I want to optimize network related disk access by avoiding
buffer copies into the main RAM. (Nic card will directly pick up
data from storage controller card.)
I do not want apps running on the host machine itself to get
adversely impacted.
I was planning to bypass buffer cache for nfs *alone* - but now as
you've told me that's not possible I am wondering what could
be other way out.
I do not mind patching nfs server or writing device driver
but I do not want to patch the kernel as I finally want it
running on a stock kernel.
---How is the general problem of "peer-to-peer" access of hardware addressed in linux? Is there a generic way hardware can pass data bypassing copies to host memory.
---Please post a copy to sara@procsys.com. I am not on the mailing list.
-Saravanan
Alan Cox wrote: > > > Is it possible to bypass the buffer cache completely > > for a block device driver? What should I do? > > No > > > Is it possible to bypass only for some processes and > > let thro for others? (This is the final behaviour I want.) > > It is up to the process to specify O_SYNC on the files it wishes to sync. > Even then the Linux kernel will do read caching. The page cache shares pages > between user space and the kernel when running binaries for example. > > Also regardless of how good the cache on your custom device is the main memory > is referenced without a copy in read only cases so will be faster than a copy > even from static ram, or a DMA. > > Alan > >> I've a custom hardware with lot of inbuilt cache. >> I want to write a block device driver for it and i >> *want to bypass the buffer cache*. >> I intend to run filesystems on top of it. >> >> I filled in the "read" and "write" functions with my own versions >> and bypassed "block_read" and "block_write". >> I could "fdisk" but mount failed complaining about "request_fn". >> >> I filled in the "request_fn" and now, it appears to me >> that all io happens through "request_fn". For instance, >> "printk" lines in "read" and "write" print >> nothing. >> >> If i check the pid in "request_fn" it is that >> of "kflushd" or "bdflush". >> >>--- >> >> Is it possible to bypass the buffer cache completely >> for a block device driver? What should I do? >> >> Is it possible to bypass only for some processes and >> let thro for others? (This is the final behaviour I want.) >> >> Where should I look for more about block device >> drivers and their interaction with buffer cache (i.e. apart from >> code in terms of design, structure etc.,). >> >>--- >> >> I am not on the mailing list, so please post a copy to >> "sara@procsys.com" too. >> >>--- >> >> Thanks in advance >> >> Saravanan >>
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/