src/afs/afs_osi_vm.c

   1 /*
   2  * Copyright 2000, International Business Machines Corporation and others.
   3  * All Rights Reserved.
   4  *
   5  * This software has been released under the terms of the IBM Public
   6  * License.  For details, see the LICENSE file in the top-level source
   7  * directory or online at http://www.openafs.org/dl/license10.html
   8  */
   9
  10 #include <afsconfig.h>
  11 #include "afs/param.h"
  12
  13
  14 #include "afs/sysincludes.h"    /* Standard vendor system headers */
  15 #include "afsincludes.h"        /* Afs-based standard headers */
  16 #include "afs/afs_stats.h"      /* afs statistics */
  17 #ifdef AFS_AIX_ENV
  18 #include <sys/adspace.h>        /* for vm_att(), vm_det() */
  19 #endif
  20
  21 int
  22 osi_Active(struct vcache *avc)
  23 {
  24     AFS_STATCNT(osi_Active);
  25 #if defined(AFS_AIX_ENV) || defined(AFS_SUN5_ENV) || (AFS_LINUX20_ENV) || defined(AFS_DARWIN_ENV) || defined(AFS_XBSD_ENV)
  26     if ((avc->opens > 0) || (avc->f.states & CMAPPED))
  27         return 1;               /* XXX: Warning, verify this XXX  */
  28 #elif defined(AFS_SGI_ENV)
  29     if ((avc->opens > 0) || AFS_VN_MAPPED(AFSTOV(avc)))
  30         return 1;
  31 #else
  32     if (avc->opens > 0 || (AFSTOV(avc)->v_flag & VTEXT))
  33         return (1);
  34 #endif
  35     return 0;
  36 }
  37
  38 /* this call, unlike osi_FlushText, is supposed to discard caches that may
  39    contain invalid information if a file is written remotely, but that may
  40    contain valid information that needs to be written back if the file is
  41    being written locally.  It doesn't subsume osi_FlushText, since the latter
  42    function may be needed to flush caches that are invalidated by local writes.
  43
  44    avc->pvnLock is already held, avc->lock is guaranteed not to be held (by
  45    us, of course).
  46 */
  47 void
  48 osi_FlushPages(struct vcache *avc, afs_ucred_t *credp)
  49 {
  50     afs_hyper_t origDV;
  51 #if defined(AFS_CACHE_BYPASS)
  52     /* The optimization to check DV under read lock below is identical a
  53      * change in CITI cache bypass work.  The problem CITI found in 1999
  54      * was that this code and background daemon doing prefetching competed
  55      * for the vcache entry shared lock.  It's not clear to me from the
  56      * tech report, but it looks like CITI fixed the general prefetch code
  57      * path as a bonus when experimenting on prefetch for cache bypass, see
  58      * citi-tr-01-3.
  59      */
  60 #endif
  61     if (vType(avc) == VDIR) {
  62         /* not applicable to directories; they're never mapped or stored in
  63          * pages */
  64         return;
  65     }
  66     ObtainReadLock(&avc->lock);
  67     /* If we've already purged this version, or if we're the ones
  68      * writing this version, don't flush it (could lose the
  69      * data we're writing). */
  70     if ((hcmp((avc->f.m.DataVersion), (avc->mapDV)) <= 0)
  71         || ((avc->execsOrWriters > 0) && afs_DirtyPages(avc))) {
  72         ReleaseReadLock(&avc->lock);
  73         return;
  74     }
  75     ReleaseReadLock(&avc->lock);
  76     ObtainWriteLock(&avc->lock, 10);
  77     /* Check again */
  78     if ((hcmp((avc->f.m.DataVersion), (avc->mapDV)) <= 0)
  79         || ((avc->execsOrWriters > 0) && afs_DirtyPages(avc))) {
  80         ReleaseWriteLock(&avc->lock);
  81         return;
  82     }
  83
  84     /* At this point, you might think that we can skip trying to flush pages
  85      * if mapDV is zero, since a file with a zero DV will not have any data in
  86      * it. However, some platforms (notably Linux 2.6.22+) will keep a page
  87      * full of zeroes around for an empty file. So play it safe and always
  88      * flush pages. */
  89
  90     AFS_STATCNT(osi_FlushPages);
  91     hset(origDV, avc->f.m.DataVersion);
  92     afs_Trace3(afs_iclSetp, CM_TRACE_FLUSHPAGES, ICL_TYPE_POINTER, avc,
  93                ICL_TYPE_INT32, origDV.low, ICL_TYPE_INT32, avc->f.m.Length);
  94
  95     ReleaseWriteLock(&avc->lock);
  96 #ifndef AFS_FBSD70_ENV
  97     AFS_GUNLOCK();
  98 #endif
  99     osi_VM_FlushPages(avc, credp);
 100 #ifndef AFS_FBSD70_ENV
 101     AFS_GLOCK();
 102 #endif
 103     ObtainWriteLock(&avc->lock, 88);
 104
 105     /* do this last, and to original version, since stores may occur
 106      * while executing above PUTPAGE call */
 107     hset(avc->mapDV, origDV);
 108     ReleaseWriteLock(&avc->lock);
 109 }
 110
 111 #ifdef  AFS_TEXT_ENV
 112
 113 /* This call is supposed to flush all caches that might be invalidated
 114  * by either a local write operation or a write operation done on
 115  * another client.  This call may be called repeatedly on the same
 116  * version of a file, even while a file is being written, so it
 117  * shouldn't do anything that would discard newly written data before
 118  * it is written to the file system. */
 119
 120 void
 121 osi_FlushText_really(struct vcache *vp)
 122 {
 123     afs_hyper_t fdv;            /* version before which we'll flush */
 124
 125     AFS_STATCNT(osi_FlushText);
 126     /* see if we've already flushed this data version */
 127     if (hcmp(vp->f.m.DataVersion, vp->flushDV) <= 0)
 128         return;
 129
 130     ObtainWriteLock(&afs_ftf, 317);
 131     hset(fdv, vp->f.m.DataVersion);
 132
 133     /* why this disgusting code below?
 134      *    xuntext, called by xrele, doesn't notice when it is called
 135      * with a freed text object.  Sun continually calls xrele or xuntext
 136      * without any locking, as long as VTEXT is set on the
 137      * corresponding vnode.
 138      *    But, if the text object is locked when you check the VTEXT
 139      * flag, several processes can wait in xuntext, waiting for the
 140      * text lock; when the second one finally enters xuntext's
 141      * critical region, the text object is already free, but the check
 142      * was already done by xuntext's caller.
 143      *    Even worse, it turns out that xalloc locks the text object
 144      * before reading or stating a file via the vnode layer.  Thus, we
 145      * could end up in getdcache, being asked to bring in a new
 146      * version of a file, but the corresponding text object could be
 147      * locked.  We can't flush the text object without causing
 148      * deadlock, so now we just don't try to lock the text object
 149      * unless it is guaranteed to work.  And we try to flush the text
 150      * when we need to a bit more often at the vnode layer.  Sun
 151      * really blew the vm-cache flushing interface.
 152      */
 153
 154 #if defined (AFS_HPUX_ENV)
 155     if (vp->v.v_flag & VTEXT) {
 156         xrele(vp);
 157
 158         if (vp->v.v_flag & VTEXT) {     /* still has a text object? */
 159             ReleaseWriteLock(&afs_ftf);
 160             return;
 161         }
 162     }
 163 #endif
 164
 165     /* next do the stuff that need not check for deadlock problems */
 166     mpurge(vp);
 167
 168     /* finally, record that we've done it */
 169     hset(vp->flushDV, fdv);
 170     ReleaseWriteLock(&afs_ftf);
 171
 172 }
 173 #endif /* AFS_TEXT_ENV */
 174
 175 /* ? is it moderately likely that there are dirty VM pages associated with
 176  * this vnode?
 177  *
 178  *  Prereqs:  avc must be write-locked
 179  *
 180  *  System Dependencies:  - *must* support each type of system for which
 181  *                          memory mapped files are supported, even if all
 182  *                          it does is return TRUE;
 183  *
 184  * NB:  this routine should err on the side of caution for ProcessFS to work
 185  *      correctly (or at least, not to introduce worse bugs than already exist)
 186  */
 187 #ifdef  notdef
 188 int
 189 osi_VMDirty_p(struct vcache *avc)
 190 {
 191     int dirtyPages;
 192
 193     if (avc->execsOrWriters <= 0)
 194         return 0;               /* can't be many dirty pages here, I guess */
 195
 196 #if defined (AFS_AIX32_ENV)
 197 #ifdef  notdef
 198     /* because of the level of hardware involvment with VM and all the
 199      * warnings about "This routine must be called at VMM interrupt
 200      * level", I thought it would be safest to disable interrupts while
 201      * looking at the software page fault table.  */
 202
 203     /* convert vm handle into index into array:  I think that stoinio is
 204      * always zero...  Look into this XXX  */
 205 #define VMHASH(handle) ( \
 206                         ( ((handle) & ~vmker.stoinio)  \
 207                          ^ ((((handle) & ~vmker.stoinio) & vmker.stoimask) << vmker.stoihash) \
 208                          ) & 0x000fffff)
 209
 210     if (avc->segid) {
 211         unsigned int pagef, pri, index, next;
 212
 213         index = VMHASH(avc->segid);
 214         if (scb_valid(index)) { /* could almost be an ASSERT */
 215
 216             pri = disable_ints();
 217             for (pagef = scb_sidlist(index); pagef >= 0; pagef = next) {
 218                 next = pft_sidfwd(pagef);
 219                 if (pft_modbit(pagef)) {        /* has page frame been modified? */
 220                     enable_ints(pri);
 221                     return 1;
 222                 }
 223             }
 224             enable_ints(pri);
 225         }
 226     }
 227 #undef VMHASH
 228 #endif
 229 #endif /* AFS_AIX32_ENV */
 230
 231 #if defined (AFS_SUN5_ENV)
 232     if (avc->f.states & CMAPPED) {
 233         struct page *pg;
 234         for (pg = AFSTOV(avc)->v_s.v_Pages; pg; pg = pg->p_vpnext) {
 235             if (pg->p_mod) {
 236                 return 1;
 237             }
 238         }
 239     }
 240 #endif
 241     return 0;
 242 }
 243 #endif /* notdef */
 244
 245
 246 /*
 247  * Solaris osi_ReleaseVM should not drop and re-obtain the vcache entry lock.
 248  * This leads to bad races when osi_ReleaseVM() is called from
 249  * afs_InvalidateAllSegments().
 250
 251  * We can do this because Solaris osi_VM_Truncate() doesn't care whether the
 252  * vcache entry lock is held or not.
 253  *
 254  * For other platforms, in some cases osi_VM_Truncate() doesn't care, but
 255  * there may be cases where it does care.  If so, it would be good to fix
 256  * them so they don't care.  Until then, we assume the worst.
 257  *
 258  * Locking:  the vcache entry lock is held.  It is dropped and re-obtained.
 259  */
 260 void
 261 osi_ReleaseVM(struct vcache *avc, afs_ucred_t *acred)
 262 {
 263 #ifdef  AFS_SUN5_ENV
 264     AFS_GUNLOCK();
 265     osi_VM_Truncate(avc, 0, acred);
 266     AFS_GLOCK();
 267 #else
 268     ReleaseWriteLock(&avc->lock);
 269     AFS_GUNLOCK();
 270     osi_VM_Truncate(avc, 0, acred);
 271     AFS_GLOCK();
 272     ObtainWriteLock(&avc->lock, 80);
 273 #endif
 274 }