Commit 8811930dc7 ("splice: missing user
pointer access verification") added the proper access_ok() calls to
copy_from_user_mmap_sem() which ensures we can copy the struct iovecs
from userspace to the kernel.
But we also must check whether we can access the actual memory region
pointed to by the struct iovec to fix the access checks properly.
Signed-off-by: Bastian Blank <waldi@debian.org>
Acked-by: Oliver Pinter <oliver.pntr@gmail.com>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
vmsplice_to_user() must always check the user pointer and length
with access_ok() before copying. Likewise, for the slow path of
copy_from_user_mmap_sem() we need to check that we may read from
the user region.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Cc: Wojciech Purczynski <cliph@research.coseinc.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Andre Majorel <aym-xunil@teaser.fr> points out that if we only updated
the atime when we transfer some data, we deviate from the standard
of always updating the atime. So change splice to always call
file_accessed() even if splice_direct_to_actor() didn't transfer
any data.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
A bug report on nfsd that states that since it was switched to use
splice instead of sendfile, the atime was no longer being updated
on the input file. do_generic_mapping_read() does this when accessing
the file, make splice do it for the direct splice handler.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Allow caller to pass in a release function, there might be
other resources that need releasing as well. Needed for
network receive.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
All instances of rw_verify_area() are followed by a call to
security_file_permission(), so just call the latter from the former.
Acked-by: Eric Paris <eparis@redhat.com>
Signed-off-by: James Morris <jmorris@namei.org>
Implement file posix capabilities. This allows programs to be given a
subset of root's powers regardless of who runs them, without having to use
setuid and giving the binary all of root's powers.
This version works with Kaigai Kohei's userspace tools, found at
http://www.kaigai.gr.jp/index.php. For more information on how to use this
patch, Chris Friedhoff has posted a nice page at
http://www.friedhoff.org/fscaps.html.
Changelog:
Nov 27:
Incorporate fixes from Andrew Morton
(security-introduce-file-caps-tweaks and
security-introduce-file-caps-warning-fix)
Fix Kconfig dependency.
Fix change signaling behavior when file caps are not compiled in.
Nov 13:
Integrate comments from Alexey: Remove CONFIG_ ifdef from
capability.h, and use %zd for printing a size_t.
Nov 13:
Fix endianness warnings by sparse as suggested by Alexey
Dobriyan.
Nov 09:
Address warnings of unused variables at cap_bprm_set_security
when file capabilities are disabled, and simultaneously clean
up the code a little, by pulling the new code into a helper
function.
Nov 08:
For pointers to required userspace tools and how to use
them, see http://www.friedhoff.org/fscaps.html.
Nov 07:
Fix the calculation of the highest bit checked in
check_cap_sanity().
Nov 07:
Allow file caps to be enabled without CONFIG_SECURITY, since
capabilities are the default.
Hook cap_task_setscheduler when !CONFIG_SECURITY.
Move capable(TASK_KILL) to end of cap_task_kill to reduce
audit messages.
Nov 05:
Add secondary calls in selinux/hooks.c to task_setioprio and
task_setscheduler so that selinux and capabilities with file
cap support can be stacked.
Sep 05:
As Seth Arnold points out, uid checks are out of place
for capability code.
Sep 01:
Define task_setscheduler, task_setioprio, cap_task_kill, and
task_setnice to make sure a user cannot affect a process in which
they called a program with some fscaps.
One remaining question is the note under task_setscheduler: are we
ok with CAP_SYS_NICE being sufficient to confine a process to a
cpuset?
It is a semantic change, as without fsccaps, attach_task doesn't
allow CAP_SYS_NICE to override the uid equivalence check. But since
it uses security_task_setscheduler, which elsewhere is used where
CAP_SYS_NICE can be used to override the uid equivalence check,
fixing it might be tough.
task_setscheduler
note: this also controls cpuset:attach_task. Are we ok with
CAP_SYS_NICE being used to confine to a cpuset?
task_setioprio
task_setnice
sys_setpriority uses this (through set_one_prio) for another
process. Need same checks as setrlimit
Aug 21:
Updated secureexec implementation to reflect the fact that
euid and uid might be the same and nonzero, but the process
might still have elevated caps.
Aug 15:
Handle endianness of xattrs.
Enforce capability version match between kernel and disk.
Enforce that no bits beyond the known max capability are
set, else return -EPERM.
With this extra processing, it may be worth reconsidering
doing all the work at bprm_set_security rather than
d_instantiate.
Aug 10:
Always call getxattr at bprm_set_security, rather than
caching it at d_instantiate.
[morgan@kernel.org: file-caps clean up for linux/capability.h]
[bunk@kernel.org: unexport cap_inode_killpriv]
Signed-off-by: Serge E. Hallyn <serue@us.ibm.com>
Cc: Stephen Smalley <sds@tycho.nsa.gov>
Cc: James Morris <jmorris@namei.org>
Cc: Chris Wright <chrisw@sous-sol.org>
Cc: Andrew Morgan <morgan@kernel.org>
Signed-off-by: Andrew Morgan <morgan@kernel.org>
Signed-off-by: Adrian Bunk <bunk@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
These are intended to replace prepare_write and commit_write with more
flexible alternatives that are also able to avoid the buffered write
deadlock problems efficiently (which prepare_write is unable to do).
[mark.fasheh@oracle.com: API design contributions, code review and fixes]
[akpm@linux-foundation.org: various fixes]
[dmonakhov@sw.ru: new aop block_write_begin fix]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Signed-off-by: Dmitriy Monakhov <dmonakhov@openvz.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Combine the file_ra_state members
unsigned long prev_index
unsigned int prev_offset
into
loff_t prev_pos
It is more consistent and better supports huge files.
Thanks to Peter for the nice proposal!
[akpm@linux-foundation.org: fix shift overflow]
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Nick Piggin points out that splice isn't being good about the mmap
semaphore: while two readers can nest inside each others, it does leave
a possible deadlock if a writer (ie a new mmap()) comes in during that
nesting.
Original "just move the locking" patch by Nick, replaced by one by me
based on an optimistic pagefault_disable(). And then Jens tested and
updated that patch.
Reported-by: Nick Piggin <npiggin@suse.de>
Tested-by: Jens Axboe <jens.axboe@oracle.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Fix some typos in pipe.c and splice.c.
Add pipes API to kernel-api.tmpl.
Signed-off-by: Randy Dunlap <randy.dunlap@oracle.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
If add_to_page_cache_lru() fails, the page will not be locked. But
splice jumps to an error path that does a page release and unlock,
causing a BUG() in unlock_page().
Fix this by adding one more label that just releases the page. This bug
was actually triggered on EL5 by gurudas pai <gurudas.pai@oracle.com>
using fio.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Split ondemand readahead interface into two functions. I think this makes it
a little clearer for non-readahead experts (like Rusty).
Internally they both call ondemand_readahead(), but the page argument is
changed to an obvious boolean flag.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pass real splice size to page_cache_readahead_ondemand().
The splice code works in chunks of 16 pages internally. The readahead code
should be told of the overall splice size, instead of the internal chunk size.
Otherwize bad things may happen. Imagine some 17-page random splice reads.
The code before this patch will result in two readahead calls: readahead(16);
readahead(1); That leads to one 16-page I/O and one 32-page I/O: one extra I/O
and 31 readahead miss pages.
Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Move synchronous page_cache_readahead_ondemand() call out of splice loop.
This avoids one pointless page allocation/insertion in case of non-zero
ra_pages, or many pointless readahead calls in case of zero ra_pages.
Note that if a user sets ra_pages to less than PIPE_BUFFERS=16 pages, he will
not get expected readahead behavior anyway. The splice code works in batches
of 16 pages, which can be taken as another form of synchronous readahead.
Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp> reported that he's noticed
nfsd read corruption in recent kernels, and did the hard work of
discovering that it's due to splice updating the file position twice.
This means that the next operation would start further ahead than it
should.
nfsd_vfs_read()
splice_direct_to_actor()
while(len) {
do_splice_to() [update sd->pos]
-> generic_file_splice_read() [read from sd->pos]
nfsd_direct_splice_actor()
-> __splice_from_pipe() [update sd->pos]
There's nothing wrong with the core splice code, but the direct
splicing is an addon that calls both input and output paths.
So it has to take care in locally caching offset so it remains correct.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
If the output actor doesn't transfer the full amount of data, we will
increment ppos too much. Two related bugs in there:
- We need to break out and return actor() retval if it is shorted than
what we spliced into the pipe.
- Adjust ppos only according to actor() return.
Also fix loop problem in generic_file_splice_read(), it should not keep
going when data has already been transferred.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Revalidate read/write permissions for splice(2) and vmslice(2), in case
security policy has changed since the files were opened.
Acked-by: Stephen Smalley <sds@tycho.nsa.gov>
Signed-off-by: James Morris <jmorris@namei.org>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
As per Andrew Mortons request, here's a set of documentation for
the generic pipe_buf_operations hooks, the pipe, and pipe_buffer
structures.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
The name 'pin' was badly chosen, it doesn't pin a pipe buffer
in the most commonly used sense in the kernel. So change the
name to 'confirm', after debating this issue with Hugh
Dickins a bit.
A good return from ->confirm() means that the buffer is really
there, and that the contents are good.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Also add fs/splice.c as a kerneldoc target with a smaller blurb that
should be expanded to better explain the overview of splice.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
relay needs this for proper consumption handling, and the network
receive support needs it as well to lookup the sk_buff on pipe
release.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
We need to move even more stuff into the header so that folks can use
the splice_to_pipe() implementation instead of open-coding a lot of
pipe knowledge (see relay implementation), so move to our own header
file finally.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
A bit of a cheat, it actually just copies the data to userspace. But
this makes the interface nice and symmetric and enables people to build
on splice, with room for future improvement in performance.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
For direct splicing (or private splicing), the output may not be a file.
So abstract out the handling into a specified actor function and put
the data in the splice_desc structure earlier, so we can build on top
of that.
This is the first step in better splice handling for drivers, and also
for implementing vmsplice _to_ user memory.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
We only ever set do_wakeup to non-zero if the pipe has an inode
backing, so it's pointless to check outside the pipe->inode
check.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
If the destination pipe is full and we already transferred
data, we break out instead of waiting for more pipe room.
The exit logic looks at spd->nr_pages to see if we moved
everything inside the spd container, but we decrement that
variable in the loop to decide when spd has emptied.
Instead we want to compare to the original page count in
the spd, so cache that in a local variable.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
As we have potentially dirtied more than 1 page, we should indicate as
such to the dirty page balancing. So call
balance_dirty_pages_ratelimited_nr() and pass in the approximate number
of pages we dirtied.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Original patch and description from Neil Brown <neilb@suse.de>,
merged and adapted to splice branch by me. Neils text follows:
__generic_file_splice_read() currently samples the i_size at the start
and doesn't do so again unless it needs to call ->readpage to load
a page. After ->readpage it has to re-sample i_size as a truncate
may have caused that page to be filled with zeros, and the read()
call should not see these.
However there are other activities that might cause ->readpage to be
called on a page between the time that __generic_file_splice_read()
samples i_size and when it finds that it has an uptodate page. These
include at least read-ahead and possibly another thread performing a
read
So we must sample i_size *after* it has an uptodate page. Thus the
current sampling at the start and after a read can be replaced with a
sampling before page addition into spd.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
__generic_file_splice_read's partial page check, at eof after readpage,
not only got its calculations wrong, but also reused the loff variable:
causing data corruption when splicing from a non-0 offset in the file's
last page (revealed by ext2 -b 1024 testing on a loop of a tmpfs file).
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
I've seen inode related deadlocks, so move this call outside of the
actor itself, which may hold the inode lock.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Eric Dumazet, thank you for disclosing this bug.
Readahead logic somehow fails to populate the page range with data.
It can be because
1) the readahead routine is not always called in the following lines of
fs/splice.c:
if (!loff || nr_pages > 1)
page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
2) even called, page_cache_readahead() wont guarantee the pages are there.
It wont submit readahead I/O for pages already in the radix tree, or when
(ra_pages == 0), or after 256 cache hits.
In your case, it should be because of the retried reads, which lead to
excessive cache hits, and disables readahead at some time.
And that _one_ failure of readahead blocks the whole read process.
The application receives EAGAIN and retries the read, but
__generic_file_splice_read() refuse to make progress:
- in the previous invocation, it has allocated a blank page and inserted it
into the radix tree, but never has the chance to start I/O for it: the test
of SPLICE_F_NONBLOCK goes before that.
- in the retried invocation, the readahead code will neither get out of the
cache hit mode, nor will it submit I/O for an already existing page.
Cc: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Currently if partial write has happened while ->commit_write() then page
wasn't marked as accessed and rebalanced.
Signed-off-by: Monakhov Dmitriy <dmonakhov@openvz.org>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Ocfs2 wants to implement it's own splice write actor so that it can better
manage cluster / page locks. This lets us re-use the rest of splice write
while only providing our own code where it's actually important.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Splice does not need to readpage to bring the page uptodate before writing
to it, because prepare_write will take care of that for us.
Splice is also wrong to SetPageUptodate before the page is actually uptodate.
This results in the old uninitialised memory leak. This gets fixed as a
matter of course when removing the readpage logic.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Stealing pages with splice is problematic because we cannot just insert
an uptodate page into the pagecache and hope the filesystem can take care
of it later.
We also cannot just ClearPageUptodate, then hope prepare_write does not
write anything into the page, because I don't think prepare_write gives
that guarantee.
Remove support for SPLICE_F_MOVE for now. If we really want to bring it
back, we might be able to do so with a the new filesystem buffered write
aops APIs I'm working on. If we really don't want to bring it back, then
we should decide that sooner rather than later, and remove the flag and
all the stealing infrastructure before anybody starts using it.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
- pipe/splice should use const pipe_buf_operations and file_operations
- struct pipe_inode_info has an unused field "start" : get rid of it.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Cc: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch changes struct file to use struct path instead of having
independent pointers to struct dentry and struct vfsmount, and converts all
users of f_{dentry,vfsmnt} in fs/ to use f_path.{dentry,mnt}.
Additionally, it adds two #define's to make the transition easier for users of
the f_dentry and f_vfsmnt.
Signed-off-by: Josef "Jeff" Sipek <jsipek@cs.sunysb.edu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
After the inode slimming patch that unionised i_pipe/i_bdev/i_cdev, it's
no longer enough to check for existance of ->i_pipe to verify that this
is a pipe.
Original patch from Eric Dumazet <dada1@cosmosbay.com>
Final solution suggested by Linus.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
- Consolidate page_cache_alloc
- Fix splice: only the pagecache pages and filesystem data need to use
mapping_gfp_mask.
- Fix grab_cache_page_nowait: same as splice, also honour NUMA placement.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Originally from Mark Fasheh <mark.fasheh@oracle.com>
generic_file_splice_write() does not remove S_ISUID or S_ISGID. This is
inconsistent with the way we generally write to files.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
This allows file systems to manage their own i_mutex locking while
still re-using the generic_file_splice_write() logic.
OCFS2 in particular wants this so that it can order cluster locks within
i_mutex.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
The splice_actor may be calling ->prepare_write() and ->commit_write(). We
want i_mutex on the inode being written to before calling those so that we
don't race i_size changes.
The double locking behavior is done elsewhere in splice.c, and if we
eventually want _nolock variants of generic_file_splice_write(), fs modules
might have to replicate the nasty locking code. We introduce
inode_double_lock() and inode_double_unlock() to consolidate the locking
rules into one set of functions.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Several issues noticed/fixed:
- We cannot reliably block in link_pipe() while holding both input and output
mutexes. So do preparatory checks before locking down both mutexes and doing
the link.
- The ipipe->nrbufs vs i check was bad, because we could have dropped the
ipipe lock in-between. This causes us to potentially look at unknown
buffers if we were racing with someone else reading this pipe.
Signed-off-by: Jens Axboe <axboe@suse.de>
Otherwise we could be racing with truncate/mapping removal.
Problem found/fixed by Nick Piggin <npiggin@suse.de>, logic rewritten
by me.
Signed-off-by: Jens Axboe <axboe@suse.de>
This can happen quite easily, if several processes are trying to splice
the same file at the same time. It's not a failure, it just means someone
raced with us in allocating this file page. So just dump the allocated
page and relookup the original.
Signed-off-by: Jens Axboe <axboe@suse.de>
Nick says that the current construct isn't safe. This goes back to the
original, but sets PIPE_BUF_FLAG_LRU on user pages as well as they all
seem to be on the LRU in the first place.
Signed-off-by: Jens Axboe <axboe@suse.de>
Looking at generic_file_buffered_write(), we need to unlock_page() if
prepare write fails and it isn't due to racing with truncate().
Also trim the size if ->prepare_write() fails, if we have to.
Signed-off-by: Jens Axboe <axboe@suse.de>
Currently we rely on the PIPE_BUF_FLAG_LRU flag being set correctly
to know whether we need to fiddle with page LRU state after stealing it,
however for some origins we just don't know if the page is on the LRU
list or not.
So remove PIPE_BUF_FLAG_LRU and do this check/add manually in pipe_to_file()
instead.
Signed-off-by: Jens Axboe <axboe@suse.de>
We need to use the minium of {len, PAGE_SIZE-off}, not {len, PAGE_SIZE}-off.
The latter doesn't make any sense, and could cause us to attempt negative
length transfers...
Signed-off-by: Jens Axboe <axboe@suse.de>
If SPLICE_F_GIFT is set, the user is basically giving this pages away to
the kernel. That means we can steal them for eg page cache uses instead
of copying it.
The data must be properly page aligned and also a multiple of the page size
in length.
Signed-off-by: Jens Axboe <axboe@suse.de>
The pipe ->map() method uses kmap() to virtually map the pages, which
is both slow and has known scalability issues on SMP. This patch enables
atomic copying of pipe pages, by pre-faulting data and using kmap_atomic()
instead.
lmbench bw_pipe and lat_pipe measurements agree this is a Good Thing. Here
are results from that on a UP machine with highmem (1.5GiB of RAM), running
first a UP kernel, SMP kernel, and SMP kernel patched.
Vanilla-UP:
Pipe bandwidth: 1622.28 MB/sec
Pipe bandwidth: 1610.59 MB/sec
Pipe bandwidth: 1608.30 MB/sec
Pipe latency: 7.3275 microseconds
Pipe latency: 7.2995 microseconds
Pipe latency: 7.3097 microseconds
Vanilla-SMP:
Pipe bandwidth: 1382.19 MB/sec
Pipe bandwidth: 1317.27 MB/sec
Pipe bandwidth: 1355.61 MB/sec
Pipe latency: 9.6402 microseconds
Pipe latency: 9.6696 microseconds
Pipe latency: 9.6153 microseconds
Patched-SMP:
Pipe bandwidth: 1578.70 MB/sec
Pipe bandwidth: 1579.95 MB/sec
Pipe bandwidth: 1578.63 MB/sec
Pipe latency: 9.1654 microseconds
Pipe latency: 9.2266 microseconds
Pipe latency: 9.1527 microseconds
Signed-off-by: Jens Axboe <axboe@suse.de>
The ->map() function is really expensive on highmem machines right now,
since it has to use the slower kmap() instead of kmap_atomic(). Splice
rarely needs to access the virtual address of a page, so it's a waste
of time doing it.
Introduce ->pin() to take over the responsibility of making sure the
page data is valid. ->map() is then reduced to just kmap(). That way we
can also share a most of the pipe buffer ops between pipe.c and splice.c
Signed-off-by: Jens Axboe <axboe@suse.de>
Found by Oleg Nesterov <oleg@tv-sign.ru>, fixed by me.
- Only allow full pages to go to the page cache.
- Check page != buf->page instead of using PIPE_BUF_FLAG_STOLEN.
- Remember to clear 'stolen' if add_to_page_cache() fails.
And as a cleanup on that:
- Make the bottom fall-through logic a little less convoluted. Also make
the steal path hold an extra reference to the page, so we don't have
to differentiate between stolen and non-stolen at the end.
Signed-off-by: Jens Axboe <axboe@suse.de>
- Check that page has suitable count for stealing in the regular pipes.
- pipe_to_file() assumes that the page is locked on succesful steal, so
do that in the pipe steal hook
- Missing unlock_page() in add_to_page_cache() failure.
Signed-off-by: Jens Axboe <axboe@suse.de>
Use the new find_get_pages_contig() to potentially look up the entire
splice range in one single call. This speeds up generic_file_splice_read()
quite a bit.
Signed-off-by: Jens Axboe <axboe@suse.de>
sys_splice() moves data to/from pipes with a file input/output. sys_vmsplice()
moves data to a pipe, with the input being a user address range instead.
This uses an approach suggested by Linus, where we can hold partial ranges
inside the pages[] map. Hopefully this will be useful for network
receive support as well.
Signed-off-by: Jens Axboe <axboe@suse.de>
Make the move_from_pipe() actors return number of bytes processed, then
move_from_pipe() can decide more cleverly when to move on to the next
buffer.
This fixes problems with pipe offset and differing file offset.
Signed-off-by: Jens Axboe <axboe@suse.de>
Since ->map() no longer locks the page, we need to adjust the handling
of those pages (and stealing) a little. This now passes full regressions
again.
Signed-off-by: Jens Axboe <axboe@suse.de>
- We need to adjust *ppos for writes as well.
- Copy back modified offset value if one was passed in, similar to
what sendfile does.
Signed-off-by: Jens Axboe <axboe@suse.de>
We need to ensure that we only drop a lock that is ordered last, to avoid
ABBA deadlocks with competing processes.
Signed-off-by: Jens Axboe <axboe@suse.de>
- generic_file_splice_read() more readable and correct
- Don't bail on page allocation with NONBLOCK set, just don't allow
direct blocking on IO (eg lock_page).
Signed-off-by: Jens Axboe <axboe@suse.de>
Basically an in-kernel implementation of tee, which uses splice and the
pipe buffers as an intelligent way to pass data around by reference.
Where the user space tee consumes the input and produces a stdout and
file output, this syscall merely duplicates the data inside a pipe to
another pipe. No data is copied, the output just grabs a reference to the
input pipe data.
Signed-off-by: Jens Axboe <axboe@suse.de>
We need not use ->f_pos as the offset for the file input/output. If the
user passed an offset pointer in through sys_splice(), just use that and
leave ->f_pos alone.
Signed-off-by: Jens Axboe <axboe@suse.de>
- capitalize consistently
- end sentences in one way or another
- update comment text to match the implementation
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Jens Axboe <axboe@suse.de>
Also corrects a few comments. Patch mainly from Ingo, changes by me.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Jens Axboe <axboe@suse.de>
- Kill the local variables that cache ->nrbufs, they just take up space.
- Only set do_wakeup for a real pipe. This is a big win for direct splicing.
- Kill i_mutex lock around ->f_pos update, regular io paths don't do this
either.
Signed-off-by: Jens Axboe <axboe@suse.de>
Using find_get_page() is a lot faster than find_or_create_page(). This
gets splice a lot closer to sendfile() for fd -> socket transfers.
Signed-off-by: Jens Axboe <axboe@suse.de>
It's more efficient for sendfile() emulation. Basically we cache an
internal private pipe and just use that as the intermediate area for
pages. Direct splicing is not available from sys_splice(), it is only
meant to be used for sendfile() emulation.
Additional patch from Ingo Molnar to avoid the PIPE_BUFFERS loop at
exit for the normal fast path.
Signed-off-by: Jens Axboe <axboe@suse.de>
add optional input and output offsets to sys_splice(), for seekable file
descriptors:
asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
int fd_out, loff_t __user *off_out,
size_t len, unsigned int flags);
semantics are straightforward: f_pos will be updated with the offset
provided by user-space, before the splice transfer is about to begin.
Providing a NULL offset pointer means the existing f_pos will be used
(and updated in situ). Providing an offset for a pipe results in
-ESPIPE. Providing an invalid offset pointer results in -EFAULT.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Jens Axboe <axboe@suse.de>
separate out the 'internal pipe object' abstraction, and make it
usable to splice. This cleans up and fixes several aspects of the
internal splice APIs and the pipe code:
- pipes: the allocation and freeing of pipe_inode_info is now more symmetric
and more streamlined with existing kernel practices.
- splice: small micro-optimization: less pointer dereferencing in splice
methods
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Update XFS for the ->splice_read/->splice_write changes.
Signed-off-by: Jens Axboe <axboe@suse.de>
We don't want to call into the read-ahead logic unless we are at the
start of a page, _or_ we have multiple pages to read.
Signed-off-by: Jens Axboe <axboe@suse.de>
We can get to out: with a NULL page, which we probably
don't want to be calling page_cache_release() on.
Signed-off-by: Dave Jones <davej@redhat.com>
Signed-off-by: Jens Axboe <axboe@suse.de>
Originally from Nick Piggin, just adapted to the newer branch.
You can't check PageLRU without holding zone->lru_lock. The page
release code can get away with it only because the page refcount is 0 at
that point. Also, you can't reliably remove pages from the LRU unless
the refcount is 0. Ever.
Signed-off-by: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Jens Axboe <axboe@suse.de>
Thanks to Andrew for the good explanation of why this is so. akpm writes:
If a page is under writeback and we remove it from pagecache, it's still
going to get written to disk. But the VFS no longer knows about that page,
nor that this page is about to modify disk blocks.
So there might be scenarios in which those
blocks-which-are-about-to-be-written-to get reused for something else.
When writeback completes, it'll scribble on those blocks.
This won't happen in ext2/ext3-style filesystems in normal mode because the
page has buffers and try_to_release_page() will fail.
But ext2 in nobh mode doesn't attach buffers at all - it just sticks the
page in a BIO, finds some new blocks, points the BIO at those blocks and
lets it rip.
While that write IO's in flight, someone could truncate the file. Truncate
won't block on the writeout because the page isn't in pagecache any more.
So truncate will the free the blocks from the file under the page's feet.
Then something else can reallocate those blocks. Then write data to them.
Now, the original write completes, corrupting the filesystem.
Signed-off-by: Jens Axboe <axboe@suse.de>
By cleaning up the writeback logic (killing write_one_page() and the manual
set_page_dirty()), we can get rid of ->stolen inside the pipe_buffer and
just keep it local in pipe_to_file().
This also adds dirty page balancing logic and O_SYNC handling.
Signed-off-by: Jens Axboe <axboe@suse.de>
Bastian Blank