[LU-3606] Implement fallocate() support for ldiskfs Lustre

• Type: New Feature
• Resolution: Fixed
• Priority: Major
• Fix Version/s: Lustre 2.14.0
• Labels: medium patch
• Assignee: Arshad Hussain
• Reporter: Andreas Dilger

Description

The sys_fallocate() syscall was introduced to the linux kernel in the 2.6.24 kernel. There is also an ext4_fallocate() method added in this same kernel release. This has been available in vendor kernels since RHEL 5.4.

We need to implement an fallocate() method for llite, and transport this to the OSTs to interface with the underlying OSD's fallocate() code (for ldiskfs, ZFS has no such method).

The FALLOC_FL_PUNCH_HOLE functionality of fallocate() fits very nicely with the OST_PUNCH RPC, since this was anticipated in Lustre from the beginning. The actual space reservation might also use OST_PUNCH or OST_WRITE, or we might consider a new OST_PREALLOC RPC, depending on what fits best.

Issue Links

• is cloned by: LU-14157 Implement fallocate() support for zfs Lustre
• is related to: LU-14333 ASSERTION( oinfo->loi_lvb.lvb_size >= oinfo->loi_kms ) failed
• is related to: LU-15167 fallocate does not increase quota usage
• is related to: LU-15519 fallocate does not increase projid quota usage
• is related to: LU-10370 "truncate" does not update blocks count on client
• is related to: LU-8699 Further fsx improvements
• is related to: LU-13155 "lfs migrate" and "lfs mirror" should fallocate() the target file
• is related to: LU-17055 Implement FALLOC_FL_ZERO_RANGE
• is related to: LU-17326 Implement FALLOC_FL_COLLAPSE_RANGE and FALLOC_FL_INSERT_RANGE
• is related to: LUDOC-487 Document llite.*.fallocate_zero_blocks tunable
• is related to: LU-14382 Implement fallocate() support at MDT
• is related to: LU-10810 SEEK_HOLE and SEEK_DATA support for lseek
• is related to: LU-14287 Add 'fallocate' to racer
• is related to: LU-14160 Implement fallocate FALLOCATE_FL_PUNCH_HOLE support

Comments

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/9275/
Subject: LU-3606 fallocate: Implement fallocate preallocate operation
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 48457868a02ae5a0407f304c9028f7e80e7fb574

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/37277/
Subject: LU-3606 lustre: Reserve OST_FALLOCATE(fallocate) opcode
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 46a11df089c92f0d18e6f3224eca76abfcd0ac67

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/37273/
Subject: LU-3606 fsx: Add fallocate operation to fsx
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 853d180121a6e313ffe678bbe5c107d7afc601d5

[Gerrit Updater]

Arshad Hussain ([e-mail]) uploaded a new patch: https://review.whamcloud.com/37277
Subject: LU-3606 lustre: Reserve fallocate opcode
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: a52c11c2c0935ce1b741dfde703bbaeb93c7c499

[Gerrit Updater]

Arshad Hussain ([e-mail]) uploaded a new patch: https://review.whamcloud.com/37273
Subject: LU-3606 fsx: Add fallocate operation to fsx
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 2269c370864bd42a5fbf0608afecba0ec8f010db

[Andreas Dilger]

Yes, please reuse the same Change-Id. That will allow comparing the new patch against the old one, and the old one is hopelessly outdated and not useful.

[Abrar-ahmed (Inactive)]

[e-mail] I worked on the patch and have it working, including newer kernel versions. Most review comments against last patch addressed. Can post it against the same change-id if that works.

[Andreas Dilger]

Jinshan, can you see what is needed to get this patch across the finish line? Swap il is no longer available to work on it.

[Jinshan Xiong (Inactive)]

I will track the state of this ticket.

[Swapnil Pimpale (Inactive)]

I have uploaded the first patchset here: http://review.whamcloud.com/#/c/9275/

[Andreas Dilger]

The fallocate() API has been added to newer versions of the Lustre kernel to provide both persistent space reservation (block preallocation for a file, possibly beyond the file size, without having to write zeroes to the whole file), and for the reverse operation of hole punching (freeing allocated blocks in the middle or end of a file). However, Lustre predates these APIs and has not yet added support for them. Being able to preallocate space for a file is very useful for HPC applications that know the size of the output file in advance, and helps Lustre make better allocation decisions based on the file size.

This project will touch all of the layers of the Lustre IO stack, from the Linux client VFS interface, through the file layout to the RPC interface, over to the storage server, and then interfacing with the back-end filesystems to use the space reservation (ext4 only) and hole punch (both ext4 and ZFS) methods. As a stretch goal, an ioctl() pass-through interface from the server mountpoint to the underlying filesystem would allow direct passing of FITRIM ioctl commands to the storage so that SSD-based filesystems can optimize their internal space management.