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 vger.kernel.org (vger.kernel.org [23.128.96.18]) by smtp.lore.kernel.org (Postfix) with ESMTP id 4EA2BECAAA1 for ; Fri, 9 Sep 2022 19:07:55 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S229728AbiIITHy (ORCPT ); Fri, 9 Sep 2022 15:07:54 -0400 Received: from lindbergh.monkeyblade.net ([23.128.96.19]:60342 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S229612AbiIITHw (ORCPT ); Fri, 9 Sep 2022 15:07:52 -0400 Received: from mail-ej1-x62f.google.com (mail-ej1-x62f.google.com [IPv6:2a00:1450:4864:20::62f]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id F0907135D73 for ; Fri, 9 Sep 2022 12:07:49 -0700 (PDT) Received: by mail-ej1-x62f.google.com with SMTP id lz22so6201370ejb.3 for ; Fri, 09 Sep 2022 12:07:49 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=szeredi.hu; s=google; h=cc:to:subject:message-id:date:from:in-reply-to:references :mime-version:from:to:cc:subject:date; bh=ziomUJ8TI35yh7PdR/qXu8p5sGWmK9UoF6cvJadWBWM=; b=MghIFancLOtXf747iECRdKBzUXqylczjEZGOA1R7N6m/gpq/sCI7pYp7I4Jci7lxKJ t24/2Rb1/qORtBmzaWEnbdA5zSbozm0QzJIVqHSXXgLxdyJV8W8fAAKZQrZk/9dJtpTj cPdEI0DFeVQ0UKvDPBsgeQBJ3XQdTKqicXlBA= X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=cc:to:subject:message-id:date:from:in-reply-to:references :mime-version:x-gm-message-state:from:to:cc:subject:date; bh=ziomUJ8TI35yh7PdR/qXu8p5sGWmK9UoF6cvJadWBWM=; b=tLQEoTS2jdtCPCZc6wDg4L5LmucWT5kirX/8eUfVTFX4/ua/YJy0FZJWq5TnQ7aWYr gzSJ/EFx2W+fOuPsTpVeiTzzr1ltsh88mxRrsqk7+vyGwxjz0CaviN34fo9uxISUNMZt hnuhouF3dh/bUdjttsMSKDaU9Ecfkjk+oE4MEhQrtVAs7f0WhupTbdI3Oc8jrmqBi6hw RbSAVUAwrmKObSrEnMR0Hp+AG17t7sczWR1Ayo0p54yS4AAW+v8GBM/VYUL15bERUQoq HudpOT5DeMja+uZcH7c4PbkjG87GlCZwCrB2qfc6e4Gj9hF88fFiTqtrwOk2fb+FVYAe 5Jxw== X-Gm-Message-State: ACgBeo0Kw1bCf3V9S95L6OtDMcNSdQ8RrNz70E/APEbC3boRrNU+QfKO oIKe/W9+AzuPTUWXT0hVjVXzJUStH51wUphR1jP+xA== X-Google-Smtp-Source: AA6agR72ua4PlvlShquFN7gQ7j8SbgHAY5/KRLp84F1m+JgTROGrfY286jrRS+xZ3vWVctXGYa54vDU/P9ZptiMXeGw= X-Received: by 2002:a17:907:6d11:b0:730:a382:d5ba with SMTP id sa17-20020a1709076d1100b00730a382d5bamr10662487ejc.371.1662750468489; Fri, 09 Sep 2022 12:07:48 -0700 (PDT) MIME-Version: 1.0 References: <20210125153057.3623715-1-balsini@android.com> <20210125153057.3623715-4-balsini@android.com> In-Reply-To: From: Miklos Szeredi Date: Fri, 9 Sep 2022 21:07:37 +0200 Message-ID: Subject: Re: [PATCH RESEND V12 3/8] fuse: Definitions and ioctl for passthrough To: Amir Goldstein Cc: Alessio Balsini , Peng Tao , Akilesh Kailash , Antonio SJ Musumeci , David Anderson , Giuseppe Scrivano , Jann Horn , Jens Axboe , Martijn Coenen , Palmer Dabbelt , Paul Lawrence , Stefano Duo , Zimuzo Ezeozue , wuyan , fuse-devel , kernel-team , "linux-fsdevel@vger.kernel.org" Content-Type: text/plain; charset="UTF-8" Precedence: bulk List-ID: X-Mailing-List: linux-fsdevel@vger.kernel.org On Thu, 8 Sept 2022 at 17:36, Amir Goldstein wrote: > > Hi Alessio and Miklos, > > Some time has passed.. and I was thinking of picking up these patches. > > On Mon, Mar 1, 2021 at 7:05 PM Alessio Balsini wrote: > > > > On Fri, Feb 19, 2021 at 09:40:21AM +0100, Miklos Szeredi wrote: > > > On Fri, Feb 19, 2021 at 8:05 AM Peng Tao wrote: > > > > > > > > On Wed, Feb 17, 2021 at 9:41 PM Miklos Szeredi wrote: > > > > > > > > What I think would be useful is to have an explicit > > > > > FUSE_DEV_IOC_PASSTHROUGH_CLOSE ioctl, that would need to be called > > > > > once the fuse server no longer needs this ID. If this turns out to > > > > > be a performance problem, we could still add the auto-close behavior > > > > > with an explicit FOPEN_PASSTHROUGH_AUTOCLOSE flag later. > > > > Hi Miklos, > > > > > > > > W/o auto closing, what happens if user space daemon forgets to call > > > > FUSE_DEV_IOC_PASSTHROUGH_CLOSE? Do we keep the ID alive somewhere? > > > > > > Kernel would keep the ID open until explicit close or fuse connection > > > is released. > > > > > > There should be some limit on the max open files referenced through > > > ID's, though. E.g. inherit RLIMIT_NOFILE from mounting task. > > > > > > Thanks, > > > Miklos > > > > I like the idea of FUSE_DEV_IOC_PASSTHROUGH_CLOSE to revoke the > > passthrough access, that is something I was already working on. What I > > had in mind was simply to break that 1:1 connection between fuse_file > > and lower filp setting a specific fuse_file::passthrough::filp to NULL, > > but this is slightly different from what you mentioned. > > > > I don't like the idea of switching between passthrough and server mid-life > of an open file. > > There are consequences related to syncing the attribute cache of the kernel > and the server that I don't even want to think about. > > > AFAIU you are suggesting to allocate one ID for each lower fs file > > opened with passthrough within a connection, and maybe using idr_find at > > every read/write/mmap operation to check if passthrough is enabled on > > that file. Something similar to fuse2_map_get(). > > This way the fuse server can pass the same ID to one or more > > fuse_file(s). > > FUSE_DEV_IOC_PASSTHROUGH_CLOSE would idr_remove the ID, so idr_find > > would fail, preventing the use of passthrough on that ID. CMIIW. > > > > I don't think that FUSE_DEV_IOC_PASSTHROUGH_CLOSE should remove the ID. > We can use a refcount for the mapping and FUSE_DEV_IOC_PASSTHROUGH_CLOSE > just drops the initial server's refcount. > > Implementing revoke for an existing mapping is something completely different. > It can be done, not even so hard, but I don't think it should be part of this > series and in any case revoke will not remove the ID. > > > After FUSE_DEV_IOC_PASSTHROUGH_CLOSE(ID) it may happen that if some > > fuse_file(s) storing that ID are still open and the same ID is reclaimed > > in a new idr_alloc, this would lead to mismatching lower fs filp being > > used by our fuse_file(s). So also the ID stored in the fuse_file(s) > > must be invalidated to prevent future uses of deallocated IDs. > > Obtaining a refcount on FOPEN_PASSTHROUGH will solve that. > > > > > Would it make sense to have a list of fuse_files using the same ID, that > > must be traversed at FUSE_DEV_IOC_PASSTHROUGH_CLOSE time? > > Negative values (maybe -ENOENT) might be used to mark IDs as invalid, > > and tested before idr_find at read/write/mmap to avoid the idr_find > > complexity in case passthrough is disabled for that file. > > > > What do you think? > > > > As I wrote above, this sounds unnecessarily complicated. > > Miklos, > > Do you agree with my interpretation of > FUSE_DEV_IOC_PASSTHROUGH_CLOSE? We need to deal with the case of too many open files. The server could manage this, but then we do need to handle the case when a cached mapping disappears, i.e: client opens file [time passes] cached passthrough fd gets evicted to make room for other passthrough I/O [time passes] new I/O request comes in need to reestablish passthrough fd before finishing I/O The way I think of this is that a passthrough mapping is assigned at open time, which is cached (which may have the lifetime longer than the open file, but shorter as well). When FUSE_DEV_IOC_PASSTHROUGH_CLOSE and there are cached mapping referring to this particular handle, then those mappings need to be purged. On a new I/O request, the mapping will need to be reestablished by sending a FUSE_MAP request, which triggers FUSE_DEV_IOC_PASSTHROUGH_OPEN. One other question that's nagging me is how to "unhide" these pseudo-fds. Could we create a kernel thread for each fuse sb which has normal file-table for these? This would would allow inspecting state through /proc/$KTHREDID/fd, etc.. Thanks, Miklos