From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from bombadil.infradead.org (bombadil.infradead.org [198.137.202.133]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id AC7F9EEC2AD for ; Mon, 23 Feb 2026 23:15:10 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; q=dns/txt; c=relaxed/relaxed; d=lists.infradead.org; s=bombadil.20210309; h=Sender: Content-Transfer-Encoding:Content-Type:List-Subscribe:List-Help:List-Post: List-Archive:List-Unsubscribe:List-Id:In-Reply-To:MIME-Version:References: Message-ID:Subject:Cc:To:From:Date:Reply-To:Content-ID:Content-Description: Resent-Date:Resent-From:Resent-Sender:Resent-To:Resent-Cc:Resent-Message-ID: List-Owner; bh=5JLWNjnEKbpCD/XKZGE+q6Gpw/MH/46gVWeZwTRBoJ0=; b=hkRPIzlKg9twGM 2qU2C7dEK3c7U9wUX9hjELV7s2X2fhGuqwwbxrGKqdR8je2pmg6VHYp/5pUnedKPyW6GRxaJ+eaFr z4en2PfQbP9i5AOWR/n5yMhkke5nygZn0J2P0EDlxcV8l08vNPugN+ZTdEDDwSy9p/1B8CSzWTEnL 9A958ni1YSIYwOWso1X3mza0bvV/Z/zq4hxRolOaa1T9gNtW8adL3jQcjTW5RObPdbGt5TQ2pYeNb 1GnAMwYce7V5nylRGGPlb06zyBbivr+cnP73U11TGTUwFqQG+Tahl7boby2L7uw+V1RFv0Q6OyT8E cDsyxbPbOl3rURzxQSuA==; Received: from localhost ([::1] helo=bombadil.infradead.org) by bombadil.infradead.org with esmtp (Exim 4.98.2 #2 (Red Hat Linux)) id 1vuf8w-00000001ADK-09I4; Mon, 23 Feb 2026 23:15:10 +0000 Received: from mail-wm1-x32b.google.com ([2a00:1450:4864:20::32b]) by bombadil.infradead.org with esmtps (Exim 4.98.2 #2 (Red Hat Linux)) id 1vuf8s-00000001ACe-1WSP for linux-phy@lists.infradead.org; Mon, 23 Feb 2026 23:15:08 +0000 Received: by mail-wm1-x32b.google.com with SMTP id 5b1f17b1804b1-4836fc075d2so5693045e9.0 for ; Mon, 23 Feb 2026 15:15:05 -0800 (PST) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20230601; t=1771888504; x=1772493304; darn=lists.infradead.org; h=in-reply-to:content-disposition:mime-version:references:message-id :subject:cc:to:from:date:from:to:cc:subject:date:message-id:reply-to; bh=tKTS42YBOWM388O2yeAIptWildo0XwTbAk8by2Z0Q8w=; b=PcsT+85hQ3SJPpcAG1QVmzXNnZC2+yvy+DNo/ZFT9kjz3MOSJX5pnb314pwWzFBpae 0QQC1gXzRZT5dO0k4dHa50g/lzuXmq+EWdkc2R/y2JmxjzNzipBvrwJFFL3LoWI0E7c0 oYaCGFxq1JVnM5V9QGJtRZMfxTLPvXbMJ0/QAEIsfI6L8XDDRuIKIA44X60iz/elJZcX vC2c4syH3p+wips0wggB1oM72IDhIQNKGG6i9qLUmjfXSSLGy2dbPYOr1ZNP/NCTqScz w/B0QKaf7VRliAmKRjjbJKaAr+ZOxfsxDdAvPKWpzB4z22XSsuJlY1s0GmvFYjJpqsYl yt7Q== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20230601; t=1771888504; x=1772493304; h=in-reply-to:content-disposition:mime-version:references:message-id :subject:cc:to:from:date:x-gm-gg:x-gm-message-state:from:to:cc :subject:date:message-id:reply-to; bh=tKTS42YBOWM388O2yeAIptWildo0XwTbAk8by2Z0Q8w=; b=etyI3uisslzTfkywgMNbTj7lhjwi4qWIV2t9eScB9b203pX5VApxQHQIo905n2W1ki I/ziSEwbvhyjx4PVia3niYjKebxtq2a7226eshfrkpd6FmDZKHogcjxBNe6cw4CRATyd bA9E9jd5ZW734qoFuKT56NzwG2DL+Fp/vzTXHxXrUVl+NpvkEF+kGjv2s99vUazVk8vi yfdjx+cY4MswdV4J0KEAttCiv8jyZurYELBrLYAqf9NYLU3NXGVxq1HT5z5ghDme+gqM 7JDivSX7JswdjoPc4pMvPH5nTpJ2G+bRM2+DyLi+skqRTTGyCxHJ73wIY4hTl2WDSMRi mtww== X-Forwarded-Encrypted: i=1; AJvYcCU+lH39ctwCBMEp5MWlHPxrMDDalwPDKAr6KWOevoepKPKeIoxR6ZgzXB0dZ/shaLYF/Usq5WQgNlI=@lists.infradead.org X-Gm-Message-State: AOJu0YxYrT/2AK6GJ5sV+Wf88qqa6Gsm3jqG3/yUf85jU2GX/+ig8k3O aN5Xv1xqkauwZSg4Ac8ys+nIn3wm7C4aTwODAyWu9wFl8owkSLiYlc7w X-Gm-Gg: AZuq6aIgfddqerGJoxS98faAEccm6FjgDwZqp7Geos2enoccfJpjm7dPjrSypbN3JG3 ydJ5XmwbJt8zbJcemJzLHV+HPklUp20KwGVO3ZRSsKRuA1DHS8xjqHqhd0V8ccpQmK54Awt98/Y 9BLaS+qn801lgdMypuvWULyNQyjWsfCBdXIYBMalzchDHbhoirNE2LDvts5laJ9VjdkU+6/UPgC ortpShiHUn8aLcJjeqX0yzfuIaWccm9A8t40zNCTba6247OnoukVwYiGkpOonxr6Xelnu0lv/WN x7kaWfZ3Q//uuGPkVeUwvKV8DLDfYJhGk5YiT+iYnVW492mP3IhDse9N+nsC1kNuvtDGQA4s1DN ndCd+MtZSWMTPQQHOMoboWEb/mp3xYKFwjoaL8xfwIG7RZi3W3VMwBgDoGJdNdmd94Ppb5k2vfC jHFu1M956lkeaFKdY= X-Received: by 2002:a05:600c:3b28:b0:483:7d93:9f8a with SMTP id 5b1f17b1804b1-483a95a75ebmr100676895e9.1.1771888503732; Mon, 23 Feb 2026 15:15:03 -0800 (PST) Received: from skbuf ([2a02:2f04:d809:4b00:a289:4e95:a726:c01d]) by smtp.gmail.com with ESMTPSA id 5b1f17b1804b1-483b8919db6sm4133545e9.31.2026.02.23.15.15.02 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Mon, 23 Feb 2026 15:15:03 -0800 (PST) Date: Tue, 24 Feb 2026 01:15:00 +0200 From: Vladimir Oltean To: Dmitry Torokhov Cc: Vinod Koul , Kishon Vijay Abraham I , Neil Armstrong , "Rafael J. Wysocki" , Geert Uytterhoeven , Johan Hovold , Claudiu Beznea , "Dr. David Alan Gilbert" , Peter Griffin , Dmitry Baryshkov , Krzysztof Kozlowski , Zijun Hu , linux-phy@lists.infradead.org, linux-kernel@vger.kernel.org Subject: Re: [PATCH] phy: core: fix potential UAF in of_phy_simple_xlate() Message-ID: <20260223231500.zeffezslctqamhp7@skbuf> References: MIME-Version: 1.0 Content-Disposition: inline In-Reply-To: X-CRM114-Version: 20100106-BlameMichelson ( TRE 0.8.0 (BSD) ) MR-646709E3 X-CRM114-CacheID: sfid-20260223_151506_469289_F1157EBC X-CRM114-Status: GOOD ( 48.61 ) X-BeenThere: linux-phy@lists.infradead.org X-Mailman-Version: 2.1.34 Precedence: list List-Id: Linux Phy Mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Sender: "linux-phy" Errors-To: linux-phy-bounces+linux-phy=archiver.kernel.org@lists.infradead.org Hi Dmitry, On Fri, Feb 20, 2026 at 05:01:50PM -0800, Dmitry Torokhov wrote: > On Thu, Feb 19, 2026 at 04:11:37PM -0800, Dmitry Torokhov wrote: > > On Thu, Feb 19, 2026 at 03:57:11PM -0800, Dmitry Torokhov wrote: > > > The implementation put_device()s located device and then uses > > > container_of() on the pointer. The device may disappear by that time, > > > resulting in UAF. > > > > > > Fix the problem by keeping the reference to the framer device, > > > avoiding getting an extra reference to it in framer_get(), and making > > > sure to drop the reference in error path when we fail to get the module. > > > > Hmm, I was too rash. There are bunch of other xlate functions that need > > to be updated to take the reference. > > So I am convinced that xlate functions need to bump up the reference to > phy devices they return. The question is how to deal with the ones that > do not. I can either convert them in the same patch (the changes are > quite mechanical) or we can do the whole song and dance, introduce a > flag, set it up in converted xlate functions, have the core respect it, > and then remove it from xlates and from the core when it is all done. > > Please let me know. You have hit on a weak spot of the Generic PHY framework and I would like to encourage you to follow through with patches for this finding (although it won't be exactly trivial, I think it's doable with some determination). On your direct question: It will impact downstream in subtle and unpleasant ways if you change the of_xlate() semantics w.r.t. reference counting, but the code will still compile just as before, i.e. when looking just at your downstream driver (what 99% of developers do) there will be no obvious way of knowing that something in the API has changed. I would avoid doing that. But first the problem. Too little has been said about the problem, and too much about the solution. We can't find a good solution if we don't call out the problem properly first. The phy_get() function follows a "lookup-then-get" approach for PHY providers, rather than "atomic-lookup-and-get". Namely, the phy_provider->of_xlate() caller, which is _of_phy_get(), returns a struct phy * with its underlying device reference count nominally at 1. All callers of _of_phy_get() do later call get_device() to bump this refcount for each PHY consumer, but it is "too late" in the sense that a window has been opened where the PHY provider driver can unbind, and the reference count of &phy->dev can drop to 0 and it can be freed. Immediately after being created by phy_create(), the &phy->dev has a refcount of 1, and only the action of its provider driver (calling phy_destroy() directly or through devres, on driver unbind) can drop that refcount to 0. As long as the driver doesn't do that, the &phy->dev refcount is not in danger. Then the PHY is exposed to the outside world using of_phy_provider_register(), where the fact that the provider driver can concurrently unregister starts being a problem. Therefore, I would say that the problem is that consumers, aka phy_get() callers, can get a phy with a &phy->dev refcount that is not already bumped by the time they are handed over that PHY. Mechanically, PHY lookup happens under &phy_provider_mutex, and I believe it would be sufficient to call get_device() under this lock, in order for consumers to get PHYs with their reference bumped. Why? Let's consider what a PHY provider does when unbinding. It runs the PHY registration and creation processes in reverse, i.e. - it calls of_phy_provider_unregister() directly or through devres - it calls phy_destroy() directly or through devres Since of_phy_provider_unregister() also acquires &phy_provider_mutex, this acts like a synchronization barrier. There are 2 cases: Consumer Provider phy_get() -> mutex_lock(&phy_provider_mutex) -> get_device(&phy->dev) -> 2 -> mutex_unlock(&phy_provider_mutex) of_phy_provider_unregister() -> mutex_lock(&phy_provider_mutex) -> removes phy from list -> mutex_unlock(&phy_provider_mutex) phy_destroy() -> device_unregister(&phy->dev) -> device_del(dev); -> put_device(dev); // -> 1 and the consumer remains with a "zombie" PHY device (more later) Or Consumer Provider phy_get() of_phy_provider_unregister() -> mutex_lock(&phy_provider_mutex) -> removes phy from list -> mutex_unlock(&phy_provider_mutex) -> mutex_lock(&phy_provider_mutex) -> doesn't find the PHY device! It was unpublished even if not freed -> mutex_unlock(&phy_provider_mutex) phy_destroy() -> device_unregister(&phy->dev) -> device_del(dev); -> put_device(dev); // -> 1 and the consumer won't be allowed to even find the PHY device in this case. So this is why get_device() under phy_provider_mutex actually helps. Now, it is much less important whether you push the get_device() to the actual of_xlate() PHY vendor implementation or not. For simplicity sake, I suggest you don't do that, but instead keep it in the core, to preserve driver API semantics (patch not even compile-tested): -- >8 -- diff --git a/drivers/phy/phy-core.c b/drivers/phy/phy-core.c index 21aaf2f76e53..c50e38f057a8 100644 --- a/drivers/phy/phy-core.c +++ b/drivers/phy/phy-core.c @@ -122,17 +122,19 @@ EXPORT_SYMBOL_GPL(phy_remove_lookup); static struct phy *phy_find(struct device *dev, const char *con_id) { const char *dev_id = dev_name(dev); - struct phy_lookup *p, *pl = NULL; + struct phy *phy = NULL; + struct phy_lookup *p; mutex_lock(&phy_provider_mutex); list_for_each_entry(p, &phys, node) if (!strcmp(p->dev_id, dev_id) && !strcmp(p->con_id, con_id)) { - pl = p; + phy = p->phy; + get_device(&phy->dev); break; } mutex_unlock(&phy_provider_mutex); - return pl ? pl->phy : ERR_PTR(-ENODEV); + return phy ? phy : ERR_PTR(-ENODEV); } static struct phy_provider *of_phy_provider_lookup(struct device_node *node) @@ -649,6 +651,7 @@ static struct phy *_of_phy_get(struct device_node *np, int index) } phy = phy_provider->of_xlate(phy_provider->dev, &args); + get_device(&phy->dev); out_put_module: module_put(phy_provider->owner); @@ -682,10 +685,10 @@ struct phy *of_phy_get(struct device_node *np, const char *con_id) if (IS_ERR(phy)) return phy; - if (!try_module_get(phy->ops->owner)) + if (!try_module_get(phy->ops->owner)) { + put_device(&phy->dev); return ERR_PTR(-EPROBE_DEFER); - - get_device(&phy->dev); + } return phy; } @@ -803,10 +806,10 @@ struct phy *phy_get(struct device *dev, const char *string) if (IS_ERR(phy)) return phy; - if (!try_module_get(phy->ops->owner)) + if (!try_module_get(phy->ops->owner)) { + put_device(&phy->dev); return ERR_PTR(-EPROBE_DEFER); - - get_device(&phy->dev); + } link = device_link_add(dev, &phy->dev, DL_FLAG_STATELESS); if (!link) @@ -969,11 +972,10 @@ struct phy *devm_of_phy_get_by_index(struct device *dev, struct device_node *np, if (!try_module_get(phy->ops->owner)) { devres_free(ptr); + put_device(&phy->dev); return ERR_PTR(-EPROBE_DEFER); } - get_device(&phy->dev); - *ptr = phy; devres_add(dev, ptr); -- >8 -- The patch above leaves a few loose ends. The most obvious is of_phy_simple_xlate(), which has that put_device() to balance out class_find_device_by_of_node() - which bumps the device refcount to 2. It "looks" wrong but it is consistent with vendor implementations of of_xlate(), which also provide a phy->dev refcount of 1. And as explained, the refcount never _actually_ drops to 0 with the above patch. I would actually address _only_ of_phy_simple_xlate(), for cosmetics sake. Instead of devm_of_phy_provider_register(..., of_phy_simple_xlate), I would offer a new helper, devm_of_phy_simple_provider_register(), which would internally use a callback that doesn't drop the refcount (say __of_phy_simple_xlate() for lack of imagination). I would convert vendor PHY drivers one by one to use this (it would be valid at any time to use either the old or the new method). You'd notice that most of the of_phy_simple_xlate() occurrences go away, except for freescale/phy-fsl-lynx-28g.c which calls this function directly from its own lynx_28g_xlate(). It will still have to keep calling it, and for refcount equalization purposes that weird put_device() will have to continue to exist. Just leave a comment as to why that is. But there are more important loose ends still. I mentioned that "zombie" device. We've solved the memory safety issue, but it's possible for consumers to hold onto a phy whose provider has disappeared. The refcount of &phy->dev hasn't dropped to 0, so technically it's a valid object, but from PHY API perspective, it's still possible to call phy_init(), phy_power_on() and friends on this PHY, and the Generic PHY core will be happy to further call into the phy->ops->init(), phy->ops->power_on() etc. But the driver has unbound, so it should really be left alone. If we fix the UAF but leave the zombie PHY problem, we've effectively done nothing but silence static analysis checkers, while the code path where the PHY provider unbinds effectively remains treated as poorly as before, just moving the crashes to a different place. I suspect what needs to be done here is to introduce a "bool dead" or similar, which is to be set from phy_destroy() and checked from every API call. The idea is that API functions on zombie PHYs should fail without calling into their driver. I **suppose** that try_module_get(phy->ops->owner) "tried" to avoid this situation, but it just protects against module removal, not against "echo device > /sys/bus/.../unbind". So it's absolutely incomplete and easily bypassable. Finally, I have identified one more loose end still. /** * of_phy_put() - release the PHY * @phy: the phy returned by of_phy_get() * * Releases a refcount the caller received from of_phy_get(). */ void of_phy_put(struct phy *phy) { if (!phy || IS_ERR(phy)) return; mutex_lock(&phy->mutex); if (phy->ops->release) phy->ops->release(phy); mutex_unlock(&phy->mutex); module_put(phy->ops->owner); put_device(&phy->dev); } EXPORT_SYMBOL_GPL(of_phy_put); This function is called by PHY consumers. A PHY provider can have multiple consumers (example in the case of Ethernet: a QSGMII SerDes lane has 4 MAC ports multiplexed onto it). If a single consumer calls of_phy_put() to release its reference to the SerDes lane, the phy->ops->release() method makes absolutely no sense. There are 3 remaining consumers with handles to the lane! But we aren't even telling the PHY which consumer has disappeared! It has nothing useful to do with this information. Looking at actual phy_ops implementations, what they want to know is when _all_ consumers went away, not when individual consumers did. So the phy->ops->release() call needs to be put somewhere which is executed when all consumers disappear. If I were to guess, that would be the phy_release() class callback, but this is completely untested. Sorry that this email is a bit long, it got a bit late writing it and I don't have a lot of energy left to trim it down. In summary I found 4 highly related problems: - use after free - misleading of_phy_simple_xlate() API pattern (not a functional issue) - zombie PHYs - premature release call -- linux-phy mailing list linux-phy@lists.infradead.org https://lists.infradead.org/mailman/listinfo/linux-phy