On Mon, Nov 24, 2008 at 12:17 PM, MinChan Kim <minchan.kim@xxxxxxxxx> wrote: > On Mon, Nov 24, 2008 at 11:58 AM, Peter Teoh <htmldeveloper@xxxxxxxxx> wrote: >> On Mon, Nov 24, 2008 at 7:29 AM, MinChan Kim <minchan.kim@xxxxxxxxx> wrote: >>> I'm so late :) >>> >>> On Sun, Nov 23, 2008 at 4:17 PM, Wang Yu <wangyuict@xxxxxxxxx> wrote: >>>> >>>> >>>> On Sun, Nov 23, 2008 at 2:23 AM, Rik van Riel <riel@xxxxxxxxxxx> wrote: >>>>> >>>>> Peter Teoh wrote: >>>>>> >>>>>> when a process mmap() a section of a file into its own process memory, >>>>>> the process memory will maintain a copy of the data of that section of >>>>>> the file. >>>>> >>>>> No, it does not maintain a copy. >>>>> >>>>> It mmaps the page cache pages into its own address space. >>>> >>>> >>>> According to your explanation, the flow is physical file(on disk)-->Page >>>> Cache(on memory, but in kernel space)-->Process Memory(on memory, but in >>>> user space). Is it? I am not sure.... >>>>> >>> >>> Yes. When kernel find no page mapping, it allocate new page in page >>> cache and copy from on-disk page to new page, then map the new page to >>> user space address. >>> so, Never duplication. >>> >>>>>> so...does there exists duplicated buffering? (one in kernel - >>>>>> pagecache, and one in userspace - for mmap() content of the file in >>>>>> process memory) >>>>> >>>>> No, there is no such double buffering. >>>> >>>> But what is the difference? Why linux do it? >>> >>> In case of read system call except O_DIRECT, It's duplication between >>> user buffer and page cache. >>> Read/write system call abstract page to file. so you always need user buffer. >>> Let think. If you want read some data, First of all you need some >>> space which is user buffer. >>> Mmap system call abstract page to memory. so you can handle file as >>> memory operation without user buffer that mean It don't have >>> duplication overhead. >> >> I see. So u are saying that read() will duplicate buffer between >> userspace (user buffer) and kernel (pagecache), but for mmap() >> operation, since there is no duplication, all access from user process >> will immediate trigger a context switch, to read in the data from >> ring0, right? > > yeb. More exactly, to read in the data from file fault in page fault handler. > >> Since u normally read in blocks of data using read(), and access data >> byte-wise with mmap()'s pointer, so read() is much more efficient, as >> it trigger much lesser context switches than mmap()'s way of pointer >> accessing kernel memory? >> >> Performance-wise, mmap() will perform worst off than read()? > > I am not sure which case is better. > That's because read operation have a overhead memory copy from kernel to user. > But it occur less page fault than mmap. but on-demand readahead or > hint function can increase performance in case of sequential read > read() vs mmap().....mmap() takes a page fault with every single byte of memory access, if the mmap() ptr is byte ptr, then it seemingly should be slower than read().... and a possible similar experience is here: http://lkml.org/lkml/2008/1/14/517 > Page fault VS memory copy ?? which is more overhead ?? > I think Rik can answer this question. > -- Regards, Peter Teoh Ernest Hemingway - "Never mistake motion for action." -- To unsubscribe from this list: send an email with "unsubscribe kernelnewbies" to ecartis@xxxxxxxxxxxx Please read the FAQ at http://kernelnewbies.org/FAQ
