Date: Wed, 6 May 2015 21:06:33 +0000 (UTC) From: Alexander Motin <mav@FreeBSD.org> To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-stable@freebsd.org, svn-src-stable-10@freebsd.org Subject: svn commit: r282568 - stable/10/sys/kern Message-ID: <201505062106.t46L6XcG036162@svn.freebsd.org>
next in thread | raw e-mail | index | archive | help
Author: mav Date: Wed May 6 21:06:32 2015 New Revision: 282568 URL: https://svnweb.freebsd.org/changeset/base/282568 Log: MFC r281825: Rewrite physio() to not allocate pbufs for unmapped I/O. pbufs is a limited resource, and their allocator is not SMP-scalable. So instead of always allocating pbuf to immediately convert it to bio, allocate bio just here. If buffer needs kernel mapping, then pbuf is still allocated, but used only as a source of KVA and storage for a list of held pages. On 40-core system doing many 512-byte reads from user level to array of raw SSDs this change removes huge lock congestion inside pbuf allocator. It improves peak performance from ~300K to ~1.2M IOPS. On my previous 24-core system this problem also existed, but was less serious. Modified: stable/10/sys/kern/kern_physio.c Directory Properties: stable/10/ (props changed) Modified: stable/10/sys/kern/kern_physio.c ============================================================================== --- stable/10/sys/kern/kern_physio.c Wed May 6 21:03:19 2015 (r282567) +++ stable/10/sys/kern/kern_physio.c Wed May 6 21:06:32 2015 (r282568) @@ -25,27 +25,26 @@ __FBSDID("$FreeBSD$"); #include <sys/bio.h> #include <sys/buf.h> #include <sys/conf.h> +#include <sys/malloc.h> #include <sys/proc.h> #include <sys/uio.h> +#include <geom/geom.h> #include <vm/vm.h> +#include <vm/vm_page.h> #include <vm/vm_extern.h> +#include <vm/vm_map.h> int physio(struct cdev *dev, struct uio *uio, int ioflag) { - struct buf *bp; - struct cdevsw *csw; + struct buf *pbuf; + struct bio *bp; + struct vm_page **pages; caddr_t sa; - u_int iolen; - int error, i, mapped; - - /* Keep the process UPAGES from being swapped. XXX: why ? */ - PHOLD(curproc); - - bp = getpbuf(NULL); - sa = bp->b_data; - error = 0; + u_int iolen, poff; + int error, i, npages, maxpages; + vm_prot_t prot; /* XXX: sanity check */ if(dev->si_iosize_max < PAGE_SIZE) { @@ -76,95 +75,128 @@ physio(struct cdev *dev, struct uio *uio uprintf("%s: request vectors=%d > 1; " "cannot split request\n", devtoname(dev), uio->uio_iovcnt); - - error = EFBIG; - goto doerror; + return (EFBIG); } + /* + * Keep the process UPAGES from being swapped. Processes swapped + * out while holding pbufs, used by swapper, may lead to deadlock. + */ + PHOLD(curproc); + + bp = g_alloc_bio(); + if (uio->uio_segflg != UIO_USERSPACE) { + pbuf = NULL; + pages = NULL; + } else if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) { + pbuf = NULL; + maxpages = btoc(MIN(uio->uio_resid, MAXPHYS)) + 1; + pages = malloc(sizeof(*pages) * maxpages, M_DEVBUF, M_WAITOK); + } else { + pbuf = getpbuf(NULL); + sa = pbuf->b_data; + maxpages = btoc(MAXPHYS); + pages = pbuf->b_pages; + } + prot = VM_PROT_READ; + if (uio->uio_rw == UIO_READ) + prot |= VM_PROT_WRITE; /* Less backwards than it looks */ + error = 0; for (i = 0; i < uio->uio_iovcnt; i++) { while (uio->uio_iov[i].iov_len) { - bp->b_flags = 0; + bzero(bp, sizeof(*bp)); if (uio->uio_rw == UIO_READ) { - bp->b_iocmd = BIO_READ; + bp->bio_cmd = BIO_READ; curthread->td_ru.ru_inblock++; } else { - bp->b_iocmd = BIO_WRITE; + bp->bio_cmd = BIO_WRITE; curthread->td_ru.ru_oublock++; } - bp->b_iodone = bdone; - bp->b_data = uio->uio_iov[i].iov_base; - bp->b_bcount = uio->uio_iov[i].iov_len; - bp->b_offset = uio->uio_offset; - bp->b_iooffset = uio->uio_offset; - bp->b_saveaddr = sa; - - /* Don't exceed drivers iosize limit */ - if (bp->b_bcount > dev->si_iosize_max) - bp->b_bcount = dev->si_iosize_max; - - /* - * Make sure the pbuf can map the request - * XXX: The pbuf has kvasize = MAXPHYS so a request - * XXX: larger than MAXPHYS - PAGE_SIZE must be - * XXX: page aligned or it will be fragmented. + bp->bio_offset = uio->uio_offset; + bp->bio_data = uio->uio_iov[i].iov_base; + bp->bio_length = uio->uio_iov[i].iov_len; + if (bp->bio_length > dev->si_iosize_max) + bp->bio_length = dev->si_iosize_max; + if (bp->bio_length > MAXPHYS) + bp->bio_length = MAXPHYS; + + /* + * Make sure the pbuf can map the request. + * The pbuf has kvasize = MAXPHYS, so a request + * larger than MAXPHYS - PAGE_SIZE must be + * page aligned or it will be fragmented. */ - iolen = ((vm_offset_t) bp->b_data) & PAGE_MASK; - if ((bp->b_bcount + iolen) > bp->b_kvasize) { - /* - * This device does not want I/O to be split. - */ + poff = (vm_offset_t)bp->bio_data & PAGE_MASK; + if (pbuf && bp->bio_length + poff > pbuf->b_kvasize) { if (dev->si_flags & SI_NOSPLIT) { uprintf("%s: request ptr %p is not " "on a page boundary; cannot split " "request\n", devtoname(dev), - bp->b_data); + bp->bio_data); error = EFBIG; goto doerror; } - bp->b_bcount = bp->b_kvasize; - if (iolen != 0) - bp->b_bcount -= PAGE_SIZE; + bp->bio_length = pbuf->b_kvasize; + if (poff != 0) + bp->bio_length -= PAGE_SIZE; } - bp->b_bufsize = bp->b_bcount; - bp->b_blkno = btodb(bp->b_offset); + bp->bio_bcount = bp->bio_length; + bp->bio_dev = dev; - csw = dev->si_devsw; - if (uio->uio_segflg == UIO_USERSPACE) { - if (dev->si_flags & SI_UNMAPPED) - mapped = 0; - else - mapped = 1; - if (vmapbuf(bp, mapped) < 0) { + if (pages) { + if ((npages = vm_fault_quick_hold_pages( + &curproc->p_vmspace->vm_map, + (vm_offset_t)bp->bio_data, bp->bio_length, + prot, pages, maxpages)) < 0) { error = EFAULT; goto doerror; } + if (pbuf) { + pmap_qenter((vm_offset_t)sa, + pages, npages); + bp->bio_data = sa + poff; + } else { + bp->bio_ma = pages; + bp->bio_ma_n = npages; + bp->bio_ma_offset = poff; + bp->bio_data = unmapped_buf; + bp->bio_flags |= BIO_UNMAPPED; + } } - dev_strategy_csw(dev, csw, bp); + dev->si_devsw->d_strategy(bp); if (uio->uio_rw == UIO_READ) - bwait(bp, PRIBIO, "physrd"); + biowait(bp, "physrd"); else - bwait(bp, PRIBIO, "physwr"); + biowait(bp, "physwr"); + + if (pages) { + if (pbuf) + pmap_qremove((vm_offset_t)sa, npages); + vm_page_unhold_pages(pages, npages); + } - if (uio->uio_segflg == UIO_USERSPACE) - vunmapbuf(bp); - iolen = bp->b_bcount - bp->b_resid; - if (iolen == 0 && !(bp->b_ioflags & BIO_ERROR)) + iolen = bp->bio_length - bp->bio_resid; + if (iolen == 0 && !(bp->bio_flags & BIO_ERROR)) goto doerror; /* EOF */ uio->uio_iov[i].iov_len -= iolen; uio->uio_iov[i].iov_base = (char *)uio->uio_iov[i].iov_base + iolen; uio->uio_resid -= iolen; uio->uio_offset += iolen; - if( bp->b_ioflags & BIO_ERROR) { - error = bp->b_error; + if (bp->bio_flags & BIO_ERROR) { + error = bp->bio_error; goto doerror; } } } doerror: - relpbuf(bp, NULL); + if (pbuf) + relpbuf(pbuf, NULL); + else if (pages) + free(pages, M_DEVBUF); + g_destroy_bio(bp); PRELE(curproc); return (error); }
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201505062106.t46L6XcG036162>