Hello All, I'm new to the mailing list and have some introductory questions. After starting the Eudyptula challenge, I began to read 'Linux Kernel Development - 2nd ed.' and 'Linux Device Drivers - 3rd ed.'. I think I am getting a feel for the data structures and a general picture of how the kernel works, but I have some questions that might help me connect some of the dots. I'll try to explain what I think is going on and perhaps you all can tell me what is really going on. Please feel free to provide references rather than re-explaining everything if the references already do a good job! Here I go...sorry in advance for the length! ==== Kernel Questions ==== When a user-space process is `clone()`d, it is provided a `struct task_struct`, call it `my_task`, from a slab cache that already has `struct task_struct`s allocated but free to be assigned. If `CLONE_VM` is is passed to `clone()`, then `my_task` inherits the parent task's address space (which means `my_task` is a thread), otherwise it is provided a `struct mm_struct` from another slab cache and the object is referenced in `my_task->mm`. 1) Assuming `my_task` does not inherit its parent's address space, what does `my_task`s address space have mapped initially? E.g. what entries are in `pgd`, `pmd` and `pte`? Does `my_task->mm` inherit the kernel's page tables? In particular, does `my_task` "see" 4gb of memory where the first 896mb (ZONE_NORMAL) are reserved for the kernel plus some extra VMA's in ZONE_HIGHMEM containing data specific to the process? 2) This is slightly connected to question (1). Suppose that `my_task` needs to open a file that may or may not be in main memory. Given `/path/to/some_file.txt`, the VFS traverses the dentries `/`, `path`, `to` and `some_file.txt` to check that the path is valid by sequentially doing something along these lines (of course, something like this would only happen if the dentry `some_file.txt` were not already in the dentry cache): e.g., for `/`: i) Somehow obtain the dentry structure for `/`. ii) Look up the inode via `dentry->d_inode` and use this to determine the address of `/`'s data (on disk?...at least initially) iii) The actual data associated to `/` contains a list of the files it contains. Therefore, this data must be mapped into main memory...first check if the data is in a page in the `pagecache`, otherwise it must be read from disk first. iv) Check that data contains an entry for `path` (the next dentry). v) If `path` is an entry in `/`s data, repeat at (i) for `path`. Repeat these steps to determine of the path is invalid or until the data for the file `some_file.txt` is memory mapped. Firstly, does this "algorithm" reflect how the file data is found/moved to main memory? Second, even if (i) only happens once at boot, how does the kernel initially obtain a dentry for the root directory/file?...or for any file? Are these created on-the-fly or are they stored on disk? The reason I say that this is connected to (1) is that the data needed to run `my_task`, e.g. the code itself, must be loaded into main memory before being mapped into the process's address space when the process is created. I assumed when `my_task` is created, the code is mapped into memory using the same sequence of steps...after all, the process itself is started by looking up and executing a file. 3) This has to do with how data is actually read from disk. A disk can only read data in sector sized chunks and the kernel only reads memory in block sized chunks. Therefore, the block size must be a multiple of the sector size (which is fixed by the hardware) and there is an additional limitation that a block must be smaller than a page. Suppose sectors are 512b and blocks are 1kb (which is below the upper bound for both 32/64bit x86 processors). Now, suppose `some_file.txt` is a file that is, say, 1.2kb in size and is not currently in main memory. Once the location of `some_file.txt`s data is found on disk (via 2 above), it must be read into main memory in block sized chunks. i) Can the hard disk only start reading at sector boundaries? i.e. the chunks 0-1kb, .5kb-1.5kb, etc... Or perhaps only at block boundaries (as enforced by the kernel)? 0-1kb, 1-2kb, 2-3kb, etc... Or can it start reading at any byte on the disk as long as it reads a contiguous block sized chunk? i.e. (some number of bytes into disk)-(some number of bytes into disk + 1kb) ii) Regardless of where the disk can start reading data, the size of `some_file.txt`s data requires that at least 2kb must be read from disk. Depending on the offset of the data inside a sector/block and where the disk can start reading, `some_file.txt`s data will lie inside of a 2kb block surrounded by .8kb of (possibly) random data (possibly on both sides). This data would then be moved to a page in memory and the address of the page as well as the offset of `some_file.txt`s data inside the page would then be available to the process for file operations. Is that .8kb of space, now in main memory, wasted? Or can a future block of data read from disk write over the "dead space" surrounding `some_file.txt`'s data? ==== Other ==== 1) It seems like most of the material (outside of the kernel itself) corresponds to version 2.6. Why is this? Has nothing "significant" changed with the core structure of the kernel since then? Or is it just that no one has decided to write a book about a newer version? 2) In some header files I see comments that look like: /** * struct some_struct - description * @property1: description * ... * @propertyn: description */ This style looks very similar to http://usejsdoc.org/ which is used to build html documentation from code comments. Does the kernel use a tool like this? i.e. is there html documentation generated from the code comments? ==== Thanks in advance for any info you can provide! -Jaret _______________________________________________ Kernelnewbies mailing list Kernelnewbies@xxxxxxxxxxxxxxxxx http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
