Re: Use pg_pwritev_with_retry() instead of write() in dir_open_for_write() to avoid partial writes?

Hi,

Interestingly, I also needed something like pg_pwrite_zeros() today. Exposed
via smgr, for more efficient relation extensions.

On 2022-10-27 14:54:00 +0900, Michael Paquier wrote:
> Regarding 0002, using pg_pwrite_zeros() as a routine name, as
> suggested by Thomas, sounds good to me. However, I am not really a
> fan of its dependency with PGAlignedXLogBlock, because it should be
> able to work with any buffers of any sizes, as long as the input
> buffer is aligned, shouldn't it? For example, what about
> PGAlignedBlock? So, should we make this more extensible? My guess
> would be the addition of the block size and the block pointer to the
> arguments of pg_pwrite_zeros(), in combination with a check to make
> sure that the input buffer is MAXALIGN()'d (with an Assert() rather
> than just an elog/pg_log_error?).

I don't like passing in the buffer. That leads to code like in Bharat's latest
version, where we now zero that buffer on every invocation of
pg_pwrite_zeros() - not at all cheap. And every caller has to have provisions
to provide that buffer.

The block sizes don't need to match, do they? As long as the block is properly
aligned, we can change the iov_len of the final iov to match whatever the size
is being passed in, no?

Why don't we define a

static PGAlignedBlock source_of_zeroes;

in file_utils.c, and use that in pg_pwrite_zeros(), being careful to set the
iov_len arguments correctly?

Greetings,

Andres Freund