linux-insides/mm/linux-mm-2.md

Linux kernel memory management Part 2.

Fix-Mapped Addresses and ioremap

Fix-Mapped addresses are a set of special compile-time addresses whose corresponding physical addresses do not have to be a linear address minus __START_KERNEL_map. Each fix-mapped address maps one page frame and the kernel uses them as pointers that never change their address. That is the main point of these addresses. As the comment says: to have a constant address at compile time, but to set the physical address only in the boot process. You can remember that in the earliest part, we already set the level2_fixmap_pgt:

NEXT_PAGE(level2_fixmap_pgt)
	.fill	506,8,0
	.quad	level1_fixmap_pgt - __START_KERNEL_map + _PAGE_TABLE
	.fill	5,8,0

NEXT_PAGE(level1_fixmap_pgt)
	.fill	512,8,0

As you can see level2_fixmap_pgt is right after the level2_kernel_pgt which is kernel code+data+bss. Every fix-mapped address is represented by an integer index which is defined in the fixed_addresses enum from the arch/x86/include/asm/fixmap.h. For example it contains entries for VSYSCALL_PAGE - if emulation of legacy vsyscall page is enabled, FIX_APIC_BASE for local apic, etc. In virtual memory fix-mapped area is placed in the modules area:

       +-----------+-----------------+---------------+------------------+
       |           |                 |               |                  |
       |kernel text|      kernel     |               |    vsyscalls     |
       | mapping   |       text      |    Modules    |    fix-mapped    |
       |from phys 0|       data      |               |    addresses     |
       |           |                 |               |                  |
       +-----------+-----------------+---------------+------------------+
__START_KERNEL_map   __START_KERNEL    MODULES_VADDR            0xffffffffffffffff

Base virtual address and size of the fix-mapped area are presented by the two following macro:

#define FIXADDR_SIZE	(__end_of_permanent_fixed_addresses << PAGE_SHIFT)
#define FIXADDR_START		(FIXADDR_TOP - FIXADDR_SIZE)

Here __end_of_permanent_fixed_addresses is an element of the fixed_addresses enum and as I wrote above: Every fix-mapped address is represented by an integer index which is defined in the fixed_addresses. PAGE_SHIFT determines the size of a page. For example size of the one page we can get with the 1 << PAGE_SHIFT. In our case we need to get the size of the fix-mapped area, but not only of one page, that's why we are using __end_of_permanent_fixed_addresses for getting the size of the fix-mapped area. In my case it's a little more than 536 kilobytes. In your case it might be a different number, because the size depends on amount of the fix-mapped addresses which are depends on your kernel's configuration.

The second FIXADDR_START macro just subtracts the fix-mapped area size from the last address of the fix-mapped area to get its base virtual address. FIXADDR_TOP is a rounded up address from the base address of the vsyscall space:

#define FIXADDR_TOP     (round_up(VSYSCALL_ADDR + PAGE_SIZE, 1<<PMD_SHIFT) - PAGE_SIZE)

The fixed_addresses enums are used as an index to get the virtual address by the fix_to_virt function. Implementation of this function is easy:

static __always_inline unsigned long fix_to_virt(const unsigned int idx)
{
        BUILD_BUG_ON(idx >= __end_of_fixed_addresses);
        return __fix_to_virt(idx);
}

first of all it checks that the index given for the fixed_addresses enum is not greater or equal than __end_of_fixed_addresses with the BUILD_BUG_ON macro and then returns the result of the __fix_to_virt macro:

#define __fix_to_virt(x)        (FIXADDR_TOP - ((x) << PAGE_SHIFT))

Here we shift left the given fix-mapped address index on the PAGE_SHIFT which determines size of a page as I wrote above and subtract it from the FIXADDR_TOP which is the highest address of the fix-mapped area. There is an inverse function for getting fix-mapped address from a virtual address:

static inline unsigned long virt_to_fix(const unsigned long vaddr)
{
        BUG_ON(vaddr >= FIXADDR_TOP || vaddr < FIXADDR_START);
        return __virt_to_fix(vaddr);
}

virt_to_fix takes a virtual address, checks that this address is between FIXADDR_START and FIXADDR_TOP and calls the __virt_to_fix macro which implemented as:

#define __virt_to_fix(x)        ((FIXADDR_TOP - ((x)&PAGE_MASK)) >> PAGE_SHIFT)

A PFN is simply an index within physical memory that is counted in page-sized units. PFN for a physical address could be trivially defined as (page_phys_addr >> PAGE_SHIFT);

__virt_to_fix clears the first 12 bits in the given address, subtracts it from the last address the of fix-mapped area (FIXADDR_TOP) and shifts the result right on PAGE_SHIFT which is 12. Let me explain how it works. As I already wrote we will clear the first 12 bits in the given address with x & PAGE_MASK. As we subtract this from the FIXADDR_TOP, we will get the last 12 bits of the FIXADDR_TOP which are present. We know that the first 12 bits of the virtual address represent the offset in the page frame. With the shifting it on PAGE_SHIFT we will get Page frame number which is just all bits in a virtual address besides the first 12 offset bits. Fix-mapped addresses are used in different places in the linux kernel. IDT descriptor stored there, Intel Trusted Execution Technology UUID stored in the fix-mapped area started from FIX_TBOOT_BASE index, Xen bootmap and many more... We already saw a little about fix-mapped addresses in the fifth part about of the linux kernel initialization. We use fix-mapped area in the early ioremap initialization. Let's look at it more closely and try to understand what ioremap is, how it is implemented in the kernel and how it is related to the fix-mapped addresses.

ioremap

The Linux kernel provides many different primitives to manage memory. For this moment we will touch I/O memory. Every device is controlled by reading/writing from/to its registers. For example a driver can turn off/on a device by writing to its registers or get the state of a device by reading from its registers. Besides registers, many devices have buffers where a driver can write something or read from there. As we know for this moment there are two ways to access device's registers and data buffers:

• through the I/O ports;
• mapping of the all registers to the memory address space;

In the first case every control register of a device has a number of input and output port. A device driver can read from a port and write to it with two in and out instructions which we already saw. If you want to know about currently registered port regions, you can learn about them by accessing /proc/ioports:

$ cat /proc/ioports
0000-0cf7 : PCI Bus 0000:00
  0000-001f : dma1
  0020-0021 : pic1
  0040-0043 : timer0
  0050-0053 : timer1
  0060-0060 : keyboard
  0064-0064 : keyboard
  0070-0077 : rtc0
  0080-008f : dma page reg
  00a0-00a1 : pic2
  00c0-00df : dma2
  00f0-00ff : fpu
    00f0-00f0 : PNP0C04:00
  03c0-03df : vesafb
  03f8-03ff : serial
  04d0-04d1 : pnp 00:06
  0800-087f : pnp 00:01
  0a00-0a0f : pnp 00:04
  0a20-0a2f : pnp 00:04
  0a30-0a3f : pnp 00:04
0cf8-0cff : PCI conf1
0d00-ffff : PCI Bus 0000:00
...
...
...

/proc/ioports provides information about which driver uses which address of a I/O port region. All of these memory regions, for example 0000-0cf7, were claimed with the request_region function from the include/linux/ioport.h. Actually request_region is a macro which is defined as:

#define request_region(start,n,name)   __request_region(&ioport_resource, (start), (n), (name), 0)

As we can see it takes three parameters:

• start - begin of region;
• n - length of region;
• name - name of requester.

request_region allocates an I/O port region. Very often the check_region function is called before the request_region to check that the given address range is available and the release_region function to release the memory region. request_region returns a pointer to the resource structure. The resource structure represents an abstraction for a tree-like subset of system resources. We already saw the resource structure in the fifth part of the kernel initialization process and it looks as follows:

struct resource {
        resource_size_t start;
        resource_size_t end;
        const char *name;
        unsigned long flags;
        struct resource *parent, *sibling, *child;
};

and contains start and end addresses of the resource, the name, etc. Every resource structure contains pointers to the parent, sibling and child resources. As it has a parent and a childs, it means that every subset of resources has root resource structure. For example, for I/O ports it is the ioport_resource structure:

struct resource ioport_resource = {
         .name   = "PCI IO",
         .start  = 0,
         .end    = IO_SPACE_LIMIT,
        .flags  = IORESOURCE_IO,
};
EXPORT_SYMBOL(ioport_resource);

Or for iomem, it is the iomem_resource structure:

struct resource iomem_resource = {
        .name   = "PCI mem",
        .start  = 0,
        .end    = -1,
        .flags  = IORESOURCE_MEM,
};

As I have mentioned before, request_regions is used to register I/O port regions and this macro is used in many places in the kernel. For example let's look at drivers/char/rtc.c. This source code file provides the Real Time Clock interface in the linux kernel. As every kernel module, rtc module contains module_init definition:

module_init(rtc_init);

where rtc_init is the rtc initialization function. This function is defined in the same rtc.c source code file. In the rtc_init function we can see a couple of calls to the rtc_request_region functions, which wrap request_region for example:

r = rtc_request_region(RTC_IO_EXTENT);

where rtc_request_region calls:

r = request_region(RTC_PORT(0), size, "rtc");

Here RTC_IO_EXTENT is the size of the memory region and it is 0x8, "rtc" is the name of the region and RTC_PORT is:

#define RTC_PORT(x)     (0x70 + (x))

So with the request_region(RTC_PORT(0), size, "rtc") we register a memory region that starts at 0x70 and and has a size of 0x8. Let's look at /proc/ioports:

~$ sudo cat /proc/ioports | grep rtc
0070-0077 : rtc0

So, we got it! Ok, that was it for the I/O ports. The second way to communicate with drivers is through the use of I/O memory. As I have mentioned above this works by mapping the control registers and the memory of a device to the memory address space. I/O memory is a set of contiguous addresses which are provided by a device to the CPU through a bus. None of the memory-mapped I/O addresses are used by the kernel directly. There is a special ioremap function which allows us to convert the physical address on a bus to a kernel virtual address. In other words, ioremap maps I/O physical memory regions to make them accessible from the kernel. The ioremap function takes two parameters:

• start of the memory region;
• size of the memory region;

The I/O memory mapping API provides functions to check, request and release memory regions as I/O memory. There are three functions for that:

• request_mem_region
• release_mem_region
• check_mem_region

~$ sudo cat /proc/iomem
...
...
...
be826000-be82cfff : ACPI Non-volatile Storage
be82d000-bf744fff : System RAM
bf745000-bfff4fff : reserved
bfff5000-dc041fff : System RAM
dc042000-dc0d2fff : reserved
dc0d3000-dc138fff : System RAM
dc139000-dc27dfff : ACPI Non-volatile Storage
dc27e000-deffefff : reserved
defff000-deffffff : System RAM
df000000-dfffffff : RAM buffer
e0000000-feafffff : PCI Bus 0000:00
  e0000000-efffffff : PCI Bus 0000:01
    e0000000-efffffff : 0000:01:00.0
  f7c00000-f7cfffff : PCI Bus 0000:06
    f7c00000-f7c0ffff : 0000:06:00.0
    f7c10000-f7c101ff : 0000:06:00.0
      f7c10000-f7c101ff : ahci
  f7d00000-f7dfffff : PCI Bus 0000:03
    f7d00000-f7d3ffff : 0000:03:00.0
      f7d00000-f7d3ffff : alx
...
...
...

Part of these addresses are from the call of the e820_reserve_resources function. We can find a call to this function in the arch/x86/kernel/setup.c and the function itself is defined in arch/x86/kernel/e820.c. e820_reserve_resources goes through the e820 map and inserts memory regions into the root iomem resource structure. All e820 memory regions which are inserted into the iomem resource have the following types:

static inline const char *e820_type_to_string(int e820_type)
{
	switch (e820_type) {
	case E820_RESERVED_KERN:
	case E820_RAM:	return "System RAM";
	case E820_ACPI:	return "ACPI Tables";
	case E820_NVS:	return "ACPI Non-volatile Storage";
	case E820_UNUSABLE:	return "Unusable memory";
	default:	return "reserved";
	}
}

and we can see them in the /proc/iomem (read above).

Now let's try to understand how ioremap works. We already know a little about ioremap, we saw it in the fifth part about linux kernel initialization. If you have read this part, you can remember the call of the early_ioremap_init function from the arch/x86/mm/ioremap.c. Initialization of the ioremap is split into two parts: there is the early part which we can use before the normal ioremap is available and the normal ioremap which is available after vmalloc initialization and the call of paging_init. We do not know anything about vmalloc for now, so let's consider early initialization of the ioremap. First of all early_ioremap_init checks that fixmap is aligned on page middle directory boundary:

BUILD_BUG_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));

more about BUILD_BUG_ON you can read in the first part about Linux Kernel initialization. So BUILD_BUG_ON macro raises a compilation error if the given expression is true. In the next step after this check, we can see call of the early_ioremap_setup function from the mm/early_ioremap.c. This function presents generic initialization of the ioremap. early_ioremap_setup function fills the slot_virt array with the virtual addresses of the early fixmaps. All early fixmaps are after __end_of_permanent_fixed_addresses in memory. They start at FIX_BITMAP_BEGIN (top) and end with FIX_BITMAP_END (down). Actually there are 512 temporary boot-time mappings, used by early ioremap:

#define NR_FIX_BTMAPS		64
#define FIX_BTMAPS_SLOTS	8
#define TOTAL_FIX_BTMAPS	(NR_FIX_BTMAPS * FIX_BTMAPS_SLOTS)

and early_ioremap_setup:

void __init early_ioremap_setup(void)
{
        int i;

        for (i = 0; i < FIX_BTMAPS_SLOTS; i++)
                if (WARN_ON(prev_map[i]))
                        break;

        for (i = 0; i < FIX_BTMAPS_SLOTS; i++)
                slot_virt[i] = __fix_to_virt(FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*i);
}

the slot_virt and other arrays are defined in the same source code file:

static void __iomem *prev_map[FIX_BTMAPS_SLOTS] __initdata;
static unsigned long prev_size[FIX_BTMAPS_SLOTS] __initdata;
static unsigned long slot_virt[FIX_BTMAPS_SLOTS] __initdata;

slot_virt contains the virtual addresses of the fix-mapped areas, prev_map array contains addresses of the early ioremap areas. Note that I wrote above: Actually there are 512 temporary boot-time mappings, used by early ioremap and you can see that all arrays are defined with the __initdata attribute which means that this memory will be released after the kernel initialization process. After early_ioremap_setup has finished its work, we're getting page middle directory where early ioremap begins with the early_ioremap_pmd function which just gets the base address of the page global directory and calculates the page middle directory for the given address:

static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
{
	pgd_t *base = __va(read_cr3());
	pgd_t *pgd = &base[pgd_index(addr)];
	pud_t *pud = pud_offset(pgd, addr);
	pmd_t *pmd = pmd_offset(pud, addr);
	return pmd;
}

After this we fill bm_pte (early ioremap page table entries) with zeros and call the pmd_populate_kernel function:

pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
memset(bm_pte, 0, sizeof(bm_pte));
pmd_populate_kernel(&init_mm, pmd, bm_pte);

pmd_populate_kernel takes three parameters:

• init_mm - memory descriptor of the init process (you can read about it in the previous part);
• pmd - page middle directory of the beginning of the ioremap fixmaps;
• bm_pte - early ioremap page table entries array which defined as:

static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;

The pmd_populate_kernel function is defined in the arch/x86/include/asm/pgalloc.h and populates the page middle directory (pmd) provided as an argument with the given page table entries (bm_pte):

static inline void pmd_populate_kernel(struct mm_struct *mm,
                                       pmd_t *pmd, pte_t *pte)
{
        paravirt_alloc_pte(mm, __pa(pte) >> PAGE_SHIFT);
        set_pmd(pmd, __pmd(__pa(pte) | _PAGE_TABLE));
}

where set_pmd is:

#define set_pmd(pmdp, pmd)              native_set_pmd(pmdp, pmd)

and native_set_pmd is:

static inline void native_set_pmd(pmd_t *pmdp, pmd_t pmd)
{
        *pmdp = pmd;
}

That's all. Early ioremap is ready to use. There are a couple of checks in the early_ioremap_init function, but they are not so important, anyway initialization of the ioremap is finished.

Use of early ioremap

As soon as early ioremap has been setup successfully, we can use it. It provides two functions:

• early_ioremap
• early_iounmap

for mapping/unmapping of I/O physical address to virtual address. Both functions depend on the CONFIG_MMU configuration option. Memory management unit is a special block of memory management. The main purpose of this block is the translation of physical addresses to virtual addresses. The memory management unit knows about the high-level page table addresses (pgd) from the cr3 control register. If CONFIG_MMU options is set to n, early_ioremap just returns the given physical address and early_iounmap does nothing. If CONFIG_MMU option is set to y, early_ioremap calls __early_ioremap which takes three parameters:

• phys_addr - base physical address of the I/O memory region to map on virtual addresses;
• size - size of the I/O memory region;
• prot - page table entry bits.

First of all in the __early_ioremap, we go through all early ioremap fixmap slots and search for the first free one in the prev_map array. When we found it we remember its number in the slot variable and set up size:

slot = -1;
for (i = 0; i < FIX_BTMAPS_SLOTS; i++) {
	if (!prev_map[i]) {
		slot = i;
		break;
	}
}
...
...
...
prev_size[slot] = size;
last_addr = phys_addr + size - 1;

In the next spte we can see the following code:

offset = phys_addr & ~PAGE_MASK;
phys_addr &= PAGE_MASK;
size = PAGE_ALIGN(last_addr + 1) - phys_addr;

Here we are using PAGE_MASK for clearing all bits in the phys_addr except the first 12 bits. PAGE_MASK macro is defined as:

#define PAGE_MASK       (~(PAGE_SIZE-1))

We know that size of a page is 4096 bytes or 1000000000000 in binary. PAGE_SIZE - 1 will be 111111111111, but with ~, we will get 000000000000, but as we use ~PAGE_MASK we will get 111111111111 again. On the second line we do the same but clear the first 12 bits and getting page-aligned size of the area on the third line. We getting aligned area and now we need to get the number of pages which are occupied by the new ioremap area and calculate the fix-mapped index from fixed_addresses in the next steps:

nrpages = size >> PAGE_SHIFT;
idx = FIX_BTMAP_BEGIN - NR_FIX_BTMAPS*slot;

Now we can fill fix-mapped area with the given physical addresses. On every iteration in the loop, we call the __early_set_fixmap function from the arch/x86/mm/ioremap.c, increase the given physical address by the page size which is 4096 bytes and update the addresses index and the number of pages:

while (nrpages > 0) {
	__early_set_fixmap(idx, phys_addr, prot);
	phys_addr += PAGE_SIZE;
	--idx;
    --nrpages;
}

The __early_set_fixmap function gets the page table entry (stored in the bm_pte, see above) for the given physical address with:

pte = early_ioremap_pte(addr);

In the next step of early_ioremap_pte we check the given page flags with the pgprot_val macro and call set_pte or pte_clear depending on the flags given:

if (pgprot_val(flags))
		set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
	else
		pte_clear(&init_mm, addr, pte);

As you can see above, we passed FIXMAP_PAGE_IO as flags to the __early_ioremap. FIXMPA_PAGE_IO expands to the:

(__PAGE_KERNEL_EXEC | _PAGE_NX)

flags, so we call set_pte function to set the page table entry which works in the same manner as set_pmd but for PTEs (read above about it). As we have set all PTEs in the loop, we can now take a look at the call of the __flush_tlb_one function:

__flush_tlb_one(addr);

This function is defined in arch/x86/include/asm/tlbflush.h and calls __flush_tlb_single or __flush_tlb depending on the value of cpu_has_invlpg:

static inline void __flush_tlb_one(unsigned long addr)
{
        if (cpu_has_invlpg)
                __flush_tlb_single(addr);
        else
                __flush_tlb();
}

The __flush_tlb_one function invalidates the given address in the TLB. As you just saw we updated the paging structure, but TLB is not informed of the changes, that's why we need to do it manually. There are two ways to do it. The first is to update the cr3 control register and the __flush_tlb function does this:

native_write_cr3(native_read_cr3());

The second method is to use the invlpg instruction to invalidate the TLB entry. Let's look at the __flush_tlb_one implementation. As you can see, first of all the function checks cpu_has_invlpg which is defined as:

#if defined(CONFIG_X86_INVLPG) || defined(CONFIG_X86_64)
# define cpu_has_invlpg         1
#else
# define cpu_has_invlpg         (boot_cpu_data.x86 > 3)
#endif

If a CPU supports the invlpg instruction, we call the __flush_tlb_single macro which expands to the call of __native_flush_tlb_single:

static inline void __native_flush_tlb_single(unsigned long addr)
{
        asm volatile("invlpg (%0)" ::"r" (addr) : "memory");
}

or call __flush_tlb which just updates the cr3 register as we have seen. After this step execution of the __early_set_fixmap function is finished and we can go back to the __early_ioremap implementation. When we have set up the fixmap area for the given address, we need to save the base virtual address of the I/O Re-mapped area in the prev_map using the slot index:

prev_map[slot] = (void __iomem *)(offset + slot_virt[slot]);

and return it.

The second function, early_iounmap, unmaps an I/O memory region. This function takes two parameters: base address and size of a I/O region and generally looks very similar to early_ioremap. It also goes through fixmap slots and looks for a slot with the given address. After that, it gets the index of the fixmap slot and calls __late_clear_fixmap or __early_set_fixmap depending on the after_paging_init value. It calls __early_set_fixmap with one difference to how early_ioremap does it: early_iounmap passes zero as physical address. And in the end it sets the address of the I/O memory region to NULL:

prev_map[slot] = NULL;

That's all about fixmaps and ioremap. Of course this part does not cover all features of ioremap, only early ioremap but there is also normal ioremap. But we need to know more things before we study that in more detail.

So, this is the end!

Conclusion

This is the end of the second part about linux kernel memory management. If you have questions or suggestions, ping me on twitter 0xAX, drop me an email or just create an issue.

Please note that English is not my first language and I am really sorry for any inconvenience. If you found any mistakes please send me a PR to linux-insides.

Links

• apic
• vsyscall
• Intel Trusted Execution Technology
• Xen
• Real Time Clock
• e820
• Memory management unit
• TLB
• Paging
• Linux kernel memory management Part 1.