Difference between revisions of "RK3566 EBC Reverse-Engineering"

From PINE64
Jump to navigation Jump to search
(→‎Sources: Add link to rk_eink uboot driver)
(→‎Waveform: Clarified a bit)
 
(15 intermediate revisions by 4 users not shown)
Line 3: Line 3:
Unfortunately, the driver published for this eInk interface within the BSP kernel is an assembly dump produced by gcc. Fortunately, it contains quite a bit of debug information, which we can use to reverse engineer it.
Unfortunately, the driver published for this eInk interface within the BSP kernel is an assembly dump produced by gcc. Fortunately, it contains quite a bit of debug information, which we can use to reverse engineer it.


= Sources =
== Sources ==


== Downstream ==
=== Downstream ===


The ebc driver source is [https://gitlab.com/pine64-org/quartz-bsp/rockchip-linux/-/tree/quartz64/drivers/gpu/drm/rockchip/ebc-dev available from the quartz-bsp repository].
The ebc driver source is [https://gitlab.com/pine64-org/quartz-bsp/rockchip-linux/-/tree/quartz64/drivers/gpu/drm/rockchip/ebc-dev available from the quartz-bsp repository].


The file of interest is <code>ebc_dev_v8.S</code>, which implements a DRM (Direct Rendering Manager) driver for the eInk panel.
The files of interest are <code>ebc_dev_v8.S</code>, which implements a DRM (Direct Rendering Manager) driver for the eInk panel, and the waveform files in the <code>epdlut</code> subdirectory.


There's also a completely open-source simple driver for u-boot: [https://github.com/JeffyCN/rockchip_mirrors/tree/u-boot/drivers/video/rk_eink drivers/video/rk_eink at JeffyCN/rockchip_mirrors]
There's also a simple driver for u-boot: [https://github.com/JeffyCN/rockchip_mirrors/tree/u-boot/drivers/video/rk_eink drivers/video/rk_eink at JeffyCN/rockchip_mirrors]


== In-Development Driver ==
These two drivers show the two different programming interfaces exposed by the TCON. The U-Boot driver operates in "LUT mode": it writes the waveform LUT to registers in the TCON's MMIO space, and passes buffers of ''pixel'' data to the TCON. The Linux driver operates in "direct mode" uses the LUTs to compute waveforms in software, and passes buffers of ''waveform'' data to the TCON.


[https://gitlab.com/pine64-org/quartz-bsp/linux-next/-/tree/rk356x-ebc-dev rk356x-ebc-dev] is the branch on the pine64 linux-next repository used for developing an upstreamable driver based on mainline kernel sources.
Some reversing of the downstream Linux driver has been done here: https://github.com/Ralim/ebc-dev-reverse-engineering


= Documentation =
=== Reimplementation ===


== Datasheets ==
A human-readable C reimplementation of the LUT and pixel data path [https://gitlab.com/smaeul/ebc-dev is available from smaeul here]. This provides everything needed to convert a framebuffer and a waveform data file into a series of "frames" for the panel. The new implementation includes a test suite to verify its output matches the output from the BSP assembly code. It is based on the downstream Linux driver.


=== TI TPS65185x ===
=== In-Development Driver ===
 
[https://github.com/smaeul/linux/commits/rk356x-ebc-dev rk356x-ebc-dev] is the branch used for developing an upstreamable DRM driver based on mainline kernel sources. The driver currently functions well enough to get a virtual console and Xorg running, but it does not support features like multiple waveforms.
 
== Documentation ==
 
=== Datasheets ===
 
==== EBC ====
 
The EBC TCON is documented in part 2 of the RK356x TRM.
 
==== TI TPS65185x ====


This is the PMIC used to drive the e-Ink panel, for which the downstream sources also implement a driver.
This is the PMIC used to drive the e-Ink panel, for which the downstream sources also implement a driver.
Line 27: Line 39:
https://www.ti.com/lit/ds/symlink/tps65185.pdf
https://www.ti.com/lit/ds/symlink/tps65185.pdf


== Assembly Syntax and Semantics ==
==== Waveform ====
 
The format of the waveform file is documented in the following PDF files:
 
* [https://www.waveshare.net/w/upload/c/c4/E-paper-mode-declaration.pdf E-paper-mode-declaration.pdf]
* [https://web.archive.org/web/20220516211757/https://www.waveshare.net/w/upload/archive/c/c4/20190611032540!E-paper-mode-declaration.pdf 20190611032540!E-paper-mode-declaration.pdf]
 
The latter file includes some additional information about the specification itself.  Think of it as containing more metadata, while the actual data is the same.
 
=== Utilities ===
 
The [https://github.com/fread-ink/inkwave <tt>inkwave</tt>] program is designed to parse waveform files. Currently it cannot fully handle the waveform files shipped with the PineNote. Adding support for this to the tool would be helpful.
 
=== Assembly Syntax and Semantics ===


The Syntax is GNU Assembler (GAS) syntax. [https://modexp.wordpress.com/2018/10/30/arm64-assembly/ This modexp article] provides a good introduction to the syntax, calling convention, semantics and some often used instructions.
The Syntax is GNU Assembler (GAS) syntax. [https://modexp.wordpress.com/2018/10/30/arm64-assembly/ This modexp article] provides a good introduction to the syntax, calling convention, semantics and some often used instructions.
Line 35: Line 60:
At the very least, you should read up on the registers and calling convention used.
At the very least, you should read up on the registers and calling convention used.


= Debug Information =
== Various Findings ==
 
The driver isn't really something that can be mainlined as-is once reversed, as it makes a number of questionable design decisions.
 
* It's technically a DRM subsystem driver, but doesn't really utilise what DRM provides at all.
* It seems to register a new ioctl to set buffer attributes like width and height, despite DRM more than likely having a way for clients to tell a driver what size the framebuffer should be.
* It directly interacts with the PMIC instead of going through regulators/hwmon.
 
However, reverse engineering to know how it works provides a good baseline from which we can rewrite it in a more sensible manner.
 
== Debug Information ==


Quite a bit of debug info is left in the assembly dump, including function names, file names and line numbers. We can take this to our advantage.
Quite a bit of debug info is left in the assembly dump, including function names, file names and line numbers. We can take this to our advantage.


== <code>.file ''file-number'' ''file-path''</code> ==
=== <code>.file ''file-number'' ''file-path''</code> ===


Specifies a number to reference a file by, and its path. All following code until the next <code>.file</code> or <code>.loc</code> statement are to be understood as originating from this file. This is particularly useful to understand which code has been inlined from other files, for which the source is available.
Specifies a number to reference a file by, and its path. All following code until the next <code>.file</code> or <code>.loc</code> statement are to be understood as originating from this file. This is particularly useful to understand which code has been inlined from other files, for which the source is available.


== <code>.loc ''file-number'' ''line-number'' 0</code> ==
=== <code>.loc ''file-number'' ''line-number'' 0</code> ===


Specifies that the following code is generated from <code>''line-number''</code> stemming from file number <code>''file-number''</code>. See the <code>.file</code> directive for this file number to understand which source file it came from.
Specifies that the following code is generated from <code>''line-number''</code> stemming from file number <code>''file-number''</code>. See the <code>.file</code> directive for this file number to understand which source file it came from.


== <code>.type ''function-name'', %function</code> ==
=== <code>.type ''function-name'', %function</code> ===


This tells us that the following code belongs to function <code>''function-name''</code>. You'll usually see a <code>.cfi_startproc</code>, which signifies the start of the function code, until the matching <code>.cfi_endproc</code>.
This tells us that the following code belongs to function <code>''function-name''</code>. You'll usually see a <code>.cfi_startproc</code>, which signifies the start of the function code, until the matching <code>.cfi_endproc</code>.
Line 53: Line 88:
A quick grep for <code>%function</code> shows that we are dealing with 30 functions in this file.
A quick grep for <code>%function</code> shows that we are dealing with 30 functions in this file.


== <code>.type ''struct-name'', %object</code> ==
=== <code>.type ''struct-name'', %object</code> ===


This seems to signify a definition of a C struct named <code>''struct-name''</code>.
This seems to signify a definition of a C struct named <code>''struct-name''</code>.
Line 59: Line 94:
A quick grep for <code>%object</code> shows that we are dealing with around 27 structs in this file.
A quick grep for <code>%object</code> shows that we are dealing with around 27 structs in this file.


== <code>.Ldebug_info0:</code> ==
=== <code>.Ldebug_info0:</code> ===


'''TODO:''' This seems to contain the main bulk of the DWARF debug information, including enough info to reverse full structs and function signatures.
'''TODO:''' This seems to contain the main bulk of the DWARF debug information, including enough info to reverse full structs and function signatures.


= Finding Structs and Function Signatures =
== Finding Structs and Function Signatures ==


First, we'll need to assemble the file:
First, we'll need to assemble the file:
Line 69: Line 104:
<code>aarch64-linux-gnu-gcc -c -o ebc_dev_v8.o ebc_dev_v8.S</code>
<code>aarch64-linux-gnu-gcc -c -o ebc_dev_v8.o ebc_dev_v8.S</code>


This gives us a <code>ebc_dev_v8.o</code> which we can feed into readelf:
This gives us a <code>ebc_dev_v8.o</code> which we can feed into analysis tools.
 
For both of these, keep in mind that we're only interested in stuff from ebc_dev.c, or any other files for which we don't have the source. There's no point in getting the struct description or reverse-engineering a function that we have the source code for. A lot more than ebc_dev will be in the object file due to inlining and such.
 
Also make sure that if you are looking up known struct accesses, that you use struct definitions from the BSP kernel, not from mainline. The kernel has no internal ABI for drivers!
 
=== Faster and Easier - Ghidra ===
 
Import the file into Ghidra, open the code browser. After analysis, you should be able to find structs in the "Data Type Manager" marked with an S icon. You'll also find functions in the symbol tree.
 
=== Slow and Painful - readelf/objdump ===
 
Use this if you want to manually look up dwarf symbols for some reason.


<code>readelf --debug-dump ebc_dev_v8.o</code>
<code>readelf --debug-dump ebc_dev_v8.o</code>
Line 124: Line 171:
[etc etc...]</nowiki>
[etc etc...]</nowiki>


= Reverse-Engineered Stuff =
== Reverse-Engineered Stuff ==


== Structs ==
=== Structs ===


=== ebc_info ===
==== ebc_info ====


<!-- I don't know why mediawiki scuffs the formatting here -->
<!-- I don't know why mediawiki scuffs the formatting here -->
See https://gitlab.com/smaeul/ebc-dev/-/blob/main/auto_image.h#L124, which is based on [https://gitlab.com/pine64-org/quartz-bsp/linux-next/-/commits/2de5fb11a888c37f366642544e5a53ec2faae32d the v1.04 BSP Linux driver].
==== ebc ====
See https://gitlab.com/smaeul/ebc-dev/-/blob/main/auto_image.h#L200, which is based on [https://gitlab.com/pine64-org/quartz-bsp/linux-next/-/commits/2de5fb11a888c37f366642544e5a53ec2faae32d the v1.04 BSP Linux driver].
==== rkf_waveform ====
Note: all known waveform data files are the "PVI" variant, not the "RKF" variant.


<source lang="c">
<source lang="c">
struct ebc_info {
struct rkf_waveform {
     long unsigned int ebc_buffer_phy;
     int length,
     char* ebc_buffer_vir;
     char[16] format,
     int ebc_buffer_size;
     char[16] version,
     int ebc_buf_real_size;
     char[16] timeandday,
     int direct_buf_real_size;
     char[16] panel_name,
     int is_busy_now;
     char[16] panel_info,
     int task_restart;
     char[64] full_version,
     int auto_refresh_done;
     char[64] reset_temp_list,
     char frame_total;
     char[64] gc16_temp_list,
     char frame_bw_total;
     char[64] gl16_temp_list,
     int auto_need_refresh0;
     char[64] glr16_temp_list,
     int auto_need_refresh1;
     char[64] gld16_temp_list,
     int frame_left;
     char[64] du_temp_list,
    int part_mode_count;
     char[64] a2_temp_list,
    int full_mode_num;
     uint[64] reset_list,
    int height;
     uint[64] gc16_list,
    int width;
     uint[64] gl16_list,
    int* lut_addr;
     uint[64] glr16_list,
    int buffer_need_check;
     uint[64] gld16_list,
    int ebc_irq_status;
     uint[64] du_list,
    int ebc_dsp_buf_status;
     uint[64] a2_list
    struct device* dev;
}
    struct epd_lut_data lut_data;
    struct task_struct* ebc_task;
    int* auto_image_new;
    int* auto_image_old;
    int* auto_image_bg;
    int* auto_image_cur;
    u8* auto_frame_count;
    int* auto_image_fb;
    void* direct_buffer;
    int ebc_power_status;
    int ebc_last_display;
     char* lut_ddr_vir;
    struct ebc_buf_s* prev_dsp_buf;
    struct ebc_buf_s* curr_dsp_buf;
     struct wake_lock suspend_lock;
     int wake_lock_is_set;
     int first_in;
     struct timer_list vdd_timer;
     struct timer_list frame_timer;
     struct work_struct auto_buffer_work;
     int is_early_suspend;
    int is_deep_sleep;
    int is_power_off;
    int overlay_enable;
    int overlay_start;
};
</source>
</source>


=== ebc ===
=== Enums ===


==== rkf_waveform_type ====
<source lang="c">
<source lang="c">
struct ebc {
enum rkf_waveform_type {
     struct device* dev;
     RKF_WF_RESET = 0,
     struct ebc_tcon* tcon;
     RKF_WF_DU    = 1,
     struct ebc_pmic* pmic;
     RKF_WF_GC16  = 2,
     struct ebc_panel* panel;
     RKF_WF_GL16  = 3,
     struct ebc_info* info;
     RKF_WF_GLR16 = 4,
};
    RKF_WF_GLD16 = 5,
    RKF_WF_A2    = 6
}
</source>
</source>


[[Category:Quartz64]][[Category:Rockchip RK3566]]
[[Category:Quartz64]][[Category:Rockchip RK3566]]

Latest revision as of 14:39, 19 September 2024

The RK3566 SoC, used in the Quartz64 SBC by PINE64, contains an eInk interface. This is referred to as ebc by Rockchip apparently.

Unfortunately, the driver published for this eInk interface within the BSP kernel is an assembly dump produced by gcc. Fortunately, it contains quite a bit of debug information, which we can use to reverse engineer it.

Sources

Downstream

The ebc driver source is available from the quartz-bsp repository.

The files of interest are ebc_dev_v8.S, which implements a DRM (Direct Rendering Manager) driver for the eInk panel, and the waveform files in the epdlut subdirectory.

There's also a simple driver for u-boot: drivers/video/rk_eink at JeffyCN/rockchip_mirrors

These two drivers show the two different programming interfaces exposed by the TCON. The U-Boot driver operates in "LUT mode": it writes the waveform LUT to registers in the TCON's MMIO space, and passes buffers of pixel data to the TCON. The Linux driver operates in "direct mode" uses the LUTs to compute waveforms in software, and passes buffers of waveform data to the TCON.

Some reversing of the downstream Linux driver has been done here: https://github.com/Ralim/ebc-dev-reverse-engineering

Reimplementation

A human-readable C reimplementation of the LUT and pixel data path is available from smaeul here. This provides everything needed to convert a framebuffer and a waveform data file into a series of "frames" for the panel. The new implementation includes a test suite to verify its output matches the output from the BSP assembly code. It is based on the downstream Linux driver.

In-Development Driver

rk356x-ebc-dev is the branch used for developing an upstreamable DRM driver based on mainline kernel sources. The driver currently functions well enough to get a virtual console and Xorg running, but it does not support features like multiple waveforms.

Documentation

Datasheets

EBC

The EBC TCON is documented in part 2 of the RK356x TRM.

TI TPS65185x

This is the PMIC used to drive the e-Ink panel, for which the downstream sources also implement a driver.

https://www.ti.com/lit/ds/symlink/tps65185.pdf

Waveform

The format of the waveform file is documented in the following PDF files:

The latter file includes some additional information about the specification itself. Think of it as containing more metadata, while the actual data is the same.

Utilities

The inkwave program is designed to parse waveform files. Currently it cannot fully handle the waveform files shipped with the PineNote. Adding support for this to the tool would be helpful.

Assembly Syntax and Semantics

The Syntax is GNU Assembler (GAS) syntax. This modexp article provides a good introduction to the syntax, calling convention, semantics and some often used instructions.

The ARM Architecture Reference Manual for ARMv8 should be used as reference for any instructions.

At the very least, you should read up on the registers and calling convention used.

Various Findings

The driver isn't really something that can be mainlined as-is once reversed, as it makes a number of questionable design decisions.

  • It's technically a DRM subsystem driver, but doesn't really utilise what DRM provides at all.
  • It seems to register a new ioctl to set buffer attributes like width and height, despite DRM more than likely having a way for clients to tell a driver what size the framebuffer should be.
  • It directly interacts with the PMIC instead of going through regulators/hwmon.

However, reverse engineering to know how it works provides a good baseline from which we can rewrite it in a more sensible manner.

Debug Information

Quite a bit of debug info is left in the assembly dump, including function names, file names and line numbers. We can take this to our advantage.

.file file-number file-path

Specifies a number to reference a file by, and its path. All following code until the next .file or .loc statement are to be understood as originating from this file. This is particularly useful to understand which code has been inlined from other files, for which the source is available.

.loc file-number line-number 0

Specifies that the following code is generated from line-number stemming from file number file-number. See the .file directive for this file number to understand which source file it came from.

.type function-name, %function

This tells us that the following code belongs to function function-name. You'll usually see a .cfi_startproc, which signifies the start of the function code, until the matching .cfi_endproc.

A quick grep for %function shows that we are dealing with 30 functions in this file.

.type struct-name, %object

This seems to signify a definition of a C struct named struct-name.

A quick grep for %object shows that we are dealing with around 27 structs in this file.

.Ldebug_info0:

TODO: This seems to contain the main bulk of the DWARF debug information, including enough info to reverse full structs and function signatures.

Finding Structs and Function Signatures

First, we'll need to assemble the file:

aarch64-linux-gnu-gcc -c -o ebc_dev_v8.o ebc_dev_v8.S

This gives us a ebc_dev_v8.o which we can feed into analysis tools.

For both of these, keep in mind that we're only interested in stuff from ebc_dev.c, or any other files for which we don't have the source. There's no point in getting the struct description or reverse-engineering a function that we have the source code for. A lot more than ebc_dev will be in the object file due to inlining and such.

Also make sure that if you are looking up known struct accesses, that you use struct definitions from the BSP kernel, not from mainline. The kernel has no internal ABI for drivers!

Faster and Easier - Ghidra

Import the file into Ghidra, open the code browser. After analysis, you should be able to find structs in the "Data Type Manager" marked with an S icon. You'll also find functions in the symbol tree.

Slow and Painful - readelf/objdump

Use this if you want to manually look up dwarf symbols for some reason.

readelf --debug-dump ebc_dev_v8.o

This will produce a lot of output, but we're mainly concerned with the start of the dump. We'll find things like:

 <2><101f8>: Abbrev Number: 0
 <1><101f9>: Abbrev Number: 79 (DW_TAG_subprogram)
    <101fa>   DW_AT_name        : (indirect string, offset: 0xa2b4): ebc_open
    <101fe>   DW_AT_decl_file   : 1
    <101ff>   DW_AT_decl_line   : 1377
    <10201>   DW_AT_prototyped  : 1
    <10201>   DW_AT_type        : <0xc6>
    <10205>   DW_AT_low_pc      : 0x0
    <1020d>   DW_AT_high_pc     : 0xc
    <10215>   DW_AT_frame_base  : 1 byte block: 9c 	(DW_OP_call_frame_cfa)
    <10217>   DW_AT_GNU_all_call_sites: 1
    <10217>   DW_AT_sibling     : <0x1023a>
 <2><1021b>: Abbrev Number: 88 (DW_TAG_formal_parameter)
    <1021c>   DW_AT_name        : (indirect string, offset: 0x1153): inode
    <10220>   DW_AT_decl_file   : 1
    <10221>   DW_AT_decl_line   : 1377
    <10223>   DW_AT_type        : <0x1c54>
    <10227>   DW_AT_location    : 0xd63 (location list)
 <2><1022b>: Abbrev Number: 106 (DW_TAG_formal_parameter)
    <1022c>   DW_AT_name        : (indirect string, offset: 0x8b06): file
    <10230>   DW_AT_decl_file   : 1
    <10231>   DW_AT_decl_line   : 1377
    <10233>   DW_AT_type        : <0x551f>
    <10237>   DW_AT_location    : 1 byte block: 51 	(DW_OP_reg1 (x1))

This essentially tells us the full function signature of ebc_open:

DW_TAG_subprogram tells us of a function, with DW_AT_name letting us know that this is ebc_open. DW_AT_type of 0xc6 let's us know, once we jump to <c6>, that this function's return type is a signed 32-bit integer.

The DW_TAG_formal_parameter that follow tell us of each of the parameter the function takes. The first one is called inode and is of type 0x1c54. Referencing what this type is, we find:

 <1><1c54>: Abbrev Number: 7 (DW_TAG_pointer_type)
    <1c55>   DW_AT_byte_size   : 8
    <1c56>   DW_AT_type        : <0x1970>

which in of itself goes on to reference 0x1970, and looking this one up, we'll find a struct definition:

 <1><1970>: Abbrev Number: 26 (DW_TAG_structure_type)
    <1971>   DW_AT_name        : (indirect string, offset: 0x1153): inode
    <1975>   DW_AT_byte_size   : 672
    <1977>   DW_AT_decl_file   : 31
    <1978>   DW_AT_decl_line   : 611
    <197a>   DW_AT_sibling     : <0x1c4f>
 <2><197e>: Abbrev Number: 27 (DW_TAG_member)
    <197f>   DW_AT_name        : (indirect string, offset: 0x7d00): i_mode
[etc etc...]

Reverse-Engineered Stuff

Structs

ebc_info

See https://gitlab.com/smaeul/ebc-dev/-/blob/main/auto_image.h#L124, which is based on the v1.04 BSP Linux driver.

ebc

See https://gitlab.com/smaeul/ebc-dev/-/blob/main/auto_image.h#L200, which is based on the v1.04 BSP Linux driver.

rkf_waveform

Note: all known waveform data files are the "PVI" variant, not the "RKF" variant.

struct rkf_waveform {
    int length,
    char[16] format,
    char[16] version,
    char[16] timeandday,
    char[16] panel_name,
    char[16] panel_info,
    char[64] full_version,
    char[64] reset_temp_list,
    char[64] gc16_temp_list,
    char[64] gl16_temp_list,
    char[64] glr16_temp_list,
    char[64] gld16_temp_list,
    char[64] du_temp_list,
    char[64] a2_temp_list,
    uint[64] reset_list,
    uint[64] gc16_list,
    uint[64] gl16_list,
    uint[64] glr16_list,
    uint[64] gld16_list,
    uint[64] du_list,
    uint[64] a2_list
}

Enums

rkf_waveform_type

enum rkf_waveform_type {
    RKF_WF_RESET = 0,
    RKF_WF_DU    = 1,
    RKF_WF_GC16  = 2,
    RKF_WF_GL16  = 3,
    RKF_WF_GLR16 = 4,
    RKF_WF_GLD16 = 5,
    RKF_WF_A2    = 6
}