Linux Kernel Architecture

Chapter 17: Data Synchronization

The implementation of the system call is remarkably simple. As documented in the manual page

msync(2), the system call essentially distinguishes two modes. IfMS_SYNCis set in the flags, then dirty

pages are written to disk synchronously, whileMS_ASYNCis supposed to schedule dirty data for later

writeback.

The good news is that forMS_ASYNC, no work at all is required! Since the kernel tracks the state of dirty

pages, they will be synchronized at some point by the mechanisms described in this chapter anyway.

WithMS_SYNCset, a little more work is necessary, and the code flow diagram in Figure 17-14 considers

this case solely.

Iterate

over all

vm_areas

sys_msync

find_vma

do_fsync

Figure 17-14: Code flow diagram

forsys_msyncwith the flag

MS_SYNCset.

find_vmafinds the firstvm_areainstance in the selected area.vm_area->vm_fileholds a pointer to the

fileinstance from which the mapped data originate (this is discussed in Chapter 4). Therefore,do_fsync

can be used to synchronize the file as described above.

The method is repeated for all intervals in the desired area. This is possible because the intervals are

linked by means ofvm_area->next, as discussed in Chapter 4.

17.16 Summary

Data are persistently stored on block devices but modified in RAM. This makes it necessary to synchro-

nize the contents of both from time to time, and this chapter has introduced you to the corresponding

methods. There are several system calls that allow forexplicitly requesting that some portions of memory

are written back to disk. Additionally, the kernel uses threads that perform the same job periodically to

ensure that not too much modified data exist in RAM. While it is a good thing to keep disks busy, the ker-

nel needs to ensure that not more information than the disk can handle is scheduled to be written back,

and this chapter has discussed the techniques used toavoid congestion of the block layer. Additionally,

some corrections are necessary for systems where power is scarce, and the policy changes in laptop mode

have also been examined.

You have been introduced to all technical aspects ofhow data are shuffled back and forth between block

devices and RAM by now. What is still lacking is the decision ofwhichpages are supposed to be syn-

chronized or discarded from RAM once the kernel gets short on memory, which is the subject of the next

chapter.