void
pvn_write_done(page_t *plist, int flags)
{
int dfree = 0;
int pgrec = 0;
int pgout = 0;
int pgpgout = 0;
int anonpgout = 0;
int anonfree = 0;
int fspgout = 0;
int fsfree = 0;
int execpgout = 0;
int execfree = 0;
page_t *pp;
struct cpu *cpup;
struct vnode *vp = NULL; /* for probe */
uint_t ppattr;
kmutex_t *vphm = NULL;
ASSERT((flags & B_READ) == 0);
/*
* If we are about to start paging anyway, start freeing pages.
*/
if (write_free && freemem < lotsfree + pages_before_pager &&
(flags & B_ERROR) == 0) {
flags |= B_FREE;
}
/*
* Handle each page involved in the i/o operation.
*/
while (plist != NULL) {
pp = plist;
ASSERT(PAGE_LOCKED(pp) && page_iolock_assert(pp));
page_sub(&plist, pp);
/* Kernel probe support */
if (vp == NULL)
vp = pp->p_vnode;
if (((flags & B_ERROR) == 0) && IS_VMODSORT(vp)) {
/*
* Move page to the top of the v_page list.
* Skip pages modified during IO.
*/
vphm = page_vnode_mutex(vp);
mutex_enter(vphm);
if ((pp->p_vpnext != pp) && !hat_ismod(pp)) {
page_vpsub(&vp->v_pages, pp);
page_vpadd(&vp->v_pages, pp);
}
mutex_exit(vphm);
}
if (flags & B_ERROR) {
/*
* Write operation failed. We don't want
* to destroy (or free) the page unless B_FORCE
* is set. We set the mod bit again and release
* all locks on the page so that it will get written
* back again later when things are hopefully
* better again.
* If B_INVAL and B_FORCE is set we really have
* to destroy the page.
*/
if ((flags & (B_INVAL|B_FORCE)) == (B_INVAL|B_FORCE)) {
page_io_unlock(pp);
/*LINTED: constant in conditional context*/
VN_DISPOSE(pp, B_INVAL, 0, kcred);
} else {
hat_setmod_only(pp);
page_io_unlock(pp);
page_unlock(pp);
}
} else if (flags & B_INVAL) {
/*
* XXX - Failed writes with B_INVAL set are
* not handled appropriately.
*/
page_io_unlock(pp);
/*LINTED: constant in conditional context*/
VN_DISPOSE(pp, B_INVAL, 0, kcred);
} else if (flags & B_FREE ||!hat_page_is_mapped(pp)) {
/*
* Update statistics for pages being paged out
*/
if (pp->p_vnode) {
if (IS_SWAPFSVP(pp->p_vnode)) {
anonpgout++;
} else {
if (pp->p_vnode->v_flag & VVMEXEC) {
execpgout++;
} else {
fspgout++;
}
}
}
page_io_unlock(pp);
pgout = 1;
pgpgout++;
TRACE_1(TR_FAC_VM, TR_PAGE_WS_OUT,
"page_ws_out:pp %p", pp);
/*
* The page_struct_lock need not be acquired to
* examine "p_lckcnt" and "p_cowcnt" since we'll
* have an "exclusive" lock if the upgrade succeeds.
*/
if (page_tryupgrade(pp) &&
pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
/*
* Check if someone has reclaimed the
* page. If ref and mod are not set, no
* one is using it so we can free it.
* The rest of the system is careful
* to use the NOSYNC flag to unload
* translations set up for i/o w/o
* affecting ref and mod bits.
*
* Obtain a copy of the real hardware
* mod bit using hat_pagesync(pp, HAT_DONTZERO)
* to avoid having to flush the cache.
*/
ppattr = hat_pagesync(pp, HAT_SYNC_DONTZERO |
HAT_SYNC_STOPON_MOD);
ck_refmod:
if (!(ppattr & (P_REF | P_MOD))) {
if (hat_page_is_mapped(pp)) {
/*
* Doesn't look like the page
* was modified so now we
* really have to unload the
* translations. Meanwhile
* another CPU could've
* modified it so we have to
* check again. We don't loop
* forever here because now
* the translations are gone
* and no one can get a new one
* since we have the "exclusive"
* lock on the page.
*/
(void) hat_pageunload(pp,
HAT_FORCE_PGUNLOAD);
ppattr = hat_page_getattr(pp,
P_REF | P_MOD);
goto ck_refmod;
}
/*
* Update statistics for pages being
* freed
*/
if (pp->p_vnode) {
if (IS_SWAPFSVP(pp->p_vnode)) {
anonfree++;
} else {
if (pp->p_vnode->v_flag
& VVMEXEC) {
execfree++;
} else {
fsfree++;
}
}
}
/*LINTED: constant in conditional ctx*/
VN_DISPOSE(pp, B_FREE,
(flags & B_DONTNEED), kcred);
dfree++;
} else {
page_unlock(pp);
pgrec++;
TRACE_1(TR_FAC_VM, TR_PAGE_WS_FREE,
"page_ws_free:pp %p", pp);
}
} else {
/*
* Page is either `locked' in memory
* or was reclaimed and now has a
* "shared" lock, so release it.
*/
page_unlock(pp);
}
} else {
/*
* Neither B_FREE nor B_INVAL nor B_ERROR.
* Just release locks.
*/
page_io_unlock(pp);
page_unlock(pp);
}
}
CPU_STATS_ENTER_K();
cpup = CPU; /* get cpup now that CPU cannot change */
CPU_STATS_ADDQ(cpup, vm, dfree, dfree);
CPU_STATS_ADDQ(cpup, vm, pgrec, pgrec);
CPU_STATS_ADDQ(cpup, vm, pgout, pgout);
CPU_STATS_ADDQ(cpup, vm, pgpgout, pgpgout);
CPU_STATS_ADDQ(cpup, vm, anonpgout, anonpgout);
CPU_STATS_ADDQ(cpup, vm, anonfree, anonfree);
CPU_STATS_ADDQ(cpup, vm, fspgout, fspgout);
CPU_STATS_ADDQ(cpup, vm, fsfree, fsfree);
CPU_STATS_ADDQ(cpup, vm, execpgout, execpgout);
CPU_STATS_ADDQ(cpup, vm, execfree, execfree);
CPU_STATS_EXIT_K();
/* Kernel probe */
TNF_PROBE_4(pageout, "vm pageio io", /* CSTYLED */,
tnf_opaque, vnode, vp,
tnf_ulong, pages_pageout, pgpgout,
tnf_ulong, pages_freed, dfree,
tnf_ulong, pages_reclaimed, pgrec);
}
/*
* Flags are composed of {B_ASYNC, B_INVAL, B_FREE, B_DONTNEED, B_DELWRI,
* B_TRUNC, B_FORCE}. B_DELWRI indicates that this page is part of a kluster
* operation and is only to be considered if it doesn't involve any
* waiting here. B_TRUNC indicates that the file is being truncated
* and so no i/o needs to be done. B_FORCE indicates that the page
* must be destroyed so don't try wrting it out.
*
* The caller must ensure that the page is locked. Returns 1, if
* the page should be written back (the "iolock" is held in this
* case), or 0 if the page has been dealt with or has been
* unlocked.
*/
int
pvn_getdirty(page_t *pp, int flags)
{
ASSERT((flags & (B_INVAL | B_FREE)) ?
PAGE_EXCL(pp) : PAGE_SHARED(pp));
ASSERT(PP_ISFREE(pp) == 0);
/*
* If trying to invalidate or free a logically `locked' page,
* forget it. Don't need page_struct_lock to check p_lckcnt and
* p_cowcnt as the page is exclusively locked.
*/
if ((flags & (B_INVAL | B_FREE)) && !(flags & (B_TRUNC|B_FORCE)) &&
(pp->p_lckcnt != 0 || pp->p_cowcnt != 0)) {
page_unlock(pp);
return (0);
}
/*
* Now acquire the i/o lock so we can add it to the dirty
* list (if necessary). We avoid blocking on the i/o lock
* in the following cases:
*
* If B_DELWRI is set, which implies that this request is
* due to a klustering operartion.
*
* If this is an async (B_ASYNC) operation and we are not doing
* invalidation (B_INVAL) [The current i/o or fsflush will ensure
* that the the page is written out].
*/
if ((flags & B_DELWRI) || ((flags & (B_INVAL | B_ASYNC)) == B_ASYNC)) {
if (!page_io_trylock(pp)) {
page_unlock(pp);
return (0);
}
} else {
page_io_lock(pp);
}
/*
* If we want to free or invalidate the page then
* we need to unload it so that anyone who wants
* it will have to take a minor fault to get it.
* Otherwise, we're just writing the page back so we
* need to sync up the hardwre and software mod bit to
* detect any future modifications. We clear the
* software mod bit when we put the page on the dirty
* list.
*/
if (flags & (B_INVAL | B_FREE)) {
(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
} else {
(void) hat_pagesync(pp, HAT_SYNC_ZERORM);
}
if (!hat_ismod(pp) || (flags & B_TRUNC)) {
/*
* Don't need to add it to the
* list after all.
*/
page_io_unlock(pp);
if (flags & B_INVAL) {
/*LINTED: constant in conditional context*/
VN_DISPOSE(pp, B_INVAL, 0, kcred);
} else if (flags & B_FREE) {
/*LINTED: constant in conditional context*/
VN_DISPOSE(pp, B_FREE, (flags & B_DONTNEED), kcred);
} else {
/*
* This is advisory path for the callers
* of VOP_PUTPAGE() who prefer freeing the
* page _only_ if no one else is accessing it.
* E.g. segmap_release()
*
* The above hat_ismod() check is useless because:
* (1) we may not be holding SE_EXCL lock;
* (2) we've not unloaded _all_ translations
*
* Let page_release() do the heavy-lifting.
*/
(void) page_release(pp, 1);
}
return (0);
}
/*
* Page is dirty, get it ready for the write back
* and add page to the dirty list.
*/
hat_clrrefmod(pp);
/*
* If we're going to free the page when we're done
* then we can let others try to use it starting now.
* We'll detect the fact that they used it when the
* i/o is done and avoid freeing the page.
*/
if (flags & B_FREE)
page_downgrade(pp);
TRACE_1(TR_FAC_VM, TR_PVN_GETDIRTY, "pvn_getdirty:pp %p", pp);
return (1);
}
/*
* Scan page_t's and issue I/O's for modified pages.
*
* Also coalesces consecutive small sized free pages into the next larger
* pagesize. This costs a tiny bit of time in fsflush, but will reduce time
* spent scanning on later passes and for anybody allocating large pages.
*/
static void
fsflush_do_pages()
{
vnode_t *vp;
ulong_t pcount;
hrtime_t timer = gethrtime();
ulong_t releases = 0;
ulong_t nexamined = 0;
ulong_t nlocked = 0;
ulong_t nmodified = 0;
ulong_t ncoalesce = 0;
ulong_t cnt;
int mod;
int fspage = 1;
u_offset_t offset;
uint_t szc;
page_t *coal_page = NULL; /* 1st page in group to coalesce */
uint_t coal_szc = 0; /* size code, coal_page->p_szc */
uint_t coal_cnt = 0; /* count of pages seen */
static ulong_t nscan = 0;
static pgcnt_t last_total_pages = 0;
static page_t *pp = NULL;
/*
* Check to see if total_pages has changed.
*/
if (total_pages != last_total_pages) {
last_total_pages = total_pages;
nscan = (last_total_pages * (tune.t_fsflushr))/v.v_autoup;
}
if (pp == NULL)
pp = memsegs->pages;
pcount = 0;
while (pcount < nscan) {
/*
* move to the next page, skipping over large pages
* and issuing prefetches.
*/
if (pp->p_szc && fspage == 0) {
pfn_t pfn;
pfn = page_pptonum(pp);
cnt = page_get_pagecnt(pp->p_szc);
cnt -= pfn & (cnt - 1);
} else
cnt = 1;
pp = page_nextn(pp, cnt);
prefetch_page_r((void *)pp);
ASSERT(pp != NULL);
pcount += cnt;
/*
* Do a bunch of dirty tests (ie. no locking) to determine
* if we can quickly skip this page. These tests are repeated
* after acquiring the page lock.
*/
++nexamined;
if (PP_ISSWAP(pp)) {
fspage = 0;
coal_page = NULL;
continue;
}
/*
* skip free pages too, but try coalescing them into larger
* pagesizes
*/
if (PP_ISFREE(pp)) {
/*
* skip pages with a file system identity or that
* are already maximum size
*/
fspage = 0;
szc = pp->p_szc;
if (pp->p_vnode != NULL || szc == fsf_npgsz - 1) {
coal_page = NULL;
continue;
}
/*
* If not in a coalescing candidate page or the size
* codes are different, start a new candidate.
*/
if (coal_page == NULL || coal_szc != szc) {
/*
* page must be properly aligned
*/
if ((page_pptonum(pp) & fsf_mask[szc]) != 0) {
coal_page = NULL;
continue;
}
coal_page = pp;
coal_szc = szc;
coal_cnt = 1;
continue;
}
/*
* acceptable to add this to existing candidate page
*/
++coal_cnt;
if (coal_cnt < fsf_pgcnt[coal_szc])
continue;
/*
* We've got enough pages to coalesce, so do it.
* After promoting, we clear coal_page, so it will
* take another pass to promote this to an even
* larger page.
*/
++ncoalesce;
(void) page_promote_size(coal_page, coal_szc);
coal_page = NULL;
continue;
} else {
coal_page = NULL;
}
if (PP_ISKAS(pp) ||
PAGE_LOCKED(pp) ||
pp->p_lckcnt != 0 ||
pp->p_cowcnt != 0) {
fspage = 0;
continue;
}
/*
* Reject pages that can't be "exclusively" locked.
*/
if (!page_trylock(pp, SE_EXCL))
continue;
++nlocked;
/*
* After locking the page, redo the above checks.
* Since we locked the page, leave out the PAGE_LOCKED() test.
*/
vp = pp->p_vnode;
if (PP_ISSWAP(pp) ||
PP_ISFREE(pp) ||
vp == NULL ||
PP_ISKAS(pp) ||
(vp->v_flag & VISSWAP) != 0) {
page_unlock(pp);
fspage = 0;
continue;
}
if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
page_unlock(pp);
continue;
}
fspage = 1;
ASSERT(vp->v_type != VCHR);
/*
* Check the modified bit. Leaving the bit alone in hardware.
* It will be cleared if we do the putpage.
*/
if (IS_VMODSORT(vp))
mod = hat_ismod(pp);
else
mod = hat_pagesync(pp,
HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD) & P_MOD;
if (mod) {
++nmodified;
offset = pp->p_offset;
/*
* Hold the vnode before releasing the page lock
* to prevent it from being freed and re-used by
* some other thread.
*/
VN_HOLD(vp);
page_unlock(pp);
(void) VOP_PUTPAGE(vp, offset, PAGESIZE, B_ASYNC,
kcred, NULL);
VN_RELE(vp);
} else {
/*
* Catch any pages which should be on the cache list,
* but aren't yet.
*/
if (hat_page_is_mapped(pp) == 0) {
++releases;
(void) page_release(pp, 1);
} else {
page_unlock(pp);
}
}
}
/*
* maintain statistics
* reset every million wakeups, just to avoid overflow
*/
if (++fsf_cycles == 1000000) {
fsf_cycles = 0;
fsf_total.fsf_scan = 0;
fsf_total.fsf_examined = 0;
fsf_total.fsf_locked = 0;
fsf_total.fsf_modified = 0;
fsf_total.fsf_coalesce = 0;
fsf_total.fsf_time = 0;
fsf_total.fsf_releases = 0;
} else {
fsf_total.fsf_scan += fsf_recent.fsf_scan = nscan;
fsf_total.fsf_examined += fsf_recent.fsf_examined = nexamined;
fsf_total.fsf_locked += fsf_recent.fsf_locked = nlocked;
fsf_total.fsf_modified += fsf_recent.fsf_modified = nmodified;
fsf_total.fsf_coalesce += fsf_recent.fsf_coalesce = ncoalesce;
fsf_total.fsf_time += fsf_recent.fsf_time = gethrtime() - timer;
fsf_total.fsf_releases += fsf_recent.fsf_releases = releases;
}
}
