#include <linux/module.h>
#include <linux/pagemap.h> // XXX kill me soon
#include <linux/version.h>
+#include <asm/div64.h>
#include <linux/obd_class.h>
#include <linux/lustre_fsfilt.h>
return lnb->rc;
}
+/* See if there are unallocated parts in given file region */
+static int filter_inode_has_holes(struct inode *inode, obd_size start,
+ int len)
+{
+ int j;
+#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
+ sector_t (*fs_bmap)(struct address_space *,
+ sector_t);
+#else
+ int (*fs_bmap)(struct address_space *, long);
+#endif
+ fs_bmap = inode->i_mapping->a_ops->bmap;
+ if (fs_bmap) {
+ for (j = 0; j <= len ; j++) {
+ if (!fs_bmap(inode->i_mapping, start+j)) {
+ return 1;
+ }
+ }
+ return 0;
+ } else {
+ /* Return -1 in case that caller cares about bmap availability.
+ */
+ return -1;
+ }
+}
+
+/* Grab the dirty and seen grant announcements from the incoming obdo.
+ * We will later calculate the clients new grant and return it. */
+static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa)
+{
+ struct filter_export_data *fed;
+ struct obd_device *obd = exp->exp_obd;
+ obd_size client_cached;
+ ENTRY;
+
+ if (!oa || (oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) !=
+ (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) {
+ if (oa)
+ oa->o_valid &= ~OBD_MD_FLGRANT;
+ EXIT;
+ return;
+ }
+
+ client_cached = oa->o_blocks;
+ fed = &exp->exp_filter_data;
+
+ if (client_cached > fed->fed_grant)
+ CERROR("client %s claims "LPU64" granted, > "LPU64" granted\n",
+ obd->obd_name, client_cached, fed->fed_grant);
+
+ spin_lock(&obd->obd_osfs_lock);
+ /* update our accounting now so that statfs takes it into account */
+ obd->u.filter.fo_tot_cached += client_cached - fed->fed_cached;
+ fed->fed_cached = client_cached;
+
+ /* Acknowledgement that the client has seen our published grant.
+ * If the client has met our shrinking target we can reuse its
+ * difference from the previous grant. It is reasonable to announce
+ * more dirty that cached as it tries to purge its previously granted
+ * dirty data down to its newly received target. */
+ if (fed->fed_grant_waiting && (oa->o_grant <= fed->fed_grant_sent)) {
+ if (fed->fed_grant_sent < fed->fed_grant) {
+ if (client_cached <= fed->fed_grant_sent) {
+ obd->u.filter.fo_tot_granted -=
+ fed->fed_grant - oa->o_grant;
+ CDEBUG(D_SUPER, "reduced grant from "LPU64" to "
+ LPU64", total grant now "LPU64"\n",
+ fed->fed_grant, oa->o_grant,
+ obd->u.filter.fo_tot_granted);
+ fed->fed_grant = oa->o_grant;
+ fed->fed_grant_waiting = 0;
+ }
+ } else {
+ fed->fed_grant_waiting = 0;
+ }
+ }
+ spin_unlock(&obd->obd_osfs_lock);
+ oa->o_valid &= ~(OBD_MD_FLGRANT|OBD_MD_FLBLOCKS);
+ EXIT;
+}
+
+/* Figure out how much space is available between what we've granted
+ * and what remains in the filesystem. Compensate for ext3 indirect
+ * block overhead when computing how much free space is left ungranted.
+ *
+ * Caller must hold obd_osfs_lock. */
+obd_size filter_grant_space_left(struct obd_export *exp)
+{
+ obd_size left = 0;
+ struct obd_device *obd = exp->exp_obd;
+ int blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
+ /* XXX I disabled statfs caching as it only creates extra problems now.
+ -- green*/
+ unsigned long max_age = jiffies/* - HZ*/+1;
+ struct filter_export_data *fed = &exp->exp_filter_data;
+ int rc;
+
+restat:
+ rc = fsfilt_statfs(obd, obd->u.filter.fo_sb, max_age);
+ if (rc) /* N.B. statfs can't really fail, just for correctness */
+ RETURN(0);
+
+ left = obd->obd_osfs.os_bavail << blockbits;
+ left -= (left >> (blockbits - 2)) + (left >> (2 * blockbits - 2));
+ /* We cannot afford having absolutely no space, we need some for
+ llog stuff */
+ if ( left >= PAGE_SIZE * 10)
+ left -= PAGE_SIZE * 10;
+ else
+ left = 0;
+
+ /* If fed->fed_grant_waiting is set, this means
+ obd->u.filter.fo_tot_granted does not represent actual granted
+ amount and client is supposedly actively shrinks its cache, so
+ no point in printing this warning */
+ if (left < obd->u.filter.fo_tot_granted && !fed->fed_grant_waiting)
+ CERROR("granted space "LPU64" more than available "LPU64"\n",
+ obd->u.filter.fo_tot_granted, left);
+
+ left -= min(left, obd->u.filter.fo_tot_granted);
+ if (left < FILTER_GRANT_CHUNK && time_after(jiffies,obd->obd_osfs_age)){
+ CDEBUG(D_SUPER, "fs has no space left and statfs too old\n");
+ max_age = jiffies;
+ goto restat;
+ }
+
+ CDEBUG(D_SUPER, "free: "LPU64" avail: "LPU64" grant left: "LPU64"\n",
+ obd->obd_osfs.os_bfree << blockbits,
+ obd->obd_osfs.os_bavail << blockbits, left);
+
+ return left;
+}
+
+/* When clients have dirtied as much space as they've been granted they
+ * fall through to sync writes. These sync writes haven't been expressed
+ * in grants and need to error with ENOSPC when there isn't room in the
+ * filesystem for them after grants are taken into account. However,
+ * writeback of the dirty data that was already granted space can write
+ * right on through. We have no need to stop writes that won't allocate
+ * new space, so we bmap to calculate how much this io is going to consume.
+ *
+ * Caller must hold obd_osfs_lock. */
+static int filter_check_space(struct obd_export *exp, int objcount,
+ struct fsfilt_objinfo *fso, int niocount,
+ struct niobuf_remote *rnb,
+ struct niobuf_local *lnb, obd_size *left,
+ obd_size *consumed, struct inode *inode)
+{
+ int blocksize = exp->exp_obd->u.filter.fo_sb->s_blocksize;
+ obd_size bytes, ungranted = 0;
+ int i, rc = -ENOSPC, obj, n = 0;
+
+ *consumed = 0;
+
+ for (obj = 0; obj < objcount; obj++) {
+ for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) {
+ obd_size tmp;
+
+ bytes = rnb[n].len;
+ tmp = rnb[n].offset & (blocksize - 1);
+ bytes += tmp;
+ tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1);
+ if (tmp)
+ bytes += blocksize - tmp;
+
+ if (rnb[n].flags & OBD_BRW_FROM_GRANT) {
+ *consumed += bytes;
+ rc = 0;
+ continue;
+ }
+ if (*left - *consumed >= bytes) {
+ /* if enough space, pretend it was granted */
+ exp->exp_obd->u.filter.fo_tot_granted += bytes;
+ exp->exp_filter_data.fed_grant += bytes;
+ *consumed += bytes;
+ *left -= bytes;
+ rc = 0;
+ continue;
+ }
+ spin_unlock(&exp->exp_obd->obd_osfs_lock);
+ if (!filter_inode_has_holes(inode,
+ rnb[n].offset >>
+ inode->i_blkbits,
+ rnb[n].len >>
+ inode->i_blkbits)) {
+ rc = 0;
+ } else {
+ rc = lnb[n].rc = -ENOSPC;
+ }
+ spin_lock(&exp->exp_obd->obd_osfs_lock);
+ if (rc)
+ goto leave;
+ }
+ }
+
+ CDEBUG((*consumed != 0 && ungranted != 0) ? D_ERROR : D_SUPER,
+ "consumed: "LPU64" ungranted: "LPU64"\n", *consumed, ungranted);
+
+ if (*consumed > exp->exp_filter_data.fed_grant)
+ CERROR("request sent from cache, but not enough grant ("LPU64
+ ","LPU64")\n", *consumed,
+ exp->exp_filter_data.fed_grant);
+leave:
+ return rc;
+}
+
+/* Calculate how much grant space to allocate to this client, based on how
+ * much space is currently free and how much of that is already granted.
+ *
+ * Caller must hold obd_osfs_lock. */
+static void filter_grant(struct obd_export *exp, struct obdo *oa,
+ obd_size left, obd_size from_grant)
+{
+ struct obd_device *obd = exp->exp_obd;
+ struct filter_export_data *fed = &exp->exp_filter_data;
+ obd_size grant, extra;
+ int blockbits;
+
+ blockbits = obd->u.filter.fo_sb->s_blocksize_bits;
+
+ /* if things go wrong conservatively try to clamp them from
+ * generating more dirty data until things are better on our end */
+ grant = fed->fed_cached;
+
+ extra = min(FILTER_GRANT_CHUNK, left / 2);
+
+ if (grant > fed->fed_grant) {
+ /* If client has screwed up, force basic grant until fixed */
+ CERROR("client %s cached more "LPU64" than granted "LPU64"\n",
+ exp->exp_client_uuid.uuid, fed->fed_cached,
+ fed->fed_grant);
+ grant = extra;
+ } else if (fed->fed_grant_waiting) {
+ /* KISS: only one grant change in flight at a time. We
+ * could move it in the "same direction" easily,
+ * but changing directions (e.g. grow then shrink
+ * before client ACKs) would be bad. */
+ grant = fed->fed_grant_sent;
+ } else {
+ /* grant will shrink or grow as client cache/extra changes */
+ grant = fed->fed_cached + extra;
+ }
+
+ /* If we've granted all we're willing, we have to revoke
+ * the grant covering what the client just wrote. */
+ if (left == 0) {
+ grant -= min(from_grant, grant);
+ }
+
+ if (!fed->fed_grant_waiting && grant + from_grant > left ) {
+ if (from_grant < left)
+ grant = left - from_grant;
+ else
+ grant = 0;
+ }
+
+ if (grant != fed->fed_grant) {
+ fed->fed_grant_waiting = 1;
+ fed->fed_grant_sent = grant;
+ if (grant > fed->fed_grant) {
+ obd->u.filter.fo_tot_granted += grant - fed->fed_grant;
+ fed->fed_grant = grant;
+ }
+ }
+
+ CDEBUG(D_SUPER,"cli %s cache:"LPU64" grant:"LPU64", granting:"LPU64"\n",
+ exp->exp_connection->c_remote_uuid.uuid, oa->o_blocks,
+ oa->o_grant, grant);
+ CDEBUG(D_SUPER, "fed sent:"LPU64" wt:%d grant:"LPU64"\n",
+ fed->fed_grant_sent, fed->fed_grant_waiting,
+ fed->fed_grant);
+ CDEBUG(D_SUPER, "tot cached:"LPU64" granted:"LPU64" num_exports: %d\n",
+ obd->u.filter.fo_tot_cached,
+ obd->u.filter.fo_tot_granted, obd->obd_num_exports);
+
+ oa->o_valid |= OBD_MD_FLGRANT;
+ oa->o_grant = grant;
+}
+
static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa,
int objcount, struct obd_ioobj *obj,
int niocount, struct niobuf_remote *nb,
struct niobuf_local *res,
struct obd_trans_info *oti)
{
+ struct obd_device *obd = exp->exp_obd;
struct obd_run_ctxt saved;
struct obd_ioobj *o;
struct niobuf_remote *rnb;
for (i = 0, o = obj; i < objcount; i++, o++) {
LASSERT(o->ioo_bufcnt);
- dentry = filter_oa2dentry(exp->exp_obd, oa);
+ dentry = filter_oa2dentry(obd, oa);
if (IS_ERR(dentry))
GOTO(cleanup, rc = PTR_ERR(dentry));
CDEBUG(D_INFO, "preprw_read setup: %lu jiffies\n",
(jiffies - now));
+ if (oa) {
+ spin_lock(&obd->obd_osfs_lock);
+ filter_grant(exp, oa, filter_grant_space_left(exp), 0);
+ spin_unlock(&obd->obd_osfs_lock);
+ }
+
for (i = 0, o = obj, rnb = nb, lnb = res; i < objcount; i++, o++) {
dentry = fso[i].fso_dentry;
inode = dentry->d_inode;
CDEBUG(D_INFO, "start_page_read: %lu jiffies\n",
(jiffies - now));
- lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_READ_BYTES,
- tot_bytes);
+ lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes);
while (lnb-- > res) {
rc = filter_finish_page_read(lnb);
if (rc) {
struct niobuf_local *res,
struct obd_trans_info *oti)
{
+ struct obd_device *obd = exp->exp_obd;
struct obd_run_ctxt saved;
- struct niobuf_remote *rnb;
- struct niobuf_local *lnb = NULL;
+ struct niobuf_remote *rnb = nb;
+ struct niobuf_local *lnb = res;
struct fsfilt_objinfo fso;
struct dentry *dentry;
int rc = 0, i, tot_bytes = 0;
+ obd_size consumed = 0, left;
unsigned long now = jiffies;
ENTRY;
LASSERT(objcount == 1);
LASSERT(obj->ioo_bufcnt > 0);
+ filter_grant_incoming(exp, oa);
+
memset(res, 0, niocount * sizeof(*res));
- push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
- dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr,
- obj->ioo_id);
+ push_ctxt(&saved, &obd->obd_ctxt, NULL);
+ dentry = filter_fid2dentry(obd, NULL, obj->ioo_gr, obj->ioo_id);
if (IS_ERR(dentry))
GOTO(cleanup, rc = PTR_ERR(dentry));
CDEBUG(D_INFO, "preprw_write setup: %lu jiffies\n",
(jiffies - now));
+ spin_lock(&obd->obd_osfs_lock);
+ left = filter_grant_space_left(exp);
+
+ rc = filter_check_space(exp, objcount, &fso, niocount, rnb, lnb,
+ &left, &consumed, dentry->d_inode);
+ if (oa)
+ filter_grant(exp, oa, left, consumed);
+
+ spin_unlock(&obd->obd_osfs_lock);
+
+ if (rc) {
+ f_dput(dentry);
+ GOTO(cleanup, rc);
+ }
+
for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
i++, lnb++, rnb++) {
+
+ /* If there were any granting failures, we should not have
+ come here */
+ LASSERT (lnb->rc == 0);
+
lnb->dentry = dentry;
lnb->offset = rnb->offset;
lnb->len = rnb->len;
CDEBUG(D_INFO, "start_page_write: %lu jiffies\n",
(jiffies - now));
- lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
- tot_bytes);
+ lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_WRITE_BYTES, tot_bytes);
EXIT;
cleanup:
- pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
+ pop_ctxt(&saved, &obd->obd_ctxt, NULL);
return rc;
}