Message-ID: <CAFEAcA-No3Z95+UQJZWTxDesd-z_Y5XnyHs6NMpzDo3RVOHQ4w@mail.gmail.com>
Date:   Mon, 20 Apr 2020 12:38:44 +0100
From:   Peter Maydell <peter.maydell@...aro.org>
To:     Florian Weimer <fw@...eb.enyo.de>
Cc:     Linus Walleij <linus.walleij@...aro.org>,
        "Theodore Ts'o" <tytso@....edu>,
        Andreas Dilger <adilger.kernel@...ger.ca>,
        Ext4 Developers List <linux-ext4@...r.kernel.org>,
        linux-fsdevel <linux-fsdevel@...r.kernel.org>,
        Linux API <linux-api@...r.kernel.org>,
        QEMU Developers <qemu-devel@...gnu.org>,
        Andy Lutomirski <luto@...nel.org>
Subject: Re: [PATCH] fcntl: Add 32bit filesystem mode

On Mon, 20 Apr 2020 at 12:23, Florian Weimer <fw@...eb.enyo.de> wrote:
>
> * Peter Maydell:
>
> > We open fd 3 to read '.'; we issue the new fcntl, which
> > succeeds. Then there's some unrelated stuff operating on
> > stdout. Then we do a getdents64(), but the d_off values
> > we get back are still 64 bits. The guest binary doesn't
> > like those, so it fails. My expectation was that we would
> > get back d_off values here that were in the 32 bit range.
>
> What's your file system?
>
> I think not all of them have 32-bit hashes (some of them probably
> can't, particularly in the network-based file system case).

Whoops, good point. I was testing this via lkvm, so it's
actually using a 9p filesystem... I'll see if I can figure
out how to test with an ext3 fs, which I think is the one
we most care about. It would be nice if the flag was
supported by other fses too, of course.

Appended is the QEMU patch I tested with.

thanks
-- PMM

>From 73471e01733dd1d998ff3cd41edebb4c78793193 Mon Sep 17 00:00:00 2001
From: Peter Maydell <peter.maydell@...aro.org>
Date: Mon, 20 Apr 2020 11:54:22 +0100
Subject: [RFC] linux-user: Use new F_SET_FILE_32BIT_FS fcntl for 32-bit guests

If the guest is 32 bit then there is a potential problem if the
host gives us back a 64-bit sized value that we can't fit into
the ABI the guest requires. This is a theoretical issue for many
syscalls, but a real issue for directory reads where the host
is using ext3 or ext4. There the 'offset' values retured via
the getdents syscall are hashes, and on a 64-bit system they
will always fill the full 64 bits.

Use the F_SET_FILE_32BIT_FS fcntl to tell the kernel to stick
to 32-bit sized hashes for fds used by the guest.

Signed-off-by: Peter Maydell <peter.maydell@...aro.org>
---
RFC patch because it depends on the kernel patch to provide
F_SET_FILE_32BIT_FS, which is still under discussion. All this
patch does is call the fcntl for every fd the guest opens.

 linux-user/syscall.c | 27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)

diff --git a/linux-user/syscall.c b/linux-user/syscall.c
index 674f70e70a5..8966d4881bd 100644
--- a/linux-user/syscall.c
+++ b/linux-user/syscall.c
@@ -884,6 +884,28 @@ static inline int host_to_target_sock_type(int host_type)
     return target_type;
 }

+/*
+ * If the guest is using a 32 bit ABI then we should try to ask the kernel
+ * to provide 32-bit offsets in getdents syscalls, as otherwise some
+ * filesystems will return 64-bit hash values which we can't fit into
+ * the field sizes the guest ABI mandates.
+ */
+#ifndef F_SET_FILE_32BIT_FS
+#define F_SET_FILE_32BIT_FS (1024 + 15)
+#endif
+
+static inline void request_32bit_fs(int fd)
+{
+#if HOST_LONG_BITS > TARGET_ABI_BITS
+    /*
+     * Ignore errors, which are likely due to the host kernel being too
+     * old to support this fcntl. We'll try anyway, which might or might
+     * not work, depending on the guest code and on the host filesystem.
+     */
+    fcntl(fd, F_SET_FILE_32BIT_FS);
+#endif
+}
+
 static abi_ulong target_brk;
 static abi_ulong target_original_brk;
 static abi_ulong brk_page;
@@ -7704,6 +7726,7 @@ static abi_long do_syscall1(void *cpu_env, int
num, abi_long arg1,
                                   target_to_host_bitmask(arg2,
fcntl_flags_tbl),
                                   arg3));
         fd_trans_unregister(ret);
+        request_32bit_fs(ret);
         unlock_user(p, arg1, 0);
         return ret;
 #endif
@@ -7714,6 +7737,7 @@ static abi_long do_syscall1(void *cpu_env, int
num, abi_long arg1,
                                   target_to_host_bitmask(arg3,
fcntl_flags_tbl),
                                   arg4));
         fd_trans_unregister(ret);
+        request_32bit_fs(ret);
         unlock_user(p, arg2, 0);
         return ret;
 #if defined(TARGET_NR_name_to_handle_at) && defined(CONFIG_OPEN_BY_HANDLE)
@@ -7725,6 +7749,7 @@ static abi_long do_syscall1(void *cpu_env, int
num, abi_long arg1,
     case TARGET_NR_open_by_handle_at:
         ret = do_open_by_handle_at(arg1, arg2, arg3);
         fd_trans_unregister(ret);
+        request_32bit_fs(ret);
         return ret;
 #endif
     case TARGET_NR_close:
@@ -7769,6 +7794,7 @@ static abi_long do_syscall1(void *cpu_env, int
num, abi_long arg1,
             return -TARGET_EFAULT;
         ret = get_errno(creat(p, arg2));
         fd_trans_unregister(ret);
+        request_32bit_fs(ret);
         unlock_user(p, arg1, 0);
         return ret;
 #endif
@@ -12393,6 +12419,7 @@ static abi_long do_syscall1(void *cpu_env, int
num, abi_long arg1,
         }
         ret = get_errno(memfd_create(p, arg2));
         fd_trans_unregister(ret);
+        request_32bit_fs(ret);
         unlock_user(p, arg1, 0);
         return ret;
 #endif
-- 
2.20.1

Powered by blists - more mailing lists