From 2bab0c553588a7c2a8dd59a0f20b9b5ded528274 Mon Sep 17 00:00:00 2001 From: Gleb Smirnoff Date: Fri, 8 Jan 2016 20:34:57 +0000 Subject: New sendfile(2) syscall. A joint effort of NGINX and Netflix from 2013 and up to now. The new sendfile is the code that Netflix uses to send their multiple tens of gigabits of data per second. The new implementation features asynchronous I/O, when I/O operations are launched, but not awaited to be complete. An explanation of why such behavior is beneficial compared to old one is going to be too long for a commit message, so we will skip it here. Additional features of new syscall are extra flags, which provide an application more control over data sent. The SF_NOCACHE flag tells kernel that data shouldn't be cached after it was sent. The SF_READAHEAD() macro allows to specify readahead size in pages. The new syscalls is a drop in replacement. No modifications are required to applications. One can take nginx binary for stable/10 and run it successfully on head. Although SF_NODISKIO lost its original sense, as now sendfile doesn't block, and now means something completely different (tm), using the new sendfile the old way is absolutely safe. Celebrates: Netflix global launch! Sponsored by: Nginx, Inc. Sponsored by: Netflix Relnotes: yes --- lib/libc/sys/sendfile.2 | 123 ++++++--- sys/dev/ti/if_ti.c | 4 +- sys/kern/uipc_mbuf.c | 4 + sys/kern/uipc_syscalls.c | 682 ++++++++++++++++++++++++++++++----------------- sys/sys/mbuf.h | 4 +- sys/sys/sf_buf.h | 7 + sys/sys/socket.h | 5 +- usr.bin/netstat/mbuf.c | 26 +- 8 files changed, 564 insertions(+), 291 deletions(-) diff --git a/lib/libc/sys/sendfile.2 b/lib/libc/sys/sendfile.2 index b363382af153..2b52dd9c883c 100644 --- a/lib/libc/sys/sendfile.2 +++ b/lib/libc/sys/sendfile.2 @@ -25,7 +25,7 @@ .\" .\" $FreeBSD$ .\" -.Dd January 7, 2010 +.Dd January 7, 2016 .Dt SENDFILE 2 .Os .Sh NAME @@ -46,7 +46,7 @@ The .Fn sendfile system call -sends a regular file specified by descriptor +sends a regular file or shared memory object specified by descriptor .Fa fd out a stream socket specified by descriptor .Fa s . @@ -101,32 +101,55 @@ the system will write the total number of bytes sent on the socket to the variable pointed to by .Fa sbytes . .Pp -The +The least significant 16 bits of .Fa flags argument is a bitmap of these values: -.Bl -item -offset indent -.It -.Dv SF_NODISKIO . -This flag causes any -.Fn sendfile -call which would block on disk I/O to instead -return -.Er EBUSY . -Busy servers may benefit by transferring requests that would -block to a separate I/O worker thread. -.It -.Dv SF_MNOWAIT . -Do not wait for some kernel resource to become available, -in particular, -.Vt mbuf -and -.Vt sf_buf . -The flag does not make the -.Fn sendfile -syscall truly non-blocking, since other resources are still allocated -in a blocking fashion. -.It -.Dv SF_SYNC . +.Bl -tag -offset indent +.It Dv SF_NODISKIO +This flag causes +.Nm +to return +.Er EBUSY +instead of blocking when a busy page is encountered. +This rare situation can happen if some other process is now working +with the same region of the file. +It is advised to retry the operation after a short period. +.Pp +Note that in older +.Fx +versions the +.Dv SF_NODISKIO +had slightly different notion. +The flag prevented +.Nm +to run I/O operations in case if an invalid (not cached) page is encountered, +thus avoiding blocking on I/O. +Starting with +.Fx 11 +.Nm +sending files off the +.Xr ffs 7 +filesystem doesn't block on I/O +(see +.Sx IMPLEMENTATION NOTES +), so the condition no longer applies. +However, it is safe if an application utilizes +.Dv SF_NODISKIO +and on +.Er EBUSY +performs the same action as it did in +older +.Fx +versions, e.g. +.Xr aio_read 2, +.Xr read 2 +or +.Nm +in a different context. +.It Dv SF_NOCACHE +The data sent to socket will not be cached by the virtual memory system, +and will be freed directly to the pool of free pages. +.It Dv SF_SYNC .Nm sleeps until the network stack no longer references the VM pages of the file, making subsequent modifications to it safe. @@ -134,6 +157,22 @@ Please note that this is not a guarantee that the data has actually been sent. .El .Pp +The most significant 16 bits of +.Fa flags +specify amount of pages that +.Nm +may read ahead when reading the file. +A macro +.Fn SF_FLAGS +is provided to combine readahead amount and flags. +Example shows specifing readahead of 16 pages and +.Dv SF_NOCACHE +flag: +.Pp +.Bd -literal -offset indent -compact + SF_FLAGS(16, SF_NOCACHE) +.Ed +.Pp When using a socket marked for non-blocking I/O, .Fn sendfile may send fewer bytes than requested. @@ -149,6 +188,18 @@ The .Fx implementation of .Fn sendfile +doesn't block on disk I/O when it sends a file off the +.Xr ffs 7 +filesystem. +The syscall returns success before the actual I/O completes, and data +is put into the socket later unattended. +However, the order of data in the socket is preserved, so it is safe +to do further writes to the socket. +.Pp +The +.Fx +implementation of +.Fn sendfile is "zero-copy", meaning that it has been optimized so that copying of the file data is avoided. .Sh TUNING On some architectures, this system call internally uses a special @@ -232,12 +283,10 @@ The argument is not a valid socket descriptor. .It Bq Er EBUSY -Completing the entire transfer would have required disk I/O, so -it was aborted. -Partial data may have been sent. -(This error can only occur when +A busy page was encountered and .Dv SF_NODISKIO -is specified.) +had been specified. +Partial data may have been sent. .It Bq Er EFAULT An invalid address was specified for an argument. .It Bq Er EINTR @@ -310,9 +359,19 @@ first appeared in .Fx 3.0 . This manual page first appeared in .Fx 3.1 . +In +.Fx 10 +support for sending shared memory descriptors had been introduced. +In +.Fx 11 +a non-blocking implementation had been introduced. .Sh AUTHORS -The +The initial implementation of .Fn sendfile system call and this manual page were written by .An David G. Lawrence Aq Mt dg@dglawrence.com . +The +.Fx 11 +implementation was written by +.An Gleb Smirnoff Aq Mt glebius@FreeBSD.org . diff --git a/sys/dev/ti/if_ti.c b/sys/dev/ti/if_ti.c index 9b9c4a173ac1..eba7f7b01029 100644 --- a/sys/dev/ti/if_ti.c +++ b/sys/dev/ti/if_ti.c @@ -1634,7 +1634,7 @@ ti_newbuf_jumbo(struct ti_softc *sc, int idx, struct mbuf *m_old) m[i]->m_data = (void *)sf_buf_kva(sf[i]); m[i]->m_len = PAGE_SIZE; MEXTADD(m[i], sf_buf_kva(sf[i]), PAGE_SIZE, - sf_buf_mext, (void*)sf_buf_kva(sf[i]), sf[i], + sf_mext_free, (void*)sf_buf_kva(sf[i]), sf[i], 0, EXT_DISPOSABLE); m[i]->m_next = m[i+1]; } @@ -1699,7 +1699,7 @@ nobufs: if (m[i]) m_freem(m[i]); if (sf[i]) - sf_buf_mext((void *)sf_buf_kva(sf[i]), sf[i]); + sf_mext_free((void *)sf_buf_kva(sf[i]), sf[i]); } return (ENOBUFS); } diff --git a/sys/kern/uipc_mbuf.c b/sys/kern/uipc_mbuf.c index de6b313f1b43..0b0b69725316 100644 --- a/sys/kern/uipc_mbuf.c +++ b/sys/kern/uipc_mbuf.c @@ -338,6 +338,9 @@ mb_free_ext(struct mbuf *m) case EXT_SFBUF: sf_ext_free(m->m_ext.ext_arg1, m->m_ext.ext_arg2); break; + case EXT_SFBUF_NOCACHE: + sf_ext_free_nocache(m->m_ext.ext_arg1, m->m_ext.ext_arg2); + break; default: KASSERT(m->m_ext.ext_cnt != NULL, ("%s: no refcounting pointer on %p", __func__, m)); @@ -404,6 +407,7 @@ mb_dupcl(struct mbuf *n, const struct mbuf *m) switch (m->m_ext.ext_type) { case EXT_SFBUF: + case EXT_SFBUF_NOCACHE: sf_ext_ref(m->m_ext.ext_arg1, m->m_ext.ext_arg2); break; default: diff --git a/sys/kern/uipc_syscalls.c b/sys/kern/uipc_syscalls.c index c33a2cf30a22..cac698a97cae 100644 --- a/sys/kern/uipc_syscalls.c +++ b/sys/kern/uipc_syscalls.c @@ -113,15 +113,6 @@ static int getpeername1(struct thread *td, struct getpeername_args *uap, counter_u64_t sfstat[sizeof(struct sfstat) / sizeof(uint64_t)]; -/* - * sendfile(2)-related variables and associated sysctls - */ -static SYSCTL_NODE(_kern_ipc, OID_AUTO, sendfile, CTLFLAG_RW, 0, - "sendfile(2) tunables"); -static int sfreadahead = 1; -SYSCTL_INT(_kern_ipc_sendfile, OID_AUTO, readahead, CTLFLAG_RW, - &sfreadahead, 0, "Number of sendfile(2) read-ahead MAXBSIZE blocks"); - static void sfstat_init(const void *unused) { @@ -1858,13 +1849,12 @@ sf_ext_free(void *arg1, void *arg2) sf_buf_free(sf); vm_page_lock(pg); - vm_page_unwire(pg, PQ_INACTIVE); /* * Check for the object going away on us. This can * happen since we don't hold a reference to it. * If so, we're responsible for freeing the page. */ - if (pg->wire_count == 0 && pg->object == NULL) + if (vm_page_unwire(pg, PQ_INACTIVE) && pg->object == NULL) vm_page_free(pg); vm_page_unlock(pg); @@ -1877,6 +1867,43 @@ sf_ext_free(void *arg1, void *arg2) } } +/* + * Same as above, but forces the page to be detached from the object + * and go into free pool. + */ +void +sf_ext_free_nocache(void *arg1, void *arg2) +{ + struct sf_buf *sf = arg1; + struct sendfile_sync *sfs = arg2; + vm_page_t pg = sf_buf_page(sf); + + sf_buf_free(sf); + + vm_page_lock(pg); + if (vm_page_unwire(pg, PQ_NONE)) { + vm_object_t obj; + + /* Try to free the page, but only if it is cheap to. */ + if ((obj = pg->object) == NULL) + vm_page_free(pg); + else if (!vm_page_xbusied(pg) && VM_OBJECT_TRYWLOCK(obj)) { + vm_page_free(pg); + VM_OBJECT_WUNLOCK(obj); + } else + vm_page_deactivate(pg); + } + vm_page_unlock(pg); + + if (sfs != NULL) { + mtx_lock(&sfs->mtx); + KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0")); + if (--sfs->count == 0) + cv_signal(&sfs->cv); + mtx_unlock(&sfs->mtx); + } +} + /* * sendfile(2) * @@ -1974,103 +2001,252 @@ freebsd4_sendfile(struct thread *td, struct freebsd4_sendfile_args *uap) } #endif /* COMPAT_FREEBSD4 */ + /* + * How much data to put into page i of n. + * Only first and last pages are special. + */ +static inline off_t +xfsize(int i, int n, off_t off, off_t len) +{ + + if (i == 0) + return (omin(PAGE_SIZE - (off & PAGE_MASK), len)); + + if (i == n - 1 && ((off + len) & PAGE_MASK) > 0) + return ((off + len) & PAGE_MASK); + + return (PAGE_SIZE); +} + +/* + * Offset within object for i page. + */ +static inline vm_offset_t +vmoff(int i, off_t off) +{ + + if (i == 0) + return ((vm_offset_t)off); + + return (trunc_page(off + i * PAGE_SIZE)); +} + +/* + * Pretend as if we don't have enough space, subtract xfsize() of + * all pages that failed. + */ +static inline void +fixspace(int old, int new, off_t off, int *space) +{ + + KASSERT(old > new, ("%s: old %d new %d", __func__, old, new)); + + /* Subtract last one. */ + *space -= xfsize(old - 1, old, off, *space); + old--; + + if (new == old) + /* There was only one page. */ + return; + + /* Subtract first one. */ + if (new == 0) { + *space -= xfsize(0, old, off, *space); + new++; + } + + /* Rest of pages are full sized. */ + *space -= (old - new) * PAGE_SIZE; + + KASSERT(*space >= 0, ("%s: space went backwards", __func__)); +} + +/* + * Structure describing a single sendfile(2) I/O, which may consist of + * several underlying pager I/Os. + * + * The syscall context allocates the structure and initializes 'nios' + * to 1. As sendfile_swapin() runs through pages and starts asynchronous + * paging operations, it increments 'nios'. + * + * Every I/O completion calls sf_iodone(), which decrements the 'nios', and + * the syscall also calls sf_iodone() after allocating all mbufs, linking them + * and sending to socket. Whoever reaches zero 'nios' is responsible to + * call pru_ready on the socket, to notify it of readyness of the data. + */ +struct sf_io { + volatile u_int nios; + u_int error; + int npages; + struct file *sock_fp; + struct mbuf *m; + vm_page_t pa[]; +}; + +static void +sf_iodone(void *arg, vm_page_t *pg, int count, int error) +{ + struct sf_io *sfio = arg; + struct socket *so; + + for (int i = 0; i < count; i++) + vm_page_xunbusy(pg[i]); + + if (error) + sfio->error = error; + + if (!refcount_release(&sfio->nios)) + return; + + so = sfio->sock_fp->f_data; + + if (sfio->error) { + struct mbuf *m; + + /* + * I/O operation failed. The state of data in the socket + * is now inconsistent, and all what we can do is to tear + * it down. Protocol abort method would tear down protocol + * state, free all ready mbufs and detach not ready ones. + * We will free the mbufs corresponding to this I/O manually. + * + * The socket would be marked with EIO and made available + * for read, so that application receives EIO on next + * syscall and eventually closes the socket. + */ + so->so_proto->pr_usrreqs->pru_abort(so); + so->so_error = EIO; + + m = sfio->m; + for (int i = 0; i < sfio->npages; i++) + m = m_free(m); + } else { + CURVNET_SET(so->so_vnet); + (void )(so->so_proto->pr_usrreqs->pru_ready)(so, sfio->m, + sfio->npages); + CURVNET_RESTORE(); + } + + /* XXXGL: curthread */ + fdrop(sfio->sock_fp, curthread); + free(sfio, M_TEMP); +} + +/* + * Iterate through pages vector and request paging for non-valid pages. + */ static int -sendfile_readpage(vm_object_t obj, struct vnode *vp, int nd, - off_t off, int xfsize, int bsize, struct thread *td, vm_page_t *res) +sendfile_swapin(vm_object_t obj, struct sf_io *sfio, off_t off, off_t len, + int npages, int rhpages, int flags) { - vm_page_t m; - vm_pindex_t pindex; - ssize_t resid; - int error, readahead, rv; + vm_page_t *pa = sfio->pa; + int nios; - pindex = OFF_TO_IDX(off); - VM_OBJECT_WLOCK(obj); - m = vm_page_grab(obj, pindex, (vp != NULL ? VM_ALLOC_NOBUSY | - VM_ALLOC_IGN_SBUSY : 0) | VM_ALLOC_WIRED | VM_ALLOC_NORMAL); + nios = 0; + flags = (flags & SF_NODISKIO) ? VM_ALLOC_NOWAIT : 0; /* - * Check if page is valid for what we need, otherwise initiate I/O. - * - * The non-zero nd argument prevents disk I/O, instead we - * return the caller what he specified in nd. In particular, - * if we already turned some pages into mbufs, nd == EAGAIN - * and the main function send them the pages before we come - * here again and block. + * First grab all the pages and wire them. Note that we grab + * only required pages. Readahead pages are dealt with later. */ - if (m->valid != 0 && vm_page_is_valid(m, off & PAGE_MASK, xfsize)) { - if (vp == NULL) - vm_page_xunbusy(m); - VM_OBJECT_WUNLOCK(obj); - *res = m; - return (0); - } else if (nd != 0) { - if (vp == NULL) - vm_page_xunbusy(m); - error = nd; - goto free_page; + VM_OBJECT_WLOCK(obj); + for (int i = 0; i < npages; i++) { + pa[i] = vm_page_grab(obj, OFF_TO_IDX(vmoff(i, off)), + VM_ALLOC_WIRED | VM_ALLOC_NORMAL | flags); + if (pa[i] == NULL) { + npages = i; + rhpages = 0; + break; + } } - /* - * Get the page from backing store. - */ - error = 0; - if (vp != NULL) { - VM_OBJECT_WUNLOCK(obj); - readahead = sfreadahead * MAXBSIZE; + for (int i = 0; i < npages;) { + int j, a, count, rv; + + /* Skip valid pages. */ + if (vm_page_is_valid(pa[i], vmoff(i, off) & PAGE_MASK, + xfsize(i, npages, off, len))) { + vm_page_xunbusy(pa[i]); + SFSTAT_INC(sf_pages_valid); + i++; + continue; + } /* - * Use vn_rdwr() instead of the pager interface for - * the vnode, to allow the read-ahead. - * - * XXXMAC: Because we don't have fp->f_cred here, we - * pass in NOCRED. This is probably wrong, but is - * consistent with our original implementation. + * Now 'i' points to first invalid page, iterate further + * to make 'j' point at first valid after a bunch of + * invalid ones. */ - error = vn_rdwr(UIO_READ, vp, NULL, readahead, trunc_page(off), - UIO_NOCOPY, IO_NODELOCKED | IO_VMIO | ((readahead / - bsize) << IO_SEQSHIFT), td->td_ucred, NOCRED, &resid, td); - SFSTAT_INC(sf_iocnt); - VM_OBJECT_WLOCK(obj); - } else { - if (vm_pager_has_page(obj, pindex, NULL, NULL)) { - rv = vm_pager_get_pages(obj, &m, 1, NULL, NULL); - SFSTAT_INC(sf_iocnt); - if (rv != VM_PAGER_OK) { - vm_page_lock(m); - vm_page_free(m); - vm_page_unlock(m); - m = NULL; - error = EIO; + for (j = i + 1; j < npages; j++) + if (vm_page_is_valid(pa[j], vmoff(j, off) & PAGE_MASK, + xfsize(j, npages, off, len))) { + SFSTAT_INC(sf_pages_valid); + break; } - } else { - pmap_zero_page(m); - m->valid = VM_PAGE_BITS_ALL; - m->dirty = 0; + + /* + * Now we got region of invalid pages between 'i' and 'j'. + * Check that they belong to pager. They may not be there, + * which is a regular situation for shmem pager. For vnode + * pager this happens only in case of sparse file. + * + * Important feature of vm_pager_has_page() is the hint + * stored in 'a', about how many pages we can pagein after + * this page in a single I/O. + */ + while (!vm_pager_has_page(obj, OFF_TO_IDX(vmoff(i, off)), + NULL, &a) && i < j) { + pmap_zero_page(pa[i]); + pa[i]->valid = VM_PAGE_BITS_ALL; + pa[i]->dirty = 0; + vm_page_xunbusy(pa[i]); + i++; } - if (m != NULL) - vm_page_xunbusy(m); - } - if (error == 0) { - *res = m; - } else if (m != NULL) { -free_page: - vm_page_lock(m); - vm_page_unwire(m, PQ_INACTIVE); + if (i == j) + continue; /* - * See if anyone else might know about this page. If - * not and it is not valid, then free it. + * We want to pagein as many pages as possible, limited only + * by the 'a' hint and actual request. + * + * We should not pagein into already valid page, thus if + * 'j' didn't reach last page, trim by that page. + * + * When the pagein fulfils the request, also specify readahead. */ - if (m->wire_count == 0 && m->valid == 0 && !vm_page_busied(m)) - vm_page_free(m); - vm_page_unlock(m); - } - KASSERT(error != 0 || (m->wire_count > 0 && - vm_page_is_valid(m, off & PAGE_MASK, xfsize)), - ("wrong page state m %p off %#jx xfsize %d", m, (uintmax_t)off, - xfsize)); + if (j < npages) + a = min(a, j - i - 1); + count = min(a + 1, npages - i); + + refcount_acquire(&sfio->nios); + rv = vm_pager_get_pages_async(obj, pa + i, count, NULL, + i + count == npages ? &rhpages : NULL, + &sf_iodone, sfio); + KASSERT(rv == VM_PAGER_OK, ("%s: pager fail obj %p page %p", + __func__, obj, pa[i])); + + SFSTAT_INC(sf_iocnt); + SFSTAT_ADD(sf_pages_read, count); + if (i + count == npages) + SFSTAT_ADD(sf_rhpages_read, rhpages); + +#ifdef INVARIANTS + for (j = i; j < i + count && j < npages; j++) + KASSERT(pa[j] == vm_page_lookup(obj, + OFF_TO_IDX(vmoff(j, off))), + ("pa[j] %p lookup %p\n", pa[j], + vm_page_lookup(obj, OFF_TO_IDX(vmoff(j, off))))); +#endif + i += count; + nios++; + } + VM_OBJECT_WUNLOCK(obj); - return (error); + + if (nios == 0 && npages != 0) + SFSTAT_INC(sf_noiocnt); + + return (nios); } static int @@ -2178,42 +2354,37 @@ vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio, struct vnode *vp; struct vm_object *obj; struct socket *so; - struct mbuf *m; + struct mbuf *m, *mh, *mhtail; struct sf_buf *sf; - struct vm_page *pg; struct shmfd *shmfd; struct sendfile_sync *sfs; struct vattr va; - off_t off, xfsize, fsbytes, sbytes, rem, obj_size; - int error, bsize, nd, hdrlen, mnw; + off_t off, sbytes, rem, obj_size; + int error, softerr, bsize, hdrlen; - pg = NULL; obj = NULL; so = NULL; - m = NULL; + m = mh = NULL; sfs = NULL; - fsbytes = sbytes = 0; - hdrlen = mnw = 0; - rem = nbytes; - obj_size = 0; + sbytes = 0; + softerr = 0; error = sendfile_getobj(td, fp, &obj, &vp, &shmfd, &obj_size, &bsize); if (error != 0) return (error); - if (rem == 0) - rem = obj_size; error = kern_sendfile_getsock(td, sockfd, &sock_fp, &so); if (error != 0) goto out; - /* - * Do not wait on memory allocations but return ENOMEM for - * caller to retry later. - * XXX: Experimental. - */ - if (flags & SF_MNOWAIT) - mnw = 1; +#ifdef MAC + error = mac_socket_check_send(td->td_ucred, so); + if (error != 0) + goto out; +#endif + + SFSTAT_INC(sf_syscalls); + SFSTAT_ADD(sf_rhpages_requested, SF_READAHEAD(flags)); if (flags & SF_SYNC) { sfs = malloc(sizeof *sfs, M_TEMP, M_WAITOK | M_ZERO); @@ -2221,37 +2392,27 @@ vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio, cv_init(&sfs->cv, "sendfile"); } -#ifdef MAC - error = mac_socket_check_send(td->td_ucred, so); - if (error != 0) - goto out; -#endif - /* If headers are specified copy them into mbufs. */ - if (hdr_uio != NULL) { + if (hdr_uio != NULL && hdr_uio->uio_resid > 0) { hdr_uio->uio_td = td; hdr_uio->uio_rw = UIO_WRITE; - if (hdr_uio->uio_resid > 0) { - /* - * In FBSD < 5.0 the nbytes to send also included - * the header. If compat is specified subtract the - * header size from nbytes. - */ - if (kflags & SFK_COMPAT) { - if (nbytes > hdr_uio->uio_resid) - nbytes -= hdr_uio->uio_resid; - else - nbytes = 0; - } - m = m_uiotombuf(hdr_uio, (mnw ? M_NOWAIT : M_WAITOK), - 0, 0, 0); - if (m == NULL) { - error = mnw ? EAGAIN : ENOBUFS; - goto out; - } - hdrlen = m_length(m, NULL); + /* + * In FBSD < 5.0 the nbytes to send also included + * the header. If compat is specified subtract the + * header size from nbytes. + */ + if (kflags & SFK_COMPAT) { + if (nbytes > hdr_uio->uio_resid) + nbytes -= hdr_uio->uio_resid; + else + nbytes = 0; } - } + mh = m_uiotombuf(hdr_uio, M_WAITOK, 0, 0, 0); + hdrlen = m_length(mh, &mhtail); + } else + hdrlen = 0; + + rem = nbytes ? omin(nbytes, obj_size - offset) : obj_size - offset; /* * Protect against multiple writers to the socket. @@ -2272,21 +2433,13 @@ vn_sendfile(struct file *fp, int sockfd, struct uio *hdr_uio, * The outer loop checks the state and available space of the socket * and takes care of the overall progress. */ - for (off = offset; ; ) { + for (off = offset; rem > 0; ) { + struct sf_io *sfio; + vm_page_t *pa; struct mbuf *mtail; - int loopbytes; - int space; - int done; - - if ((nbytes != 0 && nbytes == fsbytes) || - (nbytes == 0 && obj_size == fsbytes)) - break; + int nios, space, npages, rhpages; mtail = NULL; - loopbytes = 0; - space = 0; - done = 0; - /* * Check the socket state for ongoing connection, * no errors and space in socket buffer. @@ -2362,49 +2515,58 @@ retry_space: VOP_UNLOCK(vp, 0); goto done; } - obj_size = va.va_size; + if (va.va_size != obj_size) { + if (nbytes == 0) + rem += va.va_size - obj_size; + else if (offset + nbytes > va.va_size) + rem -= (offset + nbytes - va.va_size); + obj_size = va.va_size; + } } + if (space > rem) + space = rem; + + npages = howmany(space + (off & PAGE_MASK), PAGE_SIZE); + + /* + * Calculate maximum allowed number of pages for readahead + * at this iteration. First, we allow readahead up to "rem". + * If application wants more, let it be, but there is no + * reason to go above MAXPHYS. Also check against "obj_size", + * since vm_pager_has_page() can hint beyond EOF. + */ + rhpages = howmany(rem + (off & PAGE_MASK), PAGE_SIZE) - npages; + rhpages += SF_READAHEAD(flags); + rhpages = min(howmany(MAXPHYS, PAGE_SIZE), rhpages); + rhpages = min(howmany(obj_size - trunc_page(off), PAGE_SIZE) - + npages, rhpages); + + sfio = malloc(sizeof(struct sf_io) + + npages * sizeof(vm_page_t), M_TEMP, M_WAITOK); + refcount_init(&sfio->nios, 1); + sfio->error = 0; + + nios = sendfile_swapin(obj, sfio, off, space, npages, rhpages, + flags); + /* * Loop and construct maximum sized mbuf chain to be bulk * dumped into socket buffer. */ - while (space > loopbytes) { - vm_offset_t pgoff; + pa = sfio->pa; + for (int i = 0; i < npages; i++) { struct mbuf *m0; /* - * Calculate the amount to transfer. - * Not to exceed a page, the EOF, - * or the passed in nbytes. + * If a page wasn't grabbed successfully, then + * trim the array. Can happen only with SF_NODISKIO. */ - pgoff = (vm_offset_t)(off & PAGE_MASK); - rem = obj_size - offset; - if (nbytes != 0) - rem = omin(rem, nbytes); - rem -= fsbytes + loopbytes; - xfsize = omin(PAGE_SIZE - pgoff, rem); - xfsize = omin(space - loopbytes, xfsize); - if (xfsize <= 0) { - done = 1; /* all data sent */ - break; - } - - /* - * Attempt to look up the page. Allocate - * if not found or wait and loop if busy. - */ - if (m != NULL) - nd = EAGAIN; /* send what we already got */ - else if ((flags & SF_NODISKIO) != 0) - nd = EBUSY; - else - nd = 0; - error = sendfile_readpage(obj, vp, nd, off, - xfsize, bsize, td, &pg); - if (error != 0) { - if (error == EAGAIN) - error = 0; /* not a real error */ + if (pa[i] == NULL) { + SFSTAT_INC(sf_busy); + fixspace(npages, i, off, &space); + npages = i; + softerr = EBUSY; break; } @@ -2417,56 +2579,59 @@ retry_space: * threads might exhaust the buffers and then * deadlock. */ - sf = sf_buf_alloc(pg, (mnw || m != NULL) ? SFB_NOWAIT : - SFB_CATCH); + sf = sf_buf_alloc(pa[i], + m != NULL ? SFB_NOWAIT : SFB_CATCH); if (sf == NULL) { SFSTAT_INC(sf_allocfail); - vm_page_lock(pg); - vm_page_unwire(pg, PQ_INACTIVE); - KASSERT(pg->object != NULL, - ("%s: object disappeared", __func__)); - vm_page_unlock(pg); + for (int j = i; j < npages; j++) { + vm_page_lock(pa[j]); + vm_page_unwire(pa[j], PQ_INACTIVE); + vm_page_unlock(pa[j]); + } if (m == NULL) - error = (mnw ? EAGAIN : EINTR); + softerr = ENOBUFS; + fixspace(npages, i, off, &space); + npages = i; break; } - /* - * Get an mbuf and set it up as having - * external storage. - */ - m0 = m_get((mnw ? M_NOWAIT : M_WAITOK), MT_DATA); - if (m0 == NULL) { - error = (mnw ? EAGAIN : ENOBUFS); - sf_ext_free(sf, NULL); - break; - } - /* - * Attach EXT_SFBUF external storage. - */ - m0->m_ext.ext_buf = (caddr_t )sf_buf_kva(sf); + m0 = m_get(M_WAITOK, MT_DATA); + m0->m_ext.ext_buf = (char *)sf_buf_kva(sf); m0->m_ext.ext_size = PAGE_SIZE; m0->m_ext.ext_arg1 = sf; m0->m_ext.ext_arg2 = sfs; - m0->m_ext.ext_type = EXT_SFBUF; + /* + * SF_NOCACHE sets the page as being freed upon send. + * However, we ignore it for the last page in 'space', + * if the page is truncated, and we got more data to + * send (rem > space), or if we have readahead + * configured (rhpages > 0). + */ + if ((flags & SF_NOCACHE) == 0 || + (i == npages - 1 && + ((off + space) & PAGE_MASK) && + (rem > space || rhpages > 0))) + m0->m_ext.ext_type = EXT_SFBUF; + else + m0->m_ext.ext_type = EXT_SFBUF_NOCACHE; m0->m_ext.ext_flags = 0; - m0->m_flags |= (M_EXT|M_RDONLY); - m0->m_data = (char *)sf_buf_kva(sf) + pgoff; - m0->m_len = xfsize; + m0->m_flags |= (M_EXT | M_RDONLY); + if (nios) + m0->m_flags |= M_NOTREADY; + m0->m_data = (char *)sf_buf_kva(sf) + + (vmoff(i, off) & PAGE_MASK); + m0->m_len = xfsize(i, npages, off, space); + + if (i == 0) + sfio->m = m0; /* Append to mbuf chain. */ if (mtail != NULL) mtail->m_next = m0; - else if (m != NULL) - m_last(m)->m_next = m0; else m = m0; mtail = m0; - /* Keep track of bits processed. */ - loopbytes += xfsize; - off += xfsize; - if (sfs != NULL) { mtx_lock(&sfs->mtx); sfs->count++; @@ -2477,49 +2642,60 @@ retry_space: if (vp != NULL) VOP_UNLOCK(vp, 0); + /* Keep track of bytes processed. */ + off += space; + rem -= space; + + /* Prepend header, if any. */ + if (hdrlen) { + mhtail->m_next = m; + m = mh; + mh = NULL; + } + + if (m == NULL) { + KASSERT(softerr, ("%s: m NULL, no error", __func__)); + error = softerr; + free(sfio, M_TEMP); + goto done; + } + /* Add the buffer chain to the socket buffer. */ - if (m != NULL) { - int mlen, err; + KASSERT(m_length(m, NULL) == space + hdrlen, + ("%s: mlen %u space %d hdrlen %d", + __func__, m_length(m, NULL), space, hdrlen)); - mlen = m_length(m, NULL); - SOCKBUF_LOCK(&so->so_snd); - if (so->so_snd.sb_state & SBS_CANTSENDMORE) { - error = EPIPE; - SOCKBUF_UNLOCK(&so->so_snd); - goto done; - } - SOCKBUF_UNLOCK(&so->so_snd); - CURVNET_SET(so->so_vnet); - /* Avoid error aliasing. */ - err = (*so->so_proto->pr_usrreqs->pru_send) - (so, 0, m, NULL, NULL, td); - CURVNET_RESTORE(); - if (err == 0) { - /* - * We need two counters to get the - * file offset and nbytes to send - * right: - * - sbytes contains the total amount - * of bytes sent, including headers. - * - fsbytes contains the total amount - * of bytes sent from the file. - */ - sbytes += mlen; - fsbytes += mlen; - if (hdrlen) { - fsbytes -= hdrlen; - hdrlen = 0; - } - } else if (error == 0) - error = err; - m = NULL; /* pru_send always consumes */ + CURVNET_SET(so->so_vnet); + if (nios == 0) { + /* + * If sendfile_swapin() didn't initiate any I/Os, + * which happens if all data is cached in VM, then + * we can send data right now without the + * PRUS_NOTREADY flag. + */ + free(sfio, M_TEMP); + error = (*so->so_proto->pr_usrreqs->pru_send) + (so, 0, m, NULL, NULL, td); + } else { + sfio->sock_fp = sock_fp; + sfio->npages = npages; + fhold(sock_fp); + error = (*so->so_proto->pr_usrreqs->pru_send) + (so, PRUS_NOTREADY, m, NULL, NULL, td); + sf_iodone(sfio, NULL, 0, 0); } + CURVNET_RESTORE(); - /* Quit outer loop on error or when we're done. */ - if (done) - break; - if (error != 0) + m = NULL; /* pru_send always consumes */ + if (error) goto done; + sbytes += space + hdrlen; + if (hdrlen) + hdrlen = 0; + if (softerr) { + error = softerr; + goto done; + } } /* @@ -2552,6 +2728,8 @@ out: fdrop(sock_fp, td); if (m) m_freem(m); + if (mh) + m_freem(mh); if (sfs != NULL) { mtx_lock(&sfs->mtx); diff --git a/sys/sys/mbuf.h b/sys/sys/mbuf.h index 8509a6ce49a2..91dd8958c087 100644 --- a/sys/sys/mbuf.h +++ b/sys/sys/mbuf.h @@ -343,12 +343,13 @@ struct mbuf { * External mbuf storage buffer types. */ #define EXT_CLUSTER 1 /* mbuf cluster */ -#define EXT_SFBUF 2 /* sendfile(2)'s sf_bufs */ +#define EXT_SFBUF 2 /* sendfile(2)'s sf_buf */ #define EXT_JUMBOP 3 /* jumbo cluster page sized */ #define EXT_JUMBO9 4 /* jumbo cluster 9216 bytes */ #define EXT_JUMBO16 5 /* jumbo cluster 16184 bytes */ #define EXT_PACKET 6 /* mbuf+cluster from packet zone */ #define EXT_MBUF 7 /* external mbuf reference (M_IOVEC) */ +#define EXT_SFBUF_NOCACHE 8 /* sendfile(2)'s sf_buf not to be cached */ #define EXT_VENDOR1 224 /* for vendor-internal use */ #define EXT_VENDOR2 225 /* for vendor-internal use */ @@ -397,6 +398,7 @@ struct mbuf { */ void sf_ext_ref(void *, void *); void sf_ext_free(void *, void *); +void sf_ext_free_nocache(void *, void *); /* * Flags indicating checksum, segmentation and other offload work to be diff --git a/sys/sys/sf_buf.h b/sys/sys/sf_buf.h index d3bb037c2445..b5970d95fd82 100644 --- a/sys/sys/sf_buf.h +++ b/sys/sys/sf_buf.h @@ -31,7 +31,14 @@ #define _SYS_SF_BUF_H_ struct sfstat { /* sendfile statistics */ + uint64_t sf_syscalls; /* times sendfile was called */ + uint64_t sf_noiocnt; /* times sendfile didn't require I/O */ uint64_t sf_iocnt; /* times sendfile had to do disk I/O */ + uint64_t sf_pages_read; /* pages read as part of a request */ + uint64_t sf_pages_valid; /* pages were valid for a request */ + uint64_t sf_rhpages_requested; /* readahead pages requested */ + uint64_t sf_rhpages_read; /* readahead pages read */ + uint64_t sf_busy; /* times aborted on a busy page */ uint64_t sf_allocfail; /* times sfbuf allocation failed */ uint64_t sf_allocwait; /* times sfbuf allocation had to wait */ }; diff --git a/sys/sys/socket.h b/sys/sys/socket.h index 18e2de10e101..ee621d9dd1e8 100644 --- a/sys/sys/socket.h +++ b/sys/sys/socket.h @@ -587,11 +587,14 @@ struct sf_hdtr { * Sendfile-specific flag(s) */ #define SF_NODISKIO 0x00000001 -#define SF_MNOWAIT 0x00000002 +#define SF_MNOWAIT 0x00000002 /* obsolete */ #define SF_SYNC 0x00000004 +#define SF_NOCACHE 0x00000010 +#define SF_FLAGS(rh, flags) (((rh) << 16) | (flags)) #ifdef _KERNEL #define SFK_COMPAT 0x00000001 +#define SF_READAHEAD(flags) ((flags) >> 16) #endif /* _KERNEL */ #endif /* __BSD_VISIBLE */ diff --git a/usr.bin/netstat/mbuf.c b/usr.bin/netstat/mbuf.c index baf7a6536399..8b95ddddd04b 100644 --- a/usr.bin/netstat/mbuf.c +++ b/usr.bin/netstat/mbuf.c @@ -326,13 +326,33 @@ mbpr(void *kvmd, u_long mbaddr) kread_counters) != 0) goto out; + xo_emit("{:sendfile-syscalls/%ju} {N:sendfile syscalls}\n", + (uintmax_t)sfstat.sf_syscalls); + xo_emit("{:sendfile-no-io/%ju} " + "{N:sendfile syscalls completed without I\\/O request}\n", + (uintmax_t)sfstat.sf_noiocnt); + xo_emit("{:sendfile-io-count/%ju} " + "{N:requests for I\\/O initiated by sendfile}\n", + (uintmax_t)sfstat.sf_iocnt); + xo_emit("{:sendfile-pages-sent/%ju} " + "{N:pages read by sendfile as part of a request}\n", + (uintmax_t)sfstat.sf_pages_read); + xo_emit("{:sendfile-pages-valid/%ju} " + "{N:pages were valid at time of a sendfile request}\n", + (uintmax_t)sfstat.sf_pages_valid); + xo_emit("{:sendfile-requested-readahead/%ju} " + "{N:pages were requested for read ahead by applications}\n", + (uintmax_t)sfstat.sf_rhpages_requested); + xo_emit("{:sendfile-readahead/%ju} " + "{N:pages were read ahead by sendfile}\n", + (uintmax_t)sfstat.sf_rhpages_read); + xo_emit("{:sendfile-busy-encounters/%ju} " + "{N:times sendfile encountered an already busy page}\n", + (uintmax_t)sfstat.sf_busy); xo_emit("{:sfbufs-alloc-failed/%ju} {N:requests for sfbufs denied}\n", (uintmax_t)sfstat.sf_allocfail); xo_emit("{:sfbufs-alloc-wait/%ju} {N:requests for sfbufs delayed}\n", (uintmax_t)sfstat.sf_allocwait); - xo_emit("{:sfbufs-io-count/%ju} " - "{N:requests for I\\/O initiated by sendfile}\n", - (uintmax_t)sfstat.sf_iocnt); out: xo_close_container("mbuf-statistics"); memstat_mtl_free(mtlp); -- cgit v1.2.3