by Martin Brandenburg

[permalink] [raw]

Subject: invalidatepage questions (was Re: [PATCH 14/19] orangefs: write range tracking)

I have some questions about my assumptions writing this code. I
definitely have something wrong.

(Given my comments in invalidatepage, it should be clear I'm not quite
confident that it's doing the right thing.)

I thought writpeage could not be called while PagePrivate is not set.
It is set during a write to reflect the changed data, then cleared in
writepage as the write is done.

It is also cleared in invalidatepage since the write is no longer
necessary. However, I see writepage calls after an invalidatepage
causing a ClearPagePrivate. These naturally hit the BUG().

Running xfstests generic/127 I see a page with a pending write from
2110 to 4096 (end of page) and then an invalidate on that page from 744
to 4096. I ClearPagePrivate since there is no longer any data to write
from that page. But the kernel doens't know that and calls writepage.

Should I ClearPageDirty too?

Should I just ClearPagePrivate and then use the fact that it's not set
as a hint not to do anything in writepage?

Something I'm not thinking of?

Martin

On Sun, Oct 07, 2018 at 11:27:31PM +0000, Martin Brandenburg wrote:
> This is necessary to ensure the uid/gid responsible for the write is
> communicated with the server. Only one uid/gid may have outstanding
> changes at a time. If another uid/gid writes while there are
> outstanding changes, the changes must be written out before the new
> data is put into the page.
>
> Signed-off-by: Martin Brandenburg <[email protected]>
> ---
> fs/orangefs/file.c | 12 +-
> fs/orangefs/inode.c | 267 ++++++++++++++++++++++++++++++----
> fs/orangefs/orangefs-kernel.h | 12 +-
> 3 files changed, 261 insertions(+), 30 deletions(-)
>
> diff --git a/fs/orangefs/file.c b/fs/orangefs/file.c
> index ba580a5c6fd2..5eda483263ae 100644
> --- a/fs/orangefs/file.c
> +++ b/fs/orangefs/file.c
> @@ -46,8 +46,8 @@ static int flush_racache(struct inode *inode)
> * Post and wait for the I/O upcall to finish
> */
> ssize_t wait_for_direct_io(enum ORANGEFS_io_type type, struct inode *inode,
> - loff_t *offset, struct iov_iter *iter,
> - size_t total_size, loff_t readahead_size)
> + loff_t *offset, struct iov_iter *iter, size_t total_size,
> + loff_t readahead_size, struct orangefs_write_request *wr)
> {
> struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
> struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
> @@ -103,6 +103,10 @@ ssize_t wait_for_direct_io(enum ORANGEFS_io_type type, struct inode *inode,
> __func__, (long)ret);
> goto out;
> }
> + if (wr) {
> + new_op->upcall.uid = from_kuid(&init_user_ns, wr->uid);
> + new_op->upcall.gid = from_kgid(&init_user_ns, wr->gid);
> + }
> }
>
> gossip_debug(GOSSIP_FILE_DEBUG,
> @@ -292,7 +296,7 @@ ssize_t do_readv_writev(enum ORANGEFS_io_type type, struct file *file,
> (int)*offset);
>
> ret = wait_for_direct_io(type, inode, offset, iter,
> - each_count, 0);
> + each_count, 0, NULL);
> gossip_debug(GOSSIP_FILE_DEBUG,
> "%s(%pU): return from wait_for_io:%d\n",
> __func__,
> @@ -434,7 +438,7 @@ static vm_fault_t orangefs_fault(struct vm_fault *vmf)
> static const struct vm_operations_struct orangefs_file_vm_ops = {
> .fault = orangefs_fault,
> .map_pages = filemap_map_pages,
> - .page_mkwrite = filemap_page_mkwrite,
> + .page_mkwrite = orangefs_page_mkwrite,
> };
>
> /*
> diff --git a/fs/orangefs/inode.c b/fs/orangefs/inode.c
> index bd2ce18453f2..5c155b259b13 100644
> --- a/fs/orangefs/inode.c
> +++ b/fs/orangefs/inode.c
> @@ -15,9 +15,11 @@
> #include "orangefs-kernel.h"
> #include "orangefs-bufmap.h"
>
> -static int orangefs_writepage(struct page *page, struct writeback_control *wbc)
> +static int orangefs_writepage_locked(struct page *page,
> + struct writeback_control *wbc)
> {
> struct inode *inode = page->mapping->host;
> + struct orangefs_write_request *wr;
> struct iov_iter iter;
> struct bio_vec bv;
> size_t len, wlen;
> @@ -26,33 +28,175 @@ static int orangefs_writepage(struct page *page, struct writeback_control *wbc)
>
> set_page_writeback(page);
>
> - off = page_offset(page);
> - len = i_size_read(inode);
> - if (off + PAGE_SIZE > len)
> - wlen = len - off;
> - else
> - wlen = PAGE_SIZE;
> + if (PagePrivate(page)) {
> + wr = (struct orangefs_write_request *)page_private(page);
> + BUG_ON(!wr);
> + if (wr->mwrite) {
> + off = page_offset(page);
> + len = i_size_read(inode);
> + if (off + PAGE_SIZE > len)
> + wlen = len - off;
> + else
> + wlen = PAGE_SIZE;
> + } else {
> + off = wr->pos;
> + wlen = wr->len;
> + len = i_size_read(inode);
> + }
> + } else {
> +/* BUG();*/
> + /* It's not private so there's nothing to write, right? */
> + printk("writepage not private!\n");
> + end_page_writeback(page);
> + return 0;
> +
> + }
>
> bv.bv_page = page;
> bv.bv_len = wlen;
> bv.bv_offset = off % PAGE_SIZE;
> - if (wlen == 0)
> - dump_stack();
> iov_iter_bvec(&iter, ITER_BVEC | WRITE, &bv, 1, wlen);
>
> ret = wait_for_direct_io(ORANGEFS_IO_WRITE, inode, &off, &iter, wlen,
> - len);
> + len, wr);
> if (ret < 0) {
> SetPageError(page);
> mapping_set_error(page->mapping, ret);
> } else {
> ret = 0;
> + if (wr) {
> + ClearPagePrivate(page);
> + kfree(wr);
> + }
> }
> end_page_writeback(page);
> - unlock_page(page);
> return ret;
> }
>
> +static int do_writepage_if_necessary(struct page *page, loff_t pos,
> + unsigned len)
> +{
> + struct orangefs_write_request *wr;
> + struct writeback_control wbc = {
> + .sync_mode = WB_SYNC_ALL,
> + .nr_to_write = 0,
> + };
> + int r;
> + if (PagePrivate(page)) {
> + wr = (struct orangefs_write_request *)page_private(page);
> + BUG_ON(!wr);
> + /*
> + * If the new request is not contiguous with the last one or if
> + * the uid or gid is different, the page must be written out
> + * before continuing.
> + */
> + if (pos + len < wr->pos || wr->pos + wr->len < pos ||
> + !uid_eq(current_fsuid(), wr->uid) ||
> + !gid_eq(current_fsgid(), wr->gid)) {
> + wbc.range_start = page_file_offset(page);
> + wbc.range_end = wbc.range_start + PAGE_SIZE - 1;
> + wait_on_page_writeback(page);
> + if (clear_page_dirty_for_io(page)) {
> + r = orangefs_writepage_locked(page, &wbc);
> + if (r)
> + return r;
> + }
> + BUG_ON(PagePrivate(page));
> + }
> + }
> + return 0;
> +}
> +
> +static int update_wr(struct page *page, loff_t pos, unsigned len, int mwrite)
> +{
> + struct orangefs_write_request *wr;
> + if (PagePrivate(page)) {
> + wr = (struct orangefs_write_request *)page_private(page);
> + BUG_ON(!wr);
> + if (mwrite) {
> + wr->mwrite = 1;
> + return 0;
> + }
> + if (pos < wr->pos) {
> + wr->len += wr->pos - pos;
> + wr->pos = pos;
> + }
> + if (pos + len > wr->pos + wr->len)
> + wr->len = pos + len - wr->pos;
> + else
> + wr->len = wr->pos + wr->len - wr->pos;
> + } else {
> + wr = kmalloc(sizeof *wr, GFP_KERNEL);
> + if (wr) {
> + wr->pos = pos;
> + wr->len = len;
> + wr->uid = current_fsuid();
> + wr->gid = current_fsgid();
> + wr->mwrite = mwrite;
> + SetPagePrivate(page);
> + set_page_private(page, (unsigned long)wr);
> + } else {
> + return -ENOMEM;
> + }
> + }
> + return 0;
> +}
> +
> +int orangefs_page_mkwrite(struct vm_fault *vmf)
> +{
> + struct page *page = vmf->page;
> + struct inode *inode = file_inode(vmf->vma->vm_file);
> + unsigned len;
> + int r;
> +
> + /* Do not write past the file size. */
> + len = i_size_read(inode) - page_file_offset(page);
> + if (len > PAGE_SIZE)
> + len = PAGE_SIZE;
> +
> + lock_page(page);
> + r = do_writepage_if_necessary(page, page_file_offset(page),
> + len);
> + if (r) {
> + r = VM_FAULT_RETRY;
> + unlock_page(vmf->page);
> + return r;
> + }
> + r = update_wr(page, page_file_offset(page), len, 1);
> + if (r) {
> + r = VM_FAULT_RETRY;
> + unlock_page(vmf->page);
> + return r;
> + }
> +
> + r = VM_FAULT_LOCKED;
> + sb_start_pagefault(inode->i_sb);
> + file_update_time(vmf->vma->vm_file);
> + if (page->mapping != inode->i_mapping) {
> + unlock_page(page);
> + r = VM_FAULT_NOPAGE;
> + goto out;
> + }
> + /*
> + * We mark the page dirty already here so that when freeze is in
> + * progress, we are guaranteed that writeback during freezing will
> + * see the dirty page and writeprotect it again.
> + */
> + set_page_dirty(page);
> + wait_for_stable_page(page);
> +out:
> + sb_end_pagefault(inode->i_sb);
> + return r;
> +}
> +
> +static int orangefs_writepage(struct page *page, struct writeback_control *wbc)
> +{
> + int r;
> + r = orangefs_writepage_locked(page, wbc);
> + unlock_page(page);
> + return r;
> +}
> +
> static int orangefs_readpage(struct file *file, struct page *page)
> {
> struct inode *inode = page->mapping->host;
> @@ -68,7 +212,7 @@ static int orangefs_readpage(struct file *file, struct page *page)
> iov_iter_bvec(&iter, ITER_BVEC | READ, &bv, 1, PAGE_SIZE);
>
> ret = wait_for_direct_io(ORANGEFS_IO_READ, inode, &off, &iter,
> - PAGE_SIZE, inode->i_size);
> + PAGE_SIZE, inode->i_size, NULL);
> /* this will only zero remaining unread portions of the page data */
> iov_iter_zero(~0U, &iter);
> /* takes care of potential aliasing */
> @@ -86,10 +230,26 @@ static int orangefs_readpage(struct file *file, struct page *page)
> return ret;
> }
>
> +static int orangefs_write_begin(struct file *file,
> + struct address_space *mapping, loff_t pos, unsigned len, unsigned flags,
> + struct page **pagep, void **fsdata)
> +{
> + int r;
> + r = simple_write_begin(file, mapping, pos, len, flags, pagep, fsdata);
> + if (r)
> + return r;
> + r = do_writepage_if_necessary(*pagep, pos, len);
> + if (r)
> + unlock_page(*pagep);
> + return r;
> +}
> +
> int orangefs_write_end(struct file *file, struct address_space *mapping,
> loff_t pos, unsigned len, unsigned copied, struct page *page, void *fsdata)
> {
> int r;
> + if (update_wr(page, pos, len, 0))
> + return -ENOMEM;
> r = simple_write_end(file, mapping, pos, len, copied, page, fsdata);
> mark_inode_dirty_sync(file_inode(file));
> return r;
> @@ -99,24 +259,68 @@ static void orangefs_invalidatepage(struct page *page,
> unsigned int offset,
> unsigned int length)
> {
> - gossip_debug(GOSSIP_INODE_DEBUG,
> - "orangefs_invalidatepage called on page %p "
> - "(offset is %u)\n",
> - page,
> - offset);
> -
> - ClearPageUptodate(page);
> - ClearPageMappedToDisk(page);
> + struct orangefs_write_request *wr;
> + /* XXX move to releasepage and call + rebase */
> + struct writeback_control wbc = {
> + .sync_mode = WB_SYNC_ALL,
> + .nr_to_write = 0,
> + };
> + int r;
> + if (PagePrivate(page)) {
> + wr = (struct orangefs_write_request *)page_private(page);
> + BUG_ON(!wr);
> +/* XXX prove */
> + if (offset == 0 && length == PAGE_SIZE) {
> + ClearPagePrivate(page);
> + kfree(wr);
> + } else if (wr->pos - page_offset(page) < offset &&
> + wr->pos - page_offset(page) + wr->len > offset + length) {
> + wbc.range_start = page_file_offset(page);
> + wbc.range_end = wbc.range_start + PAGE_SIZE - 1;
> + wait_on_page_writeback(page);
> + if (clear_page_dirty_for_io(page)) {
> + r = orangefs_writepage_locked(page, &wbc);
> + if (r)
> + return;
> + } else {
> + ClearPagePrivate(page);
> + kfree(wr);
> + }
> + } else if (wr->pos - page_offset(page) < offset &&
> + wr->pos - page_offset(page) + wr->len <= offset + length) {
> + wr->len = offset;
> + } else if (wr->pos - page_offset(page) >= offset &&
> + wr->pos - page_offset(page) + wr->len > offset + length) {
> + wr->pos += length - wr->pos + page_offset(page);
> + wr->len -= length - wr->pos + page_offset(page);
> + } else {
> + /*
> + * Invalidate range is bigger than write range but
> + * entire write range is to be invalidated.
> + */
> + ClearPagePrivate(page);
> + kfree(wr);
> + }
> + }
> return;
>
> }
>
> static int orangefs_releasepage(struct page *page, gfp_t foo)
> {
> - gossip_debug(GOSSIP_INODE_DEBUG,
> - "orangefs_releasepage called on page %p\n",
> - page);
> - return 0;
> + /*
> + * Two cases are mentioned in vfs.txt. Only one is relevant
> + * "VM finds a clean page with no active users and wants to make it a
> + * free page" However this page will not be private.
> + * "request has been made to invalidate some or all pages in an
> + * address_space" So we call orangefs_invalidatepage.
> + */
> + if (PagePrivate(page)) {
> + orangefs_invalidatepage(page, 0, PAGE_SIZE);
> + return !PagePrivate(page);
> + } else {
> + return 1;
> + }
> }
>
> static ssize_t orangefs_direct_IO(struct kiocb *iocb,
> @@ -128,16 +332,29 @@ static ssize_t orangefs_direct_IO(struct kiocb *iocb,
> ORANGEFS_IO_WRITE : ORANGEFS_IO_READ, file, &pos, iter);
> }
>
> +static int orangefs_launder_page(struct page *page)
> +{
> + int r = 0;
> + if (PagePrivate(page)) {
> + if (clear_page_dirty_for_io(page))
> + r = orangefs_writepage_locked(page, NULL);
> + return r;
> + } else {
> + return 0;
> + }
> +}
> +
> /** ORANGEFS2 implementation of address space operations */
> static const struct address_space_operations orangefs_address_operations = {
> .writepage = orangefs_writepage,
> .readpage = orangefs_readpage,
> .set_page_dirty = __set_page_dirty_nobuffers,
> - .write_begin = simple_write_begin,
> + .write_begin = orangefs_write_begin,
> .write_end = orangefs_write_end,
> .invalidatepage = orangefs_invalidatepage,
> .releasepage = orangefs_releasepage,
> .direct_IO = orangefs_direct_IO,
> + .launder_page = orangefs_launder_page,
> };
>
> static int orangefs_setattr_size(struct inode *inode, struct iattr *iattr)
> diff --git a/fs/orangefs/orangefs-kernel.h b/fs/orangefs/orangefs-kernel.h
> index e128500e33b4..2e9726d1de7d 100644
> --- a/fs/orangefs/orangefs-kernel.h
> +++ b/fs/orangefs/orangefs-kernel.h
> @@ -178,6 +178,14 @@ static inline void set_op_state_purged(struct orangefs_kernel_op_s *op)
> }
> }
>
> +struct orangefs_write_request {
> + loff_t pos;
> + unsigned len;
> + kuid_t uid;
> + kgid_t gid;
> + int mwrite;
> +};
> +
> /* per inode private orangefs info */
> struct orangefs_inode_s {
> struct orangefs_object_kref refn;
> @@ -341,6 +349,8 @@ void fsid_key_table_finalize(void);
> /*
> * defined in inode.c
> */
> +int orangefs_page_mkwrite(struct vm_fault *);
> +
> struct inode *orangefs_new_inode(struct super_block *sb,
> struct inode *dir,
> int mode,
> @@ -382,7 +392,7 @@ bool __is_daemon_in_service(void);
> * defined in file.c
> */
> ssize_t wait_for_direct_io(enum ORANGEFS_io_type, struct inode *, loff_t *,
> - struct iov_iter *, size_t, loff_t);
> + struct iov_iter *, size_t, loff_t, struct orangefs_write_request *);
> ssize_t do_readv_writev(enum ORANGEFS_io_type, struct file *, loff_t *,
> struct iov_iter *);
>
> --
> 2.19.0
>