Message-Id: <1476957132-27818-2-git-send-email-wagi@monom.org>
Date:   Thu, 20 Oct 2016 11:52:07 +0200
From:   Daniel Wagner <wagi@...om.org>
To:     linux-kernel@...r.kernel.org
Cc:     Ming Lei <ming.lei@...onical.com>,
        "Luis R . Rodriguez" <mcgrof@...nel.org>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        "Srivatsa S . Bhat" <srivatsa.bhat@...ux.vnet.ibm.com>,
        "Rafael J . Wysocki" <rjw@...k.pl>,
        Daniel Vetter <daniel.vetter@...el.com>,
        Takashi Iwai <tiwai@...e.com>,
        Bjorn Andersson <bjorn.andersson@...aro.org>,
        Arend van Spriel <arend.vanspriel@...adcom.com>,
        Daniel Wagner <daniel.wagner@...-carit.de>
Subject: [PATCH v6 1/6] firmware: move umh locking code into fw_load_from_user_helper()

From: Daniel Wagner <daniel.wagner@...-carit.de>

When we load the firmware directly we don't need to take the umh
lock(1). So move this part inside fw_load_from_user_helper which is only
available when CONFIG_FW_LOADER_USER_HELPER is set.

This avoids a dependency on firmware_loading_timeout() even when
!CONFIG_FW_LOADER_USER_HELPER. firmware_loading_timeout() will be moved
into the CONFIG_FW_LOADER_USER_HELPER section later.

The usermodehelper locking code was added by b298d289c792 ("PM / Sleep:
Fix freezer failures due to racy usermodehelper_is_disabled()").

(1) Luis analized this:

The original goal of the usermode helper lock came can be traced back in
time via a144c6a ("PM: Print a warning if firmware is requested when
task are frozen) when Rafael noticed drivers were calling to request
firmware on *resume* calls! Why would that be an issue? It was because
of the stupid delays incurred on resume when firmware *was not found*
__due__ to the stupid user mode helper timeout delay and people
believing this often meant users confusing resume *stalling* for resume
was *broken! Later b298d289c7
("PM / Sleep: Fix freezer failures due to racy
usermodehelper_is_disabled()") addressed races. That code in turn was
later massaged into shape to better accept direct FS loading via commit
4e0c92d015235 ("firmware: Refactoring for splitting user-mode helper
code").

So for a while we've been nagging driver developers to not add
request_firmware() calls on resume calls. In fact the
drivers/base/firmware_class.c code *kills* firmware UMH calls when we go
to suspend for the firmware cache, as such even on suspend its a stupid
idea to use the UMH, not only because it can stall suspend but *also*
because we kill these UMH calls.

[...]

So, as I see it the user mode helper lock should have *nothing* to do
with the race between reading files from the kernel and having a path
ready. If it was *used* for that, that was papering over the real
issue. Its no different of a hack for instance as if a driver using
request_firmware() tried to use async probe to avoid the same race. Yes
it will help, but no, it does not mean it is deterministically safe.

Reviewing commit 247bc03742545 ("PM / Sleep: Mitigate race between the freezer
and request_firmware()") which originally extended umh state machine from
just being enabled/disabled, with the concepts of UMH_ENABLED, UMH_FREEZING,
UMH_DISABLED -- its goal was to prevent UMH uses during suspend. So -- the
"UMH lock" on firmware was actually added to help avoid races between freezing
and request_firmware(). We should not re-use UMH status notifiers when the
firmware UMH is disabled for the same concepts -- if we needed such a concept
then we should take this out from UMH code and generalize it.

Signed-off-by: Daniel Wagner <daniel.wagner@...-carit.de>
Cc: Ming Lei <ming.lei@...onical.com>
Cc: Luis R. Rodriguez <mcgrof@...nel.org>
Cc: Greg Kroah-Hartman <gregkh@...uxfoundation.org>
---
 drivers/base/firmware_class.c | 52 +++++++++++++++++++++++--------------------
 1 file changed, 28 insertions(+), 24 deletions(-)

diff --git a/drivers/base/firmware_class.c b/drivers/base/firmware_class.c
index 960f8f7c7aa2..d4fee06b67f3 100644
--- a/drivers/base/firmware_class.c
+++ b/drivers/base/firmware_class.c
@@ -981,13 +981,38 @@ static int fw_load_from_user_helper(struct firmware *firmware,
 				    unsigned int opt_flags, long timeout)
 {
 	struct firmware_priv *fw_priv;
+	int ret;
+
+	timeout = firmware_loading_timeout();
+	if (opt_flags & FW_OPT_NOWAIT) {
+		timeout = usermodehelper_read_lock_wait(timeout);
+		if (!timeout) {
+			dev_dbg(device, "firmware: %s loading timed out\n",
+				name);
+			return -EBUSY;
+		}
+	} else {
+		ret = usermodehelper_read_trylock();
+		if (WARN_ON(ret)) {
+			dev_err(device, "firmware: %s will not be loaded\n",
+				name);
+			return ret;
+		}
+	}
 
 	fw_priv = fw_create_instance(firmware, name, device, opt_flags);
-	if (IS_ERR(fw_priv))
-		return PTR_ERR(fw_priv);
+	if (IS_ERR(fw_priv)) {
+		ret = PTR_ERR(fw_priv);
+		goto release_lock;
+	}
 
 	fw_priv->buf = firmware->priv;
-	return _request_firmware_load(fw_priv, opt_flags, timeout);
+	ret = _request_firmware_load(fw_priv, opt_flags, timeout);
+
+release_lock:
+	usermodehelper_read_unlock();
+
+	return ret;
 }
 
 #ifdef CONFIG_PM_SLEEP
@@ -1150,25 +1175,6 @@ _request_firmware(const struct firmware **firmware_p, const char *name,
 	if (ret <= 0) /* error or already assigned */
 		goto out;
 
-	ret = 0;
-	timeout = firmware_loading_timeout();
-	if (opt_flags & FW_OPT_NOWAIT) {
-		timeout = usermodehelper_read_lock_wait(timeout);
-		if (!timeout) {
-			dev_dbg(device, "firmware: %s loading timed out\n",
-				name);
-			ret = -EBUSY;
-			goto out;
-		}
-	} else {
-		ret = usermodehelper_read_trylock();
-		if (WARN_ON(ret)) {
-			dev_err(device, "firmware: %s will not be loaded\n",
-				name);
-			goto out;
-		}
-	}
-
 	ret = fw_get_filesystem_firmware(device, fw->priv);
 	if (ret) {
 		if (!(opt_flags & FW_OPT_NO_WARN))
@@ -1185,8 +1191,6 @@ _request_firmware(const struct firmware **firmware_p, const char *name,
 	if (!ret)
 		ret = assign_firmware_buf(fw, device, opt_flags);
 
-	usermodehelper_read_unlock();
-
  out:
 	if (ret < 0) {
 		release_firmware(fw);
-- 
2.7.4

Powered by blists - more mailing lists