/* -*- 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.
*
#include <stdlib.h>
#include <string.h>
-#include <error.h>
#include <assert.h>
#include <time.h>
#include <sys/types.h>
#include <sys/queue.h>
+#include <fcntl.h>
#include <sysio.h>
#include <fs.h>
#include <inode.h>
#include <file.h>
+#undef LIST_HEAD
+
#include "llite_lib.h"
-int llu_iop_iodone(struct ioctx *ioctxp __IS_UNUSED)
+static int llu_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
{
- return 1;
+ struct llu_inode_info *lli = llu_i2info(inode);
+ struct lov_stripe_md *lsm = lli->lli_smd;
+ struct obd_export *exp = llu_i2obdexp(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 (lsm->lsm_stripe_count == 1)
+ RETURN(0);
+
+ /* 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);
+ LBUG();
+ }
+ LASSERT(stripe < lsm->lsm_stripe_count);
+ RETURN(stripe);
}
-/*
- * 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)
+static int llu_extent_lock_callback(struct ldlm_lock *lock,
+ struct ldlm_lock_desc *new, void *data,
+ int flag)
{
-#if 0
- struct ll_inode_info *lli = ll_i2info(inode);
+ struct lustre_handle lockh = { 0 };
int rc;
ENTRY;
- rc = ll_extent_lock_no_validate(fd, inode, lsm, mode, extent, lockh);
- if (rc != ELDLM_OK)
- RETURN(rc);
+ if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
+ LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
+ LBUG();
+ }
- /* 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);
-
- down(&lli->lli_getattr_sem);
-
- if (!test_bit(LLI_F_DID_GETATTR, &lli->lli_flags)) {
- rc = ll_inode_getattr(inode, lsm, fd ? &fd->fd_ost_och : NULL);
- if (rc == 0) {
- set_bit(LLI_F_DID_GETATTR, &lli->lli_flags);
- } else {
- /* XXX can this fail? */
- ll_extent_unlock(fd, inode, lsm, mode, lockh);
- }
+ 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: {
+ 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)
+ goto iput;
+ if (!lli->lli_smd)
+ goto iput;
+ lsm = lli->lli_smd;
+
+ stripe = llu_lock_to_stripe_offset(inode, lock);
+ kms = ldlm_extent_shift_kms(lock,
+ lsm->lsm_oinfo[stripe].loi_kms);
+ 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);
+}
+
+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);
+
+ rc = lustre_pack_reply(req, 1, &size, NULL);
+ if (rc) {
+ CERROR("lustre_pack_reply: %d\n", rc);
+ GOTO(iput, rc);
}
- up(&lli->lli_getattr_sem);
+ 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);
+__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 ost_lvb *lvb)
+{
+ struct llu_inode_info *lli = llu_i2info(inode);
+ 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;
+
+ CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", lli->lli_st_ino);
+
+ rc = obd_enqueue(sbi->ll_osc_exp, lli->lli_smd, LDLM_EXTENT, &policy,
+ LCK_PR, &flags, llu_extent_lock_callback,
+ ldlm_completion_ast, llu_glimpse_callback, inode,
+ sizeof(*lvb), lustre_swab_ost_lvb, &lockh);
+ if (rc > 0)
+ RETURN(-EIO);
+
+ lvb->lvb_size = lov_merge_size(lli->lli_smd, 0);
+ //inode->i_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
+
+ CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64"\n", lvb->lvb_size);
+
+ obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh);
+
RETURN(rc);
-#else
- return ELDLM_OK;
-#endif
}
-int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
- struct lov_stripe_md *lsm, int mode,
- struct lustre_handle *lockh)
+int llu_extent_lock(struct ll_file_data *fd, struct inode *inode,
+ struct lov_stripe_md *lsm, int mode,
+ ldlm_policy_data_t *policy, struct lustre_handle *lockh,
+ int ast_flags)
{
-#if 0
- struct ll_sb_info *sbi = ll_i2sbi(inode);
+ struct llu_sb_info *sbi = llu_i2sbi(inode);
+ struct llu_inode_info *lli = llu_i2info(inode);
int rc;
ENTRY;
+ LASSERT(lockh->cookie == 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);
- rc = obd_cancel(&sbi->ll_osc_conn, lsm, mode, lockh);
+ CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
+ lli->lli_st_ino, policy->l_extent.start, policy->l_extent.end);
+
+ rc = obd_enqueue(sbi->ll_osc_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;
+
+ if (policy->l_extent.start == 0 &&
+ policy->l_extent.end == OBD_OBJECT_EOF)
+ lli->lli_st_size = lov_merge_size(lsm, 1);
+
+ //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
RETURN(rc);
-#else
- return 0;
-#endif
}
-static int llu_brw(int cmd, struct inode *inode, struct page *page, int flags)
+#if 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)
{
+ struct llu_sb_info *sbi = llu_i2sbi(inode);
struct llu_inode_info *lli = llu_i2info(inode);
- struct lov_stripe_md *lsm = lli->lli_smd;
- struct brw_page pg;
int rc;
ENTRY;
- pg.pg = page;
- pg.off = ((obd_off)page->index) << PAGE_SHIFT;
-
- /* FIXME FIXME FIXME FIXME FIXME FIXME FIXME FIXME FIXME */
-#if 0
- if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > lli->lli_st_size))
- pg.count = lli->lli_st_size % PAGE_SIZE;
- else
-#endif
- pg.count = PAGE_SIZE;
+ LASSERT(lockh->cookie == 0);
- CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
- cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, lli->lli_st_ino,
- pg.off, pg.off);
- if (pg.count == 0) {
- LBUG();
- }
+ /* 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);
- pg.flag = flags;
+ CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
+ lli->lli_st_ino, extent->start, extent->end);
- rc = obd_brw(cmd, llu_i2obdconn(inode), lsm, 1, &pg, set, NULL);
- if (rc) {
- CERROR("error from obd_brw: rc = %d\n", rc);
- }
+ rc = obd_enqueue(sbi->ll_osc_exp, lsm, NULL, LDLM_EXTENT, extent,
+ sizeof(extent), mode, &ast_flags,
+ llu_extent_lock_callback, inode, lockh);
RETURN(rc);
}
-static int llu_prepare_write(struct inode *inode, struct page *page,
- unsigned from, unsigned to)
+/*
+ * 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,
+ int nonblock)
{
struct llu_inode_info *lli = llu_i2info(inode);
- obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
- int rc = 0;
+ struct obd_export *exp = llu_i2obdexp(inode);
+ struct ldlm_extent size_lock;
+ struct lustre_handle match_lockh = {0};
+ int flags, rc, matched;
+ int astflags = nonblock ? LDLM_FL_BLOCK_NOWAIT : 0;
ENTRY;
-#if 0
- if (!PageLocked(page))
- LBUG();
+ rc = llu_extent_lock_no_validate(fd, inode, lsm, mode, extent,
+ lockh, astflags);
+ if (rc != ELDLM_OK)
+ RETURN(rc);
- if (PageUptodate(page))
+ if (test_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags))
RETURN(0);
- //POISON(addr + from, 0xca, to - from);
+ rc = llu_inode_getattr(inode, lsm);
+ if (rc) {
+ llu_extent_unlock(fd, inode, lsm, mode, lockh);
+ RETURN(rc);
+ }
+
+ size_lock.start = lli->lli_st_size;
+ size_lock.end = OBD_OBJECT_EOF;
+
+ /* XXX I bet we should be checking the lock ignore flags.. */
+ /* FIXME use LDLM_FL_TEST_LOCK instead */
+ 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);
+
+ /* 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);
+ }
+
+ RETURN(0);
+}
#endif
- /* We're completely overwriting an existing page, so _don't_ set it up
- * to date until commit_write */
- if (from == 0 && to == PAGE_SIZE)
+
+int llu_extent_unlock(struct ll_file_data *fd, struct inode *inode,
+ struct lov_stripe_md *lsm, int mode,
+ struct lustre_handle *lockh)
+{
+ struct llu_sb_info *sbi = llu_i2sbi(inode);
+ int rc;
+ ENTRY;
+
+ /* 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);
- /* If are writing to a new page, no need to read old data.
- * the extent locking and getattr procedures in ll_file_write have
- * guaranteed that i_size is stable enough for our zeroing needs */
- if (lli->lli_st_size <= offset) {
- memset(kmap(page), 0, PAGE_SIZE);
- kunmap(page);
- GOTO(prepare_done, rc = 0);
+ rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
+
+ RETURN(rc);
+}
+
+#define LLAP_MAGIC 12346789
+
+struct ll_async_page {
+ int llap_magic;
+ void *llap_cookie;
+ int llap_queued;
+ struct page *llap_page;
+ 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;
+ ENTRY;
+
+ llap = llap_from_cookie(data);
+ if (IS_ERR(llap)) {
+ EXIT;
+ return;
}
- rc = llu_brw(OBD_BRW_READ, inode, page, 0);
+ inode = llap->llap_inode;
+ lsm = llu_i2info(inode)->lli_smd;
+ oa->o_id = lsm->lsm_object_id;
+ oa->o_valid = OBD_MD_FLID;
+ valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
+ if (cmd == OBD_BRW_WRITE)
+ valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
+
+ obdo_from_inode(oa, inode, valid_flags);
EXIT;
+}
- prepare_done:
- return rc;
+/* called for each page in a completed rpc.*/
+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_queued = 0;
+ page = llap->llap_page;
+
+ if (rc != 0) {
+ if (cmd == OBD_BRW_WRITE)
+ CERROR("writeback error on page %p index %ld: %d\n",
+ page, page->index, rc);
+ }
+ EXIT;
}
-static int llu_commit_write(struct inode *inode, struct page *page,
- unsigned from, unsigned to)
+static struct obd_async_page_ops llu_async_page_ops = {
+ .ap_make_ready = NULL,
+ .ap_refresh_count = NULL,
+ .ap_fill_obdo = llu_ap_fill_obdo,
+ .ap_completion = llu_ap_completion,
+};
+
+static
+struct llu_sysio_cookie* get_sysio_cookie(struct inode *inode, int maxpages)
{
- struct llu_inode_info *lli = llu_i2info(inode);
- loff_t size;
+ struct llu_sysio_cookie *cookie;
int rc;
+
+ OBD_ALLOC(cookie, LLU_SYSIO_COOKIE_SIZE(maxpages));
+ if (cookie == NULL)
+ goto out;
+
+ I_REF(inode);
+ cookie->lsc_inode = inode;
+ cookie->lsc_maxpages = maxpages;
+ cookie->lsc_llap = (struct ll_async_page *)(cookie + 1);
+ cookie->lsc_pages = (struct page *) (cookie->lsc_llap + maxpages);
+
+ rc = oig_init(&cookie->lsc_oig);
+ if (rc) {
+ OBD_FREE(cookie, LLU_SYSIO_COOKIE_SIZE(maxpages));
+ cookie = NULL;
+ }
+
+out:
+ 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_maxpages; i++) {
+ if (llap[i].llap_cookie)
+ obd_teardown_async_page(exp, lsm, NULL,
+ llap[i].llap_cookie);
+ }
+
+ I_RELE(cookie->lsc_inode);
+
+ oig_release(cookie->lsc_oig);
+ OBD_FREE(cookie, LLU_SYSIO_COOKIE_SIZE(cookie->lsc_maxpages));
+}
+
+static
+int llu_prep_async_io(struct llu_sysio_cookie *cookie, int cmd,
+ char *buf, loff_t pos, size_t count)
+{
+ struct llu_inode_info *lli = llu_i2info(cookie->lsc_inode);
+ struct lov_stripe_md *lsm = lli->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;
ENTRY;
-#if 0
- LASSERT(inode == file->f_dentry->d_inode);
- LASSERT(PageLocked(page));
-
- CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
- inode, page, from, to, page->index);
- CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu,from=%d,to=%d\n",
- inode->i_ino, from, to);
- /* to match full page case in prepare_write */
- SetPageUptodate(page);
- /* mark the page dirty, put it on mapping->dirty,
- * mark the inode PAGES_DIRTY, put it on sb->dirty */
- set_page_dirty(page);
-#endif
- rc = llu_brw(OBD_BRW_WRITE, inode, page, 0);
- if (rc)
- return rc;
- /* this is matched by a hack in obdo_to_inode at the moment */
- size = (((obd_off)page->index) << PAGE_SHIFT) + to;
- if (size > lli->lli_st_size)
- lli->lli_st_size = size;
+ if (!exp)
+ RETURN(-EINVAL);
+
+ /* prepare the pages array */
+ do {
+ unsigned long index, offset, bytes;
+
+ offset = (pos & ~PAGE_CACHE_MASK);
+ index = pos >> PAGE_CACHE_SHIFT;
+ bytes = PAGE_CACHE_SIZE - offset;
+ 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;
+
+ pages[npages]._offset = offset;
+ pages[npages]._count = bytes;
+
+ npages++;
+ count -= bytes;
+ pos += bytes;
+ buf += bytes;
+
+ cookie->lsc_rwcount += bytes;
+ } while (count);
+
+ cookie->lsc_npages = npages;
+
+ for (i = 0; i < npages; i++) {
+ llap[i].llap_magic = LLAP_MAGIC;
+ rc = obd_prep_async_page(exp, lsm, NULL, &pages[i],
+ (obd_off)pages[i].index << PAGE_SHIFT,
+ &llu_async_page_ops,
+ &llap[i], &llap[i].llap_cookie);
+ if (rc) {
+ llap[i].llap_cookie = NULL;
+ RETURN(rc);
+ }
+ CDEBUG(D_CACHE, "llap %p page %p cookie %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;
+
+ rc = obd_queue_group_io(exp, lsm, NULL, cookie->lsc_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)
+ RETURN(rc);
+
+ llap[i].llap_queued = 1;
+ }
RETURN(0);
-} /* ll_commit_write */
+}
-ssize_t
-llu_generic_file_write(struct inode *inode, const char *buf,
- size_t count, loff_t pos)
+static
+int llu_start_async_io(struct llu_sysio_cookie *cookie)
{
- struct page *page;
- ssize_t written;
- long status = 0;
- int err;
- unsigned bytes;
-
- if ((ssize_t) count < 0)
- return -EINVAL;
-#if 0
- down(&inode->i_sem);
-#endif
- if (pos < 0)
- return -EINVAL;
+ struct lov_stripe_md *lsm = llu_i2info(cookie->lsc_inode)->lli_smd;
+ struct obd_export *exp = llu_i2obdexp(cookie->lsc_inode);
- written = 0;
+ return obd_trigger_group_io(exp, lsm, NULL, cookie->lsc_oig);
+}
-#if 0
- remove_suid(inode);
- update_inode_times(inode);
-#endif
- do {
- unsigned long index, offset;
- char *kaddr;
-
- /*
- * Try to find the page in the cache. If it isn't there,
- * allocate a free page.
- */
- offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
- index = pos >> PAGE_CACHE_SHIFT;
- bytes = PAGE_CACHE_SIZE - offset;
- if (bytes > count) {
- bytes = count;
- }
-
- status = -ENOMEM; /* we'll assign it later anyway */
- page = __grab_cache_page(index);
- if (!page)
- break;
-
- kaddr = kmap(page);
- status = llu_prepare_write(inode, page, offset, offset+bytes);
- if (status)
- goto sync_failure;
-
- memcpy(kaddr+offset, buf, bytes);
-
- status = llu_commit_write(inode, page, offset, offset+bytes);
- if (!status)
- status = bytes;
-
- if (status >= 0) {
- written += status;
- count -= status;
- pos += status;
- buf += status;
- }
-unlock:
- kunmap(page);
- page_cache_release(page);
-
- if (status < 0)
- break;
- } while (count);
-done:
- err = written ? written : status;
+/*
+ * 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;
-#if 0
- up(&inode->i_sem);
-#endif
- return err;
-
- status = -EFAULT;
- goto unlock;
-
-sync_failure:
- /*
- * If blocksize < pagesize, prepare_write() may have instantiated a
- * few blocks outside i_size. Trim these off again.
- */
- kunmap(page);
- page_cache_release(page);
- goto done;
+ max_pages = (count >> PAGE_SHIFT) + 2;
+
+ cookie = get_sysio_cookie(inode, max_pages);
+ if (!cookie)
+ RETURN(ERR_PTR(-ENOMEM));
+
+ rc = llu_prep_async_io(cookie, cmd, buf, pos, count);
+ if (rc)
+ GOTO(out_cleanup, rc);
+
+ rc = llu_start_async_io(cookie);
+ if (rc)
+ GOTO(out_cleanup, rc);
+
+/*
+ rc = oig_wait(&oig);
+ if (rc) {
+ CERROR("file i/o error!\n");
+ rw_count = rc;
+ }
+*/
+ RETURN(cookie);
+
+out_cleanup:
+ put_sysio_cookie(cookie);
+ RETURN(ERR_PTR(rc));
}
-ssize_t llu_file_write(struct inode *inode, const struct iovec *iovec,
- size_t iovlen, loff_t pos)
+void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
+ obd_off size);
+
+struct llu_sysio_callback_args*
+llu_file_write(struct inode *inode, const struct iovec *iovec,
+ size_t iovlen, loff_t pos)
{
struct llu_inode_info *lli = llu_i2info(inode);
- struct ll_file_data *fd = lli->lli_file_data; /* XXX not ready don't use it now */
- struct lustre_handle lockh = { 0 };
+ struct ll_file_data *fd = lli->lli_file_data;
+ struct lustre_handle lockh = {0};
struct lov_stripe_md *lsm = lli->lli_smd;
- struct ldlm_extent extent;
+ struct obd_export *exp = NULL;
+ ldlm_policy_data_t policy;
+ struct llu_sysio_callback_args *lsca;
+ struct llu_sysio_cookie *cookie;
+ int astflag = (lli->lli_open_flags & O_NONBLOCK) ?
+ LDLM_FL_BLOCK_NOWAIT : 0;
ldlm_error_t err;
- ssize_t retval = 0;
+ int iovidx;
ENTRY;
/* XXX consider other types later */
if (!S_ISREG(lli->lli_st_mode))
LBUG();
-#if 0
- CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu,size="LPSZ",offset=%Ld\n",
- inode->i_ino, count, *ppos);
-
- /*
- * sleep doing some writeback work of this mount's dirty data
- * if the VM thinks we're low on memory.. other dirtying code
- * paths should think about doing this, too, but they should be
- * careful not to hold locked pages while they do so. like
- * ll_prepare_write. *cough*
- */
- ll_check_dirty(inode->i_sb);
-#endif
- while (iovlen--) {
- const char *buf = iovec[iovlen].iov_base;
- size_t count = iovec[iovlen].iov_len;
- /* POSIX, but surprised the VFS doesn't check this already */
+ LASSERT(iovlen <= MAX_IOVEC);
+
+ exp = llu_i2obdexp(inode);
+ if (exp == NULL)
+ RETURN(ERR_PTR(-EINVAL));
+
+ OBD_ALLOC(lsca, sizeof(*lsca));
+ if (!lsca)
+ RETURN(ERR_PTR(-ENOMEM));
+
+ /* FIXME optimize the following extent locking */
+ for (iovidx = 0; iovidx < iovlen; iovidx++) {
+ char *buf = (char*)iovec[iovidx].iov_base;
+ size_t count = iovec[iovidx].iov_len;
+
if (count == 0)
continue;
-#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;
- }
-#else
- extent.start = pos;
- extent.end = pos + count - 1;
-#endif
+ if (pos + count > lli->lli_maxbytes)
+ GOTO(err_out, err = -ERANGE);
+
+ /* FIXME libsysio haven't handle O_APPEND?? */
+ policy.l_extent.start = pos;
+ policy.l_extent.end = pos + count - 1;
- err = llu_extent_lock(fd, inode, lsm, LCK_PW, &extent, &lockh);
+ err = llu_extent_lock(fd, inode, lsm, LCK_PW, &policy,
+ &lockh, astflag);
if (err != ELDLM_OK)
- RETURN(-ENOLCK);
+ GOTO(err_out, err = -ENOLCK);
-#if 0
- if (!S_ISBLK(inode->i_mode) && file->f_flags & O_APPEND)
- *ppos = inode->i_size;
+ CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %llu\n",
+ lli->lli_st_ino, count, pos);
- CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
- inode->i_ino, count, *ppos);
-#endif
- retval += llu_generic_file_write(inode, buf, 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;
+ lov_increase_kms(exp, lsm, pos);
+ /* file size grow */
+ if (pos > lli->lli_st_size)
+ lli->lli_st_size = pos;
+ } else {
+ llu_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
+ GOTO(err_out, err = PTR_ERR(cookie));
+ }
+
+ /* XXX errors? */
+ err = llu_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
+ if (err)
+ CERROR("extent unlock error %d\n", err);
}
- /* XXX errors? */
- ll_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
- return(retval);
+ RETURN(lsca);
+
+err_out:
+ /* teardown all async stuff */
+ while (lsca->ncookies--) {
+ put_sysio_cookie(lsca->cookies[lsca->ncookies]);
+ }
+ OBD_FREE(lsca, sizeof(*lsca));
+
+ RETURN(ERR_PTR(err));
}
+#if 0
static void llu_update_atime(struct inode *inode)
{
-#if 0
struct llu_inode_info *lli = llu_i2info(inode);
#ifdef USE_ATIME
/* update atime, but don't explicitly write it out just this change */
inode->i_atime = CURRENT_TIME;
#endif
-#endif
}
+#endif
-static size_t llu_generic_file_read(struct inode *inode, char *buf,
- size_t count, loff_t pos)
+struct llu_sysio_callback_args*
+llu_file_read(struct inode *inode, const struct iovec *iovec,
+ size_t iovlen, loff_t pos)
{
struct llu_inode_info *lli = llu_i2info(inode);
- unsigned long index, offset;
- int error = 0;
- size_t readed = 0;
+ struct ll_file_data *fd = lli->lli_file_data;
+ struct lov_stripe_md *lsm = lli->lli_smd;
+ struct lustre_handle lockh = { 0 };
+ ldlm_policy_data_t policy;
+ struct llu_sysio_callback_args *lsca;
+ struct llu_sysio_cookie *cookie;
+ int astflag = (lli->lli_open_flags & O_NONBLOCK) ?
+ LDLM_FL_BLOCK_NOWAIT : 0;
+ __u64 kms;
+ int iovidx;
- index = pos >> PAGE_CACHE_SHIFT;
- offset = pos & ~PAGE_CACHE_MASK;
+ ldlm_error_t err;
+ ENTRY;
- do {
- struct page *page;
- unsigned long end_index, nr;
-
- end_index = lli->lli_st_size >> PAGE_CACHE_SHIFT;
-
- if (index > end_index)
- break;
- nr = PAGE_CACHE_SIZE;
- if (index == end_index) {
- nr = lli->lli_st_size & ~PAGE_CACHE_MASK;
- if (nr <= offset)
- break;
- }
-
- nr = nr - offset;
- if (nr > count)
- nr = count;
-
- page = grab_cache_page(index);
- if (!page) {
- error = -ENOMEM;
- break;
+ 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;
+
+ /* "If nbyte is 0, read() will return 0 and have no other results."
+ * -- Single Unix Spec */
+ if (count == 0)
+ continue;
+
+ policy.l_extent.start = pos;
+ policy.l_extent.end = pos + count - 1;
+
+ err = llu_extent_lock(fd, inode, lsm, LCK_PR, &policy,
+ &lockh, astflag);
+ if (err != ELDLM_OK)
+ GOTO(err_out, err = -ENOLCK);
+
+ 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 */
+ struct ost_lvb lvb;
+ if (llu_glimpse_size(inode, &lvb)) {
+ llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
+ GOTO(err_out, err = -ENOLCK);
+ }
+ lli->lli_st_size = lvb.lvb_size;
+ } else {
+ lli->lli_st_size = kms;
}
- error = llu_brw(OBD_BRW_READ, inode, page, 0);
- if (error) {
- page_cache_release(page);
+ CDEBUG(D_INFO, "Reading inode %lu, "LPSZ" bytes, offset %Ld, "
+ "i_size "LPU64"\n", lli->lli_st_ino, count, pos,
+ lli->lli_st_size);
+
+ if (pos >= lli->lli_st_size) {
+ llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
break;
- }
+ }
- memcpy(buf, kmap(page)+offset, nr);
- offset += nr;
- index += offset >> PAGE_CACHE_SHIFT;
- offset &= ~PAGE_CACHE_MASK;
- readed += nr;
- count -= nr;
+ 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));
+ }
- page_cache_release(page);
- } while (count);
+ /* XXX errors? */
+ err = llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
+ if (err)
+ CERROR("extent_unlock fail: %d\n", err);
+ }
+#if 0
+ if (readed > 0)
+ llu_update_atime(inode);
+#endif
+ RETURN(lsca);
+
+err_out:
+ /* teardown all async stuff */
+ while (lsca->ncookies--) {
+ put_sysio_cookie(lsca->cookies[lsca->ncookies]);
+ }
+ OBD_FREE(lsca, sizeof(*lsca));
- if (error)
- return error;
- return readed;
+ RETURN(ERR_PTR(err));
}
-ssize_t llu_file_read(struct inode *inode, const struct iovec *iovec,
- size_t iovlen, loff_t pos)
+int llu_iop_iodone(struct ioctx *ioctxp)
{
- 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 };
-#if 0
- struct ll_read_extent rextent;
-#else
- struct ldlm_extent extent;
-#endif
- ldlm_error_t err;
- ssize_t retval = 0;
+ struct llu_sysio_callback_args *lsca = ioctxp->ioctx_private;
+ struct llu_sysio_cookie *cookie;
+ int i, err = 0, rc = 0;
ENTRY;
- while (iovlen--) {
- char *buf = iovec[iovlen].iov_base;
- size_t count = iovec[iovlen].iov_len;
+ /* write/read(fd, buf, 0) */
+ if (!lsca) {
+ ioctxp->ioctx_cc = 0;
+ RETURN(1);
+ }
- /* "If nbyte is 0, read() will return 0 and have no other results."
- * -- Single Unix Spec */
- if (count == 0)
- RETURN(0);
+ LASSERT(!IS_ERR(lsca));
+
+ for (i = 0; i < lsca->ncookies; i++) {
+ cookie = lsca->cookies[i];
+ if (cookie) {
+ err = oig_wait(cookie->lsc_oig);
+ if (err && !rc)
+ rc = err;
+ if (!rc)
+ ioctxp->ioctx_cc += cookie->lsc_rwcount;
+ put_sysio_cookie(cookie);
+ }
+ }
-#if 0
- rextent.re_extent.start = pos;
- rextent.re_extent.end = pos + count - 1;
-#else
- extent.start = pos;
- extent.end = pos + count - 1;
-#endif
- err = llu_extent_lock(fd, inode, lsm, LCK_PR, &extent, &lockh);
- if (err != ELDLM_OK)
- RETURN(-ENOLCK);
-#if 0
- rextent.re_task = current;
- spin_lock(&lli->lli_read_extent_lock);
- list_add(&rextent.re_lli_item, &lli->lli_read_extents);
- spin_unlock(&lli->lli_read_extent_lock);
-#endif
- CDEBUG(D_INFO, "Reading inode %lu, "LPSZ" bytes, offset %Ld\n",
- lli->lli_st_ino, count, pos);
- retval = llu_generic_file_read(inode, buf, count, pos);
-#if 0
- spin_lock(&lli->lli_read_extent_lock);
- list_del(&rextent.re_lli_item);
- spin_unlock(&lli->lli_read_extent_lock);
-#endif
+ if (rc) {
+ LASSERT(rc < 0);
+ ioctxp->ioctx_cc = -1;
+ ioctxp->ioctx_errno = -rc;
}
- if (retval > 0)
- llu_update_atime(inode);
+ OBD_FREE(lsca, sizeof(*lsca));
+ ioctxp->ioctx_private = NULL;
- /* XXX errors? */
- ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
- RETURN(retval);
+ RETURN(1);
}
-
git://git.whamcloud.com / fs / lustre-release.git / blobdiff