Skip to contentSkip to conversation

Location of jump table in global_data structure

U-Boot Peter Tyser 3 participants 4 messages Aug 19, 2008 – Aug 20, 2008

lore
Share & tools
lore
#0 lore raw

Hello,
I've noticed that the jump table pointer (**jt) in the global_data structure is always the last field in the structure. When standalone applications are compiled, they hard code the jump table pointer offset into the global_data structure. When new versions of U-Boot come out which add/remove a field from the global_data structure, old standalone applications will no longer work as the location of the jt pointer has changed. I've noticed this issue when updating U-Boot from 1.3.0 to 1.3.4.

As an example from include/asm-avr32/global_data.h:

FROM VERSION 1.3.4:
typedef struct global_data {

        bd_t            *bd;
        unsigned long   flags;
        unsigned long   baudrate;
        unsigned long   stack_end;      /* highest stack address */
        unsigned long   have_console;   /* serial_init() was called */
        unsigned long   reloc_off;      /* Relocation Offset */
        unsigned long   env_addr;       /* Address of env struct */
        unsigned long   env_valid;      /* Checksum of env valid? */
        unsigned long   cpu_hz;         /* cpu core clock frequency */
        void            **jt;           /* jump table */

} gd_t;

?FROM FUTURE VERSION 1.3.5:
?typedef struct global_data {

        bd_t            *bd;
        unsigned long   flags;
        unsigned long   baudrate;
        unsigned long   stack_end;      /* highest stack address */
        unsigned long   have_console;   /* serial_init() was called */
        unsigned long   reloc_off;      /* Relocation Offset */
        unsigned long   env_addr;       /* Address of env struct */
        unsigned long   env_valid;      /* Checksum of env valid? */
        unsigned long   cpu_hz;         /* cpu core clock frequency */

====> unsigned long fancy_value; /* FANCY NEW VALUE ADDED!! */
void **jt; /* jump table */
} gd_t;

Adding fancy_value to global_data would break any standalone applications compiled for 1.3.4 since the app is now using fancy_value as its jump table pointer (where jt used to be).

One possible fix would be to move **jt to the 2nd item in global_data to prevent it moving in the future. This would break everyone's current standalone apps however:) eg:

?typedef struct  global_data {
        bd_t            *bd;
====>   void            **jt;           /* jump table */
        unsigned long   flags;
        unsigned long   baudrate;
        unsigned long   stack_end;      /* highest stack address */
        unsigned long   have_console;   /* serial_init() was called */
        unsigned long   reloc_off;      /* Relocation Offset */
        unsigned long   env_addr;       /* Address of env struct */
        unsigned long   env_valid;      /* Checksum of env valid? */
        unsigned long   cpu_hz;         /* cpu core clock frequency */

} gd_t;

Another option would be to mandate that new fields only be added after the **jt field to prevent it from moving, although this would be hard to enforce and seems a bit hokey.

Do others view this issue as a problem that should be fixed?

If others feel that the jt pointer should be moved to the 2nd item in global_data structure let me know and I can generate a patch.

Best,
Peter

#1 lore raw
Quoted 9 lines · Peter Tyser

Hello,
I've noticed that the jump table pointer (**jt) in the global_data structure is always the last field in the structure. When standalone applications are compiled, they hard code the jump table pointer offset into the global_data structure. When new versions of U-Boot come out which add/remove a field from the global_data structure, old standalone applications will no longer work as the location of the jt pointer has changed. I've noticed this issue when updating U-Boot from 1.3.0 to 1.3.4.

It seems to me to be very broken that the contents an interface definition would shift from version to version. IMHO, unless there are unassailable reasons, new values should *always* be appended to the struct so that the struct is backwards compatible to previous versions.

Maybe we need to upgrade our interface to a flattened device tree to avoid the horrible interface-as-a-struct layout problem.
<http://www.brainyquote.com/quotes/quotes/b/bernardbar181387.html> ;-)

[snip]

Quoted 11 lines

?FROM FUTURE VERSION 1.3.5:
?typedef struct global_data {

        bd_t            *bd;
        unsigned long   flags;
        unsigned long   baudrate;
        unsigned long   stack_end;      /* highest stack address */
        unsigned long   have_console;   /* serial_init() was called */
        unsigned long   reloc_off;      /* Relocation Offset */
        unsigned long   env_addr;       /* Address of env struct */
        unsigned long   env_valid;      /* Checksum of env valid? */
        unsigned long   cpu_hz;         /* cpu core clock frequency */
Quoted 3 lines

====> unsigned long fancy_value; /* FANCY NEW VALUE ADDED!! */
void **jt; /* jump table */
} gd_t;

This addition is broken IMHO.

Quoted 14 lines

One possible fix would be to move **jt to the 2nd item in global_data to prevent it moving in the future. This would break everyone's current standalone apps however:) eg:

?typedef struct  global_data {
        bd_t            *bd;
====>   void            **jt;           /* jump table */
        unsigned long   flags;
        unsigned long   baudrate;
        unsigned long   stack_end;      /* highest stack address */
        unsigned long   have_console;   /* serial_init() was called */
        unsigned long   reloc_off;      /* Relocation Offset */
        unsigned long   env_addr;       /* Address of env struct */
        unsigned long   env_valid;      /* Checksum of env valid? */
        unsigned long   cpu_hz;         /* cpu core clock frequency */
Quoted 1 line

} gd_t;

That only "fixes" the jump table reference. If someone adds fancy_value after baudrate, it still will break backwards compatibility (maybe not visibly, maybe not immediately, maybe not for a given application, but it still is broken).

Quoted 3 lines

Another option would be to mandate that new fields only be added after the **jt field to prevent it from moving, although this would be hard to enforce and seems a bit hokey.

No, only append new fields to the end of the struct (adding fields after **jt only fixes the problem for the first new field ;-). The correct rule is to never add fields in the middle of the struct.

An instructive comment should go a long way and we have some pretty eagle-eyed code reviewers on the mailing list that should go the rest of the way.

Quoted 1 line

Do others view this issue as a problem that should be fixed?

Yes.

Quoted 2 lines

If others feel that the jt pointer should be moved to the 2nd item in global_data structure let me know and I can generate a patch.

Add a comment and police it is my vote.

Quoted 2 lines

Best,
Peter

Thanks,
gvb

#2 lore raw
Quoted 9 lines
Quoted 8 lines

I've noticed that the jump table pointer (**jt) in the global_data structure is always the last field in the structure. When standalone applications are compiled, they hard code the jump table pointer offset into the global_data structure. When new versions of U-Boot come out which add/remove a field from the global_data structure, old standalone applications will no longer work as the location of the jt pointer has changed. I've noticed this issue when updating U-Boot from 1.3.0 to 1.3.4.

Quoted 8 lines

It seems to me to be very broken that the contents an interface definition would shift from version to version. IMHO, unless there are unassailable reasons, new values should *always* be appended to the struct so that the struct is backwards compatible to previous versions.

Maybe we need to upgrade our interface to a flattened device tree to avoid the horrible interface-as-a-struct layout problem.
<http://www.brainyquote.com/quotes/quotes/b/bernardbar181387.html> ;-)

Great quote, very fitting:)

Quoted 42 lines

[snip]

Quoted 11 lines

?FROM FUTURE VERSION 1.3.5:
?typedef struct global_data {

        bd_t            *bd;
        unsigned long   flags;
        unsigned long   baudrate;
        unsigned long   stack_end;      /* highest stack address */
        unsigned long   have_console;   /* serial_init() was called */
        unsigned long   reloc_off;      /* Relocation Offset */
        unsigned long   env_addr;       /* Address of env struct */
        unsigned long   env_valid;      /* Checksum of env valid? */
        unsigned long   cpu_hz;         /* cpu core clock frequency */
Quoted 3 lines

====> unsigned long fancy_value; /* FANCY NEW VALUE ADDED!! */
void **jt; /* jump table */
} gd_t;

This addition is broken IMHO.

Quoted 14 lines

One possible fix would be to move **jt to the 2nd item in global_data to prevent it moving in the future. This would break everyone's current standalone apps however:) eg:

?typedef struct  global_data {
        bd_t            *bd;
====>   void            **jt;           /* jump table */
        unsigned long   flags;
        unsigned long   baudrate;
        unsigned long   stack_end;      /* highest stack address */
        unsigned long   have_console;   /* serial_init() was called */
        unsigned long   reloc_off;      /* Relocation Offset */
        unsigned long   env_addr;       /* Address of env struct */
        unsigned long   env_valid;      /* Checksum of env valid? */
        unsigned long   cpu_hz;         /* cpu core clock frequency */
Quoted 1 line

} gd_t;

That only "fixes" the jump table reference. If someone adds fancy_value after baudrate, it still will break backwards compatibility (maybe not visibly, maybe not immediately, maybe not for a given application, but it still is broken).

Quoted 3 lines

Another option would be to mandate that new fields only be added after the **jt field to prevent it from moving, although this would be hard to enforce and seems a bit hokey.

Quoted 7 lines

No, only append new fields to the end of the struct (adding fields after **jt only fixes the problem for the first new field ;-). The correct rule is to never add fields in the middle of the struct.

An instructive comment should go a long way and we have some pretty eagle-eyed code reviewers on the mailing list that should go the rest of the way.

The one large downside of mandating that fields only be added to the end of the struct is that a field can never be removed from the global_data struct. I have to imagine fields will be removed at some point...

Quoted 8 lines
Quoted 1 line

Do others view this issue as a problem that should be fixed?

Yes.

Quoted 2 lines

If others feel that the jt pointer should be moved to the 2nd item in global_data structure let me know and I can generate a patch.

Add a comment and police it is my vote.

That's definitely an improvement, but doesn't handle both adding/removing fields from the global_data structure in a clean manner. I'd still lean towards moving the jt pointer to one of the earlier fields of the structs as well as adding a comment. Then at least the jump table portion of the API would be stable, even if accessing the "global_data fields" API wouldn't. Right now, neither API is stable:)

If anyone has any other clever ideas on improving the API, I'd be happy to investigate/implement.

Best,
Peter

#3 lore raw

Peter Tyser wrote:

*snip*

Quoted 14 lines
Quoted 6 lines
Quoted 1 line

Do others view this issue as a problem that should be fixed?

Yes.

Quoted 2 lines

If others feel that the jt pointer should be moved to the 2nd item in global_data structure let me know and I can generate a patch.

Add a comment and police it is my vote.

That's definitely an improvement, but doesn't handle both adding/removing fields from the global_data structure in a clean manner. I'd still lean towards moving the jt pointer to one of the earlier fields of the structs as well as adding a comment. Then at least the jump table portion of the API would be stable, even if accessing the "global_data fields" API wouldn't. Right now, neither API is stable:)

Quoted 2 lines

If anyone has any other clever ideas on improving the API, I'd be happy to investigate/implement.

There already is a modern API for external applications in U-Boot since 1.3.2 (but not enabled by default). Please see 'api/README' and 'api_examples' in the U-Boot tree.

So far it has been used successfully on PowerPC and ARM; there's no support for AVR arch yet, but providing the syscall entry point for it shouldn't be a big deal for someone with AVR knowledge. Let me know if you'd be willing to chew on it (I don't have AVR h/w but can help/support regarding the API itself).

kind regards,
Rafal