My access is almost purely sequential and primarily writing, so read-ahead doesn't help me. What's problematic with pread/pwrite is the lack of error channel from media errors. BSG looks very interesting. I'll look further into that today. On Thu, Sep 1, 2016 at 5:16 PM, Bart Van Assche <bart.vanassche@xxxxxxxxxxx> wrote: > On 09/01/2016 02:48 PM, Alex Austin wrote: >> >> CC'ing linux-scsi and linux-block. >> >> Also, please CC me in replies. >> >> On Thu, Sep 1, 2016 at 4:43 PM, Alex Austin <aaustin@xxxxxxxxxxxxxxxxxxx> >> wrote: >>> >>> Hello, >>> What is the most performant way to directly interface with an attached >>> hard >>> drive? I've so far used read()/write() on /dev/sd_ but I find error >>> handling >>> exceedingly difficult, as I don't always even get errors reported, even >>> if the >>> open() call includes O_DIRECT. I've also used ioctl(SG_IO), but find that >>> it's >>> extremely slow due to the lack of queuing support in the API. Is there a >>> mid-level API that will get me decent error handling while allowing >>> command >>> queuing, or do I just need to make multiple threads all doing separate >>> SG_IO >>> ioctls? > > > What the most efficient way is to interface is with a hard drive depends on > the I/O pattern. Are you aware that buffered I/O performs read-ahead? Have > you already tried to tune the read-ahead setting (blockdev --getra / > blockdev --setra)? > > BTW, if you need an example of how to queue SG_IO, you are welcome to have a > look at the fio source code. You will probably be interested in the bsg API. > See e.g. https://lwn.net/Articles/174469/. > > Bart. -- Intelligence is knowing that a tomoato is a fruit; wisdom is knowing that it doesn't go in a fruit salad; charisma is selling a tomato-based fruit salad. -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
