Difference between revisions of "Memory Management"

From 3dbrew
Jump to navigation Jump to search
m (Neobrain moved page Memory Allocation to Memory Management: This is related enough to fit into one page)
(Update FCRAM memory management from latest TuxSH kernel idb)
 
(18 intermediate revisions by 6 users not shown)
Line 1: Line 1:
 
= Types and structures =
 
= Types and structures =
 
== enum MemoryState ==
 
{| class="wikitable" border="1"
 
!  Memory state flags
 
!  Value
 
|-
 
| FREE
 
| 0
 
|-
 
| RESERVED
 
| 1
 
|-
 
| IO
 
| 2
 
|-
 
| STATIC
 
| 3
 
|-
 
| CODE
 
| 4
 
|-
 
| PRIVATE
 
| 5
 
|-
 
| SHARED
 
| 6
 
|-
 
| CONTINUOUS
 
| 7
 
|-
 
| ALIASED
 
| 8
 
|-
 
| ALIAS
 
| 9
 
|-
 
| ALIAS CODE
 
| 10
 
|-
 
| LOCKED
 
| 11
 
|}
 
  
 
== enum MemoryOperation ==
 
== enum MemoryOperation ==
Line 98: Line 56:
 
| 1
 
| 1
 
|-
 
|-
| W
+
| W (Invalid, see below)
 
| 2
 
| 2
 
|-
 
|-
Line 118: Line 76:
 
| DONTCARE
 
| DONTCARE
 
| 0x10000000
 
| 0x10000000
 +
|}
 +
 +
Specifying write-permission without read-permission is invalid, it's handled the same way as if the RW bits were not set(ARM11-MPCore doesn't support write-only memory permissions).
 +
 +
== enum MemoryState ==
 +
{| class="wikitable" border="1"
 +
!  Memory state flags
 +
!  Value
 +
|-
 +
| FREE
 +
| 0
 +
|-
 +
| RESERVED
 +
| 1
 +
|-
 +
| IO
 +
| 2
 +
|-
 +
| STATIC
 +
| 3
 +
|-
 +
| CODE
 +
| 4
 +
|-
 +
| PRIVATE
 +
| 5
 +
|-
 +
| SHARED
 +
| 6
 +
|-
 +
| CONTINUOUS
 +
| 7
 +
|-
 +
| ALIASED
 +
| 8
 +
|-
 +
| ALIAS
 +
| 9
 +
|-
 +
| ALIAS CODE
 +
| 10
 +
|-
 +
| LOCKED
 +
| 11
 
|}
 
|}
  
Line 155: Line 157:
 
!  Field
 
!  Field
 
|-
 
|-
| u32
+
| PageFlags (u32)
 
| Flags
 
| Flags
 
|}
 
|}
 
  
 
= Memory Mapping =
 
= Memory Mapping =
Line 167: Line 168:
 
ControlMemory parameter MemoryType with value 0x10003 is used for mapping the GSP [[Memory_layout|heap]]. The low 8-bits are the type: 1 is for un-mapping memory, 3 for mapping memory. Type4 is used to mirror the RW memory at Addr1, to Addr0. Type4 will return an error if Addr1 is located in read-only memory. Addr1 is not used for type1 and type3.
 
ControlMemory parameter MemoryType with value 0x10003 is used for mapping the GSP [[Memory_layout|heap]]. The low 8-bits are the type: 1 is for un-mapping memory, 3 for mapping memory. Type4 is used to mirror the RW memory at Addr1, to Addr0. Type4 will return an error if Addr1 is located in read-only memory. Addr1 is not used for type1 and type3.
  
The ARM11 kernel does not allow processes to create shared memory blocks via svcCreateMemoryBlock, when the process memorytype (from the kernel flags stored in the exheader kernel descriptor) is the application memorytype, and when addr=0. When the memorytype is not the application memorytype and addr=0, the kernel allocates new memory for the calling process and turns it into a shared memory block. When addr is non-zero, it must be located in memory which is already mapped. Furthermore, it appears that only regular heap pages (allocated using svcControlMemory op=COMMIT) are accepted as valid addrs.
+
The ARM11 kernel does not allow processes to create shared memory blocks via svcCreateMemoryBlock, when the process memorytype (from the kernel flags stored in the exheader kernel descriptor) is the application memorytype, and when addr=0. When the memorytype is not the application memorytype and addr=0, the kernel allocates new memory for the calling process and turns it into a shared memory block. When addr is non-zero, it must be located in memory which is already mapped. Furthermore, it appears that only regular heap pages (allocated using svcControlMemory op=COMMIT) are accepted as valid addrs. The addr(+size) must be >=0x00100000 and <0x14000000.
  
 
ControlProcessMemory maps memory in the specified process, this is the only SVC which allows mapping executable memory. Format of the permissions field for memory mapping SVCs: bit0=R, bit1=W, bit2=X. Type6 sets the Addr0 memory permissions to the input permissions, for already mapped memory. Type is the MemoryOperation enum, without the memory-type/memory-region. ControlProcessMemory only supports type4, type5, and type6. ControlProcessMemory does not support using the current KProcess handle alias.
 
ControlProcessMemory maps memory in the specified process, this is the only SVC which allows mapping executable memory. Format of the permissions field for memory mapping SVCs: bit0=R, bit1=W, bit2=X. Type6 sets the Addr0 memory permissions to the input permissions, for already mapped memory. Type is the MemoryOperation enum, without the memory-type/memory-region. ControlProcessMemory only supports type4, type5, and type6. ControlProcessMemory does not support using the current KProcess handle alias.
  
MapProcessMemory maps RW memory starting at address 0x00100000 in the specified KProcess, at the specified StartAddr in the current process. MapProcessMemory then maps 0x08000000 in the specified process, to StartAddr+0x7f00000 in the current process. UnmapProcessMemory unmaps the memory which was mapped by MapProcessMemory.
+
Note that with the MAP MemoryOperation, the kernel will refuse to MAP memory for the specified addr1, when addr1 was already used with another MAP operation as addr1. The kernel also doesn't allow memory to be freed via the FREE MemoryOperation, when other virtual-memory is mapped to this same memory(when the MAP MemoryOperation was used with this memory with addr1). With the MAP MemoryOperation, the memory permissions for the original buffer are also set to no-access.
  
Note that with the MAP MemoryOperation, the kernel will refuse to MAP memory for the specified addr1, when addr1 was already used with another MAP operation as addr1. The kernel also doesn't allow memory to be freed via the FREE MemoryOperation, when other virtual-memory is mapped to this same memory(when the MAP MemoryOperation was used with this memory with addr1).
+
<code>'''MapProcessMemory'''(Handle process, u32 destAddr, u32 limit)</code> maps memory from the given process into the current process. This memory is mapped with permissions RW-, regardless of the permissions for the memory under the specified process. First <code>min(limit, 0x3F00000)</code> bytes are mapped starting from <code>0x00100000</code> in the source process to <code>destAddr</code> in the current process. Then <code>min(limit - 0x7F00000, 0x6000000)</code> bytes (if more than 0) are mapped from <code>0x08000000</code> in the source process to <code>destAddr + 0x7F00000</code> in the current process. Another way to view this is that it is overlaying the two ranges <code>[0x0010_0000; 0x0400_0000]</code> and <code>[0x0800_0000; 0x0E00_0000]</code> from the source process onto <code>[destAddr - 0x100000; destAddr + limit]</code> in the current process, truncating whatever part of the mapping that doesn't fit.  This system call is used by [[RO Services]] to map the program's code and heap into ro. Memory mapped by MapProcessMemory is unmapped by UnmapProcessMemory.
  
 
= How The Kernel Allocates And Tracks Memory =
 
= How The Kernel Allocates And Tracks Memory =
  
== MemoryBlockHeader ==
+
FCRAM (128MiB for O3DS, or 256MiB for N3DS) is divided into three regions: APPLICATION, SYSTEM and BASE. A program is allowed to allocate memory in one of the three region. (For example games are always in the APPLICATION region). Inside one region, there are two kinds of memory that can be allocated: (regular) heap and linear heap.
  
Size : 0xC bytes
+
(Regular) heap is allocated starting from the end of the memory region, and growing down. They are mostly for application private use, and sometimes for software-based memory sharing (with other process using KSharedMemory, for example). They can be mapped to anywhere inside 0x08000000~0x10000000 virtual memory range upon application request. When the application requests for a block of heap, the block is not guaranteed continuous in FCRAM and the location is not specified either.
 +
 
 +
Linear heap is allocated starting from the beginning of the memory region, and growing up. They are mostly for hardware-based memory sharing (with GPU, DMA etc.), but can also be used privately as well. They can be mapped to the linear heap virtual memory region (0x14000000+ or 0x30000000+, depending on game's kernel version). When the application requests for a linear heap block, the block is always continuous in FCRAM, and the difference between physical address and virtual address is always a constant. Therefore it gives virtual to physical address convertibility, which enables communication with other hardware.
 +
 
 +
Each region in the kernel is managed by a dedicated KPageHeap object, which tracks free memory blocks. Each free block is represented with a KPageHeapBlock structure stored in FCRAM that stores the size of the free region and links to adjacent blocks. When a chunk of memory is allocated from the heap, the block is removed from the linked list and the region is cleared, erasing the header. On initialization a free block is inserted that covers the entire region.
 +
 
 +
== KPageHeapBlock ==
 +
 
 +
Size: 0xC-bytes(?) pre-v11.0, 0x18-bytes starting with v11.0.
  
 
{| class="wikitable" border="1"
 
{| class="wikitable" border="1"
Line 192: Line 201:
 
|-
 
|-
 
| 0x4
 
| 0x4
| MemoryBlockHeader*
+
| KPageHeapBlock*
 
| Next
 
| Next
 
|-
 
|-
 
| 0x8
 
| 0x8
| MemoryBlockHeader*
+
| KPageHeapBlock*
 
| Prev
 
| Prev
 +
|-
 +
| 0xC
 +
| KPageHeapBlock*
 +
| Pointer to the current memchunk. Added with v11.0?
 +
|-
 +
| 0x10
 +
| u32
 +
| Nonce, doesn't seem to be read at all except during MAC calculation. Added with v11.0. Used with the new heap [[11.0.0-33|security]] feature. A kernel state field is copied to this field before calculating the MAC. Once done, that kernel state field is subtracted by the value of the calculated MAC stored below. Since this kernel state field is initially 0x0, this field for the FCRAM APPLICATION+0 KPageHeapBlock during kernel boot is set to 0x0.
 +
|-
 +
| 0x14
 +
| u32
 +
| MAC calculated over rest of struct. Added with v11.0. Used with the new heap [[11.0.0-33|security]] feature.
 
|}
 
|}
  
 +
== KPageHeap ==
  
== RegionDescriptor ==
+
Size: 0x10-bytes pre-[[11.0.0-33|11.0.0-X]], 0x20-bytes starting with [[11.0.0-33|11.0.0-X]].
 
 
Size : 0x10 bytes
 
  
 
{| class="wikitable" border="1"
 
{| class="wikitable" border="1"
Line 212: Line 232:
 
|-
 
|-
 
| 0x0
 
| 0x0
| MemoryBlockHeader*
+
| KPageHeapBlock*
|  
+
| First
 
|-
 
|-
 
| 0x4
 
| 0x4
| MemoryBlockHeader*
+
| KPageHeapBlock*
|  
+
| Last
 
|-
 
|-
 
| 0x8
 
| 0x8
Line 226: Line 246:
 
| u32
 
| u32
 
| Region size
 
| Region size
 +
|-
 +
| 0x10
 +
| 0x10-byte block.
 +
| This is the "key" used with the kernel heap MAC implemented with [[11.0.0-33|11.0.0-X]].
 
|}
 
|}
  
 +
The KPageHeap objects are owned by the KMemoryManager. This object acts as a light wrapper that abstracts allocations and freeing of memory from the heaps and handles page reference tracking.
  
== FCRAM Region Descriptor ==
+
== KMemoryManager ==
  
Size : 0x50 bytes
+
Size: 0x50-bytes pre-[[11.0.0-33|11.0.0-X]], 0x80-bytes starting with [[11.0.0-33|11.0.0-X]].
 +
 
 +
X = KPageHeap_size*3.
  
 
{| class="wikitable" border="1"
 
{| class="wikitable" border="1"
Line 240: Line 267:
 
|-
 
|-
 
| 0x0
 
| 0x0
| RegionDescriptor
+
| KPageHeap
| RegionDescriptor for app memory
+
| KPageHeap for app memory
 
|-
 
|-
| 0x10
+
| KPageHeap_size*1
| RegionDescriptor
+
| KPageHeap
| RegionDescriptor for sys memory
+
| KPageHeap for sys memory
 
|-
 
|-
| 0x20
+
| KPageHeap_size*2
| RegionDescriptor
+
| KPageHeap
| RegionDescriptor for base memory
+
| KPageHeap for base memory
 
|-
 
|-
| 0x30
+
| X + 0x0
| u32
+
| KMemoryManager*
 
| Ptr to start of FCRAM region descriptor
 
| Ptr to start of FCRAM region descriptor
 
|-
 
|-
| 0x34
+
| X + 0x4
 
| u32
 
| u32
 
| FCRAM start
 
| FCRAM start
 
|-
 
|-
| 0x38
+
| X + 0x8
 
| u32
 
| u32
 
| FCRAM size in pages
 
| FCRAM size in pages
 
|-
 
|-
| 0x3C
+
| X + 0xC
| u32
+
| u32*
| Base memory start in FCRAM
+
| Pointer to FCRAM memory used for page reference tracking. Each u32 represents a page.
 
|-
 
|-
| 0x40
+
| X + 0x10
 
| u32
 
| u32
 
| Count of physical FCRAM used by the kernel, in bytes. (used by [[SVC|svcGetSystemInfo]])
 
| Count of physical FCRAM used by the kernel, in bytes. (used by [[SVC|svcGetSystemInfo]])
 
|-
 
|-
| 0x44
+
| X + 0x14
 
| u32
 
| u32
 
| ?
 
| ?
 
|-
 
|-
| 0x48
+
| X + 0x18
| KThread*
+
| KLightMutex
| Thread operating on region data
+
| Mutex used for thread synchronization during memory (de)allocation from the heaps.
|-
 
| 0x4C
 
| s16
 
| Error info for thread listed above
 
|-
 
| 0x4E
 
| u16
 
| Alignment
 
 
|}
 
|}
  
  
 
== Kernel Region Descriptor ==
 
== Kernel Region Descriptor ==

Latest revision as of 01:29, 31 March 2024

Types and structures[edit]

enum MemoryOperation[edit]

Memory operation Id
FREE 1
RESERVE 2
COMMIT 3
MAP 4
UNMAP 5
PROTECT 6
REGION APP 0x100
REGION SYSTEM 0x200
REGION BASE 0x300
LINEAR 0x10000

The LINEAR memory-operation indicates that the mapped physical address is always MappedVAddr+0x0C000000, thus this memory can be used for hardware devices' DMA(such as the GPU). Addr0+size for this must be within the 0x14000000-0x1C000000 range when Addr0 is non-zero(Addr1 must be zero), Addr0 isn't actually used by svcControlMemory for mapping memory: Addr0 is not used by the kernel after doing address-range checks. The kernel determines what physical-address to use by allocating memory from FCRAM(about the same way as other memory), which is then used to determine the virtual-address.

8.0.0-18 added a new memory mapping(0x30000000-0x38000000) for LINEAR memory, this replaces the original mapping for newer titles. The kernel uses the new mapping when the process memory-region is BASE, or when the process kernel-release-version field is >=0x022c(2.44 / system-version 8.0.0-18).

The input mem-region value for svcControlMemory is only used(when non-zero) when the PID is value 1, for the FIRM ARM11 "loader" module.

enum MemoryPermission[edit]

Memory permission Id
NONE 0
R 1
W (Invalid, see below) 2
RW 3
X 4
RX 5
WX 6
RWX 7
DONTCARE 0x10000000

Specifying write-permission without read-permission is invalid, it's handled the same way as if the RW bits were not set(ARM11-MPCore doesn't support write-only memory permissions).

enum MemoryState[edit]

Memory state flags Value
FREE 0
RESERVED 1
IO 2
STATIC 3
CODE 4
PRIVATE 5
SHARED 6
CONTINUOUS 7
ALIASED 8
ALIAS 9
ALIAS CODE 10
LOCKED 11

struct MemoryInfo[edit]

Type Field
u32 Base process virtual address
u32 Size
u32 Permission
enum MemoryState State

enum PageFlags[edit]

Page flags Bit
LOCKED 0
CHANGED 1

struct PageInfo[edit]

Type Field
PageFlags (u32) Flags

Memory Mapping[edit]

ControlMemory and MapMemoryBlock can be used to map memory pages, these two SVCs only support mapping execute-never R/W pages. The input permissions parameter for these SVCs must therefore be <=3, where value zero is used when un-mapping memory. Furthermore it appears that only regular heap pages can be mirrored (it won't work for TLS, stack, .data, .text, for example).

Bitmask 0xF00 for ControlMemory parameter MemoryType is the memory-type, when this is zero the memory-type is loaded from the kernel flags stored in the exheader ARM11 kernel descriptors, for the process using the SVC.

ControlMemory parameter MemoryType with value 0x10003 is used for mapping the GSP heap. The low 8-bits are the type: 1 is for un-mapping memory, 3 for mapping memory. Type4 is used to mirror the RW memory at Addr1, to Addr0. Type4 will return an error if Addr1 is located in read-only memory. Addr1 is not used for type1 and type3.

The ARM11 kernel does not allow processes to create shared memory blocks via svcCreateMemoryBlock, when the process memorytype (from the kernel flags stored in the exheader kernel descriptor) is the application memorytype, and when addr=0. When the memorytype is not the application memorytype and addr=0, the kernel allocates new memory for the calling process and turns it into a shared memory block. When addr is non-zero, it must be located in memory which is already mapped. Furthermore, it appears that only regular heap pages (allocated using svcControlMemory op=COMMIT) are accepted as valid addrs. The addr(+size) must be >=0x00100000 and <0x14000000.

ControlProcessMemory maps memory in the specified process, this is the only SVC which allows mapping executable memory. Format of the permissions field for memory mapping SVCs: bit0=R, bit1=W, bit2=X. Type6 sets the Addr0 memory permissions to the input permissions, for already mapped memory. Type is the MemoryOperation enum, without the memory-type/memory-region. ControlProcessMemory only supports type4, type5, and type6. ControlProcessMemory does not support using the current KProcess handle alias.

Note that with the MAP MemoryOperation, the kernel will refuse to MAP memory for the specified addr1, when addr1 was already used with another MAP operation as addr1. The kernel also doesn't allow memory to be freed via the FREE MemoryOperation, when other virtual-memory is mapped to this same memory(when the MAP MemoryOperation was used with this memory with addr1). With the MAP MemoryOperation, the memory permissions for the original buffer are also set to no-access.

MapProcessMemory(Handle process, u32 destAddr, u32 limit) maps memory from the given process into the current process. This memory is mapped with permissions RW-, regardless of the permissions for the memory under the specified process. First min(limit, 0x3F00000) bytes are mapped starting from 0x00100000 in the source process to destAddr in the current process. Then min(limit - 0x7F00000, 0x6000000) bytes (if more than 0) are mapped from 0x08000000 in the source process to destAddr + 0x7F00000 in the current process. Another way to view this is that it is overlaying the two ranges [0x0010_0000; 0x0400_0000] and [0x0800_0000; 0x0E00_0000] from the source process onto [destAddr - 0x100000; destAddr + limit] in the current process, truncating whatever part of the mapping that doesn't fit. This system call is used by RO Services to map the program's code and heap into ro. Memory mapped by MapProcessMemory is unmapped by UnmapProcessMemory.

How The Kernel Allocates And Tracks Memory[edit]

FCRAM (128MiB for O3DS, or 256MiB for N3DS) is divided into three regions: APPLICATION, SYSTEM and BASE. A program is allowed to allocate memory in one of the three region. (For example games are always in the APPLICATION region). Inside one region, there are two kinds of memory that can be allocated: (regular) heap and linear heap.

(Regular) heap is allocated starting from the end of the memory region, and growing down. They are mostly for application private use, and sometimes for software-based memory sharing (with other process using KSharedMemory, for example). They can be mapped to anywhere inside 0x08000000~0x10000000 virtual memory range upon application request. When the application requests for a block of heap, the block is not guaranteed continuous in FCRAM and the location is not specified either.

Linear heap is allocated starting from the beginning of the memory region, and growing up. They are mostly for hardware-based memory sharing (with GPU, DMA etc.), but can also be used privately as well. They can be mapped to the linear heap virtual memory region (0x14000000+ or 0x30000000+, depending on game's kernel version). When the application requests for a linear heap block, the block is always continuous in FCRAM, and the difference between physical address and virtual address is always a constant. Therefore it gives virtual to physical address convertibility, which enables communication with other hardware.

Each region in the kernel is managed by a dedicated KPageHeap object, which tracks free memory blocks. Each free block is represented with a KPageHeapBlock structure stored in FCRAM that stores the size of the free region and links to adjacent blocks. When a chunk of memory is allocated from the heap, the block is removed from the linked list and the region is cleared, erasing the header. On initialization a free block is inserted that covers the entire region.

KPageHeapBlock[edit]

Size: 0xC-bytes(?) pre-v11.0, 0x18-bytes starting with v11.0.

Offset Type Description
0x0 u32 Size in pages
0x4 KPageHeapBlock* Next
0x8 KPageHeapBlock* Prev
0xC KPageHeapBlock* Pointer to the current memchunk. Added with v11.0?
0x10 u32 Nonce, doesn't seem to be read at all except during MAC calculation. Added with v11.0. Used with the new heap security feature. A kernel state field is copied to this field before calculating the MAC. Once done, that kernel state field is subtracted by the value of the calculated MAC stored below. Since this kernel state field is initially 0x0, this field for the FCRAM APPLICATION+0 KPageHeapBlock during kernel boot is set to 0x0.
0x14 u32 MAC calculated over rest of struct. Added with v11.0. Used with the new heap security feature.

KPageHeap[edit]

Size: 0x10-bytes pre-11.0.0-X, 0x20-bytes starting with 11.0.0-X.

Offset Type Description
0x0 KPageHeapBlock* First
0x4 KPageHeapBlock* Last
0x8 u32 Region start
0xC u32 Region size
0x10 0x10-byte block. This is the "key" used with the kernel heap MAC implemented with 11.0.0-X.

The KPageHeap objects are owned by the KMemoryManager. This object acts as a light wrapper that abstracts allocations and freeing of memory from the heaps and handles page reference tracking.

KMemoryManager[edit]

Size: 0x50-bytes pre-11.0.0-X, 0x80-bytes starting with 11.0.0-X.

X = KPageHeap_size*3.

Offset Type Description
0x0 KPageHeap KPageHeap for app memory
KPageHeap_size*1 KPageHeap KPageHeap for sys memory
KPageHeap_size*2 KPageHeap KPageHeap for base memory
X + 0x0 KMemoryManager* Ptr to start of FCRAM region descriptor
X + 0x4 u32 FCRAM start
X + 0x8 u32 FCRAM size in pages
X + 0xC u32* Pointer to FCRAM memory used for page reference tracking. Each u32 represents a page.
X + 0x10 u32 Count of physical FCRAM used by the kernel, in bytes. (used by svcGetSystemInfo)
X + 0x14 u32 ?
X + 0x18 KLightMutex Mutex used for thread synchronization during memory (de)allocation from the heaps.


Kernel Region Descriptor[edit]