Hi,
Maybe someone here can quickly point me in the right direction for
this oddity that we noticed.
On newer kernels, it looks like the task io accounting is not
incrementing the read_bytes when reading from a NFS mount? This was
definitely working on v5.16 downwards, but has not been working since
v5.18 up to v6.2 (I haven't tested v5.17 yet).
If I read from a local filesystem, then the read_bytes for that PID is
incremented as expected.
If I read over NFS using directIO, then the read_bytes is also
correctly incremented for that PID. It's just when reading normally
without directIO that it is not.
The write_bytes and rchar are also still both correct in all situations.
I have checked the kernel config and I'm fairly sure I have all the
right things enabled:
CONFIG_TASK_XACCT=y
CONFIG_TASK_IO_ACCOUNTING=y
CONFIG_TASKSTATS=y
Unless there was some extra config introduced specific to the nfs
client in later kernels that I missed?
Cheers,
Daire
On Fri, Feb 17, 2023 at 6:56 AM Daire Byrne <[email protected]> wrote:
>
> Hi,
>
> Maybe someone here can quickly point me in the right direction for
> this oddity that we noticed.
>
> On newer kernels, it looks like the task io accounting is not
> incrementing the read_bytes when reading from a NFS mount? This was
> definitely working on v5.16 downwards, but has not been working since
> v5.18 up to v6.2 (I haven't tested v5.17 yet).
>
> If I read from a local filesystem, then the read_bytes for that PID is
> incremented as expected.
>
> If I read over NFS using directIO, then the read_bytes is also
> correctly incremented for that PID. It's just when reading normally
> without directIO that it is not.
>
> The write_bytes and rchar are also still both correct in all situations.
>
> I have checked the kernel config and I'm fairly sure I have all the
> right things enabled:
>
> CONFIG_TASK_XACCT=y
> CONFIG_TASK_IO_ACCOUNTING=y
> CONFIG_TASKSTATS=y
>
> Unless there was some extra config introduced specific to the nfs
> client in later kernels that I missed?
>
> Cheers,
>
> Daire
>
Daire,
Thanks for the report.
Willy,
Question for you at the bottom of this.
First, here's what looks to be the candidate changes between these versions:
$ git log --oneline v5.16..v5.18 fs/nfs/read.c
89c2be8a9516 NFS: discard NFS_RPC_SWAPFLAGS and RPC_TASK_ROOTCREDS
fc1c5abfca7e NFS: Rename fscache read and write pages functions
8786fde8421c Convert NFS from readpages to readahead
16f2f4e679cf nfs: Implement cache I/O by accessing the cache directly
I would be this is due to this patch, which went into 5.18:
8786fde8421c Convert NFS from readpages to readahead
And the hunks that now call readahead_page vs read_cache_pages:
@@ -397,14 +396,16 @@ int nfs_readpage(struct file *file, struct page *page)
return ret;
}
-int nfs_readpages(struct file *file, struct address_space *mapping,
- struct list_head *pages, unsigned nr_pages)
+void nfs_readahead(struct readahead_control *ractl)
{
+ unsigned int nr_pages = readahead_count(ractl);
+ struct file *file = ractl->file;
struct nfs_readdesc desc;
- struct inode *inode = mapping->host;
+ struct inode *inode = ractl->mapping->host;
+ struct page *page;
int ret;
- trace_nfs_aop_readahead(inode, lru_to_page(pages), nr_pages);
+ trace_nfs_aop_readahead(inode, readahead_pos(ractl), nr_pages);
nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
ret = -ESTALE;
@@ -422,14 +423,18 @@ int nfs_readpages(struct file *file, struct
address_space *mapping,
nfs_pageio_init_read(&desc.pgio, inode, false,
&nfs_async_read_completion_ops);
- ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
+ while ((page = readahead_page(ractl)) != NULL) {
+ ret = readpage_async_filler(&desc, page);
+ put_page(page);
+ if (ret)
+ break;
+ }
nfs_pageio_complete_read(&desc.pgio);
put_nfs_open_context(desc.ctx);
out:
trace_nfs_aop_readahead_done(inode, nr_pages, ret);
- return ret;
}
and this hunk:
@@ -397,14 +396,16 @@ int nfs_readpage(struct file *file, struct page *page)
return ret;
}
-int nfs_readpages(struct file *file, struct address_space *mapping,
- struct list_head *pages, unsigned nr_pages)
+void nfs_readahead(struct readahead_control *ractl)
{
+ unsigned int nr_pages = readahead_count(ractl);
+ struct file *file = ractl->file;
struct nfs_readdesc desc;
- struct inode *inode = mapping->host;
+ struct inode *inode = ractl->mapping->host;
+ struct page *page;
int ret;
- trace_nfs_aop_readahead(inode, lru_to_page(pages), nr_pages);
+ trace_nfs_aop_readahead(inode, readahead_pos(ractl), nr_pages);
nfs_inc_stats(inode, NFSIOS_VFSREADPAGES);
ret = -ESTALE;
@@ -422,14 +423,18 @@ int nfs_readpages(struct file *file, struct
address_space *mapping,
nfs_pageio_init_read(&desc.pgio, inode, false,
&nfs_async_read_completion_ops);
- ret = read_cache_pages(mapping, pages, readpage_async_filler, &desc);
+ while ((page = readahead_page(ractl)) != NULL) {
+ ret = readpage_async_filler(&desc, page);
+ put_page(page);
+ if (ret)
+ break;
+ }
nfs_pageio_complete_read(&desc.pgio);
put_nfs_open_context(desc.ctx);
out:
trace_nfs_aop_readahead_done(inode, nr_pages, ret);
- return ret;
}
int __init nfs_init_readpagecache(void)
In v5.16 we had this call to task_io_account_read(PAGE_SIZE); on line 109
of read_cache_pages();
76 /**
77 * read_cache_pages - populate an address space with some pages &
start reads against them
78 * @mapping: the address_space
79 * @pages: The address of a list_head which contains the target
pages. These
80 * pages have their ->index populated and are otherwise uninitialised.
81 * @filler: callback routine for filling a single page.
82 * @data: private data for the callback routine.
83 *
84 * Hides the details of the LRU cache etc from the filesystems.
85 *
86 * Returns: %0 on success, error return by @filler otherwise
87 */
88 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
89 int (*filler)(void *, struct page *), void *data)
90 {
91 struct page *page;
92 int ret = 0;
93
94 while (!list_empty(pages)) {
95 page = lru_to_page(pages);
96 list_del(&page->lru);
97 if (add_to_page_cache_lru(page, mapping, page->index,
98 readahead_gfp_mask(mapping))) {
99 read_cache_pages_invalidate_page(mapping, page);
100 continue;
101 }
102 put_page(page);
103
104 ret = filler(data, page);
105 if (unlikely(ret)) {
106 read_cache_pages_invalidate_pages(mapping, pages);
107 break;
108 }
109 task_io_account_read(PAGE_SIZE);
But there's no call to task_io_account_read() anymore in the new
readahead code paths that I could tell,
but maybe I'm missing something.
Willy,
Does each caller of readahead_page() now need to call
task_io_account_read() or should we add that into
readahead_page() or maybe inside read_pages()?
On Fri, Feb 17, 2023 at 09:08:27AM -0500, David Wysochanski wrote:
> On Fri, Feb 17, 2023 at 6:56 AM Daire Byrne <[email protected]> wrote:
> > On newer kernels, it looks like the task io accounting is not
> > incrementing the read_bytes when reading from a NFS mount? This was
> > definitely working on v5.16 downwards, but has not been working since
> > v5.18 up to v6.2 (I haven't tested v5.17 yet).
>
> In v5.16 we had this call to task_io_account_read(PAGE_SIZE); on line 109
> of read_cache_pages();
>
> But there's no call to task_io_account_read() anymore in the new
> readahead code paths that I could tell,
> but maybe I'm missing something.
>
> Willy,
> Does each caller of readahead_page() now need to call
> task_io_account_read() or should we add that into
> readahead_page() or maybe inside read_pages()?
I think the best way is to mimic what the block layer does as closely as
possible. Unless we can pull it out of the block layer & all
filesystems and put it in the VFS (which we can't; the VFS doesn't
know which blocks are recorded by the filesystem as holes and will not
result in I/O).
The block layer does it as part of the BIO submission path (and also
counts PGPGIN and PGPGOUT, which no network filesystems seem to do?)
You're more familiar with the NFS code than I am, so you probably
have a better idea than __nfs_pageio_add_request().
On Fri, Feb 17, 2023 at 9:36 AM Matthew Wilcox <[email protected]> wrote:
>
> On Fri, Feb 17, 2023 at 09:08:27AM -0500, David Wysochanski wrote:
> > On Fri, Feb 17, 2023 at 6:56 AM Daire Byrne <[email protected]> wrote:
> > > On newer kernels, it looks like the task io accounting is not
> > > incrementing the read_bytes when reading from a NFS mount? This was
> > > definitely working on v5.16 downwards, but has not been working since
> > > v5.18 up to v6.2 (I haven't tested v5.17 yet).
> >
> > In v5.16 we had this call to task_io_account_read(PAGE_SIZE); on line 109
> > of read_cache_pages();
> >
> > But there's no call to task_io_account_read() anymore in the new
> > readahead code paths that I could tell,
> > but maybe I'm missing something.
> >
> > Willy,
> > Does each caller of readahead_page() now need to call
> > task_io_account_read() or should we add that into
> > readahead_page() or maybe inside read_pages()?
>
> I think the best way is to mimic what the block layer does as closely as
> possible. Unless we can pull it out of the block layer & all
> filesystems and put it in the VFS (which we can't; the VFS doesn't
> know which blocks are recorded by the filesystem as holes and will not
> result in I/O).
>
Caller, ok. I see, that makes sense.
Looks like cifs has a call to task_io_account_read() after data has been
received. Also netfs has a call as well at the bottom of
netfs_rreq_unlock_folios().
Both seems to be _after_ data has been received, but I'm not sure that's
correct.
> The block layer does it as part of the BIO submission path (and also
> counts PGPGIN and PGPGOUT, which no network filesystems seem to do?)
> You're more familiar with the NFS code than I am, so you probably
> have a better idea than __nfs_pageio_add_request().
>
Hmm, yes does the block layer's accounting take into account actual
bytes read or just submitted? Maybe I need to look at this closer
but at first glance it looks like these numbers may sometimes be
incremented when actual data is received and others are incremented
when the submission happens.
As to the right location in NFS, the function you mention isn't a bad
idea, but maybe not the right location. Looking in nfs_file_direct_read()
we have the accounting at IO submission time, appears to be the
same as the block layer. Also since my NFS netfs conversion patches
are still outstanding, I'll have to somehow take the netfs call into account
if fscache is enabled. So the right place is looking like somewhere
in nfs_read_folio() and nfs_readahead().
On Fri, Feb 17, 2023 at 10:47 AM David Wysochanski <[email protected]> wrote:
>
> On Fri, Feb 17, 2023 at 9:36 AM Matthew Wilcox <[email protected]> wrote:
> >
> > On Fri, Feb 17, 2023 at 09:08:27AM -0500, David Wysochanski wrote:
> > > On Fri, Feb 17, 2023 at 6:56 AM Daire Byrne <[email protected]> wrote:
> > > > On newer kernels, it looks like the task io accounting is not
> > > > incrementing the read_bytes when reading from a NFS mount? This was
> > > > definitely working on v5.16 downwards, but has not been working since
> > > > v5.18 up to v6.2 (I haven't tested v5.17 yet).
> > >
> > > In v5.16 we had this call to task_io_account_read(PAGE_SIZE); on line 109
> > > of read_cache_pages();
> > >
> > > But there's no call to task_io_account_read() anymore in the new
> > > readahead code paths that I could tell,
> > > but maybe I'm missing something.
> > >
> > > Willy,
> > > Does each caller of readahead_page() now need to call
> > > task_io_account_read() or should we add that into
> > > readahead_page() or maybe inside read_pages()?
> >
> > I think the best way is to mimic what the block layer does as closely as
> > possible. Unless we can pull it out of the block layer & all
> > filesystems and put it in the VFS (which we can't; the VFS doesn't
> > know which blocks are recorded by the filesystem as holes and will not
> > result in I/O).
> >
> Caller, ok. I see, that makes sense.
> Looks like cifs has a call to task_io_account_read() after data has been
> received. Also netfs has a call as well at the bottom of
> netfs_rreq_unlock_folios().
> Both seems to be _after_ data has been received, but I'm not sure that's
> correct.
>
> > The block layer does it as part of the BIO submission path (and also
> > counts PGPGIN and PGPGOUT, which no network filesystems seem to do?)
> > You're more familiar with the NFS code than I am, so you probably
> > have a better idea than __nfs_pageio_add_request().
> >
> Hmm, yes does the block layer's accounting take into account actual
> bytes read or just submitted? Maybe I need to look at this closer
> but at first glance it looks like these numbers may sometimes be
> incremented when actual data is received and others are incremented
> when the submission happens.
>
> As to the right location in NFS, the function you mention isn't a bad
> idea, but maybe not the right location. Looking in nfs_file_direct_read()
> we have the accounting at IO submission time, appears to be the
> same as the block layer. Also since my NFS netfs conversion patches
> are still outstanding, I'll have to somehow take the netfs call into account
> if fscache is enabled. So the right place is looking like somewhere
> in nfs_read_folio() and nfs_readahead().
I should have read the kernel docs. From
Documentation/filesystems/proc.rst
1744 read_bytes
1745 ^^^^^^^^^^
1746
1747 I/O counter: bytes read
1748 Attempt to count the number of bytes which this process really did cause to
1749 be fetched from the storage layer. Done at the submit_bio() level, so it is
1750 accurate for block-backed filesystems. <please add status regarding NFS and
1751 CIFS at a later time>
So it looks like NFS directIO (and non-direct, prior to v5,18) did the
same thing
as the block layer and is consistent with the definition.
Fix would be just add a call to task_io_account_read() inside nfs_read_folio()
and nfs_readahead().
On Fri, Feb 17, 2023 at 10:47:01AM -0500, David Wysochanski wrote:
> On Fri, Feb 17, 2023 at 9:36 AM Matthew Wilcox <[email protected]> wrote:
> >
> > On Fri, Feb 17, 2023 at 09:08:27AM -0500, David Wysochanski wrote:
> > > On Fri, Feb 17, 2023 at 6:56 AM Daire Byrne <[email protected]> wrote:
> > > > On newer kernels, it looks like the task io accounting is not
> > > > incrementing the read_bytes when reading from a NFS mount? This was
> > > > definitely working on v5.16 downwards, but has not been working since
> > > > v5.18 up to v6.2 (I haven't tested v5.17 yet).
> > >
> > > In v5.16 we had this call to task_io_account_read(PAGE_SIZE); on line 109
> > > of read_cache_pages();
> > >
> > > But there's no call to task_io_account_read() anymore in the new
> > > readahead code paths that I could tell,
> > > but maybe I'm missing something.
> > >
> > > Willy,
> > > Does each caller of readahead_page() now need to call
> > > task_io_account_read() or should we add that into
> > > readahead_page() or maybe inside read_pages()?
> >
> > I think the best way is to mimic what the block layer does as closely as
> > possible. Unless we can pull it out of the block layer & all
> > filesystems and put it in the VFS (which we can't; the VFS doesn't
> > know which blocks are recorded by the filesystem as holes and will not
> > result in I/O).
> >
> Caller, ok. I see, that makes sense.
> Looks like cifs has a call to task_io_account_read() after data has been
> received. Also netfs has a call as well at the bottom of
> netfs_rreq_unlock_folios().
> Both seems to be _after_ data has been received, but I'm not sure that's
> correct.
It's probably correct, just different from the block layer. I don't
have any special insight here, just an inclination to be as similar
as possible.
> > The block layer does it as part of the BIO submission path (and also
> > counts PGPGIN and PGPGOUT, which no network filesystems seem to do?)
> > You're more familiar with the NFS code than I am, so you probably
> > have a better idea than __nfs_pageio_add_request().
> >
> Hmm, yes does the block layer's accounting take into account actual
> bytes read or just submitted? Maybe I need to look at this closer
> but at first glance it looks like these numbers may sometimes be
> incremented when actual data is received and others are incremented
> when the submission happens.
>
> As to the right location in NFS, the function you mention isn't a bad
> idea, but maybe not the right location. Looking in nfs_file_direct_read()
> we have the accounting at IO submission time, appears to be the
> same as the block layer. Also since my NFS netfs conversion patches
> are still outstanding, I'll have to somehow take the netfs call into account
> if fscache is enabled. So the right place is looking like somewhere
> in nfs_read_folio() and nfs_readahead().
Yes, we don't want to double-count either fscache or direct I/O.
I'm Maybe Dave as opinions about where we should be accounting it --
I'm not sure that netfs is the right place to do it. Maybe it should
be in each network filesystem instead of in netfs?
On Fri, Feb 17, 2023 at 10:59 AM Matthew Wilcox <[email protected]> wrote:
>
> On Fri, Feb 17, 2023 at 10:47:01AM -0500, David Wysochanski wrote:
> > On Fri, Feb 17, 2023 at 9:36 AM Matthew Wilcox <[email protected]> wrote:
> > >
> > > On Fri, Feb 17, 2023 at 09:08:27AM -0500, David Wysochanski wrote:
> > > > On Fri, Feb 17, 2023 at 6:56 AM Daire Byrne <[email protected]> wrote:
> > > > > On newer kernels, it looks like the task io accounting is not
> > > > > incrementing the read_bytes when reading from a NFS mount? This was
> > > > > definitely working on v5.16 downwards, but has not been working since
> > > > > v5.18 up to v6.2 (I haven't tested v5.17 yet).
> > > >
> > > > In v5.16 we had this call to task_io_account_read(PAGE_SIZE); on line 109
> > > > of read_cache_pages();
> > > >
> > > > But there's no call to task_io_account_read() anymore in the new
> > > > readahead code paths that I could tell,
> > > > but maybe I'm missing something.
> > > >
> > > > Willy,
> > > > Does each caller of readahead_page() now need to call
> > > > task_io_account_read() or should we add that into
> > > > readahead_page() or maybe inside read_pages()?
> > >
> > > I think the best way is to mimic what the block layer does as closely as
> > > possible. Unless we can pull it out of the block layer & all
> > > filesystems and put it in the VFS (which we can't; the VFS doesn't
> > > know which blocks are recorded by the filesystem as holes and will not
> > > result in I/O).
> > >
> > Caller, ok. I see, that makes sense.
> > Looks like cifs has a call to task_io_account_read() after data has been
> > received. Also netfs has a call as well at the bottom of
> > netfs_rreq_unlock_folios().
> > Both seems to be _after_ data has been received, but I'm not sure that's
> > correct.
>
> It's probably correct, just different from the block layer. I don't
> have any special insight here, just an inclination to be as similar
> as possible.
>
> > > The block layer does it as part of the BIO submission path (and also
> > > counts PGPGIN and PGPGOUT, which no network filesystems seem to do?)
> > > You're more familiar with the NFS code than I am, so you probably
> > > have a better idea than __nfs_pageio_add_request().
> > >
> > Hmm, yes does the block layer's accounting take into account actual
> > bytes read or just submitted? Maybe I need to look at this closer
> > but at first glance it looks like these numbers may sometimes be
> > incremented when actual data is received and others are incremented
> > when the submission happens.
> >
> > As to the right location in NFS, the function you mention isn't a bad
> > idea, but maybe not the right location. Looking in nfs_file_direct_read()
> > we have the accounting at IO submission time, appears to be the
> > same as the block layer. Also since my NFS netfs conversion patches
> > are still outstanding, I'll have to somehow take the netfs call into account
> > if fscache is enabled. So the right place is looking like somewhere
> > in nfs_read_folio() and nfs_readahead().
>
> Yes, we don't want to double-count either fscache or direct I/O.
> I'm Maybe Dave as opinions about where we should be accounting it --
> I'm not sure that netfs is the right place to do it. Maybe it should
> be in each network filesystem instead of in netfs?
>
David - can you comment on whether you agree the call to
task_io_account_read() inside netfs_rreq_unlock_folios() is wrong? I
think based on the discussion here, the definition in the kernel docs,
and the fact we have to record against the proper PID, the accounting
should be done in the caller of the netfs interfaces not inside netfs
unlock path. It looks like afs may rely on netfs IO accounting but
can it be moved?