Viewing: rw26.c
// SPDX-License-Identifier: GPL-2.0
/*
* Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
* Copyright (c) 2011, 2017, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
*
* Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
*/
#include <linux/buffer_head.h>
#include <linux/errno.h>
#include <linux/fs.h>
#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/mpage.h>
#include <linux/pagemap.h>
#include <linux/string.h>
#include <linux/unistd.h>
#include <linux/writeback.h>
#include <linux/migrate.h>
#include <lustre_compat/linux/uio.h>
#define DEBUG_SUBSYSTEM S_LLITE
#include "llite_internal.h"
#ifdef HAVE_INVALIDATE_FOLIO
/**
* ll_invalidate_folio() - Implements Linux VM address_space::invalidate_folio()
* method. This method is called when the folio is truncated from a file, either
* as a result of explicit truncate, or when inode is removed from memory
* (as a result of final iput(), umount, or memory pressure induced icache
* shrinking).
* @folio: Pointer to folio struct (collection of pages)
* @offset: Starting offset in bytes
* @len: length of folio to be invalidated
*
* [0, off] bytes of the folio remain valid (this is for a case of non-page
* aligned truncate). Lustre leaves partially truncated folios in the cache,
* relying on struct inode::i_size to limit further accesses.
*/
static void ll_invalidate_folio(struct folio *folio, size_t offset, size_t len)
{
struct inode *inode;
struct lu_env *env;
struct cl_page *page;
struct cl_object *obj;
LASSERT(!folio_test_writeback(folio));
LASSERT(folio_test_locked(folio));
if (!(offset == 0 && len == folio_size(folio)) &&
!folio_test_large(folio))
return;
/* Drop the pages from the folio */
env = cl_env_percpu_get();
LASSERT(!IS_ERR(env));
inode = folio_inode(folio);
obj = ll_i2info(inode)->lli_clob;
if (obj != NULL) {
int n, npgs = folio_nr_pages(folio);
for (n = 0; n < npgs; n++) {
struct page *vmpage = folio_page(folio, n);
LASSERT(PageLocked(vmpage));
LASSERT(!PageWriteback(vmpage));
page = cl_vmpage_page(vmpage, obj);
if (page != NULL) {
cl_page_delete(env, page);
cl_page_put(env, page);
}
}
} else {
LASSERT(!folio_get_private(folio));
}
cl_env_percpu_put(env);
}
#else
/**
* ll_invalidatepage() - Implements Linux VM address_space::invalidatepage()
* method. This method is called when the page is truncate from a file, either
* as a result of explicit truncate, or when inode is removed from memory
* (as a result of final iput(), umount, or memory pressure induced icache
* shrinking).
* @vmpage: pointer to struct page (single page)
* @offset: Starting offset in bytes
* @length: Length to release
*
* [0, offset] bytes of the page remain valid (this is for a case of not-page
* aligned truncate). Lustre leaves partially truncated page in the cache,
* relying on struct inode::i_size to limit further accesses.
*/
static void ll_invalidatepage(struct page *vmpage,
unsigned int offset, unsigned int length)
{
struct inode *inode;
struct lu_env *env;
struct cl_page *page;
struct cl_object *obj;
LASSERT(PageLocked(vmpage));
LASSERT(!PageWriteback(vmpage));
/*
* It is safe to not check anything in invalidatepage/releasepage
* below because they are run with page locked and all our io is
* happening with locked page too
*/
if (offset == 0 && length == PAGE_SIZE) {
/* See the comment in ll_releasepage() */
env = cl_env_percpu_get();
LASSERT(!IS_ERR(env));
inode = vmpage->mapping->host;
obj = ll_i2info(inode)->lli_clob;
if (obj != NULL) {
page = cl_vmpage_page(vmpage, obj);
if (page != NULL) {
cl_page_delete(env, page);
cl_page_put(env, page);
}
} else
LASSERT(vmpage->private == 0);
cl_env_percpu_put(env);
}
if (CFS_FAIL_PRECHECK(OBD_FAIL_LLITE_PAGE_INVALIDATE_PAUSE)) {
unlock_page(vmpage);
CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_PAGE_INVALIDATE_PAUSE,
cfs_fail_val);
lock_page(vmpage);
}
}
#endif
static bool do_release_page(struct page *vmpage, gfp_t wait)
{
struct address_space *mapping;
struct cl_object *obj;
struct cl_page *page;
struct lu_env *env;
int result = 0;
ENTRY;
LASSERT(PageLocked(vmpage));
if (PageWriteback(vmpage) || PageDirty(vmpage))
RETURN(0);
mapping = vmpage->mapping;
if (mapping == NULL)
RETURN(1);
obj = ll_i2info(mapping->host)->lli_clob;
if (obj == NULL)
RETURN(1);
page = cl_vmpage_page(vmpage, obj);
if (page == NULL)
RETURN(1);
env = cl_env_percpu_get();
LASSERT(!IS_ERR(env));
if (!cl_page_in_use(page)) {
result = 1;
cl_page_delete(env, page);
}
/* To use percpu env array, the call path can not be rescheduled;
* otherwise percpu array will be messed if ll_releaspage() called
* again on the same CPU.
*
* If this page holds the last refc of cl_object, the following
* call path may cause reschedule:
* cl_page_put -> cl_page_free -> cl_object_put ->
* lu_object_put -> lu_object_free -> lov_delete_raid0.
*
* However, the kernel can't get rid of this inode until all pages have
* been cleaned up. Now that we hold page lock here, it's pretty safe
* that we won't get into object delete path.
*/
LASSERT(cl_object_refc(obj) > 1);
cl_page_put(env, page);
cl_env_percpu_put(env);
RETURN(result);
}
#ifdef HAVE_AOPS_RELEASE_FOLIO
static bool ll_release_folio(struct folio *folio, gfp_t wait)
{
struct page *vmpage = folio_page(folio, 0);
/* folio_nr_pages(folio) == 1 is fixed with grab_cache_page* */
BUG_ON(folio_nr_pages(folio) != 1);
return do_release_page(vmpage, wait);
}
#else /* !HAVE_AOPS_RELEASE_FOLIO */
#ifdef HAVE_RELEASEPAGE_WITH_INT
#define RELEASEPAGE_ARG_TYPE int
#else
#define RELEASEPAGE_ARG_TYPE gfp_t
#endif
static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
{
return do_release_page(vmpage, gfp_mask);
}
#endif /* HAVE_AOPS_RELEASE_FOLIO */
/*
* Lustre could relax a bit for alignment, io count is not
* necessary page alignment.
*/
bool ll_iov_iter_is_unaligned(struct iov_iter *i)
{
size_t orig_size = i->count;
size_t count = orig_size & ~PAGE_MASK;
unsigned long res;
if (iov_iter_count(i) & ~PAGE_MASK)
return true;
if (!iov_iter_is_aligned(i, ~PAGE_MASK, 0))
return true;
if (!count)
return iov_iter_alignment(i) & ~PAGE_MASK;
if (orig_size > PAGE_SIZE) {
iov_iter_truncate(i, orig_size - count);
res = iov_iter_alignment(i);
iov_iter_reexpand(i, orig_size);
return res & ~PAGE_MASK;
}
res = iov_iter_alignment(i);
/* start address is page aligned */
if ((res & ~PAGE_MASK) == orig_size)
return false;
return res & ~PAGE_MASK;
}
static int
ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io, size_t size,
int rw, struct inode *inode, struct cl_sub_dio *sdio)
{
struct cl_dio_pages *cdp = &sdio->csd_dio_pages;
struct cl_sync_io *anchor = &sdio->csd_sync;
struct cl_object *obj = io->ci_obj;
struct cl_page *page;
int iot = rw == READ ? CRT_READ : CRT_WRITE;
loff_t offset = cdp->cdp_file_offset;
ssize_t rc = 0;
unsigned int i = 0;
ENTRY;
while (size > 0) {
size_t from = offset & ~PAGE_MASK;
size_t to = min(from + size, PAGE_SIZE);
page = cl_page_find(env, obj, offset >> PAGE_SHIFT,
cdp->cdp_pages[i], CPT_TRANSIENT);
if (IS_ERR(page))
GOTO(out, rc = PTR_ERR(page));
LASSERT(page->cp_type == CPT_TRANSIENT);
page->cp_sync_io = anchor;
if (inode && IS_ENCRYPTED(inode)) {
/* In case of Direct IO on encrypted file, we need to
* add a reference to the inode on the cl_page.
* This info is required by llcrypt to proceed
* to encryption/decryption.
* This is safe because we know these pages are private
* to the thread doing the Direct IO.
*/
page->cp_inode = inode;
}
cdp->cdp_cl_pages[i] = page;
/*
* Call page clip for incomplete pages, to set range of bytes
* in the page and to tell transfer formation engine to send
* the page even if it is beyond KMS (ie, don't trim IO to KMS)
*/
if (from != 0 || to != PAGE_SIZE)
cl_page_clip(env, page, from, to);
i++;
offset += to - from;
size -= to - from;
}
/* on success, we should hit every page in the cdp and have no bytes
* left in 'size'
*/
LASSERT(i == cdp->cdp_page_count);
LASSERT(size == 0);
atomic_add(cdp->cdp_page_count, &anchor->csi_sync_nr);
/*
* Avoid out-of-order execution of adding inflight
* modifications count and io submit.
*/
smp_mb();
rc = cl_dio_submit_rw(env, io, iot, cdp);
if (rc != 0) {
atomic_add(-cdp->cdp_page_count,
&anchor->csi_sync_nr);
for (i = 0; i < cdp->cdp_page_count; i++) {
page = cdp->cdp_cl_pages[i];
page->cp_sync_io = NULL;
}
}
out:
/* cleanup of the page array is handled by cl_sub_dio_end, so there's
* no work to do on error here
*/
RETURN(rc);
}
/* This is the maximum size of a single O_DIRECT request, based on the
* 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. For 4kB PAGE_SIZE this is
* up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc.
*/
#define MAX_DIO_SIZE ((KMALLOC_MAX_SIZE / sizeof(struct brw_page) * PAGE_SIZE) & \
~((size_t)DT_MAX_BRW_SIZE - 1))
static ssize_t ll_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
{
struct ll_cl_context *lcc;
const struct lu_env *env;
struct cl_io *io;
struct file *file = iocb->ki_filp;
struct inode *inode = file->f_mapping->host;
struct cl_dio_aio *ll_dio_aio;
struct cl_sub_dio *sdio;
size_t bytes = iov_iter_count(iter);
ssize_t tot_bytes = 0, result = 0;
ssize_t bytes_at_drain = 0;
loff_t file_offset = iocb->ki_pos;
int rw = iov_iter_rw(iter);
bool sync_submit = false;
bool unaligned;
struct vvp_io *vio;
ssize_t rc2;
ENTRY;
if (file_offset & ~PAGE_MASK)
unaligned = true;
else
unaligned = ll_iov_iter_is_unaligned(iter);
lcc = ll_cl_find(inode);
if (lcc == NULL)
RETURN(-EIO);
env = lcc->lcc_env;
LASSERT(!IS_ERR(env));
vio = vvp_env_io(env);
io = lcc->lcc_io;
LASSERT(io != NULL);
CDEBUG(D_VFSTRACE,
"VFS Op:inode="DFID"(%p), size=%zd (max %lu), offset=%lld=%#llx, pages %zd (max %lu)%s%s%s%s\n",
PFID(ll_inode2fid(inode)), inode, bytes, MAX_DIO_SIZE,
file_offset, file_offset,
(bytes >> PAGE_SHIFT) + !!(bytes & ~PAGE_MASK),
MAX_DIO_SIZE >> PAGE_SHIFT,
io->ci_dio_lock ? ", locked" : ", lockless",
io->ci_parallel_dio ? ", parallel" : "",
unaligned ? ", unaligned" : "",
io->ci_hybrid_switched ? ", hybrid" : "");
/* Check EOF by ourselves.
* For parity IO, use ci_parity_eof which is calculated from RAID
* geometry in the LOV layer.
*/
if (rw == READ) {
loff_t eof = io->ci_parity_io ? io->ci_parity_eof :
i_size_read(inode);
if (file_offset >= eof)
RETURN(0);
}
/* if one part of an I/O is unaligned, just handle all of it that way -
* otherwise we create significant complexities with managing the iovec
* in different ways, etc, all for very marginal benefits
*/
if (unaligned)
io->ci_unaligned_dio = true;
if (io->ci_unaligned_dio)
unaligned = true;
ll_dio_aio = io->ci_dio_aio;
LASSERT(ll_dio_aio);
LASSERT(ll_dio_aio->cda_iocb == iocb);
/* unaligned DIO support can be turned off, so is it on? */
if (unaligned && !ll_sbi_has_unaligned_dio(ll_i2sbi(inode)))
RETURN(-EINVAL);
/* unaligned AIO is not supported - see LU-18032 */
if (unaligned && ll_dio_aio->cda_is_aio)
RETURN(-EINVAL);
/* the requirement to not return EIOCBQUEUED for pipes (see bottom of
* this function) plays havoc with the unaligned I/O lifecycle, so
* don't allow unaligned I/O on pipes.
*
* Additionally, pipe iterators don't have user pages that can be
* pinned for DIO - iov_iter_get_pages_alloc2() will fail or return 0
* for pipes, so reject all pipe iterators for DIO and fall back to
* buffered I/O.
*/
if (iov_iter_is_pipe(iter))
RETURN(0);
/* returning 0 here forces the remaining I/O through buffered I/O
* while returning -EINVAL stops the I/O from continuing
*/
/* Unpatched older servers which cannot safely support unaligned DIO
* should abort here
*/
if (unaligned && !cl_io_top(io)->ci_allow_unaligned_dio)
RETURN(0);
/* We cannot do parallel submission of sub-I/Os - for AIO or regular
* DIO - unless lockless because it causes us to release the lock
* early.
*
* There are also several circumstances in which we must disable
* parallel DIO, so we check if it is enabled.
*
* The check for "is_sync_kiocb" excludes AIO, which does not need to
* be disabled in these situations.
*/
if (io->ci_dio_lock || (is_sync_kiocb(iocb) && !io->ci_parallel_dio))
sync_submit = true;
while (iov_iter_count(iter)) {
struct cl_dio_pages *cdp;
bytes = min_t(size_t, iov_iter_count(iter), MAX_DIO_SIZE);
/* Cap sub_dio size for drain+retry testing */
if (CFS_FAIL_PRECHECK(OBD_FAIL_LLITE_DIO_DRAIN_RETRY))
bytes = min_t(size_t, bytes, PAGE_SIZE);
/* For parity IO, use ci_parity_eof which is calculated from
* RAID geometry in the LOV layer.
*/
if (rw == READ) {
loff_t eof = io->ci_parity_io ? io->ci_parity_eof :
i_size_read(inode);
if (file_offset >= eof)
break;
if (file_offset + bytes > eof)
bytes = eof - file_offset;
}
/* if we are doing sync_submit, then we free this below,
* otherwise it is freed on the final call to cl_sync_io_note
* (either in this function or from a ptlrpcd daemon)
*/
sdio = cl_sub_dio_alloc(ll_dio_aio, iter, rw == WRITE,
unaligned, sync_submit);
if (!sdio)
GOTO(out, result = -ENOMEM);
cdp = &sdio->csd_dio_pages;
cdp->cdp_file_offset = file_offset;
result = cl_dio_pages_init(env, ll_dio_aio->cda_obj, cdp,
iter, rw, bytes, file_offset,
unaligned);
if (unlikely(result <= 0)) {
bool retry = (result == -ENOMEM && unaligned
&& tot_bytes > bytes_at_drain);
/* Note the failed sub_dio. When retrying,
* pass rc=0 so the alloc ENOMEM doesn't
* poison the parent anchor's sync_rc.
*/
cl_sync_io_note(env, &sdio->csd_sync,
retry ? 0 : result);
if (sync_submit) {
LASSERT(sdio->csd_creator_free);
cl_sub_dio_free(sdio);
}
if (retry) {
/* ENOMEM but we have in-flight sub_dios
* holding pages. Drain them to reclaim
* pages, then retry.
*
* By calling cl_sync_io_wait_recycle,
* cl_dio_aio_end runs — but unaligned
* DIO is never AIO, so it won't
* prematurely complete to userspace.
*/
LASSERT(!ll_dio_aio->cda_is_aio);
rc2 = cl_sync_io_wait_recycle(env,
&ll_dio_aio->cda_sync, 0, 0);
if (rc2 < 0)
GOTO(out, result = rc2);
bytes_at_drain = tot_bytes;
result = 0;
CDEBUG(D_VFSTRACE,
"DIO pool ENOMEM, drained at %zd bytes, retrying\n",
tot_bytes);
continue;
}
GOTO(out, result);
}
/* now we have the actual bytes, so store it in the sdio */
bytes = result;
sdio->csd_bytes = bytes;
result = ll_direct_rw_pages(env, io, bytes, rw, inode, sdio);
/* if the i/o was unsuccessful, we zero the number of bytes to
* copy back. Note that partial I/O completion isn't possible
* here - I/O either completes or fails. So there's no need to
* handle short I/O here by changing 'count' with the result
* from ll_direct_rw_pages.
*
* This must be done before we release the reference
* immediately below, because releasing the reference allows
* i/o completion (and copyback to userspace, if unaligned) to
* start.
*/
if (result != 0)
sdio->csd_bytes = 0;
/* We've submitted pages and can now remove the extra
* reference for that
*/
cl_sync_io_note(env, &sdio->csd_sync, result);
if (sync_submit) {
rc2 = cl_sync_io_wait(env, &sdio->csd_sync,
0);
if (result == 0 && rc2)
result = rc2;
LASSERT(sdio->csd_creator_free);
cl_sub_dio_free(sdio);
}
if (unlikely(result < 0))
GOTO(out, result);
iov_iter_advance(iter, bytes);
tot_bytes += bytes;
file_offset += bytes;
CDEBUG(D_VFSTRACE,
"result %zd tot_bytes %zd count %zd file_offset %lld\n",
result, tot_bytes, bytes, file_offset);
}
out:
if (rw == WRITE)
vio->u.readwrite.vui_written += tot_bytes;
else
vio->u.readwrite.vui_read += tot_bytes;
/* AIO is not supported on pipes, so we cannot return EIOCBQEUED like
* we normally would for both DIO and AIO here
*/
if (result == 0 && !iov_iter_is_pipe(iter))
result = -EIOCBQUEUED;
RETURN(result);
}
/**
* ll_prepare_partial_page() - Prepare partially written-to page for a write.
* @env: execution environment for this thread
* @io: pointer to the client I/O structure
* @pg: owned when passed in and disowned when it returns non-zero result to
* the caller
* @file: file structure associated with the page
*
* Return:
* * %0: Success (Ready for read/write)
* * %-ERRNO: Failure
*/
static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
struct cl_page *pg, struct file *file)
{
struct cl_attr *attr = vvp_env_new_attr(env);
struct cl_object *obj = io->ci_obj;
loff_t offset = cl_page_index(pg) << PAGE_SHIFT;
int result;
ENTRY;
cl_object_attr_lock(obj);
result = cl_object_attr_get(env, obj, attr);
cl_object_attr_unlock(obj);
if (result) {
cl_page_disown(env, io, pg);
GOTO(out, result);
}
/*
* If are writing to a new page, no need to read old data.
* The extent locking will have updated the KMS, and for our
* purposes here we can treat it like i_size.
*/
if (attr->cat_kms <= offset) {
char *kaddr = kmap_local_page(pg->cp_vmpage);
memset(kaddr, 0, PAGE_SIZE);
kunmap_local(kaddr);
GOTO(out, result = 0);
}
if (pg->cp_defer_uptodate) {
pg->cp_ra_used = 1;
GOTO(out, result = 0);
}
result = ll_io_read_page(env, io, pg, file);
if (result)
GOTO(out, result);
/* ll_io_read_page() disowns the page */
result = cl_page_own(env, io, pg);
if (!result) {
if (!PageUptodate(cl_page_vmpage(pg))) {
cl_page_disown(env, io, pg);
result = -EIO;
}
} else if (result == -ENOENT) {
/* page was truncated */
result = -EAGAIN;
}
EXIT;
out:
return result;
}
static int ll_tiny_write_begin(struct page *vmpage, struct address_space *mapping)
{
/* Page must be present, up to date, dirty, and not in writeback. */
if (!vmpage || !PageUptodate(vmpage) || !PageDirty(vmpage) ||
PageWriteback(vmpage) || vmpage->mapping != mapping)
return -ENODATA;
return 0;
}
/*
* write_begin is responsible for allocating page cache pages to be used
* to hold data for buffered i/o on the 'write' path.
* Called by generic_perform_write() to allocate one page [or one folio]
*/
static int ll_write_begin(
#ifdef HAVE_WRITE_BEGIN_KIOCB
const struct kiocb *kiocb,
#else
struct file *file,
#endif
struct address_space *mapping,
loff_t pos, unsigned int len,
#ifdef HAVE_GRAB_CACHE_PAGE_WRITE_BEGIN_WITH_FLAGS
unsigned int flags,
#endif
struct wbe_folio **foliop, void **fsdata)
{
struct ll_cl_context *lcc = NULL;
const struct lu_env *env = NULL;
struct vvp_io *vio;
struct cl_io *io = NULL;
struct cl_page *cl_page = NULL;
#ifdef HAVE_WRITE_BEGIN_KIOCB
struct file *file = kiocb->ki_filp;
#endif
struct inode *inode = file_inode(file);
struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
pgoff_t index = pos >> PAGE_SHIFT;
struct page *vmpage = NULL;
unsigned from = pos & (PAGE_SIZE - 1);
unsigned to = from + len;
int result = 0;
ENTRY;
CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
lcc = ll_cl_find(inode);
if (lcc == NULL) {
/* do not allocate a page, only find & lock */
vmpage = find_lock_page(mapping, index);
result = ll_tiny_write_begin(vmpage, mapping);
GOTO(out, result);
}
env = lcc->lcc_env;
io = lcc->lcc_io;
vio = vvp_env_io(env);
if (iocb_ki_flags_check(vio->vui_iocb, IOCB_DIRECT)) {
/* direct IO failed because it couldn't clean up cached pages,
* this causes a problem for mirror write because the cached
* page may belong to another mirror, which will result in
* problem submitting the I/O. */
if (io->ci_designated_mirror > 0)
GOTO(out, result = -EBUSY);
/**
* Direct write can fall back to buffered read, but DIO is done
* with lockless i/o, and buffered requires LDLM locking, so
* in this case we must restart without lockless.
*/
if (!io->ci_dio_lock) {
io->ci_dio_lock = 1;
io->ci_need_restart = 1;
GOTO(out, result = -ENOLCK);
}
}
again:
/* To avoid deadlock, try to lock page first. */
vmpage = grab_cache_page_nowait(mapping, index);
if (unlikely(vmpage == NULL ||
PageDirty(vmpage) || PageWriteback(vmpage))) {
struct vvp_io *vio = vvp_env_io(env);
struct cl_page_list *plist = &vio->u.readwrite.vui_queue;
/* if the page is already in dirty cache, we have to commit
* the pages right now; otherwise, it may cause deadlock
* because it holds page lock of a dirty page and request for
* more grants. It's okay for the dirty page to be the first
* one in commit page list, though. */
if (vmpage != NULL && plist->pl_nr > 0) {
unlock_page(vmpage);
put_page(vmpage);
vmpage = NULL;
}
/* commit pages and then wait for page lock */
result = vvp_io_write_commit(env, io, IO_PRIO_NORMAL);
if (result < 0)
GOTO(out, result);
if (vmpage == NULL) {
vmpage = grab_cache_page_write_begin(mapping, index
#ifdef HAVE_GRAB_CACHE_PAGE_WRITE_BEGIN_WITH_FLAGS
, flags
#endif
);
if (vmpage == NULL)
GOTO(out, result = -ENOMEM);
}
}
/* page was truncated */
if (mapping != vmpage->mapping) {
CDEBUG(D_VFSTRACE, "page: %lu was truncated\n", index);
unlock_page(vmpage);
put_page(vmpage);
vmpage = NULL;
goto again;
}
cl_page = cl_page_find(env, clob, folio_index_page(vmpage), vmpage,
CPT_CACHEABLE);
if (IS_ERR(cl_page))
GOTO(out, result = PTR_ERR(cl_page));
lcc->lcc_page = cl_page;
cl_page_assume(env, io, cl_page);
if (!PageUptodate(vmpage)) {
/*
* We're completely overwriting an existing page,
* so _don't_ set it up to date until commit_write
*/
if (from == 0 && to == PAGE_SIZE) {
CL_PAGE_HEADER(D_PAGE, env, cl_page,
"full page write\n");
} else {
/* TODO: can be optimized at OSC layer to check if it
* is a lockless IO. In that case, it's not necessary
* to read the data. */
result = ll_prepare_partial_page(env, io, cl_page,
file);
if (result) {
/* vmpage should have been unlocked */
put_page(vmpage);
vmpage = NULL;
if (result == -EAGAIN)
goto again;
GOTO(out, result);
}
}
}
EXIT;
out:
if (result < 0) {
if (vmpage != NULL) {
unlock_page(vmpage);
put_page(vmpage);
}
/* On tiny_write failure, page and io are always null. */
if (!IS_ERR_OR_NULL(cl_page))
cl_page_put(env, cl_page);
if (io)
io->ci_result = result;
} else {
*foliop = wbe_page_folio(vmpage);
*fsdata = lcc;
}
RETURN(result);
}
static int ll_tiny_write_end(struct file *file, struct address_space *mapping,
loff_t pos, unsigned int len, unsigned int copied,
struct page *vmpage)
{
struct cl_page *clpage = (struct cl_page *) vmpage->private;
loff_t kms = pos+copied;
loff_t to = kms & (PAGE_SIZE-1) ? kms & (PAGE_SIZE-1) : PAGE_SIZE;
struct lu_env *env;
int rc = 0;
ENTRY;
/* This page is dirty in cache, so it should have a cl_page pointer
* set in vmpage->private.
*/
LASSERT(clpage != NULL);
if (copied == 0)
goto out;
/* env_percpu_get cannot fail */
env = cl_env_percpu_get();
/* Update the underlying size information in the OSC/LOV objects this
* page is part of.
*/
cl_page_touch(env, clpage, to);
cl_env_percpu_put(env);
out:
/* Must return page unlocked. */
unlock_page(vmpage);
RETURN(rc);
}
/* called by generic_perform_write after each page/folio is filled */
static int ll_write_end(
#ifdef HAVE_WRITE_BEGIN_KIOCB
const struct kiocb *kiocb,
#else
struct file *file,
#endif
struct address_space *mapping,
loff_t pos, unsigned len, unsigned copied,
struct wbe_folio *folio, void *fsdata)
{
struct ll_cl_context *lcc = fsdata;
const struct lu_env *env;
#ifdef HAVE_WRITE_BEGIN_KIOCB
struct file *file = kiocb->ki_filp;
#endif
struct cl_io *io;
struct vvp_io *vio;
struct cl_page *cl_page;
struct page *vmpage = wbe_folio_page(folio);
unsigned from = pos & (PAGE_SIZE - 1);
enum cl_io_priority prio = IO_PRIO_NORMAL;
bool unplug = false;
int result = 0;
ENTRY;
put_page(vmpage);
CDEBUG(D_VFSTRACE, "pos %llu, len %u, copied %u\n", pos, len, copied);
if (lcc == NULL) {
result = ll_tiny_write_end(file, mapping, pos, len, copied,
vmpage);
GOTO(out, result);
}
LASSERT(lcc != NULL);
env = lcc->lcc_env;
cl_page = lcc->lcc_page;
io = lcc->lcc_io;
vio = vvp_env_io(env);
LASSERT(cl_page_is_owned(cl_page, io));
if (copied > 0) {
struct cl_page_list *plist = &vio->u.readwrite.vui_queue;
#ifdef SB_I_CGROUPWB
struct inode *inode = file_inode(file);
struct bdi_writeback *wb;
__mark_inode_dirty(inode, I_DIRTY_PAGES);
spin_lock(&inode->i_lock);
wb = inode_to_wb(inode);
LASSERT(wb != NULL);
if (wb->dirty_exceeded) {
unplug = true;
prio = IO_PRIO_URGENT;
CDEBUG(D_IOTRACE, "wb@%pK dirty_ratelimit=%lu balanced_dirty_ratelimit=%lu dirty_exceeded=%d state=%lX last_old_flush=%lu\n",
wb, wb->dirty_ratelimit,
wb->balanced_dirty_ratelimit,
wb->dirty_exceeded, wb->state,
wb->last_old_flush);
}
spin_unlock(&inode->i_lock);
#endif
lcc->lcc_page = NULL; /* cl_page will be queued */
/* Add it into write queue */
cl_page_list_add(plist, cl_page, true);
if (plist->pl_nr == 1) /* first cl_page */
vio->u.readwrite.vui_from = from;
else
LASSERT(from == 0);
vio->u.readwrite.vui_to = from + copied;
/* To address the deadlock in balance_dirty_pages() where
* this dirty cl_page may be written back in the same thread.
*/
if (PageDirty(vmpage))
unplug = true;
/* We may have one full RPC, commit it soon */
if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
unplug = true;
CL_PAGE_DEBUG(D_VFSTRACE, env, cl_page,
"queued cl_page: %d.\n", plist->pl_nr);
} else {
cl_page_disown(env, io, cl_page);
lcc->lcc_page = NULL;
cl_page_put(env, cl_page);
/* cl_page list is not contiguous now, commit it now */
unplug = true;
}
/* the last call into ->write_begin() can unplug the queue */
if (io->u.ci_wr.wr_sync && pos + len ==
io->u.ci_rw.crw_pos + io->u.ci_rw.crw_bytes)
unplug = true;
if (unplug)
result = vvp_io_write_commit(env, io, prio);
if (result < 0)
io->ci_result = result;
out:
RETURN(result >= 0 ? copied : result);
}
#ifdef CONFIG_MIGRATION
static int ll_migrate_folio(struct address_space *mapping,
struct folio_migr *newpage, struct folio_migr *page,
enum migrate_mode mode)
{
/* Always fail page migration until we have a proper implementation */
return -EIO;
}
#endif
const struct address_space_operations ll_aops = {
#ifdef HAVE_DIRTY_FOLIO
.dirty_folio = filemap_dirty_folio,
#else
.set_page_dirty = __set_page_dirty_nobuffers,
#endif
#ifdef HAVE_INVALIDATE_FOLIO
.invalidate_folio = ll_invalidate_folio,
#else
.invalidatepage = ll_invalidatepage,
#endif
#ifdef HAVE_AOPS_READ_FOLIO
.read_folio = ll_read_folio,
#else
.readpage = ll_readpage,
#endif
#ifdef HAVE_AOPS_RELEASE_FOLIO
.release_folio = ll_release_folio,
#else
.releasepage = (void *)ll_releasepage,
#endif
.direct_IO = ll_direct_IO,
.writepages = ll_writepages,
.write_begin = ll_write_begin,
.write_end = ll_write_end,
#ifdef CONFIG_MIGRATION
.migrate_folio = ll_migrate_folio,
#endif
};
