* along with Lustre; if not, write to the Free Software
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
*/
-#ifdef HAVE_KERNEL_CONFIG_H
+
+#ifndef AUTOCONF_INCLUDED
#include <linux/config.h>
#endif
#include <linux/kernel.h>
#include <linux/writeback.h>
#include <linux/stat.h>
#include <asm/uaccess.h>
-#include <asm/segment.h>
+#ifdef HAVE_SEGMENT_H
+# include <asm/segment.h>
+#endif
#include <linux/mm.h>
#include <linux/pagemap.h>
#include <linux/smp_lock.h>
#define DEBUG_SUBSYSTEM S_LLITE
+//#include <lustre_mdc.h>
#include <lustre_lite.h>
#include "llite_internal.h"
#include <linux/lustre_compat25.h>
#else
static void ll_invalidatepage(struct page *page, unsigned long offset)
{
- if (offset)
- return;
- if (PagePrivate(page))
+ if (offset == 0 && PagePrivate(page))
ll_removepage(page);
}
#endif
-
-static int ll_releasepage(struct page *page, gfp_t gfp_mask)
+static int ll_releasepage(struct page *page, int gfp_mask)
{
if (PagePrivate(page))
ll_removepage(page);
return 1;
}
+static int ll_set_page_dirty(struct page *page)
+{
+ struct ll_async_page *llap;
+ ENTRY;
+
+ llap = llap_from_page(page, LLAP_ORIGIN_UNKNOWN);
+ if (IS_ERR(llap))
+ RETURN(PTR_ERR(llap));
+
+ llap_write_pending(page->mapping->host, llap);
+ RETURN(__set_page_dirty_nobuffers(page));
+}
+
#define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
static inline int ll_get_user_pages(int rw, unsigned long user_addr,
size_t size, struct page ***pages)
{
int result = -ENOMEM;
- unsigned long page_count;
+ int page_count;
/* set an arbitrary limit to prevent arithmetic overflow */
if (size > MAX_DIRECTIO_SIZE) {
page_count = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
page_count -= user_addr >> CFS_PAGE_SHIFT;
- OBD_ALLOC_GFP(*pages, page_count * sizeof(**pages), GFP_KERNEL);
+ OBD_ALLOC_WAIT(*pages, page_count * sizeof(**pages));
if (*pages) {
down_read(¤t->mm->mmap_sem);
result = get_user_pages(current, current->mm, user_addr,
page_count, (rw == READ), 0, *pages,
NULL);
up_read(¤t->mm->mmap_sem);
+ if (result < 0)
+ OBD_FREE(*pages, page_count * sizeof(**pages));
}
return result;
OBD_FREE(pages, npages * sizeof(*pages));
}
-static ssize_t ll_direct_IO_26_seg(int rw, struct file *file,
+static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
struct address_space *mapping,
- struct inode *inode,
- struct lov_stripe_md *lsm,
- unsigned long user_addr, size_t size,
- loff_t file_offset, struct page **pages,
- int page_count)
+ struct obd_info *oinfo,
+ struct ptlrpc_request_set *set,
+ size_t size, loff_t file_offset,
+ struct page **pages, int page_count)
{
struct brw_page *pga;
- struct obdo oa;
int i, rc = 0;
size_t length;
ENTRY;
OBD_ALLOC(pga, sizeof(*pga) * page_count);
- if (!pga)
+ if (!pga) {
+ CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
+ (int)sizeof(*pga), page_count);
RETURN(-ENOMEM);
+ }
for (i = 0, length = size; length > 0;
length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
pga[i].pg = pages[i];
pga[i].off = file_offset;
/* To the end of the page, or the length, whatever is less */
- pga[i].count = min_t(int, CFS_PAGE_SIZE -(file_offset & ~CFS_PAGE_MASK),
+ pga[i].count = min_t(int, CFS_PAGE_SIZE -
+ (file_offset & ~CFS_PAGE_MASK),
length);
pga[i].flag = 0;
if (rw == READ)
POISON_PAGE(pages[i], 0x0d);
}
- ll_inode_fill_obdo(inode, rw, &oa);
-
- if (rw == WRITE)
- lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
- LPROC_LL_DIRECT_WRITE, size);
- else
- lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
- LPROC_LL_DIRECT_READ, size);
- rc = obd_brw_rqset(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
- ll_i2obdexp(inode), &oa, lsm, page_count, pga, NULL);
- if (rc == 0) {
+ rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
+ ll_i2dtexp(inode), oinfo, page_count,
+ pga, NULL, set);
+ if (rc == 0)
rc = size;
- if (rw == WRITE) {
- lov_stripe_lock(lsm);
- obd_adjust_kms(ll_i2obdexp(inode), lsm, file_offset, 0);
- lov_stripe_unlock(lsm);
- }
- }
OBD_FREE(pga, sizeof(*pga) * page_count);
RETURN(rc);
}
+/* This is the maximum size of a single O_DIRECT request, based on a 128kB
+ * kmalloc limit. We need to fit all of the brw_page structs, each one
+ * representing PAGE_SIZE worth of user data, into a single buffer, and
+ * then truncate this to be a full-sized RPC. This is 22MB for 4kB pages. */
+#define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
+ ~(PTLRPC_MAX_BRW_SIZE - 1))
static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
const struct iovec *iov, loff_t file_offset,
unsigned long nr_segs)
{
struct file *file = iocb->ki_filp;
+ struct inode *inode = file->f_mapping->host;
ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
- struct ll_inode_info *lli = ll_i2info(file->f_mapping->host);
- unsigned long seg;
+ struct ll_inode_info *lli = ll_i2info(inode);
+ struct lov_stripe_md *lsm = lli->lli_smd;
+ struct ptlrpc_request_set *set;
+ struct obd_info oinfo;
+ struct obdo oa;
+ unsigned long seg = 0;
+ size_t size = MAX_DIO_SIZE;
+ int opc;
ENTRY;
if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
RETURN(-EBADF);
- /* FIXME: io smaller than CFS_PAGE_SIZE is broken on ia64 ??? */
- if ((file_offset & (~CFS_PAGE_MASK)) || (count & ~CFS_PAGE_MASK))
+ /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
+ if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
RETURN(-EINVAL);
+ CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
+ "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
+ inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
+ file_offset, file_offset, count >> CFS_PAGE_SHIFT,
+ MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
+
+ if (rw == WRITE) {
+ ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
+ opc = CAPA_OPC_OSS_WRITE;
+ llap_write_pending(inode, NULL);
+ } else {
+ ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
+ opc = CAPA_OPC_OSS_RW;
+ }
+
/* Check that all user buffers are aligned as well */
for (seg = 0; seg < nr_segs; seg++) {
- unsigned long user_addr = (unsigned long)iov[seg].iov_base;
- size_t size = iov[seg].iov_len;
- if ((user_addr & ~CFS_PAGE_MASK) || (size & ~CFS_PAGE_MASK))
+ if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
+ (iov[seg].iov_len & ~CFS_PAGE_MASK))
RETURN(-EINVAL);
}
- seg = 0;
- while ((seg < nr_segs) && (tot_bytes >= 0)) {
- const struct iovec *vec = &iov[seg++];
- unsigned long user_addr = (unsigned long)vec->iov_base;
- size_t size = vec->iov_len;
- struct page **pages;
- int page_count;
- ssize_t result;
-
- page_count = ll_get_user_pages(rw, user_addr, size, &pages);
- if (page_count < 0) {
- ll_free_user_pages(pages, 0, 0);
- if (tot_bytes > 0)
- break;
- return page_count;
- }
+ set = ptlrpc_prep_set();
+ if (set == NULL)
+ RETURN(-ENOMEM);
- result = ll_direct_IO_26_seg(rw, file, file->f_mapping,
- file->f_mapping->host,
- lli->lli_smd, user_addr, size,
- file_offset, pages, page_count);
- ll_free_user_pages(pages, page_count, rw == READ);
+ ll_inode_fill_obdo(inode, rw, &oa);
+ oinfo.oi_oa = &oa;
+ oinfo.oi_md = lsm;
+ oinfo.oi_capa = ll_osscapa_get(inode, opc);
- if (result <= 0) {
- if (tot_bytes > 0)
+ /* need locking between buffered and direct access. and race with
+ *size changing by concurrent truncates and writes. */
+ if (rw == READ)
+ LOCK_INODE_MUTEX(inode);
+
+ for (seg = 0; seg < nr_segs; seg++) {
+ size_t iov_left = iov[seg].iov_len;
+ unsigned long user_addr = (unsigned long)iov[seg].iov_base;
+
+ if (rw == READ) {
+ if (file_offset >= inode->i_size)
break;
- return result;
+ if (file_offset + iov_left > inode->i_size)
+ iov_left = inode->i_size - file_offset;
}
- tot_bytes += result;
- file_offset += result;
- if (result < size)
- break;
+ while (iov_left > 0) {
+ struct page **pages;
+ int page_count;
+ ssize_t result;
+
+ page_count = ll_get_user_pages(rw, user_addr,
+ min(size, iov_left),
+ &pages);
+ LASSERT(page_count != 0);
+ if (page_count > 0) {
+ result = ll_direct_IO_26_seg(rw, inode,
+ file->f_mapping,
+ &oinfo, set,
+ min(size,iov_left),
+ file_offset, pages,
+ page_count);
+ ll_free_user_pages(pages, page_count, rw==READ);
+ } else {
+ result = 0;
+ }
+ if (page_count < 0 || result <= 0) {
+ /* If we can't allocate a large enough buffer
+ * for the request, shrink it to a smaller
+ * PAGE_SIZE multiple and try again.
+ * We should always be able to kmalloc for a
+ * page worth of page pointers = 4MB on i386. */
+ if ((page_count == -ENOMEM||result == -ENOMEM)&&
+ size > (CFS_PAGE_SIZE / sizeof(*pages)) *
+ CFS_PAGE_SIZE) {
+ size = ((((size / 2) - 1) |
+ ~CFS_PAGE_MASK) + 1) &
+ CFS_PAGE_MASK;
+ CDEBUG(D_VFSTRACE, "DIO size now %u\n",
+ (int)size);
+ continue;
+ }
+
+ if (tot_bytes <= 0)
+ tot_bytes = page_count < 0 ? page_count : result;
+ GOTO(out, tot_bytes);
+ }
+
+ tot_bytes += result;
+ file_offset += result;
+ iov_left -= result;
+ user_addr += result;
+ }
+ }
+out:
+ if (rw == READ)
+ UNLOCK_INODE_MUTEX(inode);
+
+ if (tot_bytes > 0) {
+ int rc;
+
+ rc = ptlrpc_set_wait(set);
+ if (rc) {
+ tot_bytes = rc;
+ } else if (rw == WRITE) {
+ lov_stripe_lock(lsm);
+ obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
+ lov_stripe_unlock(lsm);
+ }
}
- return tot_bytes;
+
+ capa_put(oinfo.oi_capa);
+ ptlrpc_set_destroy(set);
+ RETURN(tot_bytes);
}
struct address_space_operations ll_aops = {
.direct_IO = ll_direct_IO_26,
.writepage = ll_writepage_26,
.writepages = generic_writepages,
- .set_page_dirty = __set_page_dirty_nobuffers,
+ .set_page_dirty = ll_set_page_dirty,
.sync_page = NULL,
.prepare_write = ll_prepare_write,
.commit_write = ll_commit_write,
.releasepage = ll_releasepage,
.bmap = NULL
};
-
-static int wait_on_page_locked_range(struct address_space *mapping,
- pgoff_t start, pgoff_t end)
-{
- pgoff_t index;
- struct page *page;
- int ret = 0;
-
- if (end < start)
- return 0;
-
- for(index = start; index < end; index++) {
- page = find_get_page(mapping, index);
- if (page == NULL)
- continue;
-
- wait_on_page_locked(page);
- if (PageError(page))
- ret = -EIO;
- cond_resched();
- }
-
- /* Check for outstanding write errors */
- if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
- ret = -ENOSPC;
- if (test_and_clear_bit(AS_EIO, &mapping->flags))
- ret = -EIO;
-
- return ret;
-}
-
-int ll_sync_page_range(struct inode *inode, struct address_space *mapping,
- loff_t pos, size_t count)
-{
- pgoff_t start = pos >> CFS_PAGE_SHIFT;
- pgoff_t end = (pos + count - 1) >> CFS_PAGE_SHIFT;
- struct writeback_control wbc;
- int ret;
-
- wbc.sync_mode = WB_SYNC_ALL;
- wbc.nr_to_write = mapping->nrpages * 2;
- wbc.start = start;
- wbc.end = end;
- ret = generic_writepages(mapping, &wbc);
-
- if (ret == 0)
- ret = wait_on_page_locked_range(mapping, start, end);
- return ret;
-}
-