Burak Kaan Karaçay================================================= Refactoring in order to reduce Git’s global state =================================================
Personal Info: --------------
Name: Burak Kaan Karaçay (he/him) Email: bkkaracay@gmail.com Education: UG Sophomore, Marmara University GitHub: https://github.com/bkkaracay Timezone: UTC+3 (Istanbul, Turkey)
My Patches: -----------
+ (Microproject) t2003: modernize path existence checks using test
helpers
- Thread:
https://lore.kernel.org/git/20260208202809.270523-1-bkkaracay@gmail.com/T/
- Thread v2:
https://lore.kernel.org/git/20260209112444.1268765-1-bkkaracay@gmail.com/T/
- Status: Merged to master
- Commit Hash: 168d575719d944759964e004d17a3282b0f883d5+ [PATCH 0/2] mailmap: reduce global state
- Thread:
https://lore.kernel.org/git/20260219125954.3539324-1-bkkaracay@gmail.com/T/
- Status: Merged to master
- Commit Hash: 2d843a2d3d6c2d5e7861e6aa99743d15d36746b9Relevant Experience: --------------------
I am currently developing my own programming language as a hobby project, writing a zero-dependency interpreter for it in C. While it is still a work in progress, I have completed the core front-end pipeline. Building this project has given me practical experience with C programming, data structures and modular software architecture.
+ To support potential future multithreading, I avoided global variables in my own project. Instead, I pass state via local contexts.
+ I implemented an arena allocator (memory pool) to reduce malloc system call overhead, prevent memory fragmentation and ensure cache locality.
+ I used techniques like string interning and Pratt parsing.
My project is available on my GitHub profile [1]. If you would like to take a look at the code, 'src/main.c' is a good starting point.
Project Abstract: -----------------
Git was originally designed as a short-lived CLI tool, where relying on global variables was highly practical. Over time, the need to embed Git into other projects and applications emerged. Today, these global variables are a huge roadblock to the libification of git, as they make it impossible to properly handle multiple repositories within a single process or safely support multi-threading.
This project aims to reduce this reliance by migrating global variables from 'environment.c' into appropriate locations. This effort will support the libification goal and modernize Git's internal structure.
Technical Approach: -------------------
The core challenge of this project is choosing the correct parsing strategy more than relocating globals. The codebase currently offers two migration strategies for global state removal.
Currently, globals are loaded eagerly via 'repo_config()'. The modern 'repo_config_values()' API provides a safe and straightforward way to eagerly load variables and reduce global count. However, eager-loading parses all configurations upfront, including unnecessary ones. Users may encounter fatal configuration errors that are entirely unrelated to the command they are executing [2].
On the contrary, lazy-loading postpones the parsing process until the variable is strictly required, preventing unrelated configuration errors. However, it is significantly trickier to migrate. If a misformatted configuration triggers a 'die()' in the middle of the execution, it risks causing data corruption. Moreover, lazy-loading changes the timing of error reporting and struggles to replicate eager-loading behavior when multiple configuration keys affect a single variable [3].
If lazy-loading is considered safe for variable, git provides two APIs depending on the performance requirements:
* The 'repo_config_get*' function set is suitable for variables * accessed infrequently because of underlying string hashing costs. It * is important to use this API to not bloat the 'struct repo_settings' * [2]. * For frequently accessed variables, caching them within 'struct * repo_settings' is preferred, as it amortizes hash costs and provides * direct memory access speed.
There is no silver bullet solution for migrating globals. Because transitioning these variables require a deep understanding about the codebase, communication with mentors and the community is essential.
About Gentle Reading: ---------------------
Current config readers rely on 'die()' to handle error cases. While pragmatic for cli-tools, fatal exits are unacceptable for a library, as they will crash the host process. Building upon Derrick Stolee's recent introduction of gentle parsing functions [4], I propose implementing '_maybe' variants for core configuration readers. Since removing all 'die()' calls is inevitable for libification, sooner or later config readers will be purged from 'die()' calls. Utilizing the gentle functions for newly migrated global variables will reduce the future amount of work.
Applying this gentle API to widely used functions risks creating unreviewable patches and merge conflicts. To solve this, I plan to use a function wrapper approach, similar to the strategy used in early the_repository migrations [5]. However, the_repository changes are more mechanical work compared to the gentle transition. In complex call stacks, a gentle transition risks causing a regression or a scope creep. Utilizing the "normal" config helpers will be helpful in these conditions.
Another possible roadblock in the transition is the magic numbers in error reporting. Some of the functions in Git use -1 and 1 to inform callers about two different error cases or situations. Introducing a third hard-coded number to tell callers to stop the Git process for a misformatted config would be a poor design choice. Furthermore, adopting a standardized error structure like enum git_error_code is a step toward git's ongoing libification efforts, as it enables external callers consuming the API to handle errors programmatically.
Availability: -------------
I plan to dedicate 40+ hours per week to this project during my active coding period. However, I want to be completely transparent about my university's academic calendar to set realistic expectations.
In Turkey, the university summer break begins in July and ends in late September. During May and June, my schedule will be heavily occupied by final exams and major group project deadlines. For this reason, my availability during these two months will be limited to around 10-15 hours per week. I will use this time to stay active on the mailing list, participate in architectural discussions and submit smaller, preparatory patches.
To ensure the highest quality of work, I propose utilizing GSoC's officially supported flexible timeline. I am completely free during July, August, and September (with no summer school or internships). During these three months, I will dedicate 40+ hours per week entirely to git.
Community Bonding (May 1 - May 24): - Analyze environment.c and create a detailed mitigation plan for each variable. - Discuss the plan with mentors to identify potential roadblocks or edge cases. - Submit a patch about 'enum git_error_code' to start community discussion. - Set up a blog to share bi-weekly updates throughout the project.
Phase 1 (May 25 - June 28): - Introduce the '_maybe' versions of the config readers and write tests for them. - Begin mitigating "low-hanging" globals. To avoid wasting time while waiting for reviews, start drafting subsequent patches concurrently. - Publish the first progress reports on the blog.
Phase 2 (June 29 - September 15): - Discuss globals with mentors where mitigations might cause behavioral changes. - Shift focus to the more complex cases, specifically those involving eager-lazy or '_maybe' transitions. - Continue publishing regular blog updates.
Phase 3 (September 16 - September 30): - Act as a buffer period to respond to final feedback on patches currently under review. - Complete the final project report and publish it on the blog.
References: -----------
[1] https://github.com/bkkaracay/caret [2] https://lore.kernel.org/git/xmqq1pk3lmu3.fsf@gitster.g/ [3] https://lore.kernel.org/git/23428022-ab13-4a3e-90ed-ff91ef93f051@gmail.com/ [4] https://lore.kernel.org/all/pull.2044.v3.git.1771849615.gitgitgadget@gmail.com/ [5] https://lore.kernel.org/git/20260109213021.2546-2-l.s.r@web.de/
---
Thanks to everyone for their time and guidance. I'm really excited about the possibility of working on this project, and any feedback to make this proposal better is deeply appreciated.
