gpfs_prealloc() support to preallocate any bytes of an new created file without zero-out data

Most of recent AI workload saves raw data in cloud object storage. When GPU is going to training partial of raw data, these data will be loaded and saved in high performance file system, such as Storage Scale.

In ADAS and some of other AI workloads, file size of these raw data are mostly several KB to several hundreds of KB. When we get these data from cloud object storage and write it to GPFS file system, GPFS will create a new file, and allocate a data block (default 4MB) to save these several hundreds of KB data and then shrink the file size. There are lots of allocation map related operation for these kinds of files.

In current gpfs_prealloc API implementations:

https://www.ibm.com/docs/en/storage-scale/5.2.3?topic=interfaces-gpfs-prealloc-subroutine

"The preallocated disk storage begins at the specified offset in the file and extends for the specified number of bytes, rounded up to a GPFS block boundary. Existing data is not modified. Reading any of the preallocated blocks returns zeroes."

From our sanity test in a 5.2.3 cluster, gpfs_prealloc will always zero-out preallocate blocks before we write data into them which make some I/O amplification.

Since we already know the file size and how many data we need to save to GPFS file system, we would like to propose a method for this API to make it support to preallocate any bytes of data for a new created file (don't rounded up to block boundary), and don't zero out these bytes. Hope we can use this enhanced API to save some metadata operations in GPFS.