/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
* vim:expandtab:shiftwidth=8:tabstop=8:
*
- * Lustre Light Super operations
+ * Lustre Light block IO
*
- * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
+ * Copyright (c) 2002-2004 Cluster File Systems, Inc.
*
* This file is part of Lustre, http://www.lustre.org.
*
#include <time.h>
#include <sys/types.h>
#include <sys/queue.h>
+#include <fcntl.h>
+#include <sys/uio.h>
+#ifdef HAVE_XTIO_H
+#include <xtio.h>
+#endif
#include <sysio.h>
-#include <fs.h>
#include <mount.h>
#include <inode.h>
+#ifdef HAVE_FILE_H
#include <file.h>
+#endif
+
+#undef LIST_HEAD
#include "llite_lib.h"
-#if 0
-void llu_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
- struct ldlm_lock *lock)
+struct llu_io_group
{
- clear_bit(LLI_F_HAVE_SIZE_LOCK, &(llu_i2info(inode)->lli_flags));
-#if 0
- struct ldlm_extent *extent = &lock->l_extent;
- unsigned long start, end, count, skip, i, j;
- struct page *page;
- int ret;
- ENTRY;
-
- CDEBUG(D_INODE, "obdo %lu inode %p ["LPU64"->"LPU64"] size: %llu\n",
- inode->i_ino, inode, extent->start, extent->end, inode->i_size);
-
- start = extent->start >> PAGE_CACHE_SHIFT;
- count = ~0;
- skip = 0;
- end = (extent->end >> PAGE_CACHE_SHIFT) + 1;
- if ((end << PAGE_CACHE_SHIFT) < extent->end)
- end = ~0;
- if (lsm->lsm_stripe_count > 1) {
- struct {
- char name[16];
- struct ldlm_lock *lock;
- struct lov_stripe_md *lsm;
- } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
- __u32 stripe;
- __u32 vallen = sizeof(stripe);
- int rc;
-
- /* get our offset in the lov */
- rc = obd_get_info(ll_i2obdconn(inode), sizeof(key),
- &key, &vallen, &stripe);
- if (rc != 0) {
- CERROR("obd_get_info: rc = %d\n", rc);
- LBUG();
- }
- LASSERT(stripe < lsm->lsm_stripe_count);
-
- count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
- skip = (lsm->lsm_stripe_count - 1) * count;
- start += (start/count * skip) + (stripe * count);
- if (end != ~0)
- end += (end/count * skip) + (stripe * count);
- }
-
- i = (inode->i_size + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
- if (end >= i)
- clear_bit(LLI_F_HAVE_SIZE_LOCK, &(ll_i2info(inode)->lli_flags));
- if (i < end)
- end = i;
-
- CDEBUG(D_INODE, "start: %lu j: %lu count: %lu skip: %lu end: %lu\n",
- start, start % count, count, skip, end);
-
- /* start writeback on dirty pages in the extent when its PW */
- for (i = start, j = start % count;
- lock->l_granted_mode == LCK_PW && i < end; j++, i++) {
- if (j == count) {
- i += skip;
- j = 0;
- }
- /* its unlikely, but give us a chance to bail when we're out */
- PGCACHE_WRLOCK(inode->i_mapping);
- if (list_empty(&inode->i_mapping->dirty_pages)) {
- CDEBUG(D_INODE, "dirty list empty\n");
- PGCACHE_WRUNLOCK(inode->i_mapping);
- break;
- }
- PGCACHE_WRUNLOCK(inode->i_mapping);
-
- if (need_resched())
- schedule();
+ struct obd_io_group *lig_oig;
+ struct inode *lig_inode;
+ int lig_maxpages;
+ int lig_npages;
+ __u64 lig_rwcount;
+ struct ll_async_page *lig_llap;
+ struct page *lig_pages;
+};
- /* always do a getattr for the first person to pop out of lock
- * acquisition.. the DID_GETATTR flag and semaphore serialize
- * this initial race. we used to make a decision based on whether
- * the lock was matched or acquired, but the matcher could win the
- * waking race with the first issuer so that was no good..
- */
- if (test_bit(LLI_F_DID_GETATTR, &lli->lli_flags))
- RETURN(ELDLM_OK);
+#define LLU_IO_GROUP_SIZE(x) \
+ (sizeof(struct llu_io_group) + \
+ sizeof(struct ll_async_page) * (x) + \
+ sizeof(struct page) * (x))
- down(&lli->lli_getattr_sem);
+struct llu_io_session
+{
+ struct inode *lis_inode;
+ int lis_cmd;
+ int lis_max_groups;
+ int lis_ngroups;
+ struct llu_io_group *lis_groups[0];
+};
+#define LLU_IO_SESSION_SIZE(x) \
+ (sizeof(struct llu_io_session) + (x) * 2 * sizeof(void *))
- if (!test_bit(LLI_F_DID_GETATTR, &lli->lli_flags)) {
- rc = ll_inode_getattr(inode, lsm);
- if (rc == 0) {
- set_bit(LLI_F_DID_GETATTR, &lli->lli_flags);
- } else {
- unlock_page(page);
- }
- page_cache_release(page);
- }
+typedef ssize_t llu_file_piov_t(const struct iovec *iovec, int iovlen,
+ _SYSIO_OFF_T pos, ssize_t len,
+ void *private);
- /* our locks are page granular thanks to osc_enqueue, we invalidate the
- * whole page. */
- LASSERT((extent->start & ~PAGE_CACHE_MASK) == 0);
- LASSERT(((extent->end+1) & ~PAGE_CACHE_MASK) == 0);
- for (i = start, j = start % count ; i < end ; j++, i++) {
- if ( j == count ) {
- i += skip;
- j = 0;
- }
- PGCACHE_WRLOCK(inode->i_mapping);
- if (list_empty(&inode->i_mapping->dirty_pages) &&
- list_empty(&inode->i_mapping->clean_pages) &&
- list_empty(&inode->i_mapping->locked_pages)) {
- CDEBUG(D_INODE, "nothing left\n");
- PGCACHE_WRUNLOCK(inode->i_mapping);
- break;
- }
- PGCACHE_WRUNLOCK(inode->i_mapping);
- if (need_resched())
- schedule();
- page = find_get_page(inode->i_mapping, i);
- if (page == NULL)
- continue;
- CDEBUG(D_INODE, "dropping page %p at %lu\n", page, page->index);
- lock_page(page);
- if (page->mapping) /* might have raced */
-#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
- truncate_complete_page(page);
-#else
- truncate_complete_page(page->mapping, page);
-#endif
- unlock_page(page);
- page_cache_release(page);
- }
- EXIT;
-#endif
-}
-
-int llu_lock_callback(struct ldlm_lock *lock, struct ldlm_lock_desc *new,
- void *data, int flag)
+static int llu_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
{
- struct inode *inode = data;
struct llu_inode_info *lli = llu_i2info(inode);
- struct lustre_handle lockh = {0};
+ struct lov_stripe_md *lsm = lli->lli_smd;
+ struct obd_export *exp = llu_i2dtexp(inode);
+ struct {
+ char name[16];
+ struct ldlm_lock *lock;
+ struct lov_stripe_md *lsm;
+ } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
+ __u32 stripe, vallen = sizeof(stripe);
int rc;
ENTRY;
- if (inode == NULL)
- LBUG();
+ if (lsm->lsm_stripe_count == 1)
+ RETURN(0);
- switch (flag) {
- case LDLM_CB_BLOCKING:
- ldlm_lock2handle(lock, &lockh);
- rc = ldlm_cli_cancel(&lockh);
- if (rc != ELDLM_OK)
- CERROR("ldlm_cli_cancel failed: %d\n", rc);
- break;
- case LDLM_CB_CANCELING: {
- /* FIXME: we could be given 'canceling intents' so that we
- * could know to write-back or simply throw away the pages
- * based on if the cancel comes from a desire to, say,
- * read or truncate.. */
- llu_pgcache_remove_extent(inode, lli->lli_smd, lock);
- break;
- }
- default:
- LBUG();
+ /* get our offset in the lov */
+ rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
+ if (rc != 0) {
+ CERROR("obd_get_info: rc = %d\n", rc);
+ RETURN(rc);
}
-
- RETURN(0);
+ LASSERT(stripe < lsm->lsm_stripe_count);
+ RETURN(stripe);
}
-#endif
static int llu_extent_lock_callback(struct ldlm_lock *lock,
struct ldlm_lock_desc *new, void *data,
struct lustre_handle lockh = { 0 };
int rc;
ENTRY;
-
if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
LBUG();
}
-
+
switch (flag) {
case LDLM_CB_BLOCKING:
ldlm_lock2handle(lock, &lockh);
CERROR("ldlm_cli_cancel failed: %d\n", rc);
break;
case LDLM_CB_CANCELING: {
- struct inode *inode = llu_inode_from_lock(lock);
+ struct inode *inode;
struct llu_inode_info *lli;
-
+ struct lov_stripe_md *lsm;
+ __u32 stripe;
+ __u64 kms;
+
+ /* This lock wasn't granted, don't try to evict pages */
+ if (lock->l_req_mode != lock->l_granted_mode)
+ RETURN(0);
+
+ inode = llu_inode_from_lock(lock);
if (!inode)
RETURN(0);
lli= llu_i2info(inode);
- if (!lli) {
- I_RELE(inode);
- RETURN(0);
- }
- if (!lli->lli_smd) {
- I_RELE(inode);
- RETURN(0);
+ if (!lli)
+ goto iput;
+ if (!lli->lli_smd)
+ goto iput;
+ lsm = lli->lli_smd;
+
+ stripe = llu_lock_to_stripe_offset(inode, lock);
+ if (stripe < 0) {
+ CERROR("lock on inode without such object\n");
+ break;
}
-
-/*
- ll_pgcache_remove_extent(inode, lli->lli_smd, lock);
- iput(inode);
-*/
+#warning "fix l_lock() using here!"
+// l_lock(&lock->l_resource->lr_namespace->ns_lock);
+ kms = ldlm_extent_shift_kms(lock,
+ lsm->lsm_oinfo[stripe].loi_kms);
+// l_unlock(&lock->l_resource->lr_namespace->ns_lock);
+ if (lsm->lsm_oinfo[stripe].loi_kms != kms)
+ LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
+ lsm->lsm_oinfo[stripe].loi_kms, kms);
+ lsm->lsm_oinfo[stripe].loi_kms = kms;
+iput:
I_RELE(inode);
break;
}
default:
LBUG();
}
-
+
RETURN(0);
}
-int llu_extent_lock_no_validate(struct ll_file_data *fd,
- struct inode *inode,
- struct lov_stripe_md *lsm,
- int mode,
- struct ldlm_extent *extent,
- struct lustre_handle *lockh,
- int ast_flags)
+static int llu_glimpse_callback(struct ldlm_lock *lock, void *reqp)
+{
+ struct ptlrpc_request *req = reqp;
+ struct inode *inode = llu_inode_from_lock(lock);
+ struct llu_inode_info *lli;
+ struct ost_lvb *lvb;
+ int rc, size = sizeof(*lvb), stripe = 0;
+ ENTRY;
+
+ if (inode == NULL)
+ GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
+ lli = llu_i2info(inode);
+ if (lli == NULL)
+ GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
+ if (lli->lli_smd == NULL)
+ GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
+
+ /* First, find out which stripe index this lock corresponds to. */
+ if (lli->lli_smd->lsm_stripe_count > 1) {
+ stripe = llu_lock_to_stripe_offset(inode, lock);
+ if (stripe < 0) {
+ CWARN("lock on inode without such object\n");
+ GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
+ }
+ }
+
+ rc = lustre_pack_reply(req, 1, &size, NULL);
+ if (rc) {
+ CERROR("lustre_pack_reply: %d\n", rc);
+ GOTO(iput, rc);
+ }
+
+ lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
+ lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
+
+ LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
+ lli->lli_st_size, stripe, lvb->lvb_size);
+ iput:
+ I_RELE(inode);
+ out:
+ /* These errors are normal races, so we don't want to fill the console
+ * with messages by calling ptlrpc_error() */
+ if (rc == -ELDLM_NO_LOCK_DATA)
+ lustre_pack_reply(req, 0, NULL, NULL);
+
+ req->rq_status = rc;
+ return rc;
+}
+
+__u64 lov_merge_size(struct lov_stripe_md *lsm, int kms_only);
+__u64 lov_merge_blocks(struct lov_stripe_md *lsm);
+__u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
+
+/* NB: lov_merge_size will prefer locally cached writes if they extend the
+ * file (because it prefers KMS over RSS when larger) */
+int llu_glimpse_size(struct inode *inode)
{
- struct llu_sb_info *sbi = llu_i2sbi(inode);
struct llu_inode_info *lli = llu_i2info(inode);
- int rc;
+ struct llu_sb_info *sbi = llu_i2sbi(inode);
+ ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
+ struct lustre_handle lockh = { 0 };
+ int rc, flags = LDLM_FL_HAS_INTENT;
ENTRY;
- LASSERT(lockh->cookie == 0);
+ CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", lli->lli_st_ino);
-#if 0
- /* XXX phil: can we do this? won't it screw the file size up? */
- if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
- (sbi->ll_flags & LL_SBI_NOLCK))
- RETURN(0);
-#endif
+ rc = obd_enqueue(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, &policy,
+ LCK_PR, &flags, llu_extent_lock_callback,
+ ldlm_completion_ast, llu_glimpse_callback, inode,
+ sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
+ if (rc) {
+ CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
+ RETURN(rc > 0 ? -EIO : rc);
+ }
- CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
- lli->lli_st_ino, extent->start, extent->end);
+ lli->lli_st_size = lov_merge_size(lli->lli_smd, 0);
+ lli->lli_st_blocks = lov_merge_blocks(lli->lli_smd);
+ //lli->lli_st_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
+
+ CDEBUG(D_DLMTRACE, "glimpse: size: %llu, blocks: %lu\n",
+ lli->lli_st_size, lli->lli_st_blocks);
- rc = obd_enqueue(sbi->ll_osc_exp, lsm, NULL, LDLM_EXTENT, extent,
- sizeof(extent), mode, &ast_flags,
- llu_extent_lock_callback, inode, lockh);
+ obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
RETURN(rc);
}
-/*
- * this grabs a lock and manually implements behaviour that makes it look like
- * the OST is returning the file size with each lock acquisition.
- */
int llu_extent_lock(struct ll_file_data *fd, struct inode *inode,
struct lov_stripe_md *lsm, int mode,
- struct ldlm_extent *extent, struct lustre_handle *lockh)
+ ldlm_policy_data_t *policy, struct lustre_handle *lockh,
+ int ast_flags)
{
+ struct llu_sb_info *sbi = llu_i2sbi(inode);
struct llu_inode_info *lli = llu_i2info(inode);
- struct obd_export *exp = llu_i2obdexp(inode);
- struct ldlm_extent size_lock;
- struct lustre_handle match_lockh = {0};
- int flags, rc, matched;
+ int rc;
ENTRY;
- rc = llu_extent_lock_no_validate(fd, inode, lsm, mode, extent, lockh, 0);
- if (rc != ELDLM_OK)
- RETURN(rc);
+ LASSERT(lockh->cookie == 0);
- if (test_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags))
+ /* XXX phil: can we do this? won't it screw the file size up? */
+ if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
+ (sbi->ll_flags & LL_SBI_NOLCK))
RETURN(0);
- rc = llu_inode_getattr(inode, lsm);
- if (rc) {
- llu_extent_unlock(fd, inode, lsm, mode, lockh);
- RETURN(rc);
- }
+ CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
+ lli->lli_st_ino, policy->l_extent.start, policy->l_extent.end);
- size_lock.start = lli->lli_st_size;
- size_lock.end = OBD_OBJECT_EOF;
+ rc = obd_enqueue(sbi->ll_dt_exp, lsm, LDLM_EXTENT, policy, mode,
+ &ast_flags, llu_extent_lock_callback,
+ ldlm_completion_ast, llu_glimpse_callback, inode,
+ sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
+ if (rc > 0)
+ rc = -EIO;
- /* XXX I bet we should be checking the lock ignore flags.. */
- flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
- matched = obd_match(exp, lsm, LDLM_EXTENT, &size_lock,
- sizeof(size_lock), LCK_PR, &flags, inode,
- &match_lockh);
+ if (policy->l_extent.start == 0 &&
+ policy->l_extent.end == OBD_OBJECT_EOF)
+ lli->lli_st_size = lov_merge_size(lsm, 1);
- /* hey, alright, we hold a size lock that covers the size we
- * just found, its not going to change for a while.. */
- if (matched == 1) {
- set_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags);
- obd_cancel(exp, lsm, LCK_PR, &match_lockh);
- }
+ //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
- RETURN(0);
+ RETURN(rc);
}
int llu_extent_unlock(struct ll_file_data *fd, struct inode *inode,
struct llu_sb_info *sbi = llu_i2sbi(inode);
int rc;
ENTRY;
-#if 0
+
/* XXX phil: can we do this? won't it screw the file size up? */
if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
(sbi->ll_flags & LL_SBI_NOLCK))
RETURN(0);
-#endif
- rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
+
+ rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
RETURN(rc);
}
struct inode *llap_inode;
};
-static struct ll_async_page *llap_from_cookie(void *cookie)
-{
- struct ll_async_page *llap = cookie;
- if (llap->llap_magic != LLAP_MAGIC)
- return ERR_PTR(-EINVAL);
- return llap;
-};
-
static void llu_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
{
struct ll_async_page *llap;
struct inode *inode;
struct lov_stripe_md *lsm;
- obd_flag valid_flags;
+ obd_valid valid_flags;
ENTRY;
- llap = llap_from_cookie(data);
- if (IS_ERR(llap)) {
- EXIT;
- return;
- }
-
+ llap = LLAP_FROM_COOKIE(data);
inode = llap->llap_inode;
lsm = llu_i2info(inode)->lli_smd;
}
/* called for each page in a completed rpc.*/
-static void llu_ap_completion(void *data, int cmd, int rc)
+static void llu_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
{
struct ll_async_page *llap;
struct page *page;
- llap = llap_from_cookie(data);
- if (IS_ERR(llap)) {
- EXIT;
- return;
- }
-
+ llap = LLAP_FROM_COOKIE(data);
llap->llap_queued = 0;
page = llap->llap_page;
.ap_completion = llu_ap_completion,
};
-static
-struct llu_sysio_cookie* get_sysio_cookie(struct inode *inode, int npages)
-{
- struct llu_sysio_cookie *cookie;
-
- OBD_ALLOC(cookie, LLU_SYSIO_COOKIE_SIZE(npages));
- if (cookie) {
- I_REF(inode);
- cookie->lsc_inode = inode;
- cookie->lsc_npages = npages;
- cookie->lsc_llap = (struct ll_async_page *)(cookie + 1);
- cookie->lsc_pages = (struct page *) (cookie->lsc_llap + npages);
-
- osic_init(&cookie->lsc_osic);
- }
-
- return cookie;
-}
-
-static
-void put_sysio_cookie(struct llu_sysio_cookie *cookie)
-{
- struct lov_stripe_md *lsm = llu_i2info(cookie->lsc_inode)->lli_smd;
- struct obd_export *exp = llu_i2obdexp(cookie->lsc_inode);
- struct ll_async_page *llap = cookie->lsc_llap;
- int i;
-
- for (i = 0; i< cookie->lsc_npages; i++) {
- if (llap[i].llap_cookie)
- obd_teardown_async_page(exp, lsm, NULL,
- llap[i].llap_cookie);
- }
-
- I_RELE(cookie->lsc_inode);
-
- OBD_FREE(cookie, LLU_SYSIO_COOKIE_SIZE(cookie->lsc_npages));
-}
-
-static
-int llu_prep_async_io(struct llu_sysio_cookie *cookie, int cmd,
- char *buf, loff_t pos, size_t count)
+static int llu_queue_pio(int cmd, struct llu_io_group *group,
+ char *buf, size_t count, loff_t pos)
{
- struct lov_stripe_md *lsm = llu_i2info(cookie->lsc_inode)->lli_smd;
- struct obd_export *exp = llu_i2obdexp(cookie->lsc_inode);
- struct page *pages = cookie->lsc_pages;
- struct ll_async_page *llap = cookie->lsc_llap;
- int i, rc, npages = 0;
+ struct llu_inode_info *lli = llu_i2info(group->lig_inode);
+ struct lov_stripe_md *lsm = lli->lli_smd;
+ struct obd_export *exp = llu_i2dtexp(group->lig_inode);
+ struct page *pages = &group->lig_pages[group->lig_npages];
+ struct ll_async_page *llap = &group->lig_llap[group->lig_npages];
+ int i, rc, npages = 0, ret_bytes = 0;
ENTRY;
if (!exp)
RETURN(-EINVAL);
- cookie->lsc_rwcount = count;
-
/* prepare the pages array */
do {
unsigned long index, offset, bytes;
if (bytes > count)
bytes = count;
+ /* prevent read beyond file range */
+ if ((cmd == OBD_BRW_READ) &&
+ (pos + bytes) >= lli->lli_st_size) {
+ if (pos >= lli->lli_st_size)
+ break;
+ bytes = lli->lli_st_size - pos;
+ }
+
/* prepare page for this index */
pages[npages].index = index;
pages[npages].addr = buf - offset;
count -= bytes;
pos += bytes;
buf += bytes;
+
+ group->lig_rwcount += bytes;
+ ret_bytes += bytes;
} while (count);
+ group->lig_npages += npages;
+
for (i = 0; i < npages; i++) {
llap[i].llap_magic = LLAP_MAGIC;
rc = obd_prep_async_page(exp, lsm, NULL, &pages[i],
&llu_async_page_ops,
&llap[i], &llap[i].llap_cookie);
if (rc) {
+ LASSERT(rc < 0);
llap[i].llap_cookie = NULL;
RETURN(rc);
}
- CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n",
+ CDEBUG(D_CACHE, "llap %p page %p group %p obj off "LPU64"\n",
&llap[i], &pages[i], llap[i].llap_cookie,
(obd_off)pages[i].index << PAGE_SHIFT);
pages[i].private = (unsigned long)&llap[i];
llap[i].llap_page = &pages[i];
- llap[i].llap_inode = cookie->lsc_inode;
+ llap[i].llap_inode = group->lig_inode;
- rc = obd_queue_sync_io(exp, lsm, NULL, &cookie->lsc_osic,
- llap[i].llap_cookie, cmd,
- pages[i]._offset, pages[i]._count, 0);
- if (rc)
+ rc = obd_queue_group_io(exp, lsm, NULL, group->lig_oig,
+ llap[i].llap_cookie, cmd,
+ pages[i]._offset, pages[i]._count, 0,
+ ASYNC_READY | ASYNC_URGENT |
+ ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
+ if (rc) {
+ LASSERT(rc < 0);
RETURN(rc);
+ }
llap[i].llap_queued = 1;
}
- RETURN(0);
+ RETURN(ret_bytes);
}
static
-int llu_start_async_io(struct llu_sysio_cookie *cookie)
+struct llu_io_group * get_io_group(struct inode *inode, int maxpages)
{
- struct lov_stripe_md *lsm = llu_i2info(cookie->lsc_inode)->lli_smd;
- struct obd_export *exp = llu_i2obdexp(cookie->lsc_inode);
+ struct llu_io_group *group;
+ int rc;
- return obd_trigger_sync_io(exp, lsm, NULL, &cookie->lsc_osic);
-}
+ OBD_ALLOC(group, LLU_IO_GROUP_SIZE(maxpages));
+ if (!group)
+ return ERR_PTR(-ENOMEM);
-/*
- * read/write a continuous buffer for an inode (zero-copy)
- */
-struct llu_sysio_cookie*
-llu_rw(int cmd, struct inode *inode, char *buf, size_t count, loff_t pos)
-{
- struct llu_sysio_cookie *cookie;
- int max_pages, rc;
- ENTRY;
+ I_REF(inode);
+ group->lig_inode = inode;
+ group->lig_maxpages = maxpages;
+ group->lig_llap = (struct ll_async_page *)(group + 1);
+ group->lig_pages = (struct page *) (group->lig_llap + maxpages);
- max_pages = (count >> PAGE_SHIFT) + 2;
+ rc = oig_init(&group->lig_oig);
+ if (rc) {
+ OBD_FREE(group, LLU_IO_GROUP_SIZE(maxpages));
+ return ERR_PTR(rc);
+ }
- cookie = get_sysio_cookie(inode, max_pages);
- if (!cookie)
- RETURN(ERR_PTR(-ENOMEM));
+ return group;
+}
- rc = llu_prep_async_io(cookie, cmd, buf, pos, count);
- if (rc)
- GOTO(out_cleanup, rc);
+static int max_io_pages(ssize_t len, int iovlen)
+{
+ return (((len + PAGE_SIZE -1) / PAGE_SIZE) + 2 + iovlen - 1);
+}
- rc = llu_start_async_io(cookie);
- if (rc)
- GOTO(out_cleanup, rc);
+static
+void put_io_group(struct llu_io_group *group)
+{
+ struct lov_stripe_md *lsm = llu_i2info(group->lig_inode)->lli_smd;
+ struct obd_export *exp = llu_i2dtexp(group->lig_inode);
+ struct ll_async_page *llap = group->lig_llap;
+ int i;
-/*
- rc = osic_wait(&osic);
- if (rc) {
- CERROR("file i/o error!\n");
- rw_count = rc;
+ for (i = 0; i< group->lig_npages; i++) {
+ if (llap[i].llap_cookie)
+ obd_teardown_async_page(exp, lsm, NULL,
+ llap[i].llap_cookie);
}
-*/
- RETURN(cookie);
-out_cleanup:
- put_sysio_cookie(cookie);
- RETURN(ERR_PTR(rc));
+ I_RELE(group->lig_inode);
+
+ oig_release(group->lig_oig);
+ OBD_FREE(group, LLU_IO_GROUP_SIZE(group->lig_maxpages));
}
-struct llu_sysio_callback_args*
-llu_file_write(struct inode *inode, const struct iovec *iovec,
- size_t iovlen, loff_t pos)
+static
+ssize_t llu_file_prwv(const struct iovec *iovec, int iovlen,
+ _SYSIO_OFF_T pos, ssize_t len,
+ void *private)
{
+ struct llu_io_session *session = (struct llu_io_session *) private;
+ struct inode *inode = session->lis_inode;
struct llu_inode_info *lli = llu_i2info(inode);
struct ll_file_data *fd = lli->lli_file_data;
struct lustre_handle lockh = {0};
struct lov_stripe_md *lsm = lli->lli_smd;
- struct llu_sysio_callback_args *lsca;
- struct llu_sysio_cookie *cookie;
- struct ldlm_extent extent;
- ldlm_error_t err;
- int iovidx;
+ struct obd_export *exp = NULL;
+ ldlm_policy_data_t policy;
+ struct llu_io_group *iogroup;
+ int astflag = (lli->lli_open_flags & O_NONBLOCK) ?
+ LDLM_FL_BLOCK_NOWAIT : 0;
+ __u64 kms;
+ int err, is_read, lock_mode, iovidx, ret;
ENTRY;
- /* XXX consider other types later */
- if (!S_ISREG(lli->lli_st_mode))
- LBUG();
+ /* in a large iov read/write we'll be repeatedly called.
+ * so give a chance to answer cancel ast here
+ */
+ liblustre_wait_event(0);
+
+ exp = llu_i2dtexp(inode);
+ if (exp == NULL)
+ RETURN(-EINVAL);
- LASSERT(iovlen <= MAX_IOVEC);
+ if (len == 0 || iovlen == 0)
+ RETURN(0);
+
+ if (pos + len > lli->lli_maxbytes)
+ RETURN(-ERANGE);
+
+ iogroup = get_io_group(inode, max_io_pages(len, iovlen));
+ if (IS_ERR(iogroup))
+ RETURN(PTR_ERR(iogroup));
+
+ is_read = session->lis_cmd == OBD_BRW_READ;
+ lock_mode = is_read ? LCK_PR : LCK_PW;
- OBD_ALLOC(lsca, sizeof(*lsca));
- if (!lsca)
- RETURN(ERR_PTR(-ENOMEM));
+ if (!is_read && (lli->lli_open_flags & O_APPEND)) {
+ policy.l_extent.start = 0;
+ policy.l_extent.end = OBD_OBJECT_EOF;
+ } else {
+ policy.l_extent.start = pos;
+ policy.l_extent.end = pos + len - 1;
+ }
+
+ err = llu_extent_lock(fd, inode, lsm, lock_mode, &policy,
+ &lockh, astflag);
+ if (err != ELDLM_OK)
+ GOTO(err_put, err);
+
+ if (is_read) {
+ kms = lov_merge_size(lsm, 1);
+ if (policy.l_extent.end > kms) {
+ /* A glimpse is necessary to determine whether we
+ * return a short read or some zeroes at the end of
+ * the buffer */
+ if ((err = llu_glimpse_size(inode))) {
+ llu_extent_unlock(fd, inode, lsm,
+ lock_mode, &lockh);
+ GOTO(err_put, err);
+ }
+ } else {
+ lli->lli_st_size = kms;
+ }
+ } else {
+ if (lli->lli_open_flags & O_APPEND)
+ pos = lli->lli_st_size;
+ }
- /* FIXME optimize the following extent locking */
for (iovidx = 0; iovidx < iovlen; iovidx++) {
- char *buf = iovec[iovidx].iov_base;
+ char *buf = (char *) iovec[iovidx].iov_base;
size_t count = iovec[iovidx].iov_len;
- if (count == 0)
+ if (!count)
continue;
-
- /* FIXME libsysio haven't consider the open flags
- * such as O_APPEND */
-#if 0
- if (!S_ISBLK(lli->lli_st_mode) && file->f_flags & O_APPEND) {
- extent.start = 0;
- extent.end = OBD_OBJECT_EOF;
- } else {
- extent.start = *ppos;
- extent.end = *ppos + count - 1;
+ if (len < count)
+ count = len;
+ if (IS_BAD_PTR(buf) || IS_BAD_PTR(buf + count)) {
+ llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
+ GOTO(err_put, err = -EFAULT);
}
-#else
- extent.start = pos;
- extent.end = pos + count - 1;
-#endif
- err = llu_extent_lock(fd, inode, lsm, LCK_PW, &extent, &lockh);
- if (err != ELDLM_OK)
- GOTO(err_out, err = -ENOLCK);
-
- CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
- lli->lli_st_ino, count, pos);
-
- cookie = llu_rw(OBD_BRW_WRITE, inode, buf, count, pos);
- if (!IS_ERR(cookie)) {
- /* save cookie */
- lsca->cookies[lsca->ncookies++] = cookie;
- pos += count;
- /* file size grow. XXX should be done here? */
- if (pos > lli->lli_st_size) {
- lli->lli_st_size = pos;
- set_bit(LLI_F_PREFER_EXTENDED_SIZE,
- &lli->lli_flags);
- }
+ if (is_read) {
+ if (pos >= lli->lli_st_size)
+ break;
} else {
- llu_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
- GOTO(err_out, err = PTR_ERR(cookie));
+ if (pos >= lli->lli_maxbytes) {
+ llu_extent_unlock(fd, inode, lsm, lock_mode,
+ &lockh);
+ GOTO(err_put, err = -EFBIG);
+ }
+ if (pos + count >= lli->lli_maxbytes)
+ count = lli->lli_maxbytes - pos;
}
- /* XXX errors? */
- err = llu_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
- if (err)
- CERROR("extent unlock error %d\n", err);
+ ret = llu_queue_pio(session->lis_cmd, iogroup, buf, count, pos);
+ if (ret < 0) {
+ llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
+ GOTO(err_put, err = ret);
+ } else {
+ pos += ret;
+ if (!is_read) {
+ LASSERT(ret == count);
+ obd_adjust_kms(exp, lsm, pos, 0);
+ /* file size grow immediately */
+ if (pos > lli->lli_st_size)
+ lli->lli_st_size = pos;
+ }
+ len -= ret;
+ if (!len)
+ break;
+ }
}
+ LASSERT(len == 0 || is_read); /* libsysio should guarantee this */
- RETURN(lsca);
+ err = llu_extent_unlock(fd, inode, lsm, lock_mode, &lockh);
+ if (err)
+ CERROR("extent unlock error %d\n", err);
-err_out:
- /* teardown all async stuff */
- while (lsca->ncookies--) {
- put_sysio_cookie(lsca->cookies[lsca->ncookies]);
- }
- OBD_FREE(lsca, sizeof(*lsca));
+ err = obd_trigger_group_io(exp, lsm, NULL, iogroup->lig_oig);
+ if (err)
+ GOTO(err_put, err);
- RETURN(ERR_PTR(err));
+ session->lis_groups[session->lis_ngroups++] = iogroup;
+ RETURN(0);
+err_put:
+ put_io_group(iogroup);
+ RETURN((ssize_t)err);
}
-#if 0
-static void llu_update_atime(struct inode *inode)
+static
+struct llu_io_session *get_io_session(struct inode *ino, int ngroups, int cmd)
{
- struct llu_inode_info *lli = llu_i2info(inode);
+ struct llu_io_session *session;
-#ifdef USE_ATIME
- struct iattr attr;
+ OBD_ALLOC(session, LLU_IO_SESSION_SIZE(ngroups));
+ if (!session)
+ return NULL;
- attr.ia_atime = LTIME_S(CURRENT_TIME);
- attr.ia_valid = ATTR_ATIME;
+ I_REF(ino);
+ session->lis_inode = ino;
+ session->lis_max_groups = ngroups;
+ session->lis_cmd = cmd;
+ return session;
+}
- if (lli->lli_st_atime == attr.ia_atime) return;
- if (IS_RDONLY(inode)) return;
- if (IS_NOATIME(inode)) return;
+static void put_io_session(struct llu_io_session *session)
+{
+ int i;
- /* ll_inode_setattr() sets inode->i_atime from attr.ia_atime */
- llu_inode_setattr(inode, &attr, 0);
-#else
- /* update atime, but don't explicitly write it out just this change */
- inode->i_atime = CURRENT_TIME;
-#endif
+ for (i = 0; i < session->lis_ngroups; i++) {
+ if (session->lis_groups[i]) {
+ put_io_group(session->lis_groups[i]);
+ session->lis_groups[i] = NULL;
+ }
+ }
+
+ I_RELE(session->lis_inode);
+ OBD_FREE(session, LLU_IO_SESSION_SIZE(session->lis_max_groups));
}
-#endif
-struct llu_sysio_callback_args*
-llu_file_read(struct inode *inode, const struct iovec *iovec,
- size_t iovlen, loff_t pos)
+static int llu_file_rwx(struct inode *ino,
+ struct ioctx *ioctx,
+ int read)
{
- struct llu_inode_info *lli = llu_i2info(inode);
- struct ll_file_data *fd = lli->lli_file_data;
- struct lov_stripe_md *lsm = lli->lli_smd;
- struct lustre_handle lockh = { 0 };
- struct ldlm_extent extent;
- struct llu_sysio_callback_args *lsca;
- struct llu_sysio_cookie *cookie;
- int iovidx;
-
- ldlm_error_t err;
+ struct llu_io_session *session;
+ ssize_t cc;
+ int cmd = read ? OBD_BRW_READ : OBD_BRW_WRITE;
ENTRY;
- OBD_ALLOC(lsca, sizeof(*lsca));
- if (!lsca)
- RETURN(ERR_PTR(-ENOMEM));
-
- for (iovidx = 0; iovidx < iovlen; iovidx++) {
- char *buf = iovec[iovidx].iov_base;
- size_t count = iovec[iovidx].iov_len;
+ LASSERT(ioctx->ioctx_xtvlen >= 0);
+ LASSERT(ioctx->ioctx_iovlen >= 0);
- /* "If nbyte is 0, read() will return 0 and have no other results."
- * -- Single Unix Spec */
- if (count == 0)
- continue;
+ liblustre_wait_event(0);
- extent.start = pos;
- extent.end = pos + count - 1;
+ if (!ioctx->ioctx_xtvlen)
+ RETURN(0);
- err = llu_extent_lock(fd, inode, lsm, LCK_PR, &extent, &lockh);
- if (err != ELDLM_OK)
- GOTO(err_out, err = -ENOLCK);
+ /* XXX consider other types later */
+ LASSERT(S_ISREG(llu_i2info(ino)->lli_st_mode));
- CDEBUG(D_INFO, "Reading inode %lu, "LPSZ" bytes, offset %Ld\n",
- lli->lli_st_ino, count, pos);
+ session = get_io_session(ino, ioctx->ioctx_xtvlen * 2, cmd);
+ if (!session)
+ RETURN(-ENOMEM);
- cookie = llu_rw(OBD_BRW_READ, inode, buf, count, pos);
- if (!IS_ERR(cookie)) {
- /* save cookie */
- lsca->cookies[lsca->ncookies++] = cookie;
- pos += count;
- } else {
- llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
- GOTO(err_out, err = PTR_ERR(cookie));
- }
+ cc = _sysio_enumerate_extents(ioctx->ioctx_xtv, ioctx->ioctx_xtvlen,
+ ioctx->ioctx_iov, ioctx->ioctx_iovlen,
+ llu_file_prwv, session);
- /* XXX errors? */
- err = llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
- if (err)
- CERROR("extent_unlock fail: %d\n", err);
+ if (cc >= 0) {
+ LASSERT(!ioctx->ioctx_cc);
+ ioctx->ioctx_private = session;
+ RETURN(0);
+ } else {
+ put_io_session(session);
+ RETURN(cc);
}
-#if 0
- if (readed > 0)
- llu_update_atime(inode);
-#endif
- RETURN(lsca);
+}
+
+int llu_iop_read(struct inode *ino,
+ struct ioctx *ioctx)
+{
+ return llu_file_rwx(ino, ioctx, 1);
+}
+
+int llu_iop_write(struct inode *ino,
+ struct ioctx *ioctx)
+{
+ struct iattr iattr;
+ int rc;
-err_out:
- /* teardown all async stuff */
- while (lsca->ncookies--) {
- put_sysio_cookie(lsca->cookies[lsca->ncookies]);
+ memset(&iattr, 0, sizeof(iattr));
+ iattr.ia_mtime = iattr.ia_atime = CURRENT_TIME;
+ iattr.ia_valid = ATTR_MTIME | ATTR_ATIME | ATTR_RAW;
+
+ liblustre_wait_event(0);
+ rc = llu_setattr_raw(ino, &iattr);
+ if (rc) {
+ CERROR("failed to set mtime/atime during write: %d", rc);
+ /* XXX should continue or return error? */
}
- OBD_FREE(lsca, sizeof(*lsca));
- RETURN(ERR_PTR(err));
+ return llu_file_rwx(ino, ioctx, 0);
}
-int llu_iop_iodone(struct ioctx *ioctxp)
+int llu_iop_iodone(struct ioctx *ioctx)
{
- struct llu_sysio_callback_args *lsca = ioctxp->ioctx_private;
- struct llu_sysio_cookie *cookie;
+ struct llu_io_session *session;
+ struct llu_io_group *group;
int i, err = 0, rc = 0;
ENTRY;
- /* write/read(fd, buf, 0) */
- if (!lsca)
- return 1;
+ liblustre_wait_event(0);
- LASSERT(!IS_ERR(lsca));
+ session = (struct llu_io_session *) ioctx->ioctx_private;
+ LASSERT(session);
+ LASSERT(!IS_ERR(session));
- for (i = 0; i < lsca->ncookies; i++) {
- cookie = lsca->cookies[i];
- if (cookie) {
- err = osic_wait(&cookie->lsc_osic);
- if (err && !rc)
- rc = err;
+ for (i = 0; i < session->lis_ngroups; i++) {
+ group = session->lis_groups[i];
+ if (group) {
+ if (!rc) {
+ err = oig_wait(group->lig_oig);
+ if (err)
+ rc = err;
+ }
if (!rc)
- ioctxp->ioctx_cc += cookie->lsc_rwcount;
- put_sysio_cookie(cookie);
+ ioctx->ioctx_cc += group->lig_rwcount;
+ put_io_group(group);
+ session->lis_groups[i] = NULL;
}
}
- if (rc)
- ioctxp->ioctx_cc = rc;
+ if (rc) {
+ LASSERT(rc < 0);
+ ioctx->ioctx_cc = -1;
+ ioctx->ioctx_errno = -rc;
+ }
- OBD_FREE(lsca, sizeof(*lsca));
- ioctxp->ioctx_private = NULL;
+ put_io_session(session);
+ ioctx->ioctx_private = NULL;
RETURN(1);
}