On 9/12/2023 7:09 PM, Greg Kroah-Hartman wrote:
On Tue, Sep 12, 2023 at 06:44:46PM +0800, Linyu Yuan wrote:
Some UDC trace event will save usb gadget information, but it use one int
size buffer to save one bit information of usb gadget, so 19 int buffers
needed to save 19 bit fields which is not good.
Add one anonymous union which have one u32 member 'dw1' to struct
'usb_gadget', it inlclude all 19 bits and can be used by trace event
during fast assign stage to save more entries with same trace ring buffer
size.
Also move all original 19 bit fields into one anonymous struct which
inside struct 'usb_gadget'.
In order to allow trace event output stage access the bit from saved u32
'dw1', add following macro,
define USB_GADGET_BITFIELD(n, name) \
({\
union {\
struct {\
u32 sg_supported:1;\
u32 is_otg:1;\
u32 is_a_peripheral:1;\
u32 b_hnp_enable:1;\
u32 a_hnp_support:1;\
u32 a_alt_hnp_support:1;\
u32 hnp_polling_support:1;\
u32 host_request_flag:1;\
u32 quirk_ep_out_aligned_size:1;\
u32 quirk_altset_not_supp:1;\
u32 quirk_stall_not_supp:1;\
u32 quirk_zlp_not_supp:1;\
u32 quirk_avoids_skb_reserve:1;\
u32 is_selfpowered:1;\
u32 deactivated:1;\
u32 connected:1;\
u32 lpm_capable:1;\
u32 wakeup_capable:1;\
u32 wakeup_armed:1;\
} __packed;\
u32 dw1;\
} __aligned(4) __g_u_##name;\
u32 __g_##name; \
BUILD_BUG_ON(sizeof(__g_u_##name) != 4);\
__g_u_##name.dw1 = (n); __g_##name = __g_u_##name.name;\
__g_##name; })
define USB_GADGET_SG_SUPPORTED(n) USB_GADGET_BITFIELD((n), sg_supported)
...
change to use this kind of macro for all related trace files later.
I'm sorry, but that's horrible, and is NOT how you deal with bitfields
in an endian-neutral way at all.
There are much simpler, and easier, ways to do this properly.
do you mean define two sets of ordering bit field according BIG and
LITTLE ENDIAN like below ?
#if defined(__LITTLE_ENDIAN_BITFIELD)
struct {
u32 sg_supported:1;
u32 is_otg:1;
..
u32 :13;
} __packed;
#else
struct {
u32 :13;
..
u32 is_otg:1;
u32 sg_supported:1;
} __packed;
#endif
but Alan Stern have one comment, do it mean the bit position number is
not expect and we can't use it ?
@Alan Stern , BIT(0), BIT(1) is not the member we expect ?
"
As far as I
know, the C compiler does not specify that bit fields in packed
structures will be assigned starting from the low-order bit position
"
But I'm still missing the huge _WHY_ any of this is needed. You are not
showing any real advantage at all that I have noticed.
You need to step back and see if any of this is even anything that needs
to change, and if you feel it does need to change, you need to be able
to properly justify _why_ it needs to change.
indeed this is a small optimization.
I think i explain the benefit in previous version, when user not
increase system trace event buffer space,
or in lower system trace event buffer space, it allow more trace event
entries to be saved.
in normal condition, the usb request is most frequent things after
enumeration with useful operation,
so take below trace event class for explanation,
DECLARE_EVENT_CLASS(udc_log_req,
TP_PROTO(struct usb_ep *ep, struct usb_request *req, int ret),
TP_ARGS(ep, req, ret),
TP_STRUCT__entry(
__string(name, ep->name)
__field(unsigned, length)
__field(unsigned, actual)
__field(unsigned, num_sgs)
__field(unsigned, num_mapped_sgs)
__field(unsigned, stream_id)
__field(unsigned, no_interrupt)
__field(unsigned, zero)
__field(unsigned, short_not_ok)
__field(int, status)
__field(int, ret)
__field(struct usb_request *, req)
),
TP_fast_assign(
__assign_str(name, ep->name);
__entry->length = req->length;
__entry->actual = req->actual;
__entry->num_sgs = req->num_sgs;
__entry->num_mapped_sgs = req->num_mapped_sgs;
__entry->stream_id = req->stream_id;
__entry->no_interrupt = req->no_interrupt;
__entry->zero = req->zero;
__entry->short_not_ok = req->short_not_ok;
__entry->status = req->status;
__entry->ret = ret;
__entry->req = req;
),
TP_printk("%s: req %p length %d/%d sgs %d/%d stream %d %s%s%s
status %d --> %d",
__get_str(name),__entry->req, __entry->actual, __entry->length,
__entry->num_mapped_sgs, __entry->num_sgs, __entry->stream_id,
__entry->zero ? "Z" : "z",
__entry->short_not_ok ? "S" : "s",
__entry->no_interrupt ? "i" : "I",
__entry->status, __entry->ret
)
);
consider 32 bit ARCH,
without change, one trace entry size is:
4 (ring buffer event header ) + 8 (trace event header ) + 48 (trace
class header) + 9 (ep string name) = 69 bytes.
with change,
4 (ring buffer event header ) + 8 (trace event header ) + 36 (trace
class header) = 48 bytes.
consider there is 1MB trace buffer space,
without change, it can save 15196 entries,
with change, it can save 21845 entries.
good luck!
greg k-h
