Date: Mon, 26 Apr 2010 13:48:23 +0000 (UTC) From: Lawrence Stewart <lstewart@FreeBSD.org> To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: svn commit: r207223 - in head: share/man/man9 sys/kern sys/sys Message-ID: <201004261348.o3QDmN3H000249@svn.freebsd.org>
next in thread | raw e-mail | index | archive | help
Author: lstewart Date: Mon Apr 26 13:48:22 2010 New Revision: 207223 URL: http://svn.freebsd.org/changeset/base/207223 Log: - Rework the underlying ALQ storage to be a circular buffer, which amongst other things allows variable length messages to be easily supported. - Extend KPI with alq_writen() and alq_getn() to support variable length messages, which is enabled at ALQ creation time depending on the arguments passed to alq_open(). Also add variants of alq_open() and alq_post() that accept a flags argument. The KPI is still fully backwards compatible and shouldn't require any change in ALQ consumers unless they wish to utilise the new features. - Introduce the ALQ_NOACTIVATE and ALQ_ORDERED flags to allow ALQ consumers to have more control over IO scheduling and resource acquisition respectively. - Strengthen invariants checking. - Document ALQ changes in ALQ(9) man page. Sponsored by: FreeBSD Foundation Reviewed by: gnn, jeff, rpaulo, rwatson MFC after: 1 month Modified: head/share/man/man9/alq.9 head/sys/kern/kern_alq.c head/sys/sys/alq.h Modified: head/share/man/man9/alq.9 ============================================================================== --- head/share/man/man9/alq.9 Mon Apr 26 12:03:55 2010 (r207222) +++ head/share/man/man9/alq.9 Mon Apr 26 13:48:22 2010 (r207223) @@ -1,7 +1,13 @@ .\" .\" Copyright (c) 2003 Hiten Pandya <hmp@FreeBSD.org> +.\" Copyright (c) 2009-2010 The FreeBSD Foundation .\" All rights reserved. .\" +.\" Portions of this software were developed at the Centre for Advanced +.\" Internet Architectures, Swinburne University of Technology, Melbourne, +.\" Australia by Lawrence Stewart under sponsorship from the FreeBSD +.\" Foundation. +.\" .\" Redistribution and use in source and binary forms, with or without .\" modification, are permitted provided that the following conditions .\" are met: @@ -25,21 +31,34 @@ .\" .\" $FreeBSD$ .\" -.Dd May 16, 2003 +.Dd April 26, 2010 .Dt ALQ 9 .Os .Sh NAME .Nm alq , +.Nm alq_open_flags , .Nm alq_open , +.Nm alq_writen , .Nm alq_write , .Nm alq_flush , .Nm alq_close , +.Nm alq_getn , .Nm alq_get , +.Nm alq_post_flags , .Nm alq_post .Nd Asynchronous Logging Queues .Sh SYNOPSIS .In sys/alq.h .Ft int +.Fo alq_open_flags +.Fa "struct alq **app" +.Fa "const char *file" +.Fa "struct ucred *cred" +.Fa "int cmode" +.Fa "int size" +.Fa "int flags" +.Fc +.Ft int .Fo alq_open .Fa "struct alq **app" .Fa "const char *file" @@ -49,19 +68,25 @@ .Fa "int count" .Fc .Ft int -.Fn alq_write "struct alq *alq" "void *data" "int waitok" +.Fn alq_writen "struct alq *alq" "void *data" "int len" "int flags" +.Ft int +.Fn alq_write "struct alq *alq" "void *data" "int flags" .Ft void .Fn alq_flush "struct alq *alq" .Ft void .Fn alq_close "struct alq *alq" .Ft struct ale * -.Fn alq_get "struct alq *alq" "int waitok" +.Fn alq_getn "struct alq *alq" "int len" "int flags" +.Ft struct ale * +.Fn alq_get "struct alq *alq" "int flags" +.Ft void +.Fn alq_post_flags "struct alq *alq" "struct ale *ale" "int flags" .Ft void .Fn alq_post "struct alq *alq" "struct ale *ale" .Sh DESCRIPTION The .Nm -facility provides an asynchronous fixed length recording +facility provides an asynchronous fixed or variable length recording mechanism, known as Asynchronous Logging Queues. It can record to any .Xr vnode 9 , @@ -81,26 +106,37 @@ is defined as which has the following members: .Bd -literal -offset indent struct ale { - struct ale *ae_next; /* Next Entry */ - char *ae_data; /* Entry buffer */ - int ae_flags; /* Entry flags */ + intptr_t ae_bytesused; /* # bytes written to ALE. */ + char *ae_data; /* Write ptr. */ + int ae_pad; /* Unused, compat. */ }; .Ed .Pp -The -.Va ae_flags -field is for internal use, clients of the +An .Nm -interface should not modify this field. -Behaviour is undefined if this field is modified. +can be created in either fixed or variable length mode. +A variable length +.Nm +accommodates writes of varying length using +.Fn alq_writen +and +.Fn alq_getn . +A fixed length +.Nm +accommodates a fixed number of writes using +.Fn alq_write +and +.Fn alq_get , +each of fixed size (set at queue creation time). +Fixed length mode is deprecated in favour of variable length mode. .Sh FUNCTIONS The -.Fn alq_open -function creates a new logging queue. +.Fn alq_open_flags +function creates a new variable length asynchronous logging queue. The .Fa file -argument is the name of the file to open for logging; if the file does not -yet exist, +argument is the name of the file to open for logging. +If the file does not yet exist, .Fn alq_open will attempt to create it. The @@ -112,33 +148,99 @@ as the requested creation mode, to be us Consumers of this API may wish to pass .Dv ALQ_DEFAULT_CMODE , a default creation mode suitable for most applications. -The argument +The .Fa cred -specifies the credentials to use when opening and performing I/O on the file. -The size of each entry in the queue is determined by -.Fa size . +argument specifies the credentials to use when opening and performing I/O on the file. The +.Fa size +argument sets the size (in bytes) of the underlying queue. +The ALQ_ORDERED flag may be passed in via +.Fa flags +to indicate that the ordering of writer threads waiting for a busy +.Nm +to free up resources should be preserved. +.Pp +The deprecated +.Fn alq_open +function is implemented as a wrapper around +.Fn alq_open_flags +to provide backwards compatibility to consumers that have not been updated to +utilise the newer +.Fn alq_open_flags +function. +It passes all arguments through to +.Fn alq_open_flags +untouched except for +.Fa size +and +.Fa count , +and sets +.Fa flags +to 0. +To create a variable length mode +.Nm , +the +.Fa size +argument should be set to the size (in bytes) of the underlying queue and the +.Fa count +argument should be set to 0. +To create a fixed length mode +.Nm , +the +.Fa size +argument should be set to the size (in bytes) of each write and the .Fa count -argument determines the number of items to be stored in the -asynchronous queue over an approximate period of a disk -write operation. +argument should be set to the number of +.Fa size +byte chunks to reserve capacity for. .Pp The -.Fn alq_write +.Fn alq_writen function writes +.Fa len +bytes from .Fa data -to the designated queue, +to the designated variable length mode queue .Fa alq . -In the event that -.Fn alq_write -could not write the entry immediately, and +If +.Fn alq_writen +could not write the entry immediately and .Dv ALQ_WAITOK -is passed to -.Fa waitok , -then +is set in +.Fa flags , +the function will be allowed to +.Xr msleep_spin 9 +with the +.Dq Li alqwnord +or +.Dq Li alqwnres +wait message. +A write will automatically schedule the queue +.Fa alq +to be flushed to disk. +This behaviour can be controlled by passing ALQ_NOACTIVATE via +.Fa flags +to indicate that the write should not schedule +.Fa alq +to be flushed to disk. +.Pp +The deprecated +.Fn alq_write +function is implemented as a wrapper around +.Fn alq_writen +to provide backwards compatibility to consumers that have not been updated to +utilise variable length mode queues. +The function will write +.Fa size +bytes of data (where +.Fa size +was specified at queue creation time) from the +.Fa data +buffer to the +.Fa alq . +Note that it is an error to call .Fn alq_write -will be allowed to -.Xr tsleep 9 . +on a variable length mode queue. .Pp The .Fn alq_flush @@ -146,61 +248,136 @@ function is used for flushing .Fa alq to the log medium that was passed to .Fn alq_open . +If +.Fa alq +has data to flush and is not already in the process of being flushed, the +function will block doing IO. +Otherwise, the function will return immediately. .Pp The .Fn alq_close -function will close the asynchronous logging queue, -.Fa alq , +function will close the asynchronous logging queue +.Fa alq and flush all pending write requests to the log medium. It will free all resources that were previously allocated. .Pp The -.Fn alq_get -function returns the next available asynchronous logging entry -from the queue, -.Fa alq . -This function leaves the queue in a locked state, until a subsequent +.Fn alq_getn +function returns an asynchronous log entry from +.Fa alq , +initialised to point at a buffer capable of receiving +.Fa len +bytes of data. +This function leaves +.Fa alq +in a locked state, until a subsequent .Fn alq_post +or +.Fn alq_post_flags call is made. -In the event that -.Fn alq_get -could not retrieve an entry immediately, it will -.Xr tsleep 9 +If +.Fn alq_getn +could not obtain +.Fa len +bytes of buffer immediately and +.Dv ALQ_WAITOK +is set in +.Fa flags , +the function will be allowed to +.Xr msleep_spin 9 with the -.Dq Li alqget +.Dq Li alqgnord +or +.Dq Li alqgnres wait message. +The caller can choose to write less than +.Fa len +bytes of data to the returned asynchronous log entry by setting the entry's +ae_bytesused field to the number of bytes actually written. +This must be done prior to calling +.Fn alq_post . .Pp -The -.Fn alq_post -function schedules the asynchronous logging entry, -.Fa ale , -which is retrieved using the +The deprecated .Fn alq_get -function, -for writing to the asynchronous logging queue, +function is implemented as a wrapper around +.Fn alq_getn +to provide backwards compatibility to consumers that have not been updated to +utilise variable length mode queues. +The asynchronous log entry returned will be initialised to point at a buffer +capable of receiving +.Fa size +bytes of data (where +.Fa size +was specified at queue creation time). +Note that it is an error to call +.Fn alq_get +on a variable length mode queue. +.Pp +The +.Fn alq_post_flags +function schedules the asynchronous log entry +.Fa ale +(obtained from +.Fn alq_getn +or +.Fn alq_get ) +for writing to .Fa alq . -This function leaves the queue, -.Fa alq , +The ALQ_NOACTIVATE flag may be passed in via +.Fa flags +to indicate that the queue should not be immediately scheduled to be flushed to +disk. +This function leaves +.Fa alq in an unlocked state. +.Pp +The +.Fn alq_post +function is implemented as a wrapper around +.Fn alq_post_flags +to provide backwards compatibility to consumers that have not been updated to +utilise the newer +.Fn alq_post_flags +function. +It simply passes all arguments through to +.Fn alq_post_flags +untouched, and sets +.Fa flags +to 0. .Sh IMPLEMENTATION NOTES The +.Fn alq_writen +and .Fn alq_write -function is a wrapper around the +functions both perform a +.Xr bcopy 3 +from the supplied +.Fa data +buffer into the underlying +.Nm +buffer. +Performance critical code paths may wish to consider using +.Fn alq_getn +(variable length queues) or +.Fn alq_get +(fixed length queues) to avoid the extra memory copy. Note that a queue +remains locked between calls to +.Fn alq_getn +or .Fn alq_get and .Fn alq_post -functions; by using these functions separately, a call -to -.Fn bcopy -can be avoided for performance critical code paths. +or +.Fn alq_post_flags , +so this method of writing to a queue is unsuitable for situations where the +time between calls may be substantial. .Sh LOCKING -Each asynchronous queue is protected by a spin mutex. +Each asynchronous logging queue is protected by a spin mutex. .Pp Functions -.Fn alq_flush , -.Fn alq_open +.Fn alq_flush and -.Fn alq_post +.Fn alq_open may attempt to acquire an internal sleep mutex, and should consequently not be used in contexts where sleeping is not allowed. @@ -214,32 +391,36 @@ if it fails to open or else it returns 0. .Pp The +.Fn alq_writen +and .Fn alq_write -function returns +functions return .Er EWOULDBLOCK if .Dv ALQ_NOWAIT -was provided as a value to -.Fa waitok -and either the queue is full, or when the system is shutting down. +was set in +.Fa flags +and either the queue is full or the system is shutting down. .Pp The +.Fn alq_getn +and .Fn alq_get -function returns -.Dv NULL , +functions return +.Dv NULL if .Dv ALQ_NOWAIT -was provided as a value to -.Fa waitok -and either the queue is full, or when the system is shutting down. +was set in +.Fa flags +and either the queue is full or the system is shutting down. .Pp NOTE: invalid arguments to non-void functions will result in undefined behaviour. .Sh SEE ALSO -.Xr syslog 3 , -.Xr kthread 9 , +.Xr kproc 9 , .Xr ktr 9 , -.Xr tsleep 9 , +.Xr msleep_spin 9 , +.Xr syslog 3 , .Xr vnode 9 .Sh HISTORY The @@ -250,7 +431,11 @@ Asynchronous Logging Queues (ALQ) facili The .Nm facility was written by -.An Jeffrey Roberson Aq jeff@FreeBSD.org . +.An Jeffrey Roberson Aq jeff@FreeBSD.org +and extended by +.An Lawrence Stewart Aq lstewart@freebsd.org . .Pp This manual page was written by -.An Hiten Pandya Aq hmp@FreeBSD.org . +.An Hiten Pandya Aq hmp@FreeBSD.org +and revised by +.An Lawrence Stewart Aq lstewart@freebsd.org . Modified: head/sys/kern/kern_alq.c ============================================================================== --- head/sys/kern/kern_alq.c Mon Apr 26 12:03:55 2010 (r207222) +++ head/sys/kern/kern_alq.c Mon Apr 26 13:48:22 2010 (r207223) @@ -55,16 +55,23 @@ __FBSDID("$FreeBSD$"); /* Async. Logging Queue */ struct alq { + char *aq_entbuf; /* Buffer for stored entries */ int aq_entmax; /* Max entries */ int aq_entlen; /* Entry length */ - char *aq_entbuf; /* Buffer for stored entries */ + int aq_freebytes; /* Bytes available in buffer */ + int aq_buflen; /* Total length of our buffer */ + int aq_writehead; /* Location for next write */ + int aq_writetail; /* Flush starts at this location */ + int aq_wrapearly; /* # bytes left blank at end of buf */ int aq_flags; /* Queue flags */ + int aq_waiters; /* Num threads waiting for resources + * NB: Used as a wait channel so must + * not be first field in the alq struct + */ + struct ale aq_getpost; /* ALE for use by get/post */ struct mtx aq_mtx; /* Queue lock */ struct vnode *aq_vp; /* Open vnode handle */ struct ucred *aq_cred; /* Credentials of the opening thread */ - struct ale *aq_first; /* First ent */ - struct ale *aq_entfree; /* First free ent */ - struct ale *aq_entvalid; /* First ent valid for writing */ LIST_ENTRY(alq) aq_act; /* List of active queues */ LIST_ENTRY(alq) aq_link; /* List of all queues */ }; @@ -73,10 +80,14 @@ struct alq { #define AQ_ACTIVE 0x0002 /* on the active list */ #define AQ_FLUSHING 0x0004 /* doing IO */ #define AQ_SHUTDOWN 0x0008 /* Queue no longer valid */ +#define AQ_ORDERED 0x0010 /* Queue enforces ordered writes */ +#define AQ_LEGACY 0x0020 /* Legacy queue (fixed length writes) */ #define ALQ_LOCK(alq) mtx_lock_spin(&(alq)->aq_mtx) #define ALQ_UNLOCK(alq) mtx_unlock_spin(&(alq)->aq_mtx) +#define HAS_PENDING_DATA(alq) ((alq)->aq_freebytes != (alq)->aq_buflen) + static MALLOC_DEFINE(M_ALD, "ALD", "ALD"); /* @@ -205,7 +216,7 @@ ald_daemon(void) needwakeup = alq_doio(alq); ALQ_UNLOCK(alq); if (needwakeup) - wakeup(alq); + wakeup_one(alq); ALD_LOCK(); } @@ -252,6 +263,20 @@ alq_shutdown(struct alq *alq) /* Stop any new writers. */ alq->aq_flags |= AQ_SHUTDOWN; + /* + * If the ALQ isn't active but has unwritten data (possible if + * the ALQ_NOACTIVATE flag has been used), explicitly activate the + * ALQ here so that the pending data gets flushed by the ald_daemon. + */ + if (!(alq->aq_flags & AQ_ACTIVE) && HAS_PENDING_DATA(alq)) { + alq->aq_flags |= AQ_ACTIVE; + ALQ_UNLOCK(alq); + ALD_LOCK(); + ald_activate(alq); + ALD_UNLOCK(); + ALQ_LOCK(alq); + } + /* Drain IO */ while (alq->aq_flags & AQ_ACTIVE) { alq->aq_flags |= AQ_WANTED; @@ -271,7 +296,6 @@ alq_destroy(struct alq *alq) alq_shutdown(alq); mtx_destroy(&alq->aq_mtx); - free(alq->aq_first, M_ALD); free(alq->aq_entbuf, M_ALD); free(alq, M_ALD); } @@ -287,46 +311,54 @@ alq_doio(struct alq *alq) struct vnode *vp; struct uio auio; struct iovec aiov[2]; - struct ale *ale; - struct ale *alstart; int totlen; int iov; int vfslocked; + int wrapearly; + + KASSERT((HAS_PENDING_DATA(alq)), ("%s: queue empty!", __func__)); vp = alq->aq_vp; td = curthread; totlen = 0; - iov = 0; - - alstart = ale = alq->aq_entvalid; - alq->aq_entvalid = NULL; + iov = 1; + wrapearly = alq->aq_wrapearly; bzero(&aiov, sizeof(aiov)); bzero(&auio, sizeof(auio)); - do { - if (aiov[iov].iov_base == NULL) - aiov[iov].iov_base = ale->ae_data; - aiov[iov].iov_len += alq->aq_entlen; - totlen += alq->aq_entlen; - /* Check to see if we're wrapping the buffer */ - if (ale->ae_data + alq->aq_entlen != ale->ae_next->ae_data) - iov++; - ale->ae_flags &= ~AE_VALID; - ale = ale->ae_next; - } while (ale->ae_flags & AE_VALID); + /* Start the write from the location of our buffer tail pointer. */ + aiov[0].iov_base = alq->aq_entbuf + alq->aq_writetail; + + if (alq->aq_writetail < alq->aq_writehead) { + /* Buffer not wrapped. */ + totlen = aiov[0].iov_len = alq->aq_writehead - alq->aq_writetail; + } else if (alq->aq_writehead == 0) { + /* Buffer not wrapped (special case to avoid an empty iov). */ + totlen = aiov[0].iov_len = alq->aq_buflen - alq->aq_writetail - + wrapearly; + } else { + /* + * Buffer wrapped, requires 2 aiov entries: + * - first is from writetail to end of buffer + * - second is from start of buffer to writehead + */ + aiov[0].iov_len = alq->aq_buflen - alq->aq_writetail - + wrapearly; + iov++; + aiov[1].iov_base = alq->aq_entbuf; + aiov[1].iov_len = alq->aq_writehead; + totlen = aiov[0].iov_len + aiov[1].iov_len; + } alq->aq_flags |= AQ_FLUSHING; ALQ_UNLOCK(alq); - if (iov == 2 || aiov[iov].iov_base == NULL) - iov--; - auio.uio_iov = &aiov[0]; auio.uio_offset = 0; auio.uio_segflg = UIO_SYSSPACE; auio.uio_rw = UIO_WRITE; - auio.uio_iovcnt = iov + 1; + auio.uio_iovcnt = iov; auio.uio_resid = totlen; auio.uio_td = td; @@ -350,8 +382,28 @@ alq_doio(struct alq *alq) ALQ_LOCK(alq); alq->aq_flags &= ~AQ_FLUSHING; - if (alq->aq_entfree == NULL) - alq->aq_entfree = alstart; + /* Adjust writetail as required, taking into account wrapping. */ + alq->aq_writetail = (alq->aq_writetail + totlen + wrapearly) % + alq->aq_buflen; + alq->aq_freebytes += totlen + wrapearly; + + /* + * If we just flushed part of the buffer which wrapped, reset the + * wrapearly indicator. + */ + if (wrapearly) + alq->aq_wrapearly = 0; + + /* + * If we just flushed the buffer completely, reset indexes to 0 to + * minimise buffer wraps. + * This is also required to ensure alq_getn() can't wedge itself. + */ + if (!HAS_PENDING_DATA(alq)) + alq->aq_writehead = alq->aq_writetail = 0; + + KASSERT((alq->aq_writetail >= 0 && alq->aq_writetail < alq->aq_buflen), + ("%s: aq_writetail < 0 || aq_writetail >= aq_buflen", __func__)); if (alq->aq_flags & AQ_WANTED) { alq->aq_flags &= ~AQ_WANTED; @@ -376,27 +428,27 @@ SYSINIT(ald, SI_SUB_LOCK, SI_ORDER_ANY, /* * Create the queue data structure, allocate the buffer, and open the file. */ + int -alq_open(struct alq **alqp, const char *file, struct ucred *cred, int cmode, - int size, int count) +alq_open_flags(struct alq **alqp, const char *file, struct ucred *cred, int cmode, + int size, int flags) { struct thread *td; struct nameidata nd; - struct ale *ale; - struct ale *alp; struct alq *alq; - char *bufp; - int flags; + int oflags; int error; - int i, vfslocked; + int vfslocked; + + KASSERT((size > 0), ("%s: size <= 0", __func__)); *alqp = NULL; td = curthread; NDINIT(&nd, LOOKUP, NOFOLLOW | MPSAFE, UIO_SYSSPACE, file, td); - flags = FWRITE | O_NOFOLLOW | O_CREAT; + oflags = FWRITE | O_NOFOLLOW | O_CREAT; - error = vn_open_cred(&nd, &flags, cmode, 0, cred, NULL); + error = vn_open_cred(&nd, &oflags, cmode, 0, cred, NULL); if (error) return (error); @@ -407,31 +459,20 @@ alq_open(struct alq **alqp, const char * VFS_UNLOCK_GIANT(vfslocked); alq = malloc(sizeof(*alq), M_ALD, M_WAITOK|M_ZERO); - alq->aq_entbuf = malloc(count * size, M_ALD, M_WAITOK|M_ZERO); - alq->aq_first = malloc(sizeof(*ale) * count, M_ALD, M_WAITOK|M_ZERO); alq->aq_vp = nd.ni_vp; alq->aq_cred = crhold(cred); - alq->aq_entmax = count; - alq->aq_entlen = size; - alq->aq_entfree = alq->aq_first; mtx_init(&alq->aq_mtx, "ALD Queue", NULL, MTX_SPIN|MTX_QUIET); - bufp = alq->aq_entbuf; - ale = alq->aq_first; - alp = NULL; - - /* Match up entries with buffers */ - for (i = 0; i < count; i++) { - if (alp) - alp->ae_next = ale; - ale->ae_data = bufp; - alp = ale; - ale++; - bufp += size; - } - - alp->ae_next = alq->aq_first; + alq->aq_buflen = size; + alq->aq_entmax = 0; + alq->aq_entlen = 0; + + alq->aq_freebytes = alq->aq_buflen; + alq->aq_entbuf = malloc(alq->aq_buflen, M_ALD, M_WAITOK|M_ZERO); + alq->aq_writehead = alq->aq_writetail = 0; + if (flags & ALQ_ORDERED) + alq->aq_flags |= AQ_ORDERED; if ((error = ald_add(alq)) != 0) { alq_destroy(alq); @@ -443,77 +484,405 @@ alq_open(struct alq **alqp, const char * return (0); } +int +alq_open(struct alq **alqp, const char *file, struct ucred *cred, int cmode, + int size, int count) +{ + int ret; + + KASSERT((count >= 0), ("%s: count < 0", __func__)); + + if (count > 0) { + ret = alq_open_flags(alqp, file, cred, cmode, size*count, 0); + (*alqp)->aq_flags |= AQ_LEGACY; + (*alqp)->aq_entmax = count; + (*alqp)->aq_entlen = size; + } else + ret = alq_open_flags(alqp, file, cred, cmode, size, 0); + + return (ret); +} + + /* * Copy a new entry into the queue. If the operation would block either * wait or return an error depending on the value of waitok. */ int -alq_write(struct alq *alq, void *data, int waitok) +alq_writen(struct alq *alq, void *data, int len, int flags) { - struct ale *ale; + int activate, copy, ret; + void *waitchan; + + KASSERT((len > 0 && len <= alq->aq_buflen), + ("%s: len <= 0 || len > aq_buflen", __func__)); - if ((ale = alq_get(alq, waitok)) == NULL) + activate = ret = 0; + copy = len; + waitchan = NULL; + + ALQ_LOCK(alq); + + /* + * Fail to perform the write and return EWOULDBLOCK if: + * - The message is larger than our underlying buffer. + * - The ALQ is being shutdown. + * - There is insufficient free space in our underlying buffer + * to accept the message and the user can't wait for space. + * - There is insufficient free space in our underlying buffer + * to accept the message and the alq is inactive due to prior + * use of the ALQ_NOACTIVATE flag (which would lead to deadlock). + */ + if (len > alq->aq_buflen || + alq->aq_flags & AQ_SHUTDOWN || + (((flags & ALQ_NOWAIT) || (!(alq->aq_flags & AQ_ACTIVE) && + HAS_PENDING_DATA(alq))) && alq->aq_freebytes < len)) { + ALQ_UNLOCK(alq); return (EWOULDBLOCK); + } - bcopy(data, ale->ae_data, alq->aq_entlen); - alq_post(alq, ale); + /* + * If we want ordered writes and there is already at least one thread + * waiting for resources to become available, sleep until we're woken. + */ + if (alq->aq_flags & AQ_ORDERED && alq->aq_waiters > 0) { + KASSERT(!(flags & ALQ_NOWAIT), + ("%s: ALQ_NOWAIT set but incorrectly ignored!", __func__)); + alq->aq_waiters++; + msleep_spin(&alq->aq_waiters, &alq->aq_mtx, "alqwnord", 0); + alq->aq_waiters--; + } - return (0); + /* + * (ALQ_WAITOK && aq_freebytes < len) or aq_freebytes >= len, either + * enter while loop and sleep until we have enough free bytes (former) + * or skip (latter). If AQ_ORDERED is set, only 1 thread at a time will + * be in this loop. Otherwise, multiple threads may be sleeping here + * competing for ALQ resources. + */ + while (alq->aq_freebytes < len && !(alq->aq_flags & AQ_SHUTDOWN)) { + KASSERT(!(flags & ALQ_NOWAIT), + ("%s: ALQ_NOWAIT set but incorrectly ignored!", __func__)); + alq->aq_flags |= AQ_WANTED; + alq->aq_waiters++; + if (waitchan) + wakeup(waitchan); + msleep_spin(alq, &alq->aq_mtx, "alqwnres", 0); + alq->aq_waiters--; + + /* + * If we're the first thread to wake after an AQ_WANTED wakeup + * but there isn't enough free space for us, we're going to loop + * and sleep again. If there are other threads waiting in this + * loop, schedule a wakeup so that they can see if the space + * they require is available. + */ + if (alq->aq_waiters > 0 && !(alq->aq_flags & AQ_ORDERED) && + alq->aq_freebytes < len && !(alq->aq_flags & AQ_WANTED)) + waitchan = alq; + else + waitchan = NULL; + } + + /* + * If there are waiters, we need to signal the waiting threads after we + * complete our work. The alq ptr is used as a wait channel for threads + * requiring resources to be freed up. In the AQ_ORDERED case, threads + * are not allowed to concurrently compete for resources in the above + * while loop, so we use a different wait channel in this case. + */ + if (alq->aq_waiters > 0) { + if (alq->aq_flags & AQ_ORDERED) + waitchan = &alq->aq_waiters; + else + waitchan = alq; + } else + waitchan = NULL; + + /* Bail if we're shutting down. */ + if (alq->aq_flags & AQ_SHUTDOWN) { + ret = EWOULDBLOCK; + goto unlock; + } + + /* + * If we need to wrap the buffer to accommodate the write, + * we'll need 2 calls to bcopy. + */ + if ((alq->aq_buflen - alq->aq_writehead) < len) + copy = alq->aq_buflen - alq->aq_writehead; + + /* Copy message (or part thereof if wrap required) to the buffer. */ + bcopy(data, alq->aq_entbuf + alq->aq_writehead, copy); + alq->aq_writehead += copy; + + if (alq->aq_writehead >= alq->aq_buflen) { + KASSERT((alq->aq_writehead == alq->aq_buflen), + ("%s: alq->aq_writehead (%d) > alq->aq_buflen (%d)", + __func__, + alq->aq_writehead, + alq->aq_buflen)); + alq->aq_writehead = 0; + } + + if (copy != len) { + /* + * Wrap the buffer by copying the remainder of our message + * to the start of the buffer and resetting aq_writehead. + */ + bcopy(((uint8_t *)data)+copy, alq->aq_entbuf, len - copy); + alq->aq_writehead = len - copy; + } + + KASSERT((alq->aq_writehead >= 0 && alq->aq_writehead < alq->aq_buflen), + ("%s: aq_writehead < 0 || aq_writehead >= aq_buflen", __func__)); + + alq->aq_freebytes -= len; + + if (!(alq->aq_flags & AQ_ACTIVE) && !(flags & ALQ_NOACTIVATE)) { + alq->aq_flags |= AQ_ACTIVE; + activate = 1; + } + + KASSERT((HAS_PENDING_DATA(alq)), ("%s: queue empty!", __func__)); + +unlock: + ALQ_UNLOCK(alq); + + if (activate) { + ALD_LOCK(); + ald_activate(alq); + ALD_UNLOCK(); + } + + /* NB: We rely on wakeup_one waking threads in a FIFO manner. */ + if (waitchan != NULL) + wakeup_one(waitchan); + + return (ret); +} + +int +alq_write(struct alq *alq, void *data, int flags) +{ + /* Should only be called in fixed length message (legacy) mode. */ + KASSERT((alq->aq_flags & AQ_LEGACY), + ("%s: fixed length write on variable length queue", __func__)); + return (alq_writen(alq, data, alq->aq_entlen, flags)); } +/* + * Retrieve a pointer for the ALQ to write directly into, avoiding bcopy. + */ struct ale * -alq_get(struct alq *alq, int waitok) +alq_getn(struct alq *alq, int len, int flags) { - struct ale *ale; - struct ale *aln; + int contigbytes; + void *waitchan; + + KASSERT((len > 0 && len <= alq->aq_buflen), + ("%s: len <= 0 || len > alq->aq_buflen", __func__)); - ale = NULL; + waitchan = NULL; ALQ_LOCK(alq); - /* Loop until we get an entry or we're shutting down */ - while ((alq->aq_flags & AQ_SHUTDOWN) == 0 && - (ale = alq->aq_entfree) == NULL && - (waitok & ALQ_WAITOK)) { + /* + * Determine the number of free contiguous bytes. + * We ensure elsewhere that if aq_writehead == aq_writetail because + * the buffer is empty, they will both be set to 0 and therefore + * aq_freebytes == aq_buflen and is fully contiguous. + * If they are equal and the buffer is not empty, aq_freebytes will + * be 0 indicating the buffer is full. + */ + if (alq->aq_writehead <= alq->aq_writetail) + contigbytes = alq->aq_freebytes; + else { + contigbytes = alq->aq_buflen - alq->aq_writehead; + + if (contigbytes < len) { + /* + * Insufficient space at end of buffer to handle a + * contiguous write. Wrap early if there's space at + * the beginning. This will leave a hole at the end + * of the buffer which we will have to skip over when + * flushing the buffer to disk. + */ + if (alq->aq_writetail >= len || flags & ALQ_WAITOK) { + /* Keep track of # bytes left blank. */ + alq->aq_wrapearly = contigbytes; + /* Do the wrap and adjust counters. */ + contigbytes = alq->aq_freebytes = + alq->aq_writetail; + alq->aq_writehead = 0; + } + } + } + + /* *** DIFF OUTPUT TRUNCATED AT 1000 LINES ***
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201004261348.o3QDmN3H000249>