* Client Lustre Page.
*
* Author: Nikita Danilov <nikita.danilov@sun.com>
+ * Author: Jinshan Xiong <jinshan.xiong@intel.com>
*/
#define DEBUG_SUBSYSTEM S_CLASS
#include <cl_object.h>
#include "cl_internal.h"
-static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
- int radix);
+static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg);
#ifdef LIBCFS_DEBUG
# define PASSERT(env, page, expr) \
*
* This function can be used to obtain initial reference to previously
* unreferenced cached object. It can be called only if concurrent page
- * reclamation is somehow prevented, e.g., by locking page radix-tree
- * (cl_object_header::hdr->coh_page_guard), or by keeping a lock on a VM page,
+ * reclamation is somehow prevented, e.g., by keeping a lock on a VM page,
* associated with \a page.
*
* Use with care! Not exported.
RETURN(NULL);
}
-/**
- * Returns a page with given index in the given object, or NULL if no page is
- * found. Acquires a reference on \a page.
- *
- * Locking: called under cl_object_header::coh_page_guard spin-lock.
- */
-struct cl_page *cl_page_lookup(struct cl_object_header *hdr, pgoff_t index)
-{
- struct cl_page *page;
-
- LASSERT(spin_is_locked(&hdr->coh_page_guard));
-
- page = radix_tree_lookup(&hdr->coh_tree, index);
- if (page != NULL)
- cl_page_get_trust(page);
- return page;
-}
-EXPORT_SYMBOL(cl_page_lookup);
-
-/**
- * Returns a list of pages by a given [start, end] of \a obj.
- *
- * \param resched If not NULL, then we give up before hogging CPU for too
- * long and set *resched = 1, in that case caller should implement a retry
- * logic.
- *
- * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
- * crucial in the face of [offset, EOF] locks.
- *
- * Return at least one page in @queue unless there is no covered page.
- */
-int cl_page_gang_lookup(const struct lu_env *env, struct cl_object *obj,
- struct cl_io *io, pgoff_t start, pgoff_t end,
- cl_page_gang_cb_t cb, void *cbdata)
-{
- struct cl_object_header *hdr;
- struct cl_page *page;
- struct cl_page **pvec;
- const struct cl_page_slice *slice;
- const struct lu_device_type *dtype;
- pgoff_t idx;
- unsigned int nr;
- unsigned int i;
- unsigned int j;
- int res = CLP_GANG_OKAY;
- int tree_lock = 1;
- ENTRY;
-
- idx = start;
- hdr = cl_object_header(obj);
- pvec = cl_env_info(env)->clt_pvec;
- dtype = cl_object_top(obj)->co_lu.lo_dev->ld_type;
- spin_lock(&hdr->coh_page_guard);
- while ((nr = radix_tree_gang_lookup(&hdr->coh_tree, (void **)pvec,
- idx, CLT_PVEC_SIZE)) > 0) {
- int end_of_region = 0;
- idx = pvec[nr - 1]->cp_index + 1;
- for (i = 0, j = 0; i < nr; ++i) {
- page = pvec[i];
- pvec[i] = NULL;
-
- LASSERT(page->cp_type == CPT_CACHEABLE);
- if (page->cp_index > end) {
- end_of_region = 1;
- break;
- }
- if (page->cp_state == CPS_FREEING)
- continue;
-
- slice = cl_page_at_trusted(page, dtype);
- /*
- * Pages for lsm-less file has no underneath sub-page
- * for osc, in case of ...
- */
- PASSERT(env, page, slice != NULL);
-
- page = slice->cpl_page;
- /*
- * Can safely call cl_page_get_trust() under
- * radix-tree spin-lock.
- *
- * XXX not true, because @page is from object another
- * than @hdr and protected by different tree lock.
- */
- cl_page_get_trust(page);
- lu_ref_add_atomic(&page->cp_reference,
- "gang_lookup", current);
- pvec[j++] = page;
- }
-
- /*
- * Here a delicate locking dance is performed. Current thread
- * holds a reference to a page, but has to own it before it
- * can be placed into queue. Owning implies waiting, so
- * radix-tree lock is to be released. After a wait one has to
- * check that pages weren't truncated (cl_page_own() returns
- * error in the latter case).
- */
- spin_unlock(&hdr->coh_page_guard);
- tree_lock = 0;
-
- for (i = 0; i < j; ++i) {
- page = pvec[i];
- if (res == CLP_GANG_OKAY)
- res = (*cb)(env, io, page, cbdata);
- lu_ref_del(&page->cp_reference,
- "gang_lookup", current);
- cl_page_put(env, page);
- }
- if (nr < CLT_PVEC_SIZE || end_of_region)
- break;
-
- if (res == CLP_GANG_OKAY && need_resched())
- res = CLP_GANG_RESCHED;
- if (res != CLP_GANG_OKAY)
- break;
-
- spin_lock(&hdr->coh_page_guard);
- tree_lock = 1;
- }
- if (tree_lock)
- spin_unlock(&hdr->coh_page_guard);
- RETURN(res);
-}
-EXPORT_SYMBOL(cl_page_gang_lookup);
-
static void cl_page_free(const struct lu_env *env, struct cl_page *page)
{
struct cl_object *obj = page->cp_obj;
*(enum cl_page_state *)&page->cp_state = state;
}
-static struct cl_page *cl_page_alloc(const struct lu_env *env,
+struct cl_page *cl_page_alloc(const struct lu_env *env,
struct cl_object *o, pgoff_t ind, struct page *vmpage,
enum cl_page_type type)
{
if (page != NULL) {
int result = 0;
cfs_atomic_set(&page->cp_ref, 1);
- if (type == CPT_CACHEABLE) /* for radix tree */
- cfs_atomic_inc(&page->cp_ref);
page->cp_obj = o;
cl_object_get(o);
lu_object_ref_add_at(&o->co_lu, &page->cp_obj_ref, "cl_page",
result = o->co_ops->coo_page_init(env, o,
page, vmpage);
if (result != 0) {
- cl_page_delete0(env, page, 0);
+ cl_page_delete0(env, page);
cl_page_free(env, page);
page = ERR_PTR(result);
break;
}
RETURN(page);
}
+EXPORT_SYMBOL(cl_page_alloc);
/**
* Returns a cl_page with index \a idx at the object \a o, and associated with
*
* \see cl_object_find(), cl_lock_find()
*/
-static struct cl_page *cl_page_find0(const struct lu_env *env,
- struct cl_object *o,
- pgoff_t idx, struct page *vmpage,
- enum cl_page_type type,
- struct cl_page *parent)
+struct cl_page *cl_page_find(const struct lu_env *env,
+ struct cl_object *o,
+ pgoff_t idx, struct page *vmpage,
+ enum cl_page_type type)
{
struct cl_page *page = NULL;
- struct cl_page *ghost = NULL;
struct cl_object_header *hdr;
- int err;
LASSERT(type == CPT_CACHEABLE || type == CPT_TRANSIENT);
might_sleep();
* reference on it.
*/
page = cl_vmpage_page(vmpage, o);
- PINVRNT(env, page,
- ergo(page != NULL,
- cl_page_vmpage(env, page) == vmpage &&
- (void *)radix_tree_lookup(&hdr->coh_tree,
- idx) == page));
- }
-
- if (page != NULL) {
- CS_PAGE_INC(o, hit);
- RETURN(page);
+ if (page != NULL) {
+ CS_PAGE_INC(o, hit);
+ RETURN(page);
+ }
}
/* allocate and initialize cl_page */
page = cl_page_alloc(env, o, idx, vmpage, type);
- if (IS_ERR(page))
- RETURN(page);
-
- if (type == CPT_TRANSIENT) {
- if (parent) {
- LASSERT(page->cp_parent == NULL);
- page->cp_parent = parent;
- parent->cp_child = page;
- }
- RETURN(page);
- }
-
- /*
- * XXX optimization: use radix_tree_preload() here, and change tree
- * gfp mask to GFP_KERNEL in cl_object_header_init().
- */
- spin_lock(&hdr->coh_page_guard);
- err = radix_tree_insert(&hdr->coh_tree, idx, page);
- if (err != 0) {
- ghost = page;
- /*
- * Noted by Jay: a lock on \a vmpage protects cl_page_find()
- * from this race, but
- *
- * 0. it's better to have cl_page interface "locally
- * consistent" so that its correctness can be reasoned
- * about without appealing to the (obscure world of) VM
- * locking.
- *
- * 1. handling this race allows ->coh_tree to remain
- * consistent even when VM locking is somehow busted,
- * which is very useful during diagnosing and debugging.
- */
- page = ERR_PTR(err);
- CL_PAGE_DEBUG(D_ERROR, env, ghost,
- "fail to insert into radix tree: %d\n", err);
- } else {
- if (parent) {
- LASSERT(page->cp_parent == NULL);
- page->cp_parent = parent;
- parent->cp_child = page;
- }
- hdr->coh_pages++;
- }
- spin_unlock(&hdr->coh_page_guard);
-
- if (unlikely(ghost != NULL)) {
- cl_page_delete0(env, ghost, 0);
- cl_page_free(env, ghost);
- }
- RETURN(page);
-}
-
-struct cl_page *cl_page_find(const struct lu_env *env, struct cl_object *o,
- pgoff_t idx, struct page *vmpage,
- enum cl_page_type type)
-{
- return cl_page_find0(env, o, idx, vmpage, type, NULL);
+ RETURN(page);
}
EXPORT_SYMBOL(cl_page_find);
-
-struct cl_page *cl_page_find_sub(const struct lu_env *env, struct cl_object *o,
- pgoff_t idx, struct page *vmpage,
- struct cl_page *parent)
-{
- return cl_page_find0(env, o, idx, vmpage, parent->cp_type, parent);
-}
-EXPORT_SYMBOL(cl_page_find_sub);
-
static inline int cl_page_invariant(const struct cl_page *pg)
{
- struct cl_object_header *header;
struct cl_page *parent;
struct cl_page *child;
struct cl_io *owner;
*/
LINVRNT(cl_page_is_vmlocked(NULL, pg));
- header = cl_object_header(pg->cp_obj);
parent = pg->cp_parent;
child = pg->cp_child;
owner = pg->cp_owner;
ergo(owner != NULL && parent != NULL,
parent->cp_owner == pg->cp_owner->ci_parent) &&
ergo(owner != NULL && child != NULL,
- child->cp_owner->ci_parent == owner) &&
- /*
- * Either page is early in initialization (has neither child
- * nor parent yet), or it is in the object radix tree.
- */
- ergo(pg->cp_state < CPS_FREEING && pg->cp_type == CPT_CACHEABLE,
- (void *)radix_tree_lookup(&header->coh_tree,
- pg->cp_index) == pg ||
- (child == NULL && parent == NULL));
+ child->cp_owner->ci_parent == owner);
}
static void cl_page_state_set0(const struct lu_env *env,
* pages, e.g,. in a error handling cl_page_find()->cl_page_delete0()
* path. Doesn't check page invariant.
*/
-static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg,
- int radix)
+static void cl_page_delete0(const struct lu_env *env, struct cl_page *pg)
{
- struct cl_page *tmp = pg;
ENTRY;
PASSERT(env, pg, pg == cl_page_top(pg));
*/
cl_page_owner_clear(pg);
- /*
- * unexport the page firstly before freeing it so that
- * the page content is considered to be invalid.
- * We have to do this because a CPS_FREEING cl_page may
- * be NOT under the protection of a cl_lock.
- * Afterwards, if this page is found by other threads, then this
- * page will be forced to reread.
- */
- cl_page_export(env, pg, 0);
cl_page_state_set0(env, pg, CPS_FREEING);
- CL_PAGE_INVOID(env, pg, CL_PAGE_OP(cpo_delete),
+ CL_PAGE_INVOID_REVERSE(env, pg, CL_PAGE_OP(cpo_delete),
(const struct lu_env *, const struct cl_page_slice *));
- if (tmp->cp_type == CPT_CACHEABLE) {
- if (!radix)
- /* !radix means that @pg is not yet in the radix tree,
- * skip removing it.
- */
- tmp = pg->cp_child;
- for (; tmp != NULL; tmp = tmp->cp_child) {
- void *value;
- struct cl_object_header *hdr;
-
- hdr = cl_object_header(tmp->cp_obj);
- spin_lock(&hdr->coh_page_guard);
- value = radix_tree_delete(&hdr->coh_tree,
- tmp->cp_index);
- PASSERT(env, tmp, value == tmp);
- PASSERT(env, tmp, hdr->coh_pages > 0);
- hdr->coh_pages--;
- spin_unlock(&hdr->coh_page_guard);
- cl_page_put(env, tmp);
- }
- }
-
EXIT;
}
*/
void cl_page_delete(const struct lu_env *env, struct cl_page *pg)
{
- PINVRNT(env, pg, cl_page_invariant(pg));
- ENTRY;
- cl_page_delete0(env, pg, 1);
- EXIT;
+ PINVRNT(env, pg, cl_page_invariant(pg));
+ ENTRY;
+ cl_page_delete0(env, pg);
+ EXIT;
}
EXPORT_SYMBOL(cl_page_delete);
/**
- * Unmaps page from user virtual memory.
- *
- * Calls cl_page_operations::cpo_unmap() through all layers top-to-bottom. The
- * layer responsible for VM interaction has to unmap page from user space
- * virtual memory.
- *
- * \see cl_page_operations::cpo_unmap()
- */
-int cl_page_unmap(const struct lu_env *env,
- struct cl_io *io, struct cl_page *pg)
-{
- PINVRNT(env, pg, cl_page_is_owned(pg, io));
- PINVRNT(env, pg, cl_page_invariant(pg));
-
- return cl_page_invoke(env, io, pg, CL_PAGE_OP(cpo_unmap));
-}
-EXPORT_SYMBOL(cl_page_unmap);
-
-/**
* Marks page up-to-date.
*
* Call cl_page_operations::cpo_export() through all layers top-to-bottom. The
}
EXPORT_SYMBOL(cl_page_is_under_lock);
-static int page_prune_cb(const struct lu_env *env, struct cl_io *io,
- struct cl_page *page, void *cbdata)
-{
- cl_page_own(env, io, page);
- cl_page_unmap(env, io, page);
- cl_page_discard(env, io, page);
- cl_page_disown(env, io, page);
- return CLP_GANG_OKAY;
-}
-
-/**
- * Purges all cached pages belonging to the object \a obj.
- */
-int cl_pages_prune(const struct lu_env *env, struct cl_object *clobj)
-{
- struct cl_thread_info *info;
- struct cl_object *obj = cl_object_top(clobj);
- struct cl_io *io;
- int result;
-
- ENTRY;
- info = cl_env_info(env);
- io = &info->clt_io;
-
- /*
- * initialize the io. This is ugly since we never do IO in this
- * function, we just make cl_page_list functions happy. -jay
- */
- io->ci_obj = obj;
- io->ci_ignore_layout = 1;
- result = cl_io_init(env, io, CIT_MISC, obj);
- if (result != 0) {
- cl_io_fini(env, io);
- RETURN(io->ci_result);
- }
-
- do {
- result = cl_page_gang_lookup(env, obj, io, 0, CL_PAGE_EOF,
- page_prune_cb, NULL);
- if (result == CLP_GANG_RESCHED)
- cond_resched();
- } while (result != CLP_GANG_OKAY);
-
- cl_io_fini(env, io);
- RETURN(result);
-}
-EXPORT_SYMBOL(cl_pages_prune);
-
/**
* Tells transfer engine that only part of a page is to be transmitted.
*
git://git.whamcloud.com / fs / lustre-release.git / blobdiff