HIPC: Difference between revisions

== IPC Command Structure ==

This is an array of u32's, usually located in [[Thread Local Storage]].

| 0 || 15-0 || [[#Type|Type]].

| 0 || 19-16 || Number of buf X descriptors (each: 2 words).

| 0 || 23-20 || Number of buf A descriptors (each: 3 words).

| 0 || 27-24 || Number of buf B descriptors (each: 3 words).

| 0 || 31-28 || Number of buf W desciptors (each: 3 words), not observed

| 1 || 9-0 || Size of [[#Raw data section|raw data]] in u32s.

| 1 || 13-10 || Flags for buf C descriptor.

| 1 || 30-20 || ?

| 1 || 31 || Enable handle descriptor.

| ... || || [[#Handle descriptor|Handle descriptor]], if enabled.

| ... || || [[#Buffer descriptor X "Pointer"|Buf X descriptors]], each one 2 words.

| ... || || [[#Buffer descriptor A/B/W "Send"/"Receive"/"Exchange"|Buf A descriptors]], each one 3 words.

| ... || || [[#Buffer descriptor A/B/W "Send"/"Receive"/"Exchange"|Buf B descriptors]], each one 3 words.

| ... || || [[#Buffer descriptor A/B/W "Send"/"Receive"/"Exchange"|Type W descriptors]], each one 3 words.

| ... || || [[#Raw_data_section|Raw data section]] (including padding before and after aligned data section).

| ... || || [[#Buffer descriptor C "ReceiveList"|Buf C descriptors]], each one 2 words.

First two header u32's and handle descriptor (if enabled) are copied as-is from one process to the other.

=== Type ===

IPC commands can have different types which influence how the IPC server processes requests in "nn::sf::hipc::server::HipcServerSessionManagerBase::ProcessRequest".

! Value || Name

| 0 || Invalid

| 1 || [[#LegacyRequest, LegacyControl|LegacyRequest]]

| 2 || [[#Close|Close]]

| 3 || [[#LegacyRequest, LegacyControl|LegacyControl]]

| 4 || [[#Request, Control|Request]]

| 5 || [[#Request, Control|Control]]

| 6 || [5.0.0+] [[#RequestWithContext, ControlWithContext|RequestWithContext]]

| 7 || [5.0.0+] [[#RequestWithContext, ControlWithContext|ControlWithContext]]

==== Close ====

* "nn::sf::hipc::server::HipcServerSessionManager::DestroyServerSession"

* "nn::sf::hipc::CloseServerSessionHandle"

==== LegacyRequest, LegacyControl ====

* "nn::sf::hipc::detail::HipcMessageBufferAccessor::ParseHeader"

* "nn::sf::hipc::server::HipcServerSessionManager::RegisterServerSessionToWaitBase"

==== Request, Control ====

* "nn::sf::hipc::server::HipcServerSessionManager::RegisterServerSessionToWaitBase"

==== RequestWithContext, ControlWithContext ====

These are identical to normal Request and Control types, but with the additional requirement of suppling a token in their [[#Data payload|data payload]].

This token is used by "nn::sf::cmif::SetInlineContext" which has the sole purpose of saving it into the TLS in order for it to be distributed to any IPC commands that are made while processing the current command. It's unknown if this token serves any purpose or if it's just a debug-tool to figure out what IPC command caused a particular chain of commands.

=== Handle descriptor ===

There can only be one of this descriptor type. It is enabled by bit31 of the second word.

! Word || Bits || Description

| 0 || 0 || Send current PID.

| 0 || 4-1 || Number of handles to copy

| 0 || 8-5 || Number of handles to move

| ... || || 8-byte PID if enabled

| ... || || Handles to copy

| ... || || Handles to move

Sysmodules load the last u64 of rawdata when handling the PID. This is not written by kernel. For sysmodule handling:

* In some cases: these commands require a placeholder u64 value passed in the input parameters, as mentioned above. In these cases the OverwriteClientProcessId method is called to replace the value before it is used.

* In other cases: The rawdata_u64 is compared with the PID from the descriptor. On mismatch and when rawdata_u64!=0, error 0x60A is returned. The PID value passed to the cmdhandler vtable funcptr is the rawdata_u64.

Handle 0 is allowed, and just means no handle was sent.

=== Buffer descriptor X "Pointer" ===

This one is packed even worse than A, they inserted the bit38-36 of the address ''on top'' of the counter field.

Officially, the counter is known as "receive index". This one writes to the buffer described in the ReceiveList.

! Word || Bits || Description

| 0 || 5-0 || Bits 5-0 of counter.

| 0 || 8-6 || Bit 38-36 of address.

| 0 || 11-9 || Bits 11-9 of counter.

| 0 || 15-12 || Bit 35-32 of address.

| 0 || 31-16 || Size

| 1 || || Lower 32-bits of address.

=== Buffer descriptor A/B/W "Send"/"Receive"/"Exchange" ===

This packing is so unnecessarily complex.

! Word || Bits || Description

| 0 || || Lower 32-bits of size.

| 1 || || Lower 32-bits of address.

| 2 || 1-0 || Flags. Always set to 0, 1 or 3.

| 2 || 4-2 || Bit 38-36 of address.

| 2 || 27-24 || Bit 35-32 of size.

| 2 || 31-28 || Bit 35-32 of address.

A reply must not use A/B/W, svcReplyAndReceive will return 0xE801.

[[SVC|MemoryAttribute]] IsBorrowed and IsUncached are never allowed for the source address.

"Send" means buffer is sent from source process into service process.

"Receive" means that data is copied from service process into user process.

"Exchange" means both "Send" and "Receive".

==== Flags ====

Determines what [[SVC|MemoryState]] to use with the mapped memory in the sysmodule.

Used to enforce whether or not device mapping is allowed for src and dst buffers respectively.

There's a 4-bit flag in the main header controlling the behavior of C descriptors.

If it has value 0, the C descriptor functionality is disabled.

If it has value 1, there is an "inlined" C buffer after the raw data. Received data is copied to ROUND_UP(cmdbuf+raw_size+index, 16)

If it has value 2, there is a single C descriptor.

Otherwise it has (flag-2) C descriptors. In this case, index picks which C descriptor to copy received data to [instead of picking the offset into the buffer].

Data sent with this method must have MemoryState 0x4000000 mask set.

After reply, X descriptors are written to the sender containing the address, size and index that were copied to.

| 0 || || Lower 32-bits of address.

| 1 || 15-0 || Rest of address.

| 1 || 31-16 || Size

=== IPC buffers ===

Buffer descriptor A/B/... map memory into the sysmodule process. For the mapped memory in the sysmodule the permissions are: desc-A = R--, desc-B = RW-. The buffer is automatically unmapped while the kernel handles the cmdreply, the sysmodule doesn't need to specify anything in the cmdreply to trigger this.

This memory is mapped in the sysmodule to the same vaddr from the original user-process cmd-request, except with with bits >=(~28(?)) changed to a different ASLR'd region.

No user-process->sysmodule memcpy is done for outbufs, only sysmodule->user-process.

Buffer descriptors C/X are somewhat different. Rather than mapping new memory into the server process, C/X descriptors copy data between existing buffers in different processes. Each X descriptor in a message has its data copied into a C descriptor on the other side. Each C descriptor in a message is used to reserve space for the other side's X descriptors to copy into.

When the kernel processes X descriptors, it must determine where to copy the data to. If the destination used C descriptors with flags >= 3, each X descriptor from the source is matched to a C descriptor in the destination by the X descriptor's index field. If the destination used a "single" C descriptor, the data from all the X descriptors is copied into the same buffer specified by the destination's C descriptor (causing error 0xce01 if there is not enough space) and the X descriptor index is ignored. The kernel then modifies the addresses in the X descriptors to indicate where the data was copied to in the destination.

Before receiving a request, if the IPC server is expecting X descriptors, it prepares a message with a "single" C descriptor (flags=2) in its message buffer before calling svcReplyAndReceive so that X descriptors from the client have a place to copy their data to.  The usage of the flag-2 C descriptor allows the server to receive an arbitrary number of X descriptors, since they're all packed into the same buffer. If the server had used flag-3+ C descriptors, it would be limited in how many X descriptors it could receive since the X descriptors would have to be matched to distinct C descriptors. The buffer that the server's C descriptor points to is called the '''pointer buffer'''.

When the client sends X descriptors, data is copied into the server's pointer buffer. When the client sends C descriptors, no data is copied automatically. The server needs to use X descriptors to copy the data back to the client's C descriptors (using the index field to match X descriptors in the response back to the correct C descriptors).

[[File:Ipc msg buffer type a example.png|thumb|An example of an IPC message with a type 0xA buffer in it. Red is headers/descriptors, yellow is padding, and blue is data/buffer lengths. Note that the size of the u16 array for type A lengths is padded to fill up a whole word.]]

| ... || Padding to align to 16 bytes.

| ... || If sent to an object domain, a [[#Domain_message|domain message]], otherwise a [[#Data payload|data payload]]

| ... || Padding

| ... || Buffer type 0xA lengths (u16 array)

The total amount of padding within the raw data section is always 0x10 bytes. This means that if no padding is required before the message, there will be 0x10 bytes of padding after the message (before the buffer type 0xA lengths).

Because the switch has relatively low limits on the total number of sessions available to the system (Kernel slabheap limits, sysmodule handle table size limits), HIPC supports a "Domains" feature that allows multiplexing multiple service sessions through a single handle. Domains store (effectively) a mapping from u32 object id to a SharedPointer<IServiceObject> -- When messages are sent to a domain, an extra header is sent in the raw data section (before anything else) with information about what object in the domain is being acted on; responses similarly contain an additional header. Official session code implements this by just using the dispatch table for the object in the map with the appropriate ID, instead of the dispatch table the session was initialized with.

| 0 || 7-0 || Command. 1=send message, 2=close virtual handle

| 0 || 8-15 || Input object count

| 0 || 31-16 || Length of [[IPC_Marshalling#Data_payload|data payload]] in bytes.

| 1 || || Object ID (from cmd 0 in [[IPC_Marshalling#Control|Control]]).

| 2 || || Padding

| 3 || || [5.0.0+] Token for (NewRequest only)

| 4... || || [[#Data payload|Data payload]]

|-

| ... || || Input object IDs (u32s, not aligned)

Format for the extra response header for domain message:

| 0 || 31-0 || Output object count

| 1-3 || || Padding

=== Data payload ===

This is an array of u32's, but individual parameters are generally stored as u64's.

| 1 || Version as u32. 1 for NewRequest, 0 for Request.

| 2 || Command id as u64 for requests, [[Error_codes|error code]] as u64 for responses.

| 3 || [5.0.0+] Token (for NewRequest only, non-domain messages). On version < 5.0.0, domain messages or Request messages, simply Zero.

|-

| 4... || Input parameters or return values

[5.0.0+] Version was incremented from 0 to 1, and a token value was introduced into raw_data+12 (regardless of domain or not, in either case it overlaps with padding).

The input rawdata struct is generated by stable-sorting function parameters by alignment, from low to high. It is likely this is a mistake, as it generates structs with suboptimal possible padding -- Nintendo probably meant to sort from high to low (which would give minimized padding), but couldn't/can't change this without breaking backwards compatibility.

== Official marshalling code ==

The official marshalling function takes an array of (buf_ptr, size) pairs and a type-field for each such pair.

Bitmask 0x10 seems to indicate null-terminated strings.

! Type Mask || Description || Direction

| 4 + 1 || Creates a A descriptor with flags=0. || In

| 0x40 + 4 + 1 || Creates a A descriptor with flags=1. || In

| 0x80 + 4 + 1 || Creates a A descriptor with flags=3. || In

| 4 + 2 || Creates a B descriptor with flags=0. || Out

| 0x40 + 4 + 2 || Creates a B descriptor with flags=1. || Out

| 0x80 + 4 + 2 || Creates a B descriptor with flags=3. || Out

| 8 + 1 || Creates an X descriptor || In

| 8 + 2 || Creates a C descriptor, and writes the u16 size to an offset into raw data. || Out

| 0x10 + 8 + 2 || Creates a C descriptor || Out

| 0x20 + 1 || Creates both an A and X descriptor || In

|-

C and X (Pointer and ReceiveList) descriptors are backed by the "pointer buffer", a buffer in the service process. Its size is a u16, which is retrieved using the "QueryPointerBufferSize" control message. If the client code determines all buffers with flag 8 do not fit in the pointer buffer, it returns error 0x11A0B.

For buffers with flag 0x20 it creates two descriptors (A+X or B+C), but one descriptor is NULL (zero size and pointer), while the other holds the expected values. X/C descriptors are used as the non-NULL descriptor where possible, but if they don't fit in the pointer buffer, A/B descriptors are used instead. The code defers processing of type 0x20 buffers with sizes that fit in a u16 (and may therefore fit in the pointer buffer). This ensures all type 8 buffers get pointer-buffer space before any type 0x20.

(The order in which the deferred type 0x20 buffers are processed is determined by a convoluted loop.)

== Official IPC Cmd Structure ==

Official struct that is stored for each IPC command. It contains precalculated offsets for different portions of the command structure.

All offsets are given is in number of u32 words.

  u32 offset_w_descriptors; /* this is a guess */

When type == 5 you are talking to the IPC manager. These are processed by the sysmodule.

{| class="wikitable" border="1"

|-

|-

| 3 || QueryPointerBufferSize || None || u16 size

|-