diff mbox series

[3/3] hw/mem/cxl_type3: Fix problem with g_steal_pointer()

Message ID 20240304104406.59855-4-thuth@redhat.com (mailing list archive)
State New, archived
Headers show
Series cxl: Fix issues with g_steal_pointer() | expand

Commit Message

Thomas Huth March 4, 2024, 10:44 a.m. UTC 
When setting GLIB_VERSION_MAX_ALLOWED to GLIB_VERSION_2_58 or higher,
glib adds type safety checks to the g_steal_pointer() macro. This
triggers errors in the ct3_build_cdat_entries_for_mr() function which
uses the g_steal_pointer() for type-casting from one pointer type to
the other (which also looks quite weird since the local pointers have
all been declared with g_autofree though they are never freed here).
Fix it by using a proper typecast instead. For making this possible, we
have to remove the QEMU_PACKED attribute from some structs since GCC
otherwise complains that the source and destination pointer might
have different alignment restrictions. Removing the QEMU_PACKED should
be fine here since the structs are already naturally aligned. Anyway,
add some QEMU_BUILD_BUG_ON() statements to make sure that we've got
the right sizes (without padding in the structs).

Signed-off-by: Thomas Huth <thuth@redhat.com>
---
 include/hw/cxl/cxl_cdat.h |  9 ++++++---
 hw/mem/cxl_type3.c        | 24 ++++++++++++------------
 2 files changed, 18 insertions(+), 15 deletions(-)

Comments

Jonathan Cameron March 4, 2024, 3:10 p.m. UTC | #1 
On Mon,  4 Mar 2024 11:44:06 +0100
Thomas Huth <thuth@redhat.com> wrote:

> When setting GLIB_VERSION_MAX_ALLOWED to GLIB_VERSION_2_58 or higher,
> glib adds type safety checks to the g_steal_pointer() macro. This
> triggers errors in the ct3_build_cdat_entries_for_mr() function which
> uses the g_steal_pointer() for type-casting from one pointer type to
> the other (which also looks quite weird since the local pointers have
> all been declared with g_autofree though they are never freed here).
> Fix it by using a proper typecast instead. For making this possible, we
> have to remove the QEMU_PACKED attribute from some structs since GCC
> otherwise complains that the source and destination pointer might
> have different alignment restrictions. Removing the QEMU_PACKED should
> be fine here since the structs are already naturally aligned. Anyway,
> add some QEMU_BUILD_BUG_ON() statements to make sure that we've got
> the right sizes (without padding in the structs).

I missed these as well when getting rid of the false handling
of failure of g_new0 calls.

Another alternative would be to point to the head structures rather
than the containing structure - would avoid need to cast.
That might be neater?  Should I think also remove the alignment
question?


> 
> Signed-off-by: Thomas Huth <thuth@redhat.com>
> ---
>  include/hw/cxl/cxl_cdat.h |  9 ++++++---
>  hw/mem/cxl_type3.c        | 24 ++++++++++++------------
>  2 files changed, 18 insertions(+), 15 deletions(-)
> 
> diff --git a/include/hw/cxl/cxl_cdat.h b/include/hw/cxl/cxl_cdat.h
> index b44cefaad6..17a09066dc 100644
> --- a/include/hw/cxl/cxl_cdat.h
> +++ b/include/hw/cxl/cxl_cdat.h
> @@ -82,7 +82,8 @@ typedef struct CDATDsmas {
>      uint16_t reserved;
>      uint64_t DPA_base;
>      uint64_t DPA_length;
> -} QEMU_PACKED CDATDsmas;
> +} CDATDsmas;
> +QEMU_BUILD_BUG_ON(sizeof(CDATDsmas) != 24);
>  
>  /* Device Scoped Latency and Bandwidth Information Structure - CDAT Table 5 */
>  typedef struct CDATDslbis {
> @@ -95,7 +96,8 @@ typedef struct CDATDslbis {
>      uint64_t entry_base_unit;
>      uint16_t entry[3];
>      uint16_t reserved2;
> -} QEMU_PACKED CDATDslbis;
> +} CDATDslbis;
> +QEMU_BUILD_BUG_ON(sizeof(CDATDslbis) != 24);
>  
>  /* Device Scoped Memory Side Cache Information Structure - CDAT Table 6 */
>  typedef struct CDATDsmscis {
> @@ -122,7 +124,8 @@ typedef struct CDATDsemts {
>      uint16_t reserved;
>      uint64_t DPA_offset;
>      uint64_t DPA_length;
> -} QEMU_PACKED CDATDsemts;
> +} CDATDsemts;
> +QEMU_BUILD_BUG_ON(sizeof(CDATDsemts) != 24);
>  
>  /* Switch Scoped Latency and Bandwidth Information Structure - CDAT Table 9 */
>  typedef struct CDATSslbisHeader {
> diff --git a/hw/mem/cxl_type3.c b/hw/mem/cxl_type3.c
> index e8801805b9..b679dfae1c 100644
> --- a/hw/mem/cxl_type3.c
> +++ b/hw/mem/cxl_type3.c
> @@ -46,12 +46,12 @@ static void ct3_build_cdat_entries_for_mr(CDATSubHeader **cdat_table,
>                                            int dsmad_handle, MemoryRegion *mr,
>                                            bool is_pmem, uint64_t dpa_base)
>  {
> -    g_autofree CDATDsmas *dsmas = NULL;
> -    g_autofree CDATDslbis *dslbis0 = NULL;
> -    g_autofree CDATDslbis *dslbis1 = NULL;
> -    g_autofree CDATDslbis *dslbis2 = NULL;
> -    g_autofree CDATDslbis *dslbis3 = NULL;
> -    g_autofree CDATDsemts *dsemts = NULL;
> +    CDATDsmas *dsmas;
> +    CDATDslbis *dslbis0;
> +    CDATDslbis *dslbis1;
> +    CDATDslbis *dslbis2;
> +    CDATDslbis *dslbis3;
> +    CDATDsemts *dsemts;
>  
>      dsmas = g_malloc(sizeof(*dsmas));
>      *dsmas = (CDATDsmas) {
> @@ -135,12 +135,12 @@ static void ct3_build_cdat_entries_for_mr(CDATSubHeader **cdat_table,
>      };
>  
>      /* Header always at start of structure */
> -    cdat_table[CT3_CDAT_DSMAS] = g_steal_pointer(&dsmas);
> -    cdat_table[CT3_CDAT_DSLBIS0] = g_steal_pointer(&dslbis0);
> -    cdat_table[CT3_CDAT_DSLBIS1] = g_steal_pointer(&dslbis1);
> -    cdat_table[CT3_CDAT_DSLBIS2] = g_steal_pointer(&dslbis2);
> -    cdat_table[CT3_CDAT_DSLBIS3] = g_steal_pointer(&dslbis3);
> -    cdat_table[CT3_CDAT_DSEMTS] = g_steal_pointer(&dsemts);
> +    cdat_table[CT3_CDAT_DSMAS] = (CDATSubHeader *)dsmas;
Could do
    cdat_table[CT3_CDAT_DSMAS] = &dsmas->header;
etc
> +    cdat_table[CT3_CDAT_DSLBIS0] = (CDATSubHeader *)dslbis0;
> +    cdat_table[CT3_CDAT_DSLBIS1] = (CDATSubHeader *)dslbis1;
> +    cdat_table[CT3_CDAT_DSLBIS2] = (CDATSubHeader *)dslbis2;
> +    cdat_table[CT3_CDAT_DSLBIS3] = (CDATSubHeader *)dslbis3;
> +    cdat_table[CT3_CDAT_DSEMTS] = (CDATSubHeader *)dsemts;
>  }
>  
>  static int ct3_build_cdat_table(CDATSubHeader ***cdat_table, void *priv)
Thomas Huth March 5, 2024, 7:27 a.m. UTC | #2 
On 04/03/2024 16.10, Jonathan Cameron wrote:
> On Mon,  4 Mar 2024 11:44:06 +0100
> Thomas Huth <thuth@redhat.com> wrote:
> 
>> When setting GLIB_VERSION_MAX_ALLOWED to GLIB_VERSION_2_58 or higher,
>> glib adds type safety checks to the g_steal_pointer() macro. This
>> triggers errors in the ct3_build_cdat_entries_for_mr() function which
>> uses the g_steal_pointer() for type-casting from one pointer type to
>> the other (which also looks quite weird since the local pointers have
>> all been declared with g_autofree though they are never freed here).
>> Fix it by using a proper typecast instead. For making this possible, we
>> have to remove the QEMU_PACKED attribute from some structs since GCC
>> otherwise complains that the source and destination pointer might
>> have different alignment restrictions. Removing the QEMU_PACKED should
>> be fine here since the structs are already naturally aligned. Anyway,
>> add some QEMU_BUILD_BUG_ON() statements to make sure that we've got
>> the right sizes (without padding in the structs).
> 
> I missed these as well when getting rid of the false handling
> of failure of g_new0 calls.
> 
> Another alternative would be to point to the head structures rather
> than the containing structure - would avoid need to cast.
> That might be neater?  Should I think also remove the alignment
> question?

I gave it a try, but it does not help against the alignment issue, I still get:

../../devel/qemu/hw/mem/cxl_type3.c: In function 
‘ct3_build_cdat_entries_for_mr’:
../../devel/qemu/hw/mem/cxl_type3.c:138:34: error: taking address of packed 
member of ‘struct CDATDsmas’ may result in an unaligned pointer value 
[-Werror=address-of-packed-member]
   138 |     cdat_table[CT3_CDAT_DSMAS] = &dsmas->header;
       |                                  ^~~~~~~~~~~~~~

 From my experience, it's better anyway to avoid __attribute__((packed)) on 
structures unless it is really really required. At least we should avoid it 
as good as possible as long as we still support running QEMU on Sparc hosts 
(that don't support misaligned memory accesses), since otherwise you can end 
up with non-working code there, see e.g.:

  https://www.mail-archive.com/qemu-devel@nongnu.org/msg439899.html

or:

  https://gitlab.com/qemu-project/qemu/-/commit/cb89b349074310ff9eb7ebe18a

Thus I'd rather prefer to keep this patch as it is right now.

  Thomas
Jonathan Cameron March 5, 2024, 3:52 p.m. UTC | #3 
On Tue, 5 Mar 2024 08:27:52 +0100
Thomas Huth <thuth@redhat.com> wrote:

> On 04/03/2024 16.10, Jonathan Cameron wrote:
> > On Mon,  4 Mar 2024 11:44:06 +0100
> > Thomas Huth <thuth@redhat.com> wrote:
> >   
> >> When setting GLIB_VERSION_MAX_ALLOWED to GLIB_VERSION_2_58 or higher,
> >> glib adds type safety checks to the g_steal_pointer() macro. This
> >> triggers errors in the ct3_build_cdat_entries_for_mr() function which
> >> uses the g_steal_pointer() for type-casting from one pointer type to
> >> the other (which also looks quite weird since the local pointers have
> >> all been declared with g_autofree though they are never freed here).
> >> Fix it by using a proper typecast instead. For making this possible, we
> >> have to remove the QEMU_PACKED attribute from some structs since GCC
> >> otherwise complains that the source and destination pointer might
> >> have different alignment restrictions. Removing the QEMU_PACKED should
> >> be fine here since the structs are already naturally aligned. Anyway,
> >> add some QEMU_BUILD_BUG_ON() statements to make sure that we've got
> >> the right sizes (without padding in the structs).  
> > 
> > I missed these as well when getting rid of the false handling
> > of failure of g_new0 calls.
> > 
> > Another alternative would be to point to the head structures rather
> > than the containing structure - would avoid need to cast.
> > That might be neater?  Should I think also remove the alignment
> > question?  
> 
> I gave it a try, but it does not help against the alignment issue, I still get:
> 
> ../../devel/qemu/hw/mem/cxl_type3.c: In function 
> ‘ct3_build_cdat_entries_for_mr’:
> ../../devel/qemu/hw/mem/cxl_type3.c:138:34: error: taking address of packed 
> member of ‘struct CDATDsmas’ may result in an unaligned pointer value 
> [-Werror=address-of-packed-member]
>    138 |     cdat_table[CT3_CDAT_DSMAS] = &dsmas->header;
>        |                                  ^~~~~~~~~~~~~~
> 
>  From my experience, it's better anyway to avoid __attribute__((packed)) on 
> structures unless it is really really required. At least we should avoid it 
> as good as possible as long as we still support running QEMU on Sparc hosts 
> (that don't support misaligned memory accesses), since otherwise you can end 
> up with non-working code there, see e.g.:
> 
>   https://www.mail-archive.com/qemu-devel@nongnu.org/msg439899.html
> 
> or:
> 
>   https://gitlab.com/qemu-project/qemu/-/commit/cb89b349074310ff9eb7ebe18a
> 
> Thus I'd rather prefer to keep this patch as it is right now.
> 
>   Thomas
Fair enough.

Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>

>
diff mbox series

Patch

diff --git a/include/hw/cxl/cxl_cdat.h b/include/hw/cxl/cxl_cdat.h
index b44cefaad6..17a09066dc 100644
--- a/include/hw/cxl/cxl_cdat.h
+++ b/include/hw/cxl/cxl_cdat.h
@@ -82,7 +82,8 @@  typedef struct CDATDsmas {
     uint16_t reserved;
     uint64_t DPA_base;
     uint64_t DPA_length;
-} QEMU_PACKED CDATDsmas;
+} CDATDsmas;
+QEMU_BUILD_BUG_ON(sizeof(CDATDsmas) != 24);
 
 /* Device Scoped Latency and Bandwidth Information Structure - CDAT Table 5 */
 typedef struct CDATDslbis {
@@ -95,7 +96,8 @@  typedef struct CDATDslbis {
     uint64_t entry_base_unit;
     uint16_t entry[3];
     uint16_t reserved2;
-} QEMU_PACKED CDATDslbis;
+} CDATDslbis;
+QEMU_BUILD_BUG_ON(sizeof(CDATDslbis) != 24);
 
 /* Device Scoped Memory Side Cache Information Structure - CDAT Table 6 */
 typedef struct CDATDsmscis {
@@ -122,7 +124,8 @@  typedef struct CDATDsemts {
     uint16_t reserved;
     uint64_t DPA_offset;
     uint64_t DPA_length;
-} QEMU_PACKED CDATDsemts;
+} CDATDsemts;
+QEMU_BUILD_BUG_ON(sizeof(CDATDsemts) != 24);
 
 /* Switch Scoped Latency and Bandwidth Information Structure - CDAT Table 9 */
 typedef struct CDATSslbisHeader {
diff --git a/hw/mem/cxl_type3.c b/hw/mem/cxl_type3.c
index e8801805b9..b679dfae1c 100644
--- a/hw/mem/cxl_type3.c
+++ b/hw/mem/cxl_type3.c
@@ -46,12 +46,12 @@  static void ct3_build_cdat_entries_for_mr(CDATSubHeader **cdat_table,
                                           int dsmad_handle, MemoryRegion *mr,
                                           bool is_pmem, uint64_t dpa_base)
 {
-    g_autofree CDATDsmas *dsmas = NULL;
-    g_autofree CDATDslbis *dslbis0 = NULL;
-    g_autofree CDATDslbis *dslbis1 = NULL;
-    g_autofree CDATDslbis *dslbis2 = NULL;
-    g_autofree CDATDslbis *dslbis3 = NULL;
-    g_autofree CDATDsemts *dsemts = NULL;
+    CDATDsmas *dsmas;
+    CDATDslbis *dslbis0;
+    CDATDslbis *dslbis1;
+    CDATDslbis *dslbis2;
+    CDATDslbis *dslbis3;
+    CDATDsemts *dsemts;
 
     dsmas = g_malloc(sizeof(*dsmas));
     *dsmas = (CDATDsmas) {
@@ -135,12 +135,12 @@  static void ct3_build_cdat_entries_for_mr(CDATSubHeader **cdat_table,
     };
 
     /* Header always at start of structure */
-    cdat_table[CT3_CDAT_DSMAS] = g_steal_pointer(&dsmas);
-    cdat_table[CT3_CDAT_DSLBIS0] = g_steal_pointer(&dslbis0);
-    cdat_table[CT3_CDAT_DSLBIS1] = g_steal_pointer(&dslbis1);
-    cdat_table[CT3_CDAT_DSLBIS2] = g_steal_pointer(&dslbis2);
-    cdat_table[CT3_CDAT_DSLBIS3] = g_steal_pointer(&dslbis3);
-    cdat_table[CT3_CDAT_DSEMTS] = g_steal_pointer(&dsemts);
+    cdat_table[CT3_CDAT_DSMAS] = (CDATSubHeader *)dsmas;
+    cdat_table[CT3_CDAT_DSLBIS0] = (CDATSubHeader *)dslbis0;
+    cdat_table[CT3_CDAT_DSLBIS1] = (CDATSubHeader *)dslbis1;
+    cdat_table[CT3_CDAT_DSLBIS2] = (CDATSubHeader *)dslbis2;
+    cdat_table[CT3_CDAT_DSLBIS3] = (CDATSubHeader *)dslbis3;
+    cdat_table[CT3_CDAT_DSEMTS] = (CDATSubHeader *)dsemts;
 }
 
 static int ct3_build_cdat_table(CDATSubHeader ***cdat_table, void *priv)