use bindgen to auto generate ffi rust code #398
Conversation
fredchenbj
changed the title
Connor1996
changed the title
|
Please fix the conflicts |
@Connor1996 I have fixed the conficts. |
librocksdb_sys/crocksdb/crocksdb/c.h
Outdated
There was a problem hiding this comment.
bindgen generates FFI based on c.h, but why we modify c.h?
There was a problem hiding this comment.
@Connor1996 this function was added by us, so not conform to c's style.
There was a problem hiding this comment.
rocksdb's c.h has convention:
Bools have the type unsigned char (0 == false; rest == true)
There was a problem hiding this comment.
can you extract them to a separate PR?
step 1 finished. Signed-off-by: fredchenbj <[email protected]>
Signed-off-by: fredchenbj <[email protected]>
fredchenbj
force-pushed
the
fredchenbj/bindgen_ffi
branch
from
cc31d36 to
7f4b70e
Compare
|
@fredchenbj awesome changes! Some general questions:
Thanks. |
|
It introduces clang dependency, which can be a pain to fix. You can learn how to avoid it from tikv/grpc-rs. |
|
@BusyJay You mean pre-generate binding? Yeah, that's what we plan to do. For easy reviewing, we plan to do it step by step. |
| pub trait TablePropertiesCollector { | ||
| /// Will be called when a new key/value pair is inserted into the table. | ||
| fn add(&mut self, key: &[u8], value: &[u8], entry_type: DBEntryType, seq: u64, file_size: u64); | ||
| fn add(&mut self, key: &[i8], value: &[i8], entry_type: DBEntryType, seq: u64, file_size: u64); |
There was a problem hiding this comment.
for const char* in c.h is transferred to *const libc::c_char by bindgen, which requires i8. All transfer like this is for that. @Connor1996
There was a problem hiding this comment.
see references:
- rust std implementation https://doc.rust-lang.org/src/std/ffi/c_str.rs.html#975,
- https://doc.rust-lang.org/std/os/raw/type.c_char.html,
- https://stackoverflow.com/questions/47684111/make-string-type-compatible-with-arm/47684200.
i8 or u8 is only due to how C char represents in different platforms. But the data is always one-byte long. As str in Rust is represented as &[u8], so we must convert it to u8.
There was a problem hiding this comment.
got it. I would fix it.
| unsafe { fetch_str!(crocksdb_writestallinfo_cf_name(&self.0)) } | ||
| } | ||
| pub fn cur(&self) -> WriteStallCondition { | ||
| pub fn cur(&self) -> crocksdb_writestallcondition_t { |
There was a problem hiding this comment.
All public functions/traits of src/xxx.rs should be unchanged, otherwise TiKV or other users may change code dramatically. Consider rename type crocksdb_writestallcondition_t to WriteStallCondition. Same for other enums like DBEntryType.
@yiwu-arbug Thanks for your questions. for 3, the migration plan has three steps: 1) use the old crocksdb's c.h and bindgen to generate ffi rust code, and make it work for all dependency code; 2) replace with rocksdb's c.h, and make all code in c.h conform to rocksdb's style, here we would only include rocksdb's c.h and not change it at all; 3) try to use rocksdb cpp api to generate ffi rust code. This pr is only step 1. for 1, this pr has mainly three changes: 1) change crocksdb's c.h and c.cc to conform to bindgen's requirement; 2) use bindgen to generate ffi rust code; 3) change code in rust-rocksdb, including many type changes, to make those code could call the code geneated by bindgen. for 2, the generated code is in |
|
@fredchenbj for 3, can you give more details for step 1)? Does the generated code provide difference interface than the existing rust API? If so, how much work is it to update tikv? If that work is huge, is there way to split the work into small bits? Also for 3, if we will trying generating code from c++ directly, how useful step 2) would be? Currently the wrapper doesn't handle |
5 participants
refer to #381.
just use old c.h and bindgen to gen ffi's rust code. and fix to pass
cargo build.this is step 1, later replace with rocksdb's c.h to gen code.
Signed-off-by: fredchenbj [email protected]