Message-ID: <54BE4F1D.7090807@gmail.com>
Date:	Tue, 20 Jan 2015 13:50:37 +0100
From:	"Michael Kerrisk (man-pages)" <mtk.manpages@...il.com>
To:	Fam Zheng <famz@...hat.com>, linux-kernel@...r.kernel.org
CC:	mtk.manpages@...il.com, Thomas Gleixner <tglx@...utronix.de>,
	Ingo Molnar <mingo@...hat.com>,
	"H. Peter Anvin" <hpa@...or.com>, x86@...nel.org,
	Alexander Viro <viro@...iv.linux.org.uk>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Kees Cook <keescook@...omium.org>,
	Andy Lutomirski <luto@...capital.net>,
	David Herrmann <dh.herrmann@...il.com>,
	Alexei Starovoitov <ast@...mgrid.com>,
	Miklos Szeredi <mszeredi@...e.cz>,
	David Drysdale <drysdale@...gle.com>,
	Oleg Nesterov <oleg@...hat.com>,
	"David S. Miller" <davem@...emloft.net>,
	Vivek Goyal <vgoyal@...hat.com>,
	Mike Frysinger <vapier@...too.org>,
	Theodore Ts'o <tytso@....edu>,
	Heiko Carstens <heiko.carstens@...ibm.com>,
	Rasmus Villemoes <linux@...musvillemoes.dk>,
	Rashika Kheria <rashika.kheria@...il.com>,
	Hugh Dickins <hughd@...gle.com>,
	Mathieu Desnoyers <mathieu.desnoyers@...icios.com>,
	Peter Zijlstra <peterz@...radead.org>,
	linux-fsdevel@...r.kernel.org, linux-api@...r.kernel.org,
	Josh Triplett <josh@...htriplett.org>,
	Paolo Bonzini <pbonzini@...hat.com>
Subject: Re: [PATCH RFC 5/6] epoll: Add implementation for epoll_mod_wait

Hello Fam Zheng,

On 01/20/2015 10:57 AM, Fam Zheng wrote:
> This syscall is a sequence of
> 
> 1) a number of epoll_ctl calls
> 2) a epoll_pwait, with timeout enhancement.
> 
> The epoll_ctl operations are embeded so that application doesn't have to use
> separate syscalls to insert/delete/update the fds before poll. It is more
> efficient if the set of fds varies from one poll to another, which is the
> common pattern for certain applications. 

Which applications? Could we have some specific examples? This is a 
complex API, and it needs good justification.

> For example, depending on the input
> buffer status, a data reading program may decide to temporarily not polling an
> fd.
> 
> Because the enablement of batching in this interface, even that regular
> epoll_ctl call sequence, which manipulates several fds, can be optimized to one
> single epoll_ctl_wait (while specifying spec=NULL to skip the poll part).
         ^^^^^^^^^^^^^^ should be epoll_mod_wait

I think you mean to say:

    The ability to batch multiple "epoll_ctl" operations into a single call
    means that even when no wait events are requested (i.e., spec == NULL),
    poll_mod_wait() provides a performance optimization over using multiple
    epoll_ctl() calls.

Right? If yes, please amend the commit message, and this text should
also make its way into the revised man page under a heading "NOTES".

> The only complexity is returning the result of each operation.  For each
> epoll_mod_cmd in cmds, the field "error" is an output field that will be stored
> the return code *iff* the command is executed (0 for success and -errno of the
> equivalent epoll_ctl call), and will be left unchanged if the command is not
> executed because some earlier error, for example due to failure of
> copy_from_user to copy the array.
> 
> Applications can utilize this fact to do error handling: they could initialize
> all the epoll_mod_wait.error to a positive value, which is by definition not a
> possible output value from epoll_mod_wait. Then when the syscall returned, they
> know whether or not the command is executed by comparing each error with the
> init value, if they're different, they have the result of the command.
> More roughly, they can put any non-zero and not distinguish "not run" from
> failure.

The "cmds' are not executed in a specified order plus the need to
initialize the 'errors' fields to a positive value feels a bit ugly.
And indeed the whole "command list was only partially run" case
is not pretty. Am I correct to understand that if an error is found
during execution of one of the "epoll_ctl" commands in 'cmds' then
the system call will return -1 with errno set, indicating an error,
even though the epoll interest list may have changed because some
of the earlier 'cmds' executed successfully? This all seems a bit of
a headache for user space.

I have a couple of questions:

Q1. I can see that batching "epoll_ctl" commands might be useful,
since it results in fewer systems calls. But, does it really
need to be bound together with the "epoll_pwait" functionality?
(Perhaps this point was covered in previous discussions, but
neither the message accompanying this patch nor the 0/6 man page
provide a compelling rationale for the need to bind these two
operations together.)

Yes, I realize you might save a system call, but it makes for a
cumbersome API that has the above headache, and also forces the 
need for double pointer indirection in the 'spec' argument (i.e., 
spec is a pointer to an array of structures where each element
in turn includes an 'events' pointer that points to another array).

Why not a simpler API with two syscalls such as:

epoll_ctl_batch(int epfd, int flags,
                int ncmds, struct epoll_mod_cmd *cmds);

epoll_pwait1(int epfd, struct epoll_event *events, int maxevents, 
             struct timespec *timeout, int clock_id, 
             const sigset_t *sigmask, size_t sigsetsize);

This gives us much of the benefit of reducing system calls, but 
with greater simplicity. And epoll_ctl_batch() could simply return
the number of 'cmds' that were successfully executed.)

Q2. In the man page in 0/6 you said that the 'cmds' were not 
guaranteed to be executed in order. Why not? If you did provide
such a guarantee, then, when using your current epoll_mod_wait(),
user space could do the following:

1. Initialize the cmd.errors fields to zero.
2. Call epoll_ctl_mod()
3. Iterate through cmd.errors looking for the first nonzero 
   field.

> Also, timeout parameter is enhanced: timespec is used, compared to the old ms
> scalar. This provides higher precision. 

Yes, that change seemed inevitable. It slightly puzzled me at the time when
Davide Libenzi added epoll_wait() that the timeout was milliseconds, even
though pselect() already had demonstrated the need for higher precision.
I should have called it out way back then :-{.

> The parameter field in struct
> epoll_wait_spec, "clockid", also makes it possible for users to use a different
> clock than the default when it makes more sense.
> 
> Signed-off-by: Fam Zheng <famz@...hat.com>
> ---
>  fs/eventpoll.c           | 60 ++++++++++++++++++++++++++++++++++++++++++++++++
>  include/linux/syscalls.h |  5 ++++
>  2 files changed, 65 insertions(+)
> 
> diff --git a/fs/eventpoll.c b/fs/eventpoll.c
> index e7a116d..2cc22c9 100644
> --- a/fs/eventpoll.c
> +++ b/fs/eventpoll.c
> @@ -2067,6 +2067,66 @@ SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
>  			      sigmask ? &ksigmask : NULL);
>  }
>  
> +SYSCALL_DEFINE5(epoll_mod_wait, int, epfd, int, flags,
> +		int, ncmds, struct epoll_mod_cmd __user *, cmds,
> +		struct epoll_wait_spec __user *, spec)
> +{
> +	struct epoll_mod_cmd *kcmds = NULL;
> +	int i, ret = 0;
> +	int cmd_size = sizeof(struct epoll_mod_cmd) * ncmds;
> +
> +	if (flags)
> +		return -EINVAL;
> +	if (ncmds) {
> +		if (!cmds)
> +			return -EINVAL;
> +		kcmds = kmalloc(cmd_size, GFP_KERNEL);
> +		if (!kcmds)
> +			return -ENOMEM;
> +		if (copy_from_user(kcmds, cmds, cmd_size)) {
> +			ret = -EFAULT;
> +			goto out;
> +		}
> +	}
> +	for (i = 0; i < ncmds; i++) {
> +		struct epoll_event ev = (struct epoll_event) {
> +			.events = kcmds[i].events,
> +			.data = kcmds[i].data,
> +		};
> +		if (kcmds[i].flags) {
> +			kcmds[i].error = ret = -EINVAL;

To make the 'ret' change a little more obvious, maybe it's better to write

			ret = kcmds[i].error = -EINVAL;

> +			goto out;
> +		}
> +		kcmds[i].error = ret = ep_ctl_do(epfd, kcmds[i].op, kcmds[i].fd, ev);

Likewise:
		ret = kcmds[i].error = ep_ctl_do(epfd, kcmds[i].op, kcmds[i].fd, ev);

> +		if (ret)
> +			goto out;
> +	}
> +	if (spec) {
> +		sigset_t ksigmask;
> +		struct epoll_wait_spec kspec;
> +		ktime_t timeout;
> +
> +		if(copy_from_user(&kspec, spec, sizeof(struct epoll_wait_spec)))

Cosmetic point: s/if(/if (/

> +			return -EFAULT;
> +		if (kspec.sigmask) {
> +			if (kspec.sigsetsize != sizeof(sigset_t))
> +				return -EINVAL;
> +			if (copy_from_user(&ksigmask, kspec.sigmask, sizeof(ksigmask)))
> +				return -EFAULT;
> +		}
> +		timeout = timespec_to_ktime(kspec.timeout);
> +		ret = epoll_pwait_do(epfd, kspec.events, kspec.maxevents,
> +				     kspec.clockid, timeout,
> +				     kspec.sigmask ? &ksigmask : NULL);

If I understand correctly, the implementation means that the
'size_t sigsetsize' field will probably need to be exposed to 
applications. In the existing epoll_pwait() call (as in  ppoll()
and pselect()) the 'size_t sigsetsize' argument is hidden by glibc.
However, unless we expect glibc to do some structure copying to/from
a structure that hides this field, then we're going end up exposing
'size_t sigsetsize' to applications. (This could be avoided, if we
split the API as I suggest above. glibc would do the same thing 
in epoll_pwait1() that it currently does in epoll_pwait().)

Thanks,

Michael

-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists