[PATCHSET v3 0/3] Convert fs drivers to ->read_iter()

[PATCHSET v3 0/3] Convert fs drivers to ->read_iter()

[Jens Axboe]

Hi,

There are still a few users of fops->read() in the core parts of the
fs stack. Which is a shame, since it'd be nice to get rid of the
non-iterator parts of down the line, and reclaim that part of the
file_operations struct.

Outside of moving in that direction as a cleanup, using ->read_iter()
enables us to mark them with FMODE_NOWAIT. This is important for users
like io_uring, where per-IO nonblocking hints make a difference in how
efficiently IO can be done.

Those two things are my main motivation for starting this work, with
hopefully more to come down the line.

All patches have been booted and tested, and the corresponding test
cases from ltp have been run.

Since v2:
- Pass appropriate flags to get_unused_fd() and anon_inode_getfile().
  Doesn't change anything, but it does convey more closely what each
  of them look at.
- Add copy_to_iter_full() helper and use it for all conversions

 fs/signalfd.c       | 44 ++++++++++++++++++++++++++++----------------
 fs/timerfd.c        | 33 +++++++++++++++++++++++----------
 fs/userfaultfd.c    | 44 ++++++++++++++++++++++++++++----------------
 include/linux/uio.h | 10 ++++++++++
 4 files changed, 89 insertions(+), 42 deletions(-)

-- 
Jens Axboe

[PATCH 1/4] iov_iter: add copy_to_iter_full()

[Jens Axboe]

Add variant of copy_to_iter() that either copies the full amount asked
for and return success, or ensures that the iov_iter is back to where
it started on failure and returns false.

Suggested-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 include/linux/uio.h | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/include/linux/uio.h b/include/linux/uio.h
index 00cebe2b70de..9e9510672b28 100644
--- a/include/linux/uio.h
+++ b/include/linux/uio.h
@@ -197,6 +197,16 @@ size_t copy_to_iter(const void *addr, size_t bytes, struct iov_iter *i)
 	return 0;
 }
 
+static __always_inline __must_check
+bool copy_to_iter_full(const void *addr, size_t bytes, struct iov_iter *i)
+{
+	size_t copied = copy_to_iter(addr, bytes, i);
+	if (likely(copied == bytes))
+		return true;
+	iov_iter_revert(i, copied);
+	return false;
+}
+
 static __always_inline __must_check
 size_t copy_from_iter(void *addr, size_t bytes, struct iov_iter *i)
 {
-- 
2.43.0

Re: [PATCH 1/4] iov_iter: add copy_to_iter_full()

[Al Viro]

On Tue, Apr 09, 2024 at 09:22:15AM -0600, Jens Axboe wrote:
> Add variant of copy_to_iter() that either copies the full amount asked
> for and return success, or ensures that the iov_iter is back to where
> it started on failure and returns false.

FWIW, see git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs.git #work.iov_iter

There was an open-coded instance (skb_copy_linear()) that I'd converted to
that helper in the same commit; I can split it, of course, but I don't
see much point in that.

Re: [PATCH 1/4] iov_iter: add copy_to_iter_full()

[Jens Axboe]

On 4/9/24 11:06 AM, Al Viro wrote:
> On Tue, Apr 09, 2024 at 09:22:15AM -0600, Jens Axboe wrote:
>> Add variant of copy_to_iter() that either copies the full amount asked
>> for and return success, or ensures that the iov_iter is back to where
>> it started on failure and returns false.
> 
> FWIW, see git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs.git #work.iov_iter
> 
> There was an open-coded instance (skb_copy_linear()) that I'd converted to
> that helper in the same commit; I can split it, of course, but I don't
> see much point in that.

No reason to split it, I'll have a dependency regardless. I'll just pull
your branch in, so ignore patch 1 here, 2-4 will remain the same.

-- 
Jens Axboe

[PATCH 2/4] timerfd: convert to ->read_iter()

[Jens Axboe]

Switch timerfd to using fops->read_iter(), so it can support not just
O_NONBLOCK but IOCB_NOWAIT as well. With the latter, users like io_uring
interact with timerfds a lot better, as they can be driven purely
by the poll trigger.

Manually get and install the required fd, so that FMODE_NOWAIT can be
set before the file is installed into the file table.

No functional changes intended in this patch, it's purely a straight
conversion to using the read iterator method.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/timerfd.c | 33 +++++++++++++++++++++++----------
 1 file changed, 23 insertions(+), 10 deletions(-)

diff --git a/fs/timerfd.c b/fs/timerfd.c
index e9c96a0c79f1..f0d82dcbffef 100644
--- a/fs/timerfd.c
+++ b/fs/timerfd.c
@@ -262,17 +262,18 @@ static __poll_t timerfd_poll(struct file *file, poll_table *wait)
 	return events;
 }
 
-static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
-			    loff_t *ppos)
+static ssize_t timerfd_read_iter(struct kiocb *iocb, struct iov_iter *to)
 {
+	struct file *file = iocb->ki_filp;
 	struct timerfd_ctx *ctx = file->private_data;
 	ssize_t res;
 	u64 ticks = 0;
 
-	if (count < sizeof(ticks))
+	if (iov_iter_count(to) < sizeof(ticks))
 		return -EINVAL;
+
 	spin_lock_irq(&ctx->wqh.lock);
-	if (file->f_flags & O_NONBLOCK)
+	if (file->f_flags & O_NONBLOCK || iocb->ki_flags & IOCB_NOWAIT)
 		res = -EAGAIN;
 	else
 		res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks);
@@ -312,8 +313,8 @@ static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
 		ctx->ticks = 0;
 	}
 	spin_unlock_irq(&ctx->wqh.lock);
-	if (ticks)
-		res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks);
+	if (ticks && !copy_to_iter_full(&ticks, sizeof(ticks), to))
+		res = -EFAULT;
 	return res;
 }
 
@@ -384,7 +385,7 @@ static long timerfd_ioctl(struct file *file, unsigned int cmd, unsigned long arg
 static const struct file_operations timerfd_fops = {
 	.release	= timerfd_release,
 	.poll		= timerfd_poll,
-	.read		= timerfd_read,
+	.read_iter	= timerfd_read_iter,
 	.llseek		= noop_llseek,
 	.show_fdinfo	= timerfd_show,
 	.unlocked_ioctl	= timerfd_ioctl,
@@ -407,6 +408,7 @@ SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
 {
 	int ufd;
 	struct timerfd_ctx *ctx;
+	struct file *file;
 
 	/* Check the TFD_* constants for consistency.  */
 	BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC);
@@ -443,11 +445,22 @@ SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
 
 	ctx->moffs = ktime_mono_to_real(0);
 
-	ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx,
-			       O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
-	if (ufd < 0)
+	ufd = get_unused_fd_flags(flags & TFD_SHARED_FCNTL_FLAGS);
+	if (ufd < 0) {
 		kfree(ctx);
+		return ufd;
+	}
+
+	file = anon_inode_getfile("[timerfd]", &timerfd_fops, ctx,
+				    O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
+	if (IS_ERR(file)) {
+		put_unused_fd(ufd);
+		kfree(ctx);
+		return PTR_ERR(file);
+	}
 
+	file->f_mode |= FMODE_NOWAIT;
+	fd_install(ufd, file);
 	return ufd;
 }
 
-- 
2.43.0

Re: [PATCH 2/4] timerfd: convert to ->read_iter()

[Jens Axboe]

On 4/9/24 9:22 AM, Jens Axboe wrote:
> @@ -312,8 +313,8 @@ static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
>  		ctx->ticks = 0;
>  	}
>  	spin_unlock_irq(&ctx->wqh.lock);
> -	if (ticks)
> -		res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks);
> +	if (ticks && !copy_to_iter_full(&ticks, sizeof(ticks), to))
> +		res = -EFAULT;
>  	return res;
>  }

Dumb thinko here, as that should be:

if (ticks) {                                                            
	res = copy_to_iter(&ticks, sizeof(ticks), to);                  
	if (!res)                                                       
		res = -EFAULT;                                          
}            

I've updated my branch, just a heads-up. Odd how it passing testing,
guess I got stack lucky...

-- 
Jens Axboe

Re: [PATCH 2/4] timerfd: convert to ->read_iter()

[Marek Szyprowski]

Hi,

On 11.04.2024 00:27, Jens Axboe wrote:
> On 4/9/24 9:22 AM, Jens Axboe wrote:
>> @@ -312,8 +313,8 @@ static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
>>   		ctx->ticks = 0;
>>   	}
>>   	spin_unlock_irq(&ctx->wqh.lock);
>> -	if (ticks)
>> -		res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks);
>> +	if (ticks && !copy_to_iter_full(&ticks, sizeof(ticks), to))
>> +		res = -EFAULT;
>>   	return res;
>>   }
> Dumb thinko here, as that should be:
>
> if (ticks) {
> 	res = copy_to_iter(&ticks, sizeof(ticks), to);
> 	if (!res)
> 		res = -EFAULT;
> }
>
> I've updated my branch, just a heads-up. Odd how it passing testing,
> guess I got stack lucky...

The old version got its way into today's linux-next and bisecting the 
boot issues directed me here. There is nothing more to report, but I can 
confirm that the above change indeed fixes the problems observed on 
next-20240411.

Tested-by: Marek Szyprowski <m.szyprowski@samsung.com>

I hope that tomorrow's linux-next will have the correct version of this 
patch.

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland

Re: [PATCH 2/4] timerfd: convert to ->read_iter()

[Jens Axboe]

On 4/11/24 5:40 AM, Marek Szyprowski wrote:
> Hi,
> 
> On 11.04.2024 00:27, Jens Axboe wrote:
>> On 4/9/24 9:22 AM, Jens Axboe wrote:
>>> @@ -312,8 +313,8 @@ static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
>>>   		ctx->ticks = 0;
>>>   	}
>>>   	spin_unlock_irq(&ctx->wqh.lock);
>>> -	if (ticks)
>>> -		res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks);
>>> +	if (ticks && !copy_to_iter_full(&ticks, sizeof(ticks), to))
>>> +		res = -EFAULT;
>>>   	return res;
>>>   }
>> Dumb thinko here, as that should be:
>>
>> if (ticks) {
>> 	res = copy_to_iter(&ticks, sizeof(ticks), to);
>> 	if (!res)
>> 		res = -EFAULT;
>> }
>>
>> I've updated my branch, just a heads-up. Odd how it passing testing,
>> guess I got stack lucky...
> 
> The old version got its way into today's linux-next and bisecting the 
> boot issues directed me here. There is nothing more to report, but I can 
> confirm that the above change indeed fixes the problems observed on 
> next-20240411.

Yeah sorry about that :(

> Tested-by: Marek Szyprowski <m.szyprowski@samsung.com>

Thanks!

> I hope that tomorrow's linux-next will have the correct version of this 
> patch.

It should, the branches have been updated.

-- 
Jens Axboe

[PATCH 3/4] userfaultfd: convert to ->read_iter()

[Jens Axboe]

Rather than use the older style ->read() hook, use ->read_iter() so that
userfaultfd can support both O_NONBLOCK and IOCB_NOWAIT for non-blocking
read attempts.

Split the fd setup into two parts, so that userfaultfd can mark the file
mode with FMODE_NOWAIT before installing it into the process table. With
that, we can also defer grabbing the mm until we know the rest will
succeed, as the fd isn't visible before then.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/userfaultfd.c | 44 ++++++++++++++++++++++++++++----------------
 1 file changed, 28 insertions(+), 16 deletions(-)

diff --git a/fs/userfaultfd.c b/fs/userfaultfd.c
index 60dcfafdc11a..6d963402c835 100644
--- a/fs/userfaultfd.c
+++ b/fs/userfaultfd.c
@@ -31,6 +31,7 @@
 #include <linux/hugetlb.h>
 #include <linux/swapops.h>
 #include <linux/miscdevice.h>
+#include <linux/uio.h>
 
 static int sysctl_unprivileged_userfaultfd __read_mostly;
 
@@ -282,7 +283,7 @@ static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx,
 /*
  * Verify the pagetables are still not ok after having reigstered into
  * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any
- * userfault that has already been resolved, if userfaultfd_read and
+ * userfault that has already been resolved, if userfaultfd_read_iter and
  * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different
  * threads.
  */
@@ -1177,34 +1178,34 @@ static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 	return ret;
 }
 
-static ssize_t userfaultfd_read(struct file *file, char __user *buf,
-				size_t count, loff_t *ppos)
+static ssize_t userfaultfd_read_iter(struct kiocb *iocb, struct iov_iter *to)
 {
+	struct file *file = iocb->ki_filp;
 	struct userfaultfd_ctx *ctx = file->private_data;
 	ssize_t _ret, ret = 0;
 	struct uffd_msg msg;
-	int no_wait = file->f_flags & O_NONBLOCK;
 	struct inode *inode = file_inode(file);
+	bool no_wait;
 
 	if (!userfaultfd_is_initialized(ctx))
 		return -EINVAL;
 
+	no_wait = file->f_flags & O_NONBLOCK || iocb->ki_flags & IOCB_NOWAIT;
 	for (;;) {
-		if (count < sizeof(msg))
+		if (iov_iter_count(to) < sizeof(msg))
 			return ret ? ret : -EINVAL;
 		_ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode);
 		if (_ret < 0)
 			return ret ? ret : _ret;
-		if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg)))
+		_ret = !copy_to_iter_full(&msg, sizeof(msg), to);
+		if (_ret)
 			return ret ? ret : -EFAULT;
 		ret += sizeof(msg);
-		buf += sizeof(msg);
-		count -= sizeof(msg);
 		/*
 		 * Allow to read more than one fault at time but only
 		 * block if waiting for the very first one.
 		 */
-		no_wait = O_NONBLOCK;
+		no_wait = true;
 	}
 }
 
@@ -2172,7 +2173,7 @@ static const struct file_operations userfaultfd_fops = {
 #endif
 	.release	= userfaultfd_release,
 	.poll		= userfaultfd_poll,
-	.read		= userfaultfd_read,
+	.read_iter	= userfaultfd_read_iter,
 	.unlocked_ioctl = userfaultfd_ioctl,
 	.compat_ioctl	= compat_ptr_ioctl,
 	.llseek		= noop_llseek,
@@ -2192,6 +2193,7 @@ static void init_once_userfaultfd_ctx(void *mem)
 static int new_userfaultfd(int flags)
 {
 	struct userfaultfd_ctx *ctx;
+	struct file *file;
 	int fd;
 
 	BUG_ON(!current->mm);
@@ -2215,16 +2217,26 @@ static int new_userfaultfd(int flags)
 	init_rwsem(&ctx->map_changing_lock);
 	atomic_set(&ctx->mmap_changing, 0);
 	ctx->mm = current->mm;
-	/* prevent the mm struct to be freed */
-	mmgrab(ctx->mm);
+
+	fd = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS);
+	if (fd < 0)
+		goto err_out;
 
 	/* Create a new inode so that the LSM can block the creation.  */
-	fd = anon_inode_create_getfd("[userfaultfd]", &userfaultfd_fops, ctx,
+	file = anon_inode_create_getfile("[userfaultfd]", &userfaultfd_fops, ctx,
 			O_RDONLY | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL);
-	if (fd < 0) {
-		mmdrop(ctx->mm);
-		kmem_cache_free(userfaultfd_ctx_cachep, ctx);
+	if (IS_ERR(file)) {
+		put_unused_fd(fd);
+		fd = PTR_ERR(file);
+		goto err_out;
 	}
+	/* prevent the mm struct to be freed */
+	mmgrab(ctx->mm);
+	file->f_mode |= FMODE_NOWAIT;
+	fd_install(fd, file);
+	return fd;
+err_out:
+	kmem_cache_free(userfaultfd_ctx_cachep, ctx);
 	return fd;
 }
 
-- 
2.43.0

[PATCH 4/4] signalfd: convert to ->read_iter()

[Jens Axboe]

Rather than use the older style ->read() hook, use ->read_iter() so that
signalfd can support both O_NONBLOCK and IOCB_NOWAIT for non-blocking
read attempts.

Split the fd setup into two parts, so that signalfd can mark the file
mode with FMODE_NOWAIT before installing it into the process table.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/signalfd.c | 44 ++++++++++++++++++++++++++++----------------
 1 file changed, 28 insertions(+), 16 deletions(-)

diff --git a/fs/signalfd.c b/fs/signalfd.c
index e20d1484c663..4a5614442dbf 100644
--- a/fs/signalfd.c
+++ b/fs/signalfd.c
@@ -68,8 +68,7 @@ static __poll_t signalfd_poll(struct file *file, poll_table *wait)
 /*
  * Copied from copy_siginfo_to_user() in kernel/signal.c
  */
-static int signalfd_copyinfo(struct signalfd_siginfo __user *uinfo,
-			     kernel_siginfo_t const *kinfo)
+static int signalfd_copyinfo(struct iov_iter *to, kernel_siginfo_t const *kinfo)
 {
 	struct signalfd_siginfo new;
 
@@ -146,10 +145,10 @@ static int signalfd_copyinfo(struct signalfd_siginfo __user *uinfo,
 		break;
 	}
 
-	if (copy_to_user(uinfo, &new, sizeof(struct signalfd_siginfo)))
+	if (!copy_to_iter_full(&new, sizeof(struct signalfd_siginfo), to))
 		return -EFAULT;
 
-	return sizeof(*uinfo);
+	return sizeof(struct signalfd_siginfo);
 }
 
 static ssize_t signalfd_dequeue(struct signalfd_ctx *ctx, kernel_siginfo_t *info,
@@ -199,28 +198,27 @@ static ssize_t signalfd_dequeue(struct signalfd_ctx *ctx, kernel_siginfo_t *info
  * error code. The "count" parameter must be at least the size of a
  * "struct signalfd_siginfo".
  */
-static ssize_t signalfd_read(struct file *file, char __user *buf, size_t count,
-			     loff_t *ppos)
+static ssize_t signalfd_read_iter(struct kiocb *iocb, struct iov_iter *to)
 {
+	struct file *file = iocb->ki_filp;
 	struct signalfd_ctx *ctx = file->private_data;
-	struct signalfd_siginfo __user *siginfo;
-	int nonblock = file->f_flags & O_NONBLOCK;
+	size_t count = iov_iter_count(to);
 	ssize_t ret, total = 0;
 	kernel_siginfo_t info;
+	bool nonblock;
 
 	count /= sizeof(struct signalfd_siginfo);
 	if (!count)
 		return -EINVAL;
 
-	siginfo = (struct signalfd_siginfo __user *) buf;
+	nonblock = file->f_flags & O_NONBLOCK || iocb->ki_flags & IOCB_NOWAIT;
 	do {
 		ret = signalfd_dequeue(ctx, &info, nonblock);
 		if (unlikely(ret <= 0))
 			break;
-		ret = signalfd_copyinfo(siginfo, &info);
+		ret = signalfd_copyinfo(to, &info);
 		if (ret < 0)
 			break;
-		siginfo++;
 		total += ret;
 		nonblock = 1;
 	} while (--count);
@@ -246,7 +244,7 @@ static const struct file_operations signalfd_fops = {
 #endif
 	.release	= signalfd_release,
 	.poll		= signalfd_poll,
-	.read		= signalfd_read,
+	.read_iter	= signalfd_read_iter,
 	.llseek		= noop_llseek,
 };
 
@@ -265,20 +263,34 @@ static int do_signalfd4(int ufd, sigset_t *mask, int flags)
 	signotset(mask);
 
 	if (ufd == -1) {
+		struct file *file;
+
 		ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
 		if (!ctx)
 			return -ENOMEM;
 
 		ctx->sigmask = *mask;
 
+		ufd = get_unused_fd_flags(flags & O_CLOEXEC);
+		if (ufd < 0) {
+			kfree(ctx);
+			return ufd;
+		}
+
+		file = anon_inode_getfile("[signalfd]", &signalfd_fops, ctx,
+				       O_RDWR | (flags & O_NONBLOCK));
+		if (IS_ERR(file)) {
+			put_unused_fd(ufd);
+			kfree(ctx);
+			return ufd;
+		}
+		file->f_mode |= FMODE_NOWAIT;
+
 		/*
 		 * When we call this, the initialization must be complete, since
 		 * anon_inode_getfd() will install the fd.
 		 */
-		ufd = anon_inode_getfd("[signalfd]", &signalfd_fops, ctx,
-				       O_RDWR | (flags & (O_CLOEXEC | O_NONBLOCK)));
-		if (ufd < 0)
-			kfree(ctx);
+		fd_install(ufd, file);
 	} else {
 		struct fd f = fdget(ufd);
 		if (!f.file)
-- 
2.43.0