Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware loading for new boot architecture
On Thu, 26 Mar 2015 18:57:02 -0500 Rivera Jose-B46482 wrote: > > From: Kim Phillips [mailto:kim.phill...@freescale.com] > > Sent: Wednesday, March 25, 2015 4:13 PM > > > > On Tue, 24 Mar 2015 21:32:56 -0500 > > Stuart Yoder wrote: > > > > > On a system with the MC running with caches on, u-boot took 20 seconds > > > to boot. The MC took about 5 seconds of that, most of that in DPL > > > processing. > > > > > > We would welcome help analyzing u-boot boot time and where the time is > > > going. But, seriously, saving 1 ms is not going to help at all. > > > > I did a quick experiment and saved ~50ms when setting both udelays down > > to 50, which, sure, isn't a big deal given it's out of the order of > > 10sec, but it's something, and, like I said, development time for our > > users can be seriously helped if MC initialization were omitted from the > > main u-boot boot sequence, and occurred only when necessary, i.e., when > > users want to use one of the DP net interfaces. Most of the time when we > > boot today, we don't use DP net interfaces, so MC init - with or without > > DPL processing - is just a waste of our time! > > > The 50ms improvement you claim will not make any difference to save time > For development users, since 50ms is not something humans can perceive. > As Stuart said, we (the MC team) should first analyze where is the bulk of > The DPL processing by the MC fw is being spent, to see how much we can > reduce the 5 seconds spent there. That is, we should be concerned first > about where we can save big bucks, before being concerned about where we > can save one penny or two. my point is that all 10 seconds of MC processing should be removed from u-boot startup time, and moved to only when u-boot needs to use MC-based ethernet, e.g., when a tftp command is invoked. and, fwiw, the 50ms figure will improve as MC firmware performance improves. Kim ___ U-Boot mailing list U-Boot@lists.denx.de http://lists.denx.de/mailman/listinfo/u-boot
Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware loading for new boot architecture
> -Original Message- > From: Kim Phillips [mailto:kim.phill...@freescale.com] > Sent: Wednesday, March 25, 2015 4:13 PM > To: Stuart Yoder > Cc: Rivera Jose-B46482; Sun York-R58495; u-boot@lists.denx.de > Subject: Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware > loading for new boot architecture > > On Tue, 24 Mar 2015 21:32:56 -0500 > Stuart Yoder wrote: > > > Kim, > > why is this being taken off-list? Adding it back. Also, please don't > top-post. > > > I think it is premature to start focusing on saving single digit > > milliseconds in u-boot. The 35 second boot time you are seeing on > > some ls2085 boards is due to the fact that the MC was running with > > it's CPU caches off until a few weeks ago. DDR is still running at > > the slowest speed, which will affect the MC's performance. It's still > > early in hardware bringup and things are not even stable yet. > > > > I'm still convinced that the MC itself probably has significant > > performance work. > > > > The 1ms delay used polling for MC to boot has nothing to do with DPL > > size. DPL processing is a separate, later initialization step. We're > > just waiting for the MC to initialize here. The MC boot time > > shouldn't vary. My visual reading watching things boot is that the MC > > boot takes a few hundred milliseconds. I don't see what's wrong with > > a 1 ms polling delay here. > > > > On a system with the MC running with caches on, u-boot took 20 seconds > > to boot. The MC took about 5 seconds of that, most of that in DPL > > processing. > > > > We would welcome help analyzing u-boot boot time and where the time is > > going. But, seriously, saving 1 ms is not going to help at all. > > I did a quick experiment and saved ~50ms when setting both udelays down > to 50, which, sure, isn't a big deal given it's out of the order of > 10sec, but it's something, and, like I said, development time for our > users can be seriously helped if MC initialization were omitted from the > main u-boot boot sequence, and occurred only when necessary, i.e., when > users want to use one of the DP net interfaces. Most of the time when we > boot today, we don't use DP net interfaces, so MC init - with or without > DPL processing - is just a waste of our time! > The 50ms improvement you claim will not make any difference to save time For development users, since 50ms is not something humans can perceive. As Stuart said, we (the MC team) should first analyze where is the bulk of The DPL processing by the MC fw is being spent, to see how much we can reduce the 5 seconds spent there. That is, we should be concerned first about where we can save big bucks, before being concerned about where we can save one penny or two. Thanks, German > Thanks, > > Kim > > > Thanks, > > Stuart > > > > > > > > On Tue, Mar 24, 2015 at 10:35 AM, Kim Phillips > > wrote: > > > On Tue, 24 Mar 2015 10:14:39 -0500 > > > Rivera Jose-B46482 wrote: > > > > > >> > From: Kim Phillips [mailto:kim.phill...@freescale.com] > > >> > Sent: Monday, March 23, 2015 5:06 PM > > >> > > > >> > On Mon, 23 Mar 2015 16:15:56 -0500 Rivera Jose-B46482 > > >> > wrote: > > >> > > > >> > > > From: Kim Phillips [mailto:kim.phill...@freescale.com] > > >> > > > Sent: Monday, March 23, 2015 3:34 PM > > >> > > > > > >> > > > On Mon, 23 Mar 2015 15:06:11 -0500 Rivera Jose-B46482 > > >> > > > wrote: > > >> > > > > > >> > > > > > From: Kim Phillips [mailto:kim.phill...@freescale.com] > > >> > > > > > Sent: Thursday, March 19, 2015 12:53 PM > > >> > > > > > > > >> > > > > > On Thu, 19 Mar 2015 09:45:45 -0700 York Sun > > >> > > > > > wrote: > > >> > > > > > > > >> > > > > > > From: "J. German Rivera" > > >> > > > > > > > > >> > > > > > > Changed MC firmware loading to comply with the new MC > > >> > > > > > > boot > > >> > > > > > architecture. > > >> > > > > > > Flush D-cache hierarchy after loading MC images. Add > > >> > > > > > > environment variables "mcboottimeout" for MC boot > > >> >
Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware loading for new boot architecture
On Tue, 24 Mar 2015 21:32:56 -0500 Stuart Yoder wrote: > Kim, why is this being taken off-list? Adding it back. Also, please don't top-post. > I think it is premature to start focusing on saving single digit > milliseconds in u-boot. The 35 second boot time you are seeing on > some ls2085 boards is due to the fact that the MC was running with > it's CPU caches off until a few weeks ago. DDR is still running at > the slowest speed, which will affect the MC's performance. It's still > early in hardware bringup and things are not even stable yet. > > I'm still convinced that the MC itself probably has significant > performance work. > > The 1ms delay used polling for MC to boot has nothing to do with DPL > size. DPL processing is a separate, later initialization step. We're > just waiting for the MC to initialize here. The MC boot time > shouldn't vary. My visual reading watching things boot is that the MC > boot takes a few hundred milliseconds. I don't see what's wrong with > a 1 ms polling delay here. > > On a system with the MC running with caches on, u-boot took 20 seconds > to boot. The MC took about 5 seconds of that, most of that in DPL > processing. > > We would welcome help analyzing u-boot boot time and where the time is > going. But, seriously, saving 1 ms is not going to help at all. I did a quick experiment and saved ~50ms when setting both udelays down to 50, which, sure, isn't a big deal given it's out of the order of 10sec, but it's something, and, like I said, development time for our users can be seriously helped if MC initialization were omitted from the main u-boot boot sequence, and occurred only when necessary, i.e., when users want to use one of the DP net interfaces. Most of the time when we boot today, we don't use DP net interfaces, so MC init - with or without DPL processing - is just a waste of our time! Thanks, Kim > Thanks, > Stuart > > > > On Tue, Mar 24, 2015 at 10:35 AM, Kim Phillips > wrote: > > On Tue, 24 Mar 2015 10:14:39 -0500 > > Rivera Jose-B46482 wrote: > > > >> > From: Kim Phillips [mailto:kim.phill...@freescale.com] > >> > Sent: Monday, March 23, 2015 5:06 PM > >> > > >> > On Mon, 23 Mar 2015 16:15:56 -0500 > >> > Rivera Jose-B46482 wrote: > >> > > >> > > > From: Kim Phillips [mailto:kim.phill...@freescale.com] > >> > > > Sent: Monday, March 23, 2015 3:34 PM > >> > > > > >> > > > On Mon, 23 Mar 2015 15:06:11 -0500 > >> > > > Rivera Jose-B46482 wrote: > >> > > > > >> > > > > > From: Kim Phillips [mailto:kim.phill...@freescale.com] > >> > > > > > Sent: Thursday, March 19, 2015 12:53 PM > >> > > > > > > >> > > > > > On Thu, 19 Mar 2015 09:45:45 -0700 York Sun > >> > > > > > wrote: > >> > > > > > > >> > > > > > > From: "J. German Rivera" > >> > > > > > > > >> > > > > > > Changed MC firmware loading to comply with the new MC boot > >> > > > > > architecture. > >> > > > > > > Flush D-cache hierarchy after loading MC images. Add > >> > > > > > > environment variables "mcboottimeout" for MC boot timeout in > >> > > > > > > milliseconds, "mcmemsize" for MC DRAM block size. Check MC > >> > > > > > > boot status before calling flib functions. > >> > > > > > > >> > > > > > Can we just assign actual and/or optimal values for > >> > 'mcboottimeout' > >> > > > > > and 'mcmemsize' instead of making them environment variables? > >> > > > > > > >> > > > > Having environment variables gives us the flexibility if these > >> > > > > values need to be changed for a given customer configuration. The > >> > > > > actual > >> > > > > >> > > > what defines a 'customer configuration,' and how does that manifest > >> > > > itself at u-boot boot time? > >> > > A DPL (data path layout - a device-tree-like structure describing The > >> > > DPAA2 objects created at boot time and their connections) > >> > > > >> > > > Is it the amount of time it takes to load (and execute?) firmare? > >> > > Yes, bigger DPLs take longer to process by the MC. > >> > > > >> > > > Why isn't customer-specific firmware being loaded via linux? All > >> > > > u-boot needs is basic networking, pretty static > >> > > > setup: fixed numbers for both memsize & timeout. > >> > > This is not customer-specific firmware. What is customer-specific is > >> > just the DPL. > >> > > In order to have networking in u-boot, we need to load the MC firmware > >> > > in u-boot, For cases in which the target system has only DPAA2-based > >> > network interfaces. > >> > > >> > ok, for that case, when time comes for u-boot to do some DPAA2 networking > >> > arrives (i.e., we shouldn't have to be loading firmware at board boot- > >> > time), then we should load a minimal DPL for the number of singular, non- > >> > virtual/switch, etc., interfaces for that board just to tftp: this > >> > shouldn't be a big DPL at all, and its time complexity is fixed. > >> > > >> It is up to the customer to decide what kind of DPL they want to have. > > > > true, but in this case the 'customer' is the average upstream u-boo
Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware loading for new boot architecture
> -Original Message-
> From: Kim Phillips [mailto:kim.phill...@freescale.com]
> Sent: Monday, March 23, 2015 5:06 PM
> To: Rivera Jose-B46482
> Cc: Sun York-R58495; u-boot@lists.denx.de
> Subject: Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware
> loading for new boot architecture
>
> On Mon, 23 Mar 2015 16:15:56 -0500
> Rivera Jose-B46482 wrote:
>
> > > -Original Message-
> > > From: Kim Phillips [mailto:kim.phill...@freescale.com]
> > > Sent: Monday, March 23, 2015 3:34 PM
> > > To: Rivera Jose-B46482
> > > Cc: Sun York-R58495; u-boot@lists.denx.de
> > > Subject: Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC
> > > firmware loading for new boot architecture
> > >
> > > On Mon, 23 Mar 2015 15:06:11 -0500
> > > Rivera Jose-B46482 wrote:
> > >
> > > > > From: Kim Phillips [mailto:kim.phill...@freescale.com]
> > > > > Sent: Thursday, March 19, 2015 12:53 PM
> > > > >
> > > > > On Thu, 19 Mar 2015 09:45:45 -0700 York Sun
> > > > > wrote:
> > > > >
> > > > > > From: "J. German Rivera"
> > > > > >
> > > > > > Changed MC firmware loading to comply with the new MC boot
> > > > > architecture.
> > > > > > Flush D-cache hierarchy after loading MC images. Add
> > > > > > environment variables "mcboottimeout" for MC boot timeout in
> > > > > > milliseconds, "mcmemsize" for MC DRAM block size. Check MC
> > > > > > boot status before calling flib functions.
> > > > >
> > > > > Can we just assign actual and/or optimal values for
> 'mcboottimeout'
> > > > > and 'mcmemsize' instead of making them environment variables?
> > > > >
> > > > Having environment variables gives us the flexibility if these
> > > > values need to be changed for a given customer configuration. The
> > > > actual
> > >
> > > what defines a 'customer configuration,' and how does that manifest
> > > itself at u-boot boot time?
> > A DPL (data path layout - a device-tree-like structure describing The
> > DPAA2 objects created at boot time and their connections)
> >
> > > Is it the amount of time it takes to load (and execute?) firmare?
> > Yes, bigger DPLs take longer to process by the MC.
> >
> > > Why isn't customer-specific firmware being loaded via linux? All
> > > u-boot needs is basic networking, pretty static
> > > setup: fixed numbers for both memsize & timeout.
> > This is not customer-specific firmware. What is customer-specific is
> just the DPL.
> > In order to have networking in u-boot, we need to load the MC firmware
> > in u-boot, For cases in which the target system has only DPAA2-based
> network interfaces.
>
> ok, for that case, when time comes for u-boot to do some DPAA2 networking
> arrives (i.e., we shouldn't have to be loading firmware at board boot-
> time), then we should load a minimal DPL for the number of singular, non-
> virtual/switch, etc., interfaces for that board just to tftp: this
> shouldn't be a big DPL at all, and its time complexity is fixed.
>
It is up to the customer to decide what kind of DPL they want to have.
> > > > boot time of the MC and the amount of memory needed by the MC is
> > > > dependent on how big/complex is the DPL used. Also, the memory
> > > > needed by the MC needs to account for how much memory is needed
> > > > for AIOP programs, which may depend on how big/complex they are.
> > >
> > > ok, that helps (modulo not knowing what 'DPL' is), but still, the
> > > massive customer configurations should be being loaded via linux'
> > > firmware loading infrastructure: u-boot should be using a static
> > > image for u-boot's needs.
> > >
> > > > > > +static int wait_for_mc(bool booting_mc, u32 *final_reg_gsr) {
> > > > > > + u32 reg_gsr;
> > > > > > + u32 mc_fw_boot_status;
> > > > > > + unsigned long timeout_ms = get_mc_boot_timeout_ms();
> > > > > > + struct mc_ccsr_registers __iomem *mc_ccsr_regs =
> > > > > > +MC_CCSR_BASE_ADDR;
> > > > > > +
> > > > > > + dmb();
> > > > > > + debug("Polling mc_ccsr_regs->reg_gsr ...\n");
> > > > > > + assert(timeout_ms >
Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware loading for new boot architecture
On Tue, 24 Mar 2015 10:14:39 -0500
Rivera Jose-B46482 wrote:
> > From: Kim Phillips [mailto:kim.phill...@freescale.com]
> > Sent: Monday, March 23, 2015 5:06 PM
> >
> > On Mon, 23 Mar 2015 16:15:56 -0500
> > Rivera Jose-B46482 wrote:
> >
> > > > From: Kim Phillips [mailto:kim.phill...@freescale.com]
> > > > Sent: Monday, March 23, 2015 3:34 PM
> > > >
> > > > On Mon, 23 Mar 2015 15:06:11 -0500
> > > > Rivera Jose-B46482 wrote:
> > > >
> > > > > > From: Kim Phillips [mailto:kim.phill...@freescale.com]
> > > > > > Sent: Thursday, March 19, 2015 12:53 PM
> > > > > >
> > > > > > On Thu, 19 Mar 2015 09:45:45 -0700 York Sun
> > > > > > wrote:
> > > > > >
> > > > > > > From: "J. German Rivera"
> > > > > > >
> > > > > > > Changed MC firmware loading to comply with the new MC boot
> > > > > > architecture.
> > > > > > > Flush D-cache hierarchy after loading MC images. Add
> > > > > > > environment variables "mcboottimeout" for MC boot timeout in
> > > > > > > milliseconds, "mcmemsize" for MC DRAM block size. Check MC
> > > > > > > boot status before calling flib functions.
> > > > > >
> > > > > > Can we just assign actual and/or optimal values for
> > 'mcboottimeout'
> > > > > > and 'mcmemsize' instead of making them environment variables?
> > > > > >
> > > > > Having environment variables gives us the flexibility if these
> > > > > values need to be changed for a given customer configuration. The
> > > > > actual
> > > >
> > > > what defines a 'customer configuration,' and how does that manifest
> > > > itself at u-boot boot time?
> > > A DPL (data path layout - a device-tree-like structure describing The
> > > DPAA2 objects created at boot time and their connections)
> > >
> > > > Is it the amount of time it takes to load (and execute?) firmare?
> > > Yes, bigger DPLs take longer to process by the MC.
> > >
> > > > Why isn't customer-specific firmware being loaded via linux? All
> > > > u-boot needs is basic networking, pretty static
> > > > setup: fixed numbers for both memsize & timeout.
> > > This is not customer-specific firmware. What is customer-specific is
> > just the DPL.
> > > In order to have networking in u-boot, we need to load the MC firmware
> > > in u-boot, For cases in which the target system has only DPAA2-based
> > network interfaces.
> >
> > ok, for that case, when time comes for u-boot to do some DPAA2 networking
> > arrives (i.e., we shouldn't have to be loading firmware at board boot-
> > time), then we should load a minimal DPL for the number of singular, non-
> > virtual/switch, etc., interfaces for that board just to tftp: this
> > shouldn't be a big DPL at all, and its time complexity is fixed.
> >
> It is up to the customer to decide what kind of DPL they want to have.
true, but in this case the 'customer' is the average upstream u-boot
user, presumably whose DPL is the simplest ethernet use-case for
tftp'ing kernels.
> > > > > boot time of the MC and the amount of memory needed by the MC is
> > > > > dependent on how big/complex is the DPL used. Also, the memory
> > > > > needed by the MC needs to account for how much memory is needed
> > > > > for AIOP programs, which may depend on how big/complex they are.
> > > >
> > > > ok, that helps (modulo not knowing what 'DPL' is), but still, the
> > > > massive customer configurations should be being loaded via linux'
> > > > firmware loading infrastructure: u-boot should be using a static
> > > > image for u-boot's needs.
> > > >
> > > > > > > +static int wait_for_mc(bool booting_mc, u32 *final_reg_gsr) {
> > > > > > > + u32 reg_gsr;
> > > > > > > + u32 mc_fw_boot_status;
> > > > > > > + unsigned long timeout_ms = get_mc_boot_timeout_ms();
> > > > > > > + struct mc_ccsr_registers __iomem *mc_ccsr_regs =
> > > > > > > +MC_CCSR_BASE_ADDR;
> > > > > > > +
> > > > > > > + dmb();
> > > > > > > + debug("Polling mc_ccsr_regs->reg_gsr ...\n");
> > > > > > > + assert(timeout_ms > 0);
> > > > > > > + for (;;) {
> > > > > > > + udelay(1000); /* throttle polling */
> > > > > >
> > > > > > does this really need to be a whole 1ms?
> > > > >
> > > > > It is unlikely that the MC fw will boot in less than 1 ms.
> > > >
> > > > is wait_for_mc() only called for the boot command, or all commands?
> >
> > I see: there's a udelay(500) in mc_send_command(), which is too high,
> > too, IMO, but I'm not that familiar with the h/w: How long does the
> > shortest command take?
Can you answer this?
> > > > > So, checking more frequently than 1 ms is not necessary.
> > > >
> > > > yes it is, because e.g., if it takes 1.1ms we will have wasted 0.9ms
> > > > on this.
> > > >
> > > How significant is to save 0.9ms of the whole boot time?
> >
> > Why waste 0.9ms of boot time when there's no need? It already takes the
> > boards *way* too long to boot, and now I'm understanding
> > mc_send_command's delay should probably be adjusted, too.
> >
> I have not heard any complain about RDB/QDS boards taking too long to boot
> Due to this "wasteds 0.9
Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware loading for new boot architecture
> -Original Message-
> From: Kim Phillips [mailto:kim.phill...@freescale.com]
> Sent: Thursday, March 19, 2015 12:53 PM
> To: Sun York-R58495
> Cc: u-boot@lists.denx.de; Rivera Jose-B46482
> Subject: Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware
> loading for new boot architecture
>
> On Thu, 19 Mar 2015 09:45:45 -0700
> York Sun wrote:
>
> > From: "J. German Rivera"
> >
> > Changed MC firmware loading to comply with the new MC boot
> architecture.
> > Flush D-cache hierarchy after loading MC images. Add environment
> > variables "mcboottimeout" for MC boot timeout in milliseconds,
> > "mcmemsize" for MC DRAM block size. Check MC boot status before
> > calling flib functions.
>
> Can we just assign actual and/or optimal values for 'mcboottimeout'
> and 'mcmemsize' instead of making them environment variables?
>
Having environment variables gives us the flexibility if these
values need to be changed for a given customer configuration. The actual
boot time of the MC and the amount of memory needed by the MC is dependent
on how big/complex is the DPL used. Also, the memory needed by the MC
needs to account for how much memory is needed for AIOP programs,
which may depend on how big/complex they are.
> > +static int wait_for_mc(bool booting_mc, u32 *final_reg_gsr) {
> > + u32 reg_gsr;
> > + u32 mc_fw_boot_status;
> > + unsigned long timeout_ms = get_mc_boot_timeout_ms();
> > + struct mc_ccsr_registers __iomem *mc_ccsr_regs = MC_CCSR_BASE_ADDR;
> > +
> > + dmb();
> > + debug("Polling mc_ccsr_regs->reg_gsr ...\n");
> > + assert(timeout_ms > 0);
> > + for (;;) {
> > + udelay(1000); /* throttle polling */
>
> does this really need to be a whole 1ms?
It is unlikely that the MC fw will boot in less than 1 ms.
So, checking more frequently than 1 ms is not necessary.
>
> > @@ -318,14 +545,28 @@ int get_mc_boot_status(void)
> >
> > /**
> > * Return the actual size of the MC private DRAM block.
> > - *
> > - * NOTE: For now this function always returns the minimum required
> > size,
> > - * However, in the future, the actual size may be obtained from an
> > environment
> > - * variable.
> > */
>
> why?
>
See answer above.
German
___
U-Boot mailing list
U-Boot@lists.denx.de
http://lists.denx.de/mailman/listinfo/u-boot
Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware loading for new boot architecture
On Mon, 23 Mar 2015 16:15:56 -0500
Rivera Jose-B46482 wrote:
> > -Original Message-
> > From: Kim Phillips [mailto:kim.phill...@freescale.com]
> > Sent: Monday, March 23, 2015 3:34 PM
> > To: Rivera Jose-B46482
> > Cc: Sun York-R58495; u-boot@lists.denx.de
> > Subject: Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware
> > loading for new boot architecture
> >
> > On Mon, 23 Mar 2015 15:06:11 -0500
> > Rivera Jose-B46482 wrote:
> >
> > > > From: Kim Phillips [mailto:kim.phill...@freescale.com]
> > > > Sent: Thursday, March 19, 2015 12:53 PM
> > > >
> > > > On Thu, 19 Mar 2015 09:45:45 -0700
> > > > York Sun wrote:
> > > >
> > > > > From: "J. German Rivera"
> > > > >
> > > > > Changed MC firmware loading to comply with the new MC boot
> > > > architecture.
> > > > > Flush D-cache hierarchy after loading MC images. Add environment
> > > > > variables "mcboottimeout" for MC boot timeout in milliseconds,
> > > > > "mcmemsize" for MC DRAM block size. Check MC boot status before
> > > > > calling flib functions.
> > > >
> > > > Can we just assign actual and/or optimal values for 'mcboottimeout'
> > > > and 'mcmemsize' instead of making them environment variables?
> > > >
> > > Having environment variables gives us the flexibility if these values
> > > need to be changed for a given customer configuration. The actual
> >
> > what defines a 'customer configuration,' and how does that manifest
> > itself at u-boot boot time?
> A DPL (data path layout - a device-tree-like structure describing
> The DPAA2 objects created at boot time and their connections)
>
> > Is it the amount of time it takes to load
> > (and execute?) firmare?
> Yes, bigger DPLs take longer to process by the MC.
>
> > Why isn't customer-specific firmware being
> > loaded via linux? All u-boot needs is basic networking, pretty static
> > setup: fixed numbers for both memsize & timeout.
> This is not customer-specific firmware. What is customer-specific is just the
> DPL.
> In order to have networking in u-boot, we need to load the MC firmware in
> u-boot,
> For cases in which the target system has only DPAA2-based network interfaces.
ok, for that case, when time comes for u-boot to do some DPAA2
networking arrives (i.e., we shouldn't have to be loading firmware
at board boot-time), then we should load a minimal DPL for the
number of singular, non-virtual/switch, etc., interfaces for that
board just to tftp: this shouldn't be a big DPL at all, and its
time complexity is fixed.
> > > boot time of the MC and the amount of memory needed by the MC is
> > > dependent on how big/complex is the DPL used. Also, the memory needed
> > > by the MC needs to account for how much memory is needed for AIOP
> > > programs, which may depend on how big/complex they are.
> >
> > ok, that helps (modulo not knowing what 'DPL' is), but still, the massive
> > customer configurations should be being loaded via linux'
> > firmware loading infrastructure: u-boot should be using a static image
> > for u-boot's needs.
> >
> > > > > +static int wait_for_mc(bool booting_mc, u32 *final_reg_gsr) {
> > > > > + u32 reg_gsr;
> > > > > + u32 mc_fw_boot_status;
> > > > > + unsigned long timeout_ms = get_mc_boot_timeout_ms();
> > > > > + struct mc_ccsr_registers __iomem *mc_ccsr_regs =
> > > > > +MC_CCSR_BASE_ADDR;
> > > > > +
> > > > > + dmb();
> > > > > + debug("Polling mc_ccsr_regs->reg_gsr ...\n");
> > > > > + assert(timeout_ms > 0);
> > > > > + for (;;) {
> > > > > + udelay(1000); /* throttle polling */
> > > >
> > > > does this really need to be a whole 1ms?
> > >
> > > It is unlikely that the MC fw will boot in less than 1 ms.
> >
> > is wait_for_mc() only called for the boot command, or all commands?
I see: there's a udelay(500) in mc_send_command(), which is too high,
too, IMO, but I'm not that familiar with the h/w: How long does the
shortest command take?
> > > So, checking more frequently than 1 ms is not necessary.
> >
> > yes it is, because e.g., if it takes 1.1ms we will have wasted 0.9ms on
> > this.
> >
> How significant
Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware loading for new boot architecture
> -Original Message-
> From: Kim Phillips [mailto:kim.phill...@freescale.com]
> Sent: Monday, March 23, 2015 3:34 PM
> To: Rivera Jose-B46482
> Cc: Sun York-R58495; u-boot@lists.denx.de
> Subject: Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware
> loading for new boot architecture
>
> On Mon, 23 Mar 2015 15:06:11 -0500
> Rivera Jose-B46482 wrote:
>
> > > From: Kim Phillips [mailto:kim.phill...@freescale.com]
> > > Sent: Thursday, March 19, 2015 12:53 PM
> > >
> > > On Thu, 19 Mar 2015 09:45:45 -0700
> > > York Sun wrote:
> > >
> > > > From: "J. German Rivera"
> > > >
> > > > Changed MC firmware loading to comply with the new MC boot
> > > architecture.
> > > > Flush D-cache hierarchy after loading MC images. Add environment
> > > > variables "mcboottimeout" for MC boot timeout in milliseconds,
> > > > "mcmemsize" for MC DRAM block size. Check MC boot status before
> > > > calling flib functions.
> > >
> > > Can we just assign actual and/or optimal values for 'mcboottimeout'
> > > and 'mcmemsize' instead of making them environment variables?
> > >
> > Having environment variables gives us the flexibility if these values
> > need to be changed for a given customer configuration. The actual
>
> what defines a 'customer configuration,' and how does that manifest
> itself at u-boot boot time?
A DPL (data path layout - a device-tree-like structure describing
The DPAA2 objects created at boot time and their connections)
> Is it the amount of time it takes to load
> (and execute?) firmare?
Yes, bigger DPLs take longer to process by the MC.
> Why isn't customer-specific firmware being
> loaded via linux? All u-boot needs is basic networking, pretty static
> setup: fixed numbers for both memsize & timeout.
This is not customer-specific firmware. What is customer-specific is just the
DPL.
In order to have networking in u-boot, we need to load the MC firmware in
u-boot,
For cases in which the target system has only DPAA2-based network interfaces.
>
> > boot time of the MC and the amount of memory needed by the MC is
> > dependent on how big/complex is the DPL used. Also, the memory needed
> > by the MC needs to account for how much memory is needed for AIOP
> > programs, which may depend on how big/complex they are.
>
> ok, that helps (modulo not knowing what 'DPL' is), but still, the massive
> customer configurations should be being loaded via linux'
> firmware loading infrastructure: u-boot should be using a static image
> for u-boot's needs.
>
> > > > +static int wait_for_mc(bool booting_mc, u32 *final_reg_gsr) {
> > > > + u32 reg_gsr;
> > > > + u32 mc_fw_boot_status;
> > > > + unsigned long timeout_ms = get_mc_boot_timeout_ms();
> > > > + struct mc_ccsr_registers __iomem *mc_ccsr_regs =
> > > > +MC_CCSR_BASE_ADDR;
> > > > +
> > > > + dmb();
> > > > + debug("Polling mc_ccsr_regs->reg_gsr ...\n");
> > > > + assert(timeout_ms > 0);
> > > > + for (;;) {
> > > > + udelay(1000); /* throttle polling */
> > >
> > > does this really need to be a whole 1ms?
> >
> > It is unlikely that the MC fw will boot in less than 1 ms.
>
> is wait_for_mc() only called for the boot command, or all commands?
>
> > So, checking more frequently than 1 ms is not necessary.
>
> yes it is, because e.g., if it takes 1.1ms we will have wasted 0.9ms on
> this.
>
How significant is to save 0.9ms of the whole boot time?
As the comment in the code says, the intent was to throttle down the polling,
to reduce traffic on the system bus due to polling. This traffic competes with
the MC itself accessing the system bus, as it boots. Having the polling traffic
get
in the way of the MC traffic may increase the MC boot time. Too small delay
between polls may cause the MC boot time to increase more than the .9ms you
are concerned of wasting in the delay.
What value would you suggest to use for the delay instead 1000ms?
> Kim
___
U-Boot mailing list
U-Boot@lists.denx.de
http://lists.denx.de/mailman/listinfo/u-boot
Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware loading for new boot architecture
On Mon, 23 Mar 2015 15:06:11 -0500
Rivera Jose-B46482 wrote:
> > From: Kim Phillips [mailto:kim.phill...@freescale.com]
> > Sent: Thursday, March 19, 2015 12:53 PM
> >
> > On Thu, 19 Mar 2015 09:45:45 -0700
> > York Sun wrote:
> >
> > > From: "J. German Rivera"
> > >
> > > Changed MC firmware loading to comply with the new MC boot
> > architecture.
> > > Flush D-cache hierarchy after loading MC images. Add environment
> > > variables "mcboottimeout" for MC boot timeout in milliseconds,
> > > "mcmemsize" for MC DRAM block size. Check MC boot status before
> > > calling flib functions.
> >
> > Can we just assign actual and/or optimal values for 'mcboottimeout'
> > and 'mcmemsize' instead of making them environment variables?
> >
> Having environment variables gives us the flexibility if these
> values need to be changed for a given customer configuration. The actual
what defines a 'customer configuration,' and how does that manifest
itself at u-boot boot time? Is it the amount of time it takes to
load (and execute?) firmare? Why isn't customer-specific firmware
being loaded via linux? All u-boot needs is basic networking,
pretty static setup: fixed numbers for both memsize & timeout.
> boot time of the MC and the amount of memory needed by the MC is dependent
> on how big/complex is the DPL used. Also, the memory needed by the MC
> needs to account for how much memory is needed for AIOP programs,
> which may depend on how big/complex they are.
ok, that helps (modulo not knowing what 'DPL' is), but still, the
massive customer configurations should be being loaded via linux'
firmware loading infrastructure: u-boot should be using a static
image for u-boot's needs.
> > > +static int wait_for_mc(bool booting_mc, u32 *final_reg_gsr) {
> > > + u32 reg_gsr;
> > > + u32 mc_fw_boot_status;
> > > + unsigned long timeout_ms = get_mc_boot_timeout_ms();
> > > + struct mc_ccsr_registers __iomem *mc_ccsr_regs = MC_CCSR_BASE_ADDR;
> > > +
> > > + dmb();
> > > + debug("Polling mc_ccsr_regs->reg_gsr ...\n");
> > > + assert(timeout_ms > 0);
> > > + for (;;) {
> > > + udelay(1000); /* throttle polling */
> >
> > does this really need to be a whole 1ms?
>
> It is unlikely that the MC fw will boot in less than 1 ms.
is wait_for_mc() only called for the boot command, or all
commands?
> So, checking more frequently than 1 ms is not necessary.
yes it is, because e.g., if it takes 1.1ms we will have wasted 0.9ms
on this.
Kim
___
U-Boot mailing list
U-Boot@lists.denx.de
http://lists.denx.de/mailman/listinfo/u-boot
Re: [U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware loading for new boot architecture
On Thu, 19 Mar 2015 09:45:45 -0700
York Sun wrote:
> From: "J. German Rivera"
>
> Changed MC firmware loading to comply with the new MC boot architecture.
> Flush D-cache hierarchy after loading MC images. Add environment
> variables "mcboottimeout" for MC boot timeout in milliseconds,
> "mcmemsize" for MC DRAM block size. Check MC boot status before calling
> flib functions.
Can we just assign actual and/or optimal values for 'mcboottimeout'
and 'mcmemsize' instead of making them environment variables?
> +static int wait_for_mc(bool booting_mc, u32 *final_reg_gsr)
> +{
> + u32 reg_gsr;
> + u32 mc_fw_boot_status;
> + unsigned long timeout_ms = get_mc_boot_timeout_ms();
> + struct mc_ccsr_registers __iomem *mc_ccsr_regs = MC_CCSR_BASE_ADDR;
> +
> + dmb();
> + debug("Polling mc_ccsr_regs->reg_gsr ...\n");
> + assert(timeout_ms > 0);
> + for (;;) {
> + udelay(1000); /* throttle polling */
does this really need to be a whole 1ms?
> @@ -318,14 +545,28 @@ int get_mc_boot_status(void)
>
> /**
> * Return the actual size of the MC private DRAM block.
> - *
> - * NOTE: For now this function always returns the minimum required size,
> - * However, in the future, the actual size may be obtained from an
> environment
> - * variable.
> */
why?
Kim
___
U-Boot mailing list
U-Boot@lists.denx.de
http://lists.denx.de/mailman/listinfo/u-boot
[U-Boot] [PATCH 14/28] drivers/fsl-mc: Changed MC firmware loading for new boot architecture
From: "J. German Rivera"
Changed MC firmware loading to comply with the new MC boot architecture.
Flush D-cache hierarchy after loading MC images. Add environment
variables "mcboottimeout" for MC boot timeout in milliseconds,
"mcmemsize" for MC DRAM block size. Check MC boot status before calling
flib functions.
Signed-off-by: J. German Rivera
---
arch/arm/cpu/armv8/fsl-lsch3/README |8 +
drivers/net/fsl-mc/mc.c | 494 ++-
include/configs/ls2085a_common.h|6 +-
include/configs/ls2085a_emu.h | 13 +-
include/configs/ls2085a_simu.h |5 +
5 files changed, 395 insertions(+), 131 deletions(-)
diff --git a/arch/arm/cpu/armv8/fsl-lsch3/README
b/arch/arm/cpu/armv8/fsl-lsch3/README
index 99fc39a..f781620 100644
--- a/arch/arm/cpu/armv8/fsl-lsch3/README
+++ b/arch/arm/cpu/armv8/fsl-lsch3/README
@@ -33,3 +33,11 @@ is shown below considering a 32MB NOR flash device:
- > 0x_
32-MB NOR flash layout
+
+Environment Variables
+=
+mcboottimeout: MC boot timeout in milliseconds. If this variable is not defined
+ the value CONFIG_SYS_LS_MC_BOOT_TIMEOUT_MS will be assumed.
+
+mcmemsize: MC DRAM block size. If this variable is not defined, the value
+ CONFIG_SYS_LS_MC_DRAM_BLOCK_MIN_SIZE will be assumed.
diff --git a/drivers/net/fsl-mc/mc.c b/drivers/net/fsl-mc/mc.c
index 2a2b0af..c5c44bc 100644
--- a/drivers/net/fsl-mc/mc.c
+++ b/drivers/net/fsl-mc/mc.c
@@ -3,7 +3,6 @@
*
* SPDX-License-Identifier:GPL-2.0+
*/
-
#include
#include
#include
@@ -15,14 +14,64 @@
#include
#include
+#define MC_RAM_BASE_ADDR_ALIGNMENT (512UL * 1024 * 1024)
+#define MC_RAM_BASE_ADDR_ALIGNMENT_MASK(~(MC_RAM_BASE_ADDR_ALIGNMENT -
1))
+#define MC_RAM_SIZE_ALIGNMENT (256UL * 1024 * 1024)
+
+#define MC_MEM_SIZE_ENV_VAR"mcmemsize"
+#define MC_BOOT_TIMEOUT_ENV_VAR"mcboottimeout"
+
DECLARE_GLOBAL_DATA_PTR;
static int mc_boot_status;
struct fsl_mc_io *dflt_mc_io = NULL;
uint16_t dflt_dprc_handle = 0;
struct fsl_dpbp_obj *dflt_dpbp = NULL;
struct fsl_dpio_obj *dflt_dpio = NULL;
-uint16_t dflt_dpio_handle = NULL;
+uint16_t dflt_dpio_handle = 0;
+
+#ifdef DEBUG
+void dump_ram_words(const char *title, void *addr)
+{
+ int i;
+ uint32_t *words = addr;
+
+ printf("Dumping beginning of %s (%p):\n", title, addr);
+ for (i = 0; i < 16; i++)
+ printf("%#x ", words[i]);
+
+ printf("\n");
+}
+void dump_mc_ccsr_regs(struct mc_ccsr_registers __iomem *mc_ccsr_regs)
+{
+ printf("MC CCSR registers:\n"
+ "reg_gcr1 %#x\n"
+ "reg_gsr %#x\n"
+ "reg_sicbalr %#x\n"
+ "reg_sicbahr %#x\n"
+ "reg_sicapr %#x\n"
+ "reg_mcfbalr %#x\n"
+ "reg_mcfbahr %#x\n"
+ "reg_mcfapr %#x\n"
+ "reg_psr %#x\n",
+ mc_ccsr_regs->reg_gcr1,
+ mc_ccsr_regs->reg_gsr,
+ mc_ccsr_regs->reg_sicbalr,
+ mc_ccsr_regs->reg_sicbahr,
+ mc_ccsr_regs->reg_sicapr,
+ mc_ccsr_regs->reg_mcfbalr,
+ mc_ccsr_regs->reg_mcfbahr,
+ mc_ccsr_regs->reg_mcfapr,
+ mc_ccsr_regs->reg_psr);
+}
+#else
+
+#define dump_ram_words(title, addr)
+#define dump_mc_ccsr_regs(mc_ccsr_regs)
+
+#endif /* DEBUG */
+
+#ifndef CONFIG_SYS_LS_MC_FW_IN_DDR
/**
* Copying MC firmware or DPL image to DDR
*/
@@ -31,6 +80,7 @@ static int mc_copy_image(const char *title,
{
debug("%s copied to address %p\n", title, (void *)mc_ram_addr);
memcpy((void *)mc_ram_addr, (void *)image_addr, image_size);
+ flush_dcache_range(mc_ram_addr, mc_ram_addr + image_size);
return 0;
}
@@ -92,22 +142,254 @@ int parse_mc_firmware_fit_image(const void
**raw_image_addr,
return 0;
}
+#endif
+
+/*
+ * Calculates the values to be used to specify the address range
+ * for the MC private DRAM block, in the MCFBALR/MCFBAHR registers.
+ * It returns the highest 512MB-aligned address within the given
+ * address range, in '*aligned_base_addr', and the number of 256 MiB
+ * blocks in it, in 'num_256mb_blocks'.
+ */
+static int calculate_mc_private_ram_params(u64 mc_private_ram_start_addr,
+ size_t mc_ram_size,
+ u64 *aligned_base_addr,
+ u8 *num_256mb_blocks)
+{
+ u64 addr;
+ u16 num_blocks;
+
+ if (mc_ram_size % MC_RAM_SIZE_ALIGNMENT != 0) {
+ printf("fsl-mc: ERROR: invalid MC private RAM size (%lu)\n",
+ mc_ram_size);
+ return -EINVAL;
+ }
+
+ num_blocks = mc_ram_size / MC_RAM_SIZE_ALIGNMENT;
+ if (num_blocks < 1 || num_blocks > 0xff) {
+ printf("fsl-mc: ERROR: invalid MC private RAM size (%
