下它的ROM空间的资源分配。照以前的方式一样,如果资源分配失败,就让它的start置为0。End置为区间的长度。
处理完之后,进入到pci_assign_unassigned_resources()。我们希望每一个PCI设备,bios都为我们处理好了。可是总是事与愿违。进行的这里,不得不处理一下之前资源分配失败的设备了。代码如下:
void __init
pci_assign_unassigned_resources(void)
{
         struct pci_bus *bus;
 
         /* Depth first, calculate sizes and alignments of all
            subordinate buses. */
         list_for_each_entry(bus, &pci_root_buses, node) {
                   pci_bus_size_bridges(bus);
         }
         /* Depth last, allocate resources and update the hardware. */
         list_for_each_entry(bus, &pci_root_buses, node) {
                   pci_bus_assign_resources(bus);
                   pci_enable_bridges(bus);
         }
}
首先我们要处理资源分配失败的pci_bus。在上面的分析中,如果pci bus资源分配失败。就会将其所属资源的flags置为0.对于这些总线,是在第一个循环里处理的.
第一个循环,遍历挂在pci_root_buses上的所有根结点。然后调用pci_bus_size_bridges()。代码如下:
void __ref pci_bus_size_bridges(struct pci_bus *bus)
{
         struct pci_dev *dev;
         unsigned long mask, prefmask;
 
         list_for_each_entry(dev, &bus->devices, bus_list) {
                   struct pci_bus *b = dev->subordinate;
                   if (!b)
                            continue;
 
                   switch (dev->class >> 8) {
                   case PCI_CLASS_BRIDGE_CARDBUS:
                            pci_bus_size_cardbus(b);
                            break;
 
                   case PCI_CLASS_BRIDGE_PCI:
                   default:
                            pci_bus_size_bridges(b);
                            break;
                   }
         }
 
         /* The root bus? */
         if (!bus->self)
                   return;
 
         switch (bus->self->class >> 8) {
         case PCI_CLASS_BRIDGE_CARDBUS:
                   /* don't size cardbuses yet. */
                   break;
 
         case PCI_CLASS_BRIDGE_PCI:
                   pci_bridge_check_ranges(bus);
         default:
                   pbus_size_io(bus);
                   /* If the bridge supports prefetchable range, size it
                      separately. If it doesn't, or its prefetchable window
                      has already been allocated by arch code, try
                      non-prefetchable range for both types of PCI memory
                      resources. */
                   mask = IORESOURCE_MEM;
                   prefmask = IORESOURCE_MEM | IORESOURCE_PREFETCH;
                   if (pbus_size_mem(bus, prefmask, prefmask))
                            mask = prefmask; /* Success, size non-prefetch only. */
                   pbus_size_mem(bus, mask, IORESOURCE_MEM);
                   break;
         }
}
这是一个深度优先搜索算法。首先遍历总线上的所有设备,如果是pci bridge,递归调用pci_bus_size_bridges()处理下层pci bus.对于每一条不是根总线的pci bus都会经过大循环后面的处理,即对应于case PCI_CLASS_BRIDGE_PCI后面的处理.它要经过的第一个函数是pci_bridge_check_ranges().代码如下:
static void pci_bridge_check_ranges(struct pci_bus *bus)
{
         u16 io;
         u32 pmem;
         struct pci_dev *bridge = bus->self;
         struct resource *b_res;
 
         b_res = &bridge->resource[PCI_BRIDGE_RESOURCES];
         b_res[1].flags |= IORESOURCE_MEM;
 
         pci_read_config_word(bridge, PCI_IO_BASE, &io);
         if (!io) {
                   pci_write_config_word(bridge, PCI_IO_BASE, 0xf0f0);
                   pci_read_config_word(bridge, PCI_IO_BASE, &io);
                  pci_write_config_word(bridge, PCI_IO_BASE, 0x0);
        }
        if (io)
                   b_res[0].flags |= IORESOURCE_IO;
         /*  DECchip 21050 pass 2 errata: the bridge may miss an address
             disconnect boundary by one PCI data phase.
             Workaround: do not use prefetching on this device. */
         if (bridge->vendor == PCI_VENDOR_ID_DEC && bridge->device == 0x0001)
                   return;
         pci_read_config_dword(bridge, PCI_PREF_MEMORY_BASE, &pmem);
         if (!pmem) {
                   pci_write_config_dword(bridge, PCI_PREF_MEMORY_BASE,
                                                      0xfff0fff0);
                   pci_read_config_dword(bridge, PCI_PREF_MEMORY_BASE, &pmem);
                   pci_write_config_dword(bridge, PCI_PREF_MEMORY_BASE, 0x0);
         }
         if (pmem)
                   b_res[2].flags |= IORESOURCE_MEM | IORESOURCE_PREFETCH;
}
Pci bridge设备的7,8,9项是属于过滤窗口,也就是对应于总线的资源。
在这里,检查pci bus是否支持I/O和prefetch memory类型的窗口.如果不支持,则相应寄存器是只读的,而且值为0. 顺便说一句,所有的pci bus都是支持memory窗口的.如果相应类型的资源有效则给相应资源的flags置相应的标志。最后还要记得将相关寄存器清空。
我们在前面分析过.
经过这里的处理,只要判断区间存在,就会置相应的标志。那么,对于我们在上面分配资源失败的pci bus.如果区间确实存在的话。就会重新设置这个标志。
重新设置标志的操作是很重要的一个步骤.它能将前面处理时,因资源冲突的pci bus修正过来.运行到这里.怎么区分哪些pci bus是有待修正的,而哪些pci bus是正常的呢?
可能根据 resource->parent来判断.如果是正常的pci bus.该成员会指向它的父结点.如果是有待修正的pci bus.它的这个域是空的.也就是说,还没有链接到父结点.
 
回到pci_bus_size_bridges()中,经过pci_bridge_check_ranges()的处理过后,流程会转向pbus_size_io().
从字面意思看这个函数是用来计算i/o的大小.主要是用来处理有待修正的pci bus.我们来思考一下,为什么pci bus的资源会有冲突呢?可能有两个原因:
1:pci bus的起始地址冲突.在上层pci bus中,该区间已经被其它设备占用了.这种情况只需要在上层设备中偏移到特定位置就可以了.
2:pci bus的长度过长.这可能是bios在处理PCI的时候出现了错误,我们需要重新计算pci bus的长度.
而pbus_size_io()就是用来处理第二种情况的,处理资源的类型为I/O.代码如下:
static void pbus_size_io(struct pci_bus *bus)
{
         struct pci_dev *dev;
         struct resource *b_res = find_free_bus_resource(bus, IORESOURCE_IO);
         unsigned long size = 0, size1 = 0;
 
         if (!b_res)
                  return;
 
         list_for_each_entry(dev, &bus->devices, bus_list) {
                  int i;
 
                   for (i = 0; i < PCI_NUM_RESOURCES; i++) {
                            struct resource *r = &dev->resource[i];
                            unsigned long r_size;
 
                            if (r->parent || !(r->flags & IORESOURCE_IO))
                                     continue;
                            r_size = r->end - r->start + 1;
 
                            if (r_size < 0x400)
                                     /* Might be re-aligned for ISA */
                                     size += r_size;
                            else
                                     size1 += r_size;
                   }
         }
/* To be fixed in 2.5: we should have sort of HAVE_ISA
   flag in the struct pci_bus. */
#if defined(CONFIG_ISA) || defined(CONFIG_EISA)
         size = (size & 0xff) + ((size & ~0xffUL) << 2);
#endif
         size = ALIGN(size + size1, 4096);
         if (!size) {
                   b_res->flags = 0;
                   return;
         }
         /* Alignment of the IO window is always 4K */
         b_res->start = 4096;
         b_res->end = b_res->start + size - 1;
}
首先,它查看对应类型(I/O)的总线资源是否有冲突的情况.如果有冲突,则计算下层需要该资源的总数.冲突资源的长度就是下层设备所需该资源的总数.只不过,对于pci bus 的I/O资源是4K对齐的.因此总长度也是4K对齐.
 
find_free_bus_resource()就是用来检查冲突资源项的.代码如下:
static struct resource *find_free_bus_resource(struct pci_bus *bus, unsigned long type)
{
         int i;
         struct resource *r;
         unsigned long type_mask = IORESOURCE_IO | IORESOURCE_MEM |
                                       IORESOURCE_PREFETCH;
 
         for (i = 0; i < PCI_BUS_NUM_RESOURCES; i++) {
                   r = bus->resource[i];
                   if (r == &ioport_resource || r == &iomem_resource)
                            continue;
                   if (r && (r->flags & type_mask) == type && !r->parent)
                            return r;
         }
         return NULL;
}
对于资源是ioport_resource, iomem_resource的情况.表示该总线是根总线,不需要处理.注意这一句:
if (r && (r->flags & type_mask) == type && !r->parent)
不仅要求类型匹配,而且要父节点为空.父节点为空.说明之前在分配资源的时候失败,也即资源冲突.
 
因为pci_bus_size_bridges()使用的是深度优先搜索算法,它是先处理最底层的总线,然后再逐层上处理.而下层总线又是上层总线的一个设备(通过pci-pci bridge相连)所以,下层所需的资源长度就可以从下往上反应到上层总线中.最后,到最上层资源冲突的pci bus就可以知道下层需要多少长度的资源了.
 
pbus_size_mem()类似于pbus_size_io(),只是它是用来查找memory类型资源的长度.在代码中,以不同的参数两次调用了这个函数,分别是用来查找带预读的存储区间和一般的存储区间.
pbus_size_mem()和pbus_size_io()相比只是对齐因子不一样,其它大部份的处理都是一样,所以在这里不分析这个函数的代码了.请自行查阅.
 
回到pci_assign_unassigned_resources()函数中.剩余的代码如下所示:
void __init
pci_assign_unassigned_resources(void)
{
         ……
         ……
list_for_each_entry(bus, &pci_root_buses, node) {
                   pci_bus_assign_resources(bus);
                   //启用pci-pci briage的窗口机制
                   pci_enable_bridges(bus);
         }
}
遍历pci_root_buses中存放的根总线,对每条根总线调用pci_bus_assign_resources().代码如下:
void __ref pci_bus_assign_resources(struct pci_bus *bus)
{
         struct pci_bus *b;
         struct pci_dev *dev;
 
         pbus_assign_resources_sorted(bus);
 
         list_for_each_entry(dev, &bus->devices, bus_list) {
                   b = dev->subordinate;
                   if (!b)
                            continue;
 
                   pci_bus_assign_resources(b);
 
                   switch (dev->class >> 8) {
                   case PCI_CLASS_BRIDGE_PCI:
                            pci_setup_bridge(b);
                            break;
 
                   case PCI_CLASS_BRIDGE_CARDBUS:
                            pci_setup_cardbus(b);
                            break;
 
                   default:
                            printk(KERN_INFO "PCI: not setting up bridge %s "
                                   "for bus %d\n", pci_name(dev), b->number);
                            break;
                   }
         }
}
这也是一个使用深度优先搜索算法的函数.
现在是时候处理资源冲突的设备了.这是在pbus_assign_resources_sorted()中完成的,代码如下:
static void pbus_assign_resources_sorted(struct pci_bus *bus)
{
         struct pci_dev *dev;
         struct resource *res;
         struct resource_list head, *list, *tmp;
         int idx;
 
         head.next = NULL;
         //遍历该总线下的所有设备
         list_for_each_entry(dev, &bus->devices, bus_list) {
                   u16 class = dev->class >> 8;
 
                   /* Don't touch classless devices or host bridges or ioapics.  */
                   if (class == PCI_CLASS_NOT_DEFINED ||
                       class == PCI_CLASS_BRIDGE_HOST)
                            continue;
 
                   /* Don't touch ioapic devices already enabled by firmware */
                   if (class == PCI_CLASS_SYSTEM_PIC) {
                            u16 command;
                            pci_read_config_word(dev, PCI_COMMAND, &command);
                            if (command & (PCI_COMMAND_IO | PCI_COMMAND_MEMORY))
                                     continue;
                   }
 
                   //所有没有成功分配资源的resource会放到head链表中
                   pdev_sort_resources(dev, &head);
         }
 
         for (list = head.next; list;) {
                   res = list->res;
                   idx = res - &list->dev->resource[0];
                   if (pci_assign_resource(list->dev, idx)) {
                            res->start = 0;
                            res->end = 0;
                            res->flags = 0;
                   }
                   tmp = list;
                   list = list->next;
                   kfree(tmp);
         }
}
首先,它将有总线下有资源冲突的设备加到链表中,然后遍历这个链表,处理每个资源冲突的设备.
pdev_sort_resources()代码很简单,就是计算对齐因子,然后按照对齐因子大小添加到链表.在这里要特别注意.对于pci bus中的资源是要按起始地址对齐的(还记得我们在为冲突的pci bus计算I/O资源长度的时候,它的start成员就是设为4K.表示要4K对齐).这个函数代码比较简单.不再详细分析,请自己阅读.
具体的资源冲突处理过程是在pci_assign_resource()中完成的,代码如下:
int pci_assign_resource(struct pci_dev *dev, int resno)
{
         struct pci_bus *bus = dev->bus;
         struct resource *res = dev->resource + resno;
         resource_size_t size, min, align;
         int ret;
 
         //空间大小
         size = res->end - res->start + 1;
         //起点最小值
         min = (res->flags & IORESOURCE_IO) ? PCIBIOS_MIN_IO : PCIBIOS_MIN_MEM;
         /* The bridge resources are special, as their
            size != alignment. Sizing routines return
            required alignment in the "start" field. */
           
            //对于pci-pci bridge.有特殊的对齐方式
         align = (resno < PCI_BRIDGE_RESOURCES) ? size : res->start;
 
         /* First, try exact prefetching match.. */
         //首先,带IORESOURCE_PREFETCH 标志处理
        
         ret = pci_bus_alloc_resource(bus, res, size, align, min,
                                          IORESOURCE_PREFETCH,
                                          pcibios_align_resource, dev);
 
         if (ret < 0 && (res->flags & IORESOURCE_PREFETCH)) {
                   /*
                    * That failed.
                    *
                    * But a prefetching area can handle a non-prefetching
                    * window (it will just not perform as well).
                    */
                   //分配失败,不带IORESOURCE_PREFETCH标志处理
                   ret = pci_bus_alloc_resource(bus, res, size, align, min, 0,
                                                    pcibios_align_resource, dev);
         }
 
         if (ret) {
                   printk(KERN_ERR "PCI: Failed to allocate %s resource "
                            "#%d:%llx@%llx for %s\n",
                            res->flags & IORESOURCE_IO ? "I/O" : "mem",
                            resno, (unsigned long long)size,
                            (unsigned long long)res->start, pci_name(dev));
         } else if (resno < PCI_BRIDGE_RESOURCES) {
                   //如果冲突修正成功
                   pci_update_resource(dev, res, resno);
         }
 
         return ret;
}
首先计算该资源的长度和起始地址的最小值以及对齐因子.然后向父结点去申请资源.如果分配失败,那就放松条件,不带预读标志再去分配一次.如果成功,则更新设备的寄存器.如果还是失败,那就没什么办法了.
pci_bus_alloc_resource()代码如下:
int
pci_bus_alloc_resource(struct pci_bus *bus, struct resource *res,
                   resource_size_t size, resource_size_t align,
                   resource_size_t min, unsigned int type_mask,
                   void (*alignf)(void *, struct resource *, resource_size_t,
                                     resource_size_t),
                   void *alignf_data)
{
         int i, ret = -ENOMEM;
 
         type_mask |= IORESOURCE_IO | IORESOURCE_MEM;
 
         for (i = 0; i < PCI_BUS_NUM_RESOURCES; i++) {
                   struct resource *r = bus->resource[i];
                   if (!r)
                            continue;
 
                   /* type_mask must match */
                   if ((res->flags ^ r->flags) & type_mask)
                            continue;
 
                   /* We cannot allocate a non-prefetching resource
                      from a pre-fetching area */
                   if ((r->flags & IORESOURCE_PREFETCH) &&
                       !(res->flags & IORESOURCE_PREFETCH))
                            continue;
 
                   /* Ok, try it out.. */
                   ret = allocate_resource(r, res, size,
                                               r->start ? : min,
                                               -1, align,
                                               alignf, alignf_data);
                   if (ret == 0)
                            break;
         }
         return ret;
}
遍历上层总线类型相同的资源区,从中分配资源.如果分配成功退出.
allocate_resource()如下所示:
int allocate_resource(struct resource *root, struct resource *new,
                         resource_size_t size, resource_size_t min,
                         resource_size_t max, resource_size_t align,
                         void (*alignf)(void *, struct resource *,
                                          resource_size_t, resource_size_t),
                         void *alignf_data)
{
         int err;
 
         write_lock(&resource_lock);
         err = find_resource(root, new, size, min, max, align, alignf, alignf_data);
         if (err >= 0 && __request_resource(root, new))
                   err = -EBUSY;
         write_unlock(&resource_lock);
         return err;
}
先到父节点中寻找是否有符合的资源区间.如果有,则向父节点请求这个区间.
__request_resource()的代码之前我们研究了,这里不再赘述. find_resource()代码如下:
static int find_resource(struct resource *root, struct resource *new,
                             resource_size_t size, resource_size_t min,
                             resource_size_t max, resource_size_t align,
                             void (*alignf)(void *, struct resource *,
                                               resource_size_t, resource_size_t),
                             void *alignf_data)
{
         struct resource *this = root->child;
 
         new->start = root->start;
         /*
          * Skip past an allocated resource that starts at 0, since the assignment
          * of this->start - 1 to new->end below would cause an underflow.
          */
         if (this && this->start == 0) {
                   new->start = this->end + 1;
                   this = this->sibling;
         }
         for(;;) {
                   if (this)
                            new->end = this->start - 1;
                   else
                            new->end = root->end;
                   if (new->start < min)
                            new->start = min;
                   if (new->end > max)
                            new->end = max;
                  new->start = ALIGN(new->start, align);
                   if (alignf)
                            alignf(alignf_data, new, size, align);
                   if (new->start < new->end && new->end - new->start >= size - 1) {
                            new->end = new->start + size - 1;
                            return 0;
                   }
                   if (!this)
                            break;
                   new->start = this->end + 1;
                   this = this->sibling;
         }
         return -EBUSY;
}
这个函数比较简单,就是判断父节点是否有足够长的空间区间.
返回到pci_assign_resource()中,如果修正成功,就会调用pci_update_resource().因为pci设备的资源区间的起始地址改了,要将它更新到寄存器.这正是pci_update_resource()要处理的事情.函数比较简单.就不详细分析了,但这个函数中有个特别的处理,需要指出来.代码片段如下:
void
pci_update_resource(struct pci_dev *dev, struct resource *res, int resno)
{
         ……
         ……
         if (resno < 6) {
                   reg = PCI_BASE_ADDRESS_0 + 4 * resno;
         } else if (resno == PCI_ROM_RESOURCE) {
                   if (!(res->flags & IORESOURCE_ROM_ENABLE))
                            return;
                   new |= PCI_ROM_ADDRESS_ENABLE;
                   reg = dev->rom_base_reg;
         } else {
                   /* Hmm, non-standard resource. */
        
                   return;                /* kill uninitialised var warning */
         }
……
……
}
从上面的处理可以看到,它只会处理前7个资源区间,对于8.9.10区间,是pci-pci bridge的中存放的关于pci bus中的资源信息.
也就是说,这个函数只会处理pci device中的寄存器,不会处理pci bus.
 
返回到pci_bus_assign_resources()中.剩余的代码片段如下:
void __ref pci_bus_assign_resources(struct pci_bus *bus)
{
         ……
         ……
         list_for_each_entry(dev, &bus->devices, bus_list) {
                   b = dev->subordinate;
                   if (!b)
                            continue;
 
                   pci_bus_assign_resources(b);
 
                   switch (dev->class >> 8) {
                   case PCI_CLASS_BRIDGE_PCI:
                            pci_setup_bridge(b);
                            break;
 
                   case PCI_CLASS_BRIDGE_CARDBUS:
                            pci_setup_cardbus(b);
                            break;
 
                   default:
                            printk(KERN_INFO "PCI: not setting up bridge %s "
                                   "for bus %d\n", pci_name(dev), b->number);
                            break;
                   }
         }
}
处理完pbus_assign_resources_sorted()之后,遍历总线下面的pci-pci bridge(dev->subordinate不为空).对每一条下层总线,都递归调用pci_bus_assign_resources().
Switch后面的部份,是处理完下层总线,回到本层的处理了.忽略对cardbus的处理.看到了pci_setup_bridge().
我们在前面已经分析到.对于修正过后的处理,只会更新pci device的寄存器信息,而不会更新pci bus中的寄存器. pci_setup_bridge()就是更新pci bus寄存器信息的.只是不管它之前是不是资源冲突的.所有的pci bus全部都会更新一下.该函数代码比较简单.就不做分析了.
 
回到pci_assign_unassigned_resources()中.对每条根总线,我们还要经过pci_enable_bridges()的处理,这个函数是这一节分析的最后一个函数了.
代码如下:
void pci_enable_bridges(struct pci_bus *bus)
{
         struct pci_dev *dev;
         int retval;
 
         list_for_each_entry(dev, &bus->devices, bus_list) {
                   if (dev->subordinate) {
                            retval = pci_enable_device(dev);
                            pci_set_master(dev);
                            pci_enable_bridges(dev->subordinate);
                   }
         }
}
运行到这里,pci总线和设备都准备好了,现在就可以启用pci-pci bridge了.pci_enable_bridge()做的工作如下:
1:启用所有pci-pci bridge的所有I/O和内存.
2:启用pci-pci bridge的中断
3:设置pci-pci bridge 的电源管理状态
4:设置pci-pci bridge对总线的控制能力
在这里要注意了,这个函数只是启用了pci-pci bridge的相关功能.对于一般pci_dev.需要驱动程序在使用的时候启用它们.
函数比较简单,内部就不加详细分析了.
阅读(1391) | 评论(0) | 转发(0) |
给主人留下些什么吧!~~