Re: [PATCH v15 08/11] fw_cfg: handle fw_cfg_read_blob() error

From: Gabriel Somlo
Date: Wed Feb 28 2018 - 19:49:53 EST


On Thu, Mar 01, 2018 at 01:58:22AM +0200, Michael S. Tsirkin wrote:
> On Thu, Mar 01, 2018 at 01:58:01AM +0200, Michael S. Tsirkin wrote:
> > On Wed, Feb 28, 2018 at 06:25:58PM -0500, Gabriel Somlo wrote:
> > > On Wed, Feb 28, 2018 at 05:32:52PM +0200, Michael S. Tsirkin wrote:
> > > > On Wed, Feb 28, 2018 at 12:49:35PM +0100, Marc-André Lureau wrote:
> > > > > Hi
> > > > >
> > > > > On Tue, Feb 27, 2018 at 1:20 AM, Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > > > > > On Thu, Feb 15, 2018 at 10:33:09PM +0100, Marc-André Lureau wrote:
> > > > > >> fw_cfg_read_blob() may fail, but does not return error. This may lead
> > > > > >> to undefined behaviours, such as a memcmp(sig, "QEMU") on uninitilized
> > > > > >> memory.
> > > > > >
> > > > > > I don't think that's true - there's a memset there that
> > > > > > will initialize the memory. probe is likely the only
> > > > > > case where it returns a slightly incorrect data.
> > > > >
> > > > > Right, I'll update the commit message.
> > > > >
> > > > > >> Return an error if ACPI locking failed. Also, the following
> > > > > >> DMA read/write extension will add more error paths that should be
> > > > > >> handled appropriately.
> > > > > >>
> > > > > >> Signed-off-by: Marc-André Lureau <marcandre.lureau@xxxxxxxxxx>
> > > > > >> ---
> > > > > >> drivers/firmware/qemu_fw_cfg.c | 32 ++++++++++++++++++++++----------
> > > > > >> 1 file changed, 22 insertions(+), 10 deletions(-)
> > > > > >>
> > > > > >> diff --git a/drivers/firmware/qemu_fw_cfg.c b/drivers/firmware/qemu_fw_cfg.c
> > > > > >> index f6f90bef604c..5e6e5ac71dab 100644
> > > > > >> --- a/drivers/firmware/qemu_fw_cfg.c
> > > > > >> +++ b/drivers/firmware/qemu_fw_cfg.c
> > > > > >> @@ -59,8 +59,8 @@ static void fw_cfg_sel_endianness(u16 key)
> > > > > >> }
> > > > > >>
> > > > > >> /* read chunk of given fw_cfg blob (caller responsible for sanity-check) */
> > > > > >> -static void fw_cfg_read_blob(u16 key,
> > > > > >> - void *buf, loff_t pos, size_t count)
> > > > > >> +static ssize_t fw_cfg_read_blob(u16 key,
> > > > > >> + void *buf, loff_t pos, size_t count)
> > > > > >> {
> > > > > >> u32 glk = -1U;
> > > > > >> acpi_status status;
> > > > > >> @@ -73,7 +73,7 @@ static void fw_cfg_read_blob(u16 key,
> > > > > >> /* Should never get here */
> > > > > >> WARN(1, "fw_cfg_read_blob: Failed to lock ACPI!\n");
> > > > > >> memset(buf, 0, count);
> > > > > >> - return;
> > > > > >> + return -EINVAL;
> > > > > >> }
> > > > > >>
> > > > > >> mutex_lock(&fw_cfg_dev_lock);
> > > > > >
> > > > > > Wouldn't something like -EBUSY be more appropriate?
> > > > >
> > > > > In theory, it would be a general failure right? I don't think we want
> > > > > the caller to retry. I think in EINVAL fits better, but I don't think
> > > > > it matters much this or EBUSY.
> > > > >
> > > > > >> @@ -84,6 +84,7 @@ static void fw_cfg_read_blob(u16 key,
> > > > > >> mutex_unlock(&fw_cfg_dev_lock);
> > > > > >>
> > > > > >> acpi_release_global_lock(glk);
> > > > > >> + return count;
> > > > > >> }
> > > > > >>
> > > > > >> /* clean up fw_cfg device i/o */
> > > > > >> @@ -165,8 +166,9 @@ static int fw_cfg_do_platform_probe(struct platform_device *pdev)
> > > > > >> }
> > > > > >>
> > > > > >> /* verify fw_cfg device signature */
> > > > > >> - fw_cfg_read_blob(FW_CFG_SIGNATURE, sig, 0, FW_CFG_SIG_SIZE);
> > > > > >> - if (memcmp(sig, "QEMU", FW_CFG_SIG_SIZE) != 0) {
> > > > > >> + if (fw_cfg_read_blob(FW_CFG_SIGNATURE, sig,
> > > > > >> + 0, FW_CFG_SIG_SIZE) < 0 ||
> > > > > >> + memcmp(sig, "QEMU", FW_CFG_SIG_SIZE) != 0) {
> > > > > >> fw_cfg_io_cleanup();
> > > > > >> return -ENODEV;
> > > > > >> }
> > > > > >> @@ -326,8 +328,7 @@ static ssize_t fw_cfg_sysfs_read_raw(struct file *filp, struct kobject *kobj,
> > > > > >> if (count > entry->size - pos)
> > > > > >> count = entry->size - pos;
> > > > > >>
> > > > > >> - fw_cfg_read_blob(entry->select, buf, pos, count);
> > > > > >> - return count;
> > > > > >> + return fw_cfg_read_blob(entry->select, buf, pos, count);
> > > > > >> }
> > > > > >>
> > > > > >> static struct bin_attribute fw_cfg_sysfs_attr_raw = {
> > > > > >> @@ -483,7 +484,11 @@ static int fw_cfg_register_dir_entries(void)
> > > > > >> struct fw_cfg_file *dir;
> > > > > >> size_t dir_size;
> > > > > >>
> > > > > >> - fw_cfg_read_blob(FW_CFG_FILE_DIR, &files_count, 0, sizeof(files_count));
> > > > > >> + ret = fw_cfg_read_blob(FW_CFG_FILE_DIR, &files_count,
> > > > > >> + 0, sizeof(files_count));
> > > > > >> + if (ret < 0)
> > > > > >> + return ret;
> > > > > >> +
> > > > > >> count = be32_to_cpu(files_count);
> > > > > >> dir_size = count * sizeof(struct fw_cfg_file);
> > > > > >>
> > > > > >> @@ -491,7 +496,10 @@ static int fw_cfg_register_dir_entries(void)
> > > > > >> if (!dir)
> > > > > >> return -ENOMEM;
> > > > > >>
> > > > > >> - fw_cfg_read_blob(FW_CFG_FILE_DIR, dir, sizeof(files_count), dir_size);
> > > > > >> + ret = fw_cfg_read_blob(FW_CFG_FILE_DIR, dir,
> > > > > >> + sizeof(files_count), dir_size);
> > > > > >> + if (ret < 0)
> > > > > >> + goto end;
> > > > > >>
> > > > > >> for (i = 0; i < count; i++) {
> > > > > >> ret = fw_cfg_register_file(&dir[i]);
> > > > > >> @@ -499,6 +507,7 @@ static int fw_cfg_register_dir_entries(void)
> > > > > >> break;
> > > > > >> }
> > > > > >>
> > > > > >> +end:
> > > > > >> kfree(dir);
> > > > > >> return ret;
> > > > > >> }
> > > > > >> @@ -539,7 +548,10 @@ static int fw_cfg_sysfs_probe(struct platform_device *pdev)
> > > > > >> goto err_probe;
> > > > > >>
> > > > > >> /* get revision number, add matching top-level attribute */
> > > > > >> - fw_cfg_read_blob(FW_CFG_ID, &rev, 0, sizeof(rev));
> > > > > >> + err = fw_cfg_read_blob(FW_CFG_ID, &rev, 0, sizeof(rev));
> > > > > >> + if (err < 0)
> > > > > >> + goto err_probe;
> > > > > >> +
> > > > > >> fw_cfg_rev = le32_to_cpu(rev);
> > > > > >> err = sysfs_create_file(fw_cfg_top_ko, &fw_cfg_rev_attr.attr);
> > > > > >> if (err)
> > > > > >
> > > > > > I think that this is the only case where it's not doing the right thing right now in
> > > > > > that it shows 0 as the revision to the users. Is it worth failing probe
> > > > > > here? We could just skip the attribute, could we not?
> > > > >
> > > > > I think it's best to fail the probe if we have a read failure at that time.
> > > >
> > > > I'd rather we just dropped this attribute completely.
> > > > Why is it there?
> > > > Does any userspace actually need it?
> > > > Gabriel?
> > >
> > > I'd recommend keeping it: `cat /sys/firmware/qemu_fw_cfg/rev` is how
> > > you can easily tell if, e.g., dma is supported :)
>
> Does user ever care?

I'd rather not presume they *don't*

For instance, yourself: you add DMA support, then one day in the
future on some system you're on, things feel sluggish. You're curious
-- does this thing claim to support DMA? The answer is one easy `cat`
away. If you then found out someone decided you're not supposed to care,
and forced you to dig through the sources for a guess instead, you'd
probably be less than happy about it... :)

That said, if there's a better reason to leave it out than handling
some never-to-occur error condition elegantly, I can be persuaded...

As I replied to Michael earlier, leaving it out if that impossible
error condition *does* occur is OK with me (since it "never happens"
in practice... :)

Thanks,
--G

> > > If you end up with a '0' there, it's because locking ACPI with
> > > WAIT_FOREVER failed with something other than AE_NOT_CONFIGURED, which
> > > should never happen, so we're off the reservation, having morally
> > > failed the equivalent of an assertion.
> > >
> > > Perhaps memset to something other than 0 (all-Fs, so we'd get a -1 for
> > > rev)?
> > >
> > > Thanks,
> > > --Gabriel
> >
> > Why not just skip adding the attribute on this error?
> >
> > > >
> > > > > --
> > > > > Marc-André Lureau