Root/Documentation/ia64/aliasing.txt

1             MEMORY ATTRIBUTE ALIASING ON IA-64
2
3               Bjorn Helgaas
4               <bjorn.helgaas@hp.com>
5                May 4, 2006
6
7
8MEMORY ATTRIBUTES
9
10    Itanium supports several attributes for virtual memory references.
11    The attribute is part of the virtual translation, i.e., it is
12    contained in the TLB entry. The ones of most interest to the Linux
13    kernel are:
14
15    WB Write-back (cacheable)
16    UC Uncacheable
17    WC Write-coalescing
18
19    System memory typically uses the WB attribute. The UC attribute is
20    used for memory-mapped I/O devices. The WC attribute is uncacheable
21    like UC is, but writes may be delayed and combined to increase
22    performance for things like frame buffers.
23
24    The Itanium architecture requires that we avoid accessing the same
25    page with both a cacheable mapping and an uncacheable mapping[1].
26
27    The design of the chipset determines which attributes are supported
28    on which regions of the address space. For example, some chipsets
29    support either WB or UC access to main memory, while others support
30    only WB access.
31
32MEMORY MAP
33
34    Platform firmware describes the physical memory map and the
35    supported attributes for each region. At boot-time, the kernel uses
36    the EFI GetMemoryMap() interface. ACPI can also describe memory
37    devices and the attributes they support, but Linux/ia64 currently
38    doesn't use this information.
39
40    The kernel uses the efi_memmap table returned from GetMemoryMap() to
41    learn the attributes supported by each region of physical address
42    space. Unfortunately, this table does not completely describe the
43    address space because some machines omit some or all of the MMIO
44    regions from the map.
45
46    The kernel maintains another table, kern_memmap, which describes the
47    memory Linux is actually using and the attribute for each region.
48    This contains only system memory; it does not contain MMIO space.
49
50    The kern_memmap table typically contains only a subset of the system
51    memory described by the efi_memmap. Linux/ia64 can't use all memory
52    in the system because of constraints imposed by the identity mapping
53    scheme.
54
55    The efi_memmap table is preserved unmodified because the original
56    boot-time information is required for kexec.
57
58KERNEL IDENTITY MAPPINGS
59
60    Linux/ia64 identity mappings are done with large pages, currently
61    either 16MB or 64MB, referred to as "granules." Cacheable mappings
62    are speculative[2], so the processor can read any location in the
63    page at any time, independent of the programmer's intentions. This
64    means that to avoid attribute aliasing, Linux can create a cacheable
65    identity mapping only when the entire granule supports cacheable
66    access.
67
68    Therefore, kern_memmap contains only full granule-sized regions that
69    can referenced safely by an identity mapping.
70
71    Uncacheable mappings are not speculative, so the processor will
72    generate UC accesses only to locations explicitly referenced by
73    software. This allows UC identity mappings to cover granules that
74    are only partially populated, or populated with a combination of UC
75    and WB regions.
76
77USER MAPPINGS
78
79    User mappings are typically done with 16K or 64K pages. The smaller
80    page size allows more flexibility because only 16K or 64K has to be
81    homogeneous with respect to memory attributes.
82
83POTENTIAL ATTRIBUTE ALIASING CASES
84
85    There are several ways the kernel creates new mappings:
86
87    mmap of /dev/mem
88
89    This uses remap_pfn_range(), which creates user mappings. These
90    mappings may be either WB or UC. If the region being mapped
91    happens to be in kern_memmap, meaning that it may also be mapped
92    by a kernel identity mapping, the user mapping must use the same
93    attribute as the kernel mapping.
94
95    If the region is not in kern_memmap, the user mapping should use
96    an attribute reported as being supported in the EFI memory map.
97
98    Since the EFI memory map does not describe MMIO on some
99    machines, this should use an uncacheable mapping as a fallback.
100
101    mmap of /sys/class/pci_bus/.../legacy_mem
102
103    This is very similar to mmap of /dev/mem, except that legacy_mem
104    only allows mmap of the one megabyte "legacy MMIO" area for a
105    specific PCI bus. Typically this is the first megabyte of
106    physical address space, but it may be different on machines with
107    several VGA devices.
108
109    "X" uses this to access VGA frame buffers. Using legacy_mem
110    rather than /dev/mem allows multiple instances of X to talk to
111    different VGA cards.
112
113    The /dev/mem mmap constraints apply.
114
115    mmap of /proc/bus/pci/.../??.?
116
117        This is an MMIO mmap of PCI functions, which additionally may or
118    may not be requested as using the WC attribute.
119
120    If WC is requested, and the region in kern_memmap is either WC
121    or UC, and the EFI memory map designates the region as WC, then
122    the WC mapping is allowed.
123
124    Otherwise, the user mapping must use the same attribute as the
125    kernel mapping.
126
127    read/write of /dev/mem
128
129    This uses copy_from_user(), which implicitly uses a kernel
130    identity mapping. This is obviously safe for things in
131    kern_memmap.
132
133    There may be corner cases of things that are not in kern_memmap,
134    but could be accessed this way. For example, registers in MMIO
135    space are not in kern_memmap, but could be accessed with a UC
136    mapping. This would not cause attribute aliasing. But
137    registers typically can be accessed only with four-byte or
138    eight-byte accesses, and the copy_from_user() path doesn't allow
139    any control over the access size, so this would be dangerous.
140
141    ioremap()
142
143    This returns a mapping for use inside the kernel.
144
145    If the region is in kern_memmap, we should use the attribute
146    specified there.
147
148    If the EFI memory map reports that the entire granule supports
149    WB, we should use that (granules that are partially reserved
150    or occupied by firmware do not appear in kern_memmap).
151
152    If the granule contains non-WB memory, but we can cover the
153    region safely with kernel page table mappings, we can use
154    ioremap_page_range() as most other architectures do.
155
156    Failing all of the above, we have to fall back to a UC mapping.
157
158PAST PROBLEM CASES
159
160    mmap of various MMIO regions from /dev/mem by "X" on Intel platforms
161
162      The EFI memory map may not report these MMIO regions.
163
164      These must be allowed so that X will work. This means that
165      when the EFI memory map is incomplete, every /dev/mem mmap must
166      succeed. It may create either WB or UC user mappings, depending
167      on whether the region is in kern_memmap or the EFI memory map.
168
169    mmap of 0x0-0x9FFFF /dev/mem by "hwinfo" on HP sx1000 with VGA enabled
170
171      The EFI memory map reports the following attributes:
172        0x00000-0x9FFFF WB only
173        0xA0000-0xBFFFF UC only (VGA frame buffer)
174        0xC0000-0xFFFFF WB only
175
176      This mmap is done with user pages, not kernel identity mappings,
177      so it is safe to use WB mappings.
178
179      The kernel VGA driver may ioremap the VGA frame buffer at 0xA0000,
180      which uses a granule-sized UC mapping. This granule will cover some
181      WB-only memory, but since UC is non-speculative, the processor will
182      never generate an uncacheable reference to the WB-only areas unless
183      the driver explicitly touches them.
184
185    mmap of 0x0-0xFFFFF legacy_mem by "X"
186
187      If the EFI memory map reports that the entire range supports the
188      same attributes, we can allow the mmap (and we will prefer WB if
189      supported, as is the case with HP sx[12]000 machines with VGA
190      disabled).
191
192      If EFI reports the range as partly WB and partly UC (as on sx[12]000
193      machines with VGA enabled), we must fail the mmap because there's no
194      safe attribute to use.
195
196      If EFI reports some of the range but not all (as on Intel firmware
197      that doesn't report the VGA frame buffer at all), we should fail the
198      mmap and force the user to map just the specific region of interest.
199
200    mmap of 0xA0000-0xBFFFF legacy_mem by "X" on HP sx1000 with VGA disabled
201
202      The EFI memory map reports the following attributes:
203        0x00000-0xFFFFF WB only (no VGA MMIO hole)
204
205      This is a special case of the previous case, and the mmap should
206      fail for the same reason as above.
207
208    read of /sys/devices/.../rom
209
210      For VGA devices, this may cause an ioremap() of 0xC0000. This
211      used to be done with a UC mapping, because the VGA frame buffer
212      at 0xA0000 prevents use of a WB granule. The UC mapping causes
213      an MCA on HP sx[12]000 chipsets.
214
215      We should use WB page table mappings to avoid covering the VGA
216      frame buffer.
217
218NOTES
219
220    [1] SDM rev 2.2, vol 2, sec 4.4.1.
221    [2] SDM rev 2.2, vol 2, sec 4.4.6.
222

Archive Download this file



interactive