On 13 February 2010 01:37, <tytso@xxxxxxx> wrote: > On Fri, Feb 12, 2010 at 08:52:15AM +0530, Kailas Joshi wrote: >> Won't this get fixed by performing early reservations as mentioned in >> my scheme? We are reserving required credits in the path of write >> system call and these will be kept reserved until transaction commit. >> So, the journal space for allocation at commit will be guaranteed. > > Yes, if you account for these separately. One challenge is > over-estimating the needed credits will be tricky. If we go down this > path, be sure that the bonnie style write(fd, &ch, 1) in a tight loop > doesn't end up reserving a separate set of credits for each write > system call to the same block. (It can be done; if the DA block is > already instantiated, you can assume that credits have already been > reserved.) Okay >> Sorry, I didn't understand why processes need to be suspended. >> In my scheme, I am issuing magic handle only after locking the current >> transaction. AFAIK after the transaction is locked, it can receive the >> block journaling requests for already created handles(in our case, for >> already reserved journal space), and the new concurrent requests for >> journal_start() will go to the new current transaction. Since, the >> credits for locked transaction are fixed (by means of early >> reservations) we can know whether journal has enough space for the new >> journal_start(). So, as long as journal has enough space available, >> new processes need now be stalled. > > But while you are modifying blocks that need to go into the journal > via the locked (old) transaction, it's not safe to start a new > transaction and start issuing handles against the new transaction. > > Just to give one example, suppose we need to update the extent > allocation tree for an inode in the locked/committing transaction as > the delayed allocation blocks are being resolved --- and in another > process, that inode is getting truncated or unlinked, which also needs > to modify the extent allocation tree? Hilarty ensues, unless you use > a block all attempts to create a new handle (practically speaking, by > blocking all attempts to start a new transaction), until this new > delayed allocation resolution phase which you have proposed is > complete. Okay. So, basically process stalling is unavoidable as we cannot modify a buffer data in past transaction after it has been modified in current transaction. Can we restrict the scope for this blocking? Blocking on journal_start() will block all processes even though they are operating on mutually exclusive sets of metadata buffers. Can we restrict this blocking to allocation/deallocation paths by blocking in get_write_access() on specific cases(some condition on buffer)? This way, since all files will use commit-time allocation, very few(sync and direct-io mode) file operations will be stalled. I am not sure whether this is feasible or not. Please let me know more on this. Thanks & Regards, Kailas -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
