From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from smtp.kernel.org (aws-us-west-2-korg-mail-1.web.codeaurora.org [10.30.226.201]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.subspace.kernel.org (Postfix) with ESMTPS id 81FE9377566 for ; Tue, 17 Mar 2026 09:37:47 +0000 (UTC) Authentication-Results: smtp.subspace.kernel.org; arc=none smtp.client-ip=10.30.226.201 ARC-Seal:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1773740267; cv=none; b=W8Xwx3AhF02yll6Jmb3stVIRdw0LeCVQDd55uUrqf+L23cvDCmGTwvcphIp7TW8OHhcefvhAbIJrhE17yu7qI+bSqHvJVSi0Onu2JjoUKyuAuBIAgczsxrvulfj5MpkSpETA5mP3/+3fIdbGOw3SDvhSqAGv5ljmVjTHNSwL3nE= ARC-Message-Signature:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1773740267; c=relaxed/simple; bh=3908OmCkPwPpavgQLfiWA7d+Y76sSzivYABpXZDKedU=; h=Date:From:To:Cc:Subject:Message-ID:References:MIME-Version: Content-Type:Content-Disposition:In-Reply-To; b=ElJzUEDk4BhuMhvjmfhIFIn5DomWUC23/xop1B+CElXrQBAXNEJXH9RJgUeIA3Nk8GC7FB8ygk6wBWCcDdgKd3mvmPaALDDomW6mznKmWJj6ttz2uVtUP5JIYPiFzKb+kQ15ywWf3G8625rsl3ZNzRiJDBKYKvReGI8bs9f0jlQ= ARC-Authentication-Results:i=1; smtp.subspace.kernel.org; dkim=pass (2048-bit key) header.d=kernel.org header.i=@kernel.org header.b=ID4VjXHp; arc=none smtp.client-ip=10.30.226.201 Authentication-Results: smtp.subspace.kernel.org; dkim=pass (2048-bit key) header.d=kernel.org header.i=@kernel.org header.b="ID4VjXHp" Received: by smtp.kernel.org (Postfix) with ESMTPSA id 683A1C4CEF7; Tue, 17 Mar 2026 09:37:41 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=kernel.org; s=k20201202; t=1773740267; bh=3908OmCkPwPpavgQLfiWA7d+Y76sSzivYABpXZDKedU=; h=Date:From:To:Cc:Subject:References:In-Reply-To:From; b=ID4VjXHpTAT1SyuqsuYtdSP8PqxsJf6Ik/YXiROwmdjgJoJyGAKM+NUEj9VlRMrHL cn0tHf+PB4Mw74GlOpNKA1iSnuFMPbMNLsMVK2VnqvT86Qb8FhRW8GUBXSUaakgS4T G92uznoQYNWb8Asz/4FhSNJktgF4Mh0H750nj6XpMGA5XYr5TNgJXqeLbcfFpmsBkI 2MwrzN4OeCnEm3VPpe5qMP2cgzZncu5aObz5L/6StWNfNT8MuPPE+tXnwXlr7XSGaL 1BtIaRaK4uWC4wnrgDGyhdZjxMUqHyD1hE42yCJBK5j7t6yea8ZztUjF/irNJRCVTl 7A9kTOE1hn96Q== Date: Tue, 17 Mar 2026 09:37:39 +0000 From: "Lorenzo Stoakes (Oracle)" To: Sohil Mehta Cc: Breno Leitao , Andrew Morton , David Hildenbrand , Lorenzo Stoakes , Zi Yan , Baolin Wang , "Liam R. Howlett" , Nico Pache , Ryan Roberts , Dev Jain , Barry Song , Lance Yang , Vlastimil Babka , Suren Baghdasaryan , Michal Hocko , Brendan Jackman , Johannes Weiner , Mike Rapoport , linux-mm@kvack.org, linux-kernel@vger.kernel.org, usamaarif642@gmail.com, kas@kernel.org, kernel-team@meta.com, Wei Yang , "Accardi, Kristen C" Subject: Re: [PATCH v6 3/4] mm: huge_memory: refactor enabled_store() with change_enabled() Message-ID: <1b7d1bdf-c002-4543-b858-09fd2b1b73f9@lucifer.local> References: <20260311-thp_logs-v6-0-421e30d881e0@debian.org> <20260311-thp_logs-v6-3-421e30d881e0@debian.org> <322f6f2f-840e-462f-96b0-b603b9c88582@intel.com> Precedence: bulk X-Mailing-List: linux-kernel@vger.kernel.org List-Id: List-Subscribe: List-Unsubscribe: MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: On Mon, Mar 16, 2026 at 04:26:30PM -0700, Sohil Mehta wrote: > On 3/16/2026 3:12 AM, Breno Leitao wrote: > > >> I am not a mm expert and typically do not follow the mm list. Is there > >> an issue with the usage of non-atomic variants here? The commit message > >> says this uses the same pattern as set_anon_enabled_mode(). > >> > >> However, set_anon_enabled_mode() has a spinlock=>huge_anon_orders_lock > >> protecting the access. But, transparent_hugepage_flags seems to be > >> unprotected in that regard. > > > > I don't think that the atomic vs non-atomic will not help much, given > > this is a compoud operation. Independently if this is atomic or not, it > > is racy with anyone changing these fields (transparent_hugepage_flags). > > In other words, Atomic ops make each individual bit flip safe, but > > set_global_enabled_mode() and defrag_store() need to flip multiple bits > > as a group. With atomic ops, two concurrent writers can still interleave > > and leave the flags in an invalid state. > > You are right it is a compound operation. So, there is an existing issue > with two concurrent writers which can leave the flags in an invalid state. > > But, I was wondering if there is a slightly different issue now due to > the non-atomic part. Some updates could be lost depending on the timing > of the operation. > > For example, > > CPU1 does: CPU2 does: > set_global_enabled_mode("always") defrag_store("always") > > ___test_and_set_bit(): > // Trying to set bit 1 > > old = *p > // reads flags, sees defrag bit=0 > > set_bit(DEFRAG_DIRECT_FLAG) > // atomic: sets bit 3 > > > *p = old | TRANSPARENT_HUGEPAGE_FLAG; > // writes back old value with bit 1 set > // DEFRAG_DIRECT_FLAG is lost (reverted to 0) > > IIUC, this issue didn't exist before. I think it might be safer to use > the atomic test_and_set_bit() to be compatible with the older code. > Though, I'll leave it up to you as I don't have expertise here. No, it's up to the maintainers :) Given the above I think we should switch back to the atomic accessors. We can address the broader issues with this horrible code in a separate series. > > Overall, as you mentioned below, protecting transparent_hugepage_flags > with a spinlock seems like a better, long-term solution to me as well. Yeah, let's look at doing a follow up that cleans this up in general and address that then. > > > > > That said, Although I don't think this patch is making it worse, I think > > the is a racy issue here that we can make better. > > > > My suggestion is to move the rest of the helpers (defrag_store()) to use > > sysfs_match_string(), and then create a thp_flags_lock spinlock to > > protect operations against transparent_hugepage_flags. Any concern about > > this approach? > > > > Cheers, Lorenzo