| Message ID | 1359085160-28675-1-git-send-email-robherring2@gmail.com |
|---|---|
| State | Deferred |
| Delegated to: | Albert ARIBAUD |
| Headers | show |
On Thu, Jan 24, 2013 at 09:39:19PM -0600, Rob Herring wrote: > From: Rob Herring <rob.herring@calxeda.com> > > Currently, u-boot will always fixup the DT memory node on ARM. If the dtb > has correct memory information, then we don't want or need u-boot to touch > the memory node. Allow platforms to skip this by not filling in dram bank > information. > > Signed-off-by: Rob Herring <rob.herring@calxeda.com> OK. With the notion that it's best to not fiddle with already correct things, we should have a way "out" there. But this patch means that bd->bi_dram is never set which looks to have some bad side-effects for cache (see arch/arm/lib/cache-cp15.c) so can we come up with some other way out here? Thanks!
Dear Rob Herring, In message <1359085160-28675-1-git-send-email-robherring2@gmail.com> you wrote: > > Currently, u-boot will always fixup the DT memory node on ARM. If the dtb > has correct memory information, then we don't want or need u-boot to touch > the memory node. Allow platforms to skip this by not filling in dram bank > information. Well, if this is what you want, then please implement just this, and do not do much more. > void __dram_init_banksize(void) > { > +#if CONFIG_NR_DRAM_BANKS > gd->bd->bi_dram[0].start = CONFIG_SYS_SDRAM_BASE; > gd->bd->bi_dram[0].size = gd->ram_size; > +#endif > } This change will leave the memory information in struct bd_info uninitialized, so the "bdinfo" command will print just bogus data. This is not a good idea. > --- a/arch/arm/lib/bootm.c > +++ b/arch/arm/lib/bootm.c > @@ -79,6 +79,7 @@ void arch_lmb_reserve(struct lmb *lmb) > #ifdef CONFIG_OF_LIBFDT > static int fixup_memory_node(void *blob) > { > +#if CONFIG_NR_DRAM_BANKS > bd_t *bd = gd->bd; > int bank; > u64 start[CONFIG_NR_DRAM_BANKS]; > @@ -90,6 +91,9 @@ static int fixup_memory_node(void *blob) > } > > return fdt_fixup_memory_banks(blob, start, size, CONFIG_NR_DRAM_BANKS); > +#else > + return 0; > +#endif I dislike this. If you want to make a property of your device tree "immutable", then you should mark it there as such. Instead of removing the code here, U-Boot should then check for such a property and leave the value unchanged. Actually you are reaching here a point where it seems necessary that U-Boot itself is able to read the DT to configure itself correctly, i. e. the memory size information it holds should then also be extracted from the DT. Best regards, Wolfgang Denk
On 02/19/2013 01:26 PM, Wolfgang Denk wrote: > Dear Rob Herring, > > In message <1359085160-28675-1-git-send-email-robherring2@gmail.com> you wrote: >> >> Currently, u-boot will always fixup the DT memory node on ARM. If the dtb >> has correct memory information, then we don't want or need u-boot to touch >> the memory node. Allow platforms to skip this by not filling in dram bank >> information. > > Well, if this is what you want, then please implement just this, and > do not do much more. > >> void __dram_init_banksize(void) >> { >> +#if CONFIG_NR_DRAM_BANKS >> gd->bd->bi_dram[0].start = CONFIG_SYS_SDRAM_BASE; >> gd->bd->bi_dram[0].size = gd->ram_size; >> +#endif >> } > > This change will leave the memory information in struct bd_info > uninitialized, so the "bdinfo" command will print just bogus data. > > This is not a good idea. This hunk can be dropped I think. >> --- a/arch/arm/lib/bootm.c >> +++ b/arch/arm/lib/bootm.c >> @@ -79,6 +79,7 @@ void arch_lmb_reserve(struct lmb *lmb) >> #ifdef CONFIG_OF_LIBFDT >> static int fixup_memory_node(void *blob) >> { >> +#if CONFIG_NR_DRAM_BANKS >> bd_t *bd = gd->bd; >> int bank; >> u64 start[CONFIG_NR_DRAM_BANKS]; >> @@ -90,6 +91,9 @@ static int fixup_memory_node(void *blob) >> } >> >> return fdt_fixup_memory_banks(blob, start, size, CONFIG_NR_DRAM_BANKS); >> +#else >> + return 0; >> +#endif > > I dislike this. If you want to make a property of your device tree > "immutable", then you should mark it there as such. Instead of > removing the code here, U-Boot should then check for such a property > and leave the value unchanged. There is no such way defined to flag that and I'm not going to invent one. It's not that the memory node is immutable, but the default should not be "needs fixups". No one puts purposely wrong data in their DT. So therefore all data should be immutable? What happens when we need to fixup immutable data? What I would really prefer to do is like powerpc where platforms override this if they need the fixup and the default behavior is no fixup. But that would require knowing which platforms do and don't need the fixup. It only going to get harder to change that. > Actually you are reaching here a point where it seems necessary that > U-Boot itself is able to read the DT to configure itself correctly, > i. e. the memory size information it holds should then also be > extracted from the DT. I thought about doing this. That's really an orthogonal issue. The problem is you cannot adjust the amount of memory u-boot uses to be different than the amount of RAM in the /memory node. If you adjust the size in u-boot, the adjusted size is passed to the kernel no matter what. There's issues with LPAE systems having >4GB of RAM as all the size and address values are 32-bit. This could be fixed, but I'm not sure there is much value in u-boot knowing about memory above 4GB as it can't really access all of it and it would be purely informational. Rob > > Best regards, > > Wolfgang Denk >
Dear Rob Herring, In message <5123E311.1070607@gmail.com> you wrote: > > > I dislike this. If you want to make a property of your device tree > > "immutable", then you should mark it there as such. Instead of > > removing the code here, U-Boot should then check for such a property > > and leave the value unchanged. > > There is no such way defined to flag that and I'm not going to invent > one. It's not that the memory node is immutable, but the default should > not be "needs fixups". No one puts purposely wrong data in their DT. So > therefore all data should be immutable? What happens when we need to > fixup immutable data? There is for example the "status" property, which can take the values "okay" or "disabled". This is used in other places for not so unsimilar purposes. > What I would really prefer to do is like powerpc where platforms > override this if they need the fixup and the default behavior is no > fixup. But that would require knowing which platforms do and don't need > the fixup. It only going to get harder to change that. Normally we use get_ramsize() to check / verify the amount of available RAM. And on PPC, we nearly always set the ram size in the DT. I don't know why you think the default was no fixup; this is not true. It's the other way round. > > Actually you are reaching here a point where it seems necessary that > > U-Boot itself is able to read the DT to configure itself correctly, > > i. e. the memory size information it holds should then also be > > extracted from the DT. > > I thought about doing this. That's really an orthogonal issue. The > problem is you cannot adjust the amount of memory u-boot uses to be > different than the amount of RAM in the /memory node. If you adjust the > size in u-boot, the adjusted size is passed to the kernel no matter what. > > There's issues with LPAE systems having >4GB of RAM as all the size and > address values are 32-bit. This could be fixed, but I'm not sure there > is much value in u-boot knowing about memory above 4GB as it can't > really access all of it and it would be purely informational. Well, let's face it - no matter what you do, if the RAM sizes as used in U-Boot and in Linux differ, at least one of these is wrong. This is a bug that should be fixed. What you attempt to do is implementing a method to paper over such bugs. This is not a good thing to do. Best regards, Wolfgang Denk
On 02/19/2013 05:14 PM, Wolfgang Denk wrote: > Dear Rob Herring, > > In message <5123E311.1070607@gmail.com> you wrote: >> [snip] >> What I would really prefer to do is like powerpc where platforms >> override this if they need the fixup and the default behavior is no >> fixup. But that would require knowing which platforms do and don't need >> the fixup. It only going to get harder to change that. > > Normally we use get_ramsize() to check / verify the amount of > available RAM. And on PPC, we nearly always set the ram size in the > DT. I don't know why you think the default was no fixup; this is not > true. It's the other way round. No it is not. The fixup is in bootm.c for all of arm, yet appears in all the board/soc files for powerpc. It may be nearly always fixed up because every platform duplicates the fixup code, but it is not the default. It is a per platform decision on powerpc and it is not on arm. $ git grep fixup_memory arch/arm/lib/bootm.c:static int fixup_memory_node(void *blob) arch/arm/lib/bootm.c: return fdt_fixup_memory_banks(blob, start, size, CONFIG_NR_DRAM_BANKS); arch/arm/lib/bootm.c: fixup_memory_node(*of_flat_tree); arch/powerpc/cpu/74xx_7xx/cpu.c: fdt_fixup_memory(blob, (u64)bd->bi_memstart, (u64)bd->bi_memsize); arch/powerpc/cpu/mpc512x/cpu.c: fdt_fixup_memory(blob, (u64)bd->bi_memstart, (u64)bd->bi_memsize); arch/powerpc/cpu/mpc5xxx/cpu.c: fdt_fixup_memory(blob, (u64)bd->bi_memstart, (u64)bd->bi_memsize); arch/powerpc/cpu/mpc8260/cpu.c: fdt_fixup_memory(blob, (u64)bd->bi_memstart, (u64)bd->bi_memsize); arch/powerpc/cpu/mpc83xx/fdt.c: fdt_fixup_memory(blob, (u64)bd->bi_memstart, (u64)bd->bi_memsize); arch/powerpc/cpu/mpc85xx/fdt.c: fdt_fixup_memory(blob, (u64)bd->bi_memstart, (u64)bd->bi_memsize); arch/powerpc/cpu/mpc86xx/fdt.c: fdt_fixup_memory(blob, (u64)bd->bi_memstart, (u64)bd->bi_memsize); arch/powerpc/cpu/mpc8xx/fdt.c: fdt_fixup_memory(blob, (u64)bd->bi_memstart, (u64)bd->bi_memsize); arch/powerpc/cpu/ppc4xx/fdt.c: fdt_fixup_memory(blob, (u64)bd->bi_memstart, (u64)bd->bi_memsize); board/exmeritus/hww1u1a/hww1u1a.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/bsc9131rdb/bsc9131rdb.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/corenet_ds/corenet_ds.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/mpc7448hpc2/mpc7448hpc2.c: fdt_fixup_memory(blob, (u64)bd->bi_memstart, (u64)bd->bi_memsize); board/freescale/mpc8572ds/mpc8572ds.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/p1010rdb/p1010rdb.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/p1022ds/p1022ds.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/p1023rds/p1023rds.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/p1_p2_rdb/p1_p2_rdb.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/p1_p2_rdb_pc/p1_p2_rdb_pc.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/p2020come/p2020come.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/p2020ds/p2020ds.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/p2041rdb/p2041rdb.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/freescale/t4qds/t4qds.c: fdt_fixup_memory(blob, (u64)base, (u64)size); board/ipek01/ipek01.c: fdt_fixup_memory (blob, (u64) bd->bi_memstart, (u64) bd->bi_memsize); board/pdm360ng/pdm360ng.c: fdt_fixup_memory(blob, (u64)bd->bi_memstart, (u64)bd->bi_memsize); common/cmd_fdt.c: err = fdt_fixup_memory(working_fdt, addr, size); common/fdt_support.c:int fdt_fixup_memory_banks(void *blob, u64 start[], u64 size[], int banks) common/fdt_support.c:int fdt_fixup_memory(void *blob, u64 start, u64 size) common/fdt_support.c: return fdt_fixup_memory_banks(blob, &start, &size, 1); include/fdt_support.h:int fdt_fixup_memory(void *blob, u64 start, u64 size); include/fdt_support.h:int fdt_fixup_memory_banks(void *blob, u64 start[], u64 size[], int banks); >>> Actually you are reaching here a point where it seems necessary that >>> U-Boot itself is able to read the DT to configure itself correctly, >>> i. e. the memory size information it holds should then also be >>> extracted from the DT. >> >> I thought about doing this. That's really an orthogonal issue. The >> problem is you cannot adjust the amount of memory u-boot uses to be >> different than the amount of RAM in the /memory node. If you adjust the >> size in u-boot, the adjusted size is passed to the kernel no matter what. >> >> There's issues with LPAE systems having >4GB of RAM as all the size and >> address values are 32-bit. This could be fixed, but I'm not sure there >> is much value in u-boot knowing about memory above 4GB as it can't >> really access all of it and it would be purely informational. > > Well, let's face it - no matter what you do, if the RAM sizes as used > in U-Boot and in Linux differ, at least one of these is wrong. This > is a bug that should be fixed. What you attempt to do is implementing > a method to paper over such bugs. This is not a good thing to do. We'll probably just have to disagree that u-boot needs to know about memory it can't access on 32-bit PAE system unless you want to implement paging. Rob
Dear Rob, In message <20130219231428.44C8A200536@gemini.denx.de> I wrote: > > > There is no such way defined to flag that and I'm not going to invent > > one. It's not that the memory node is immutable, but the default should > > not be "needs fixups". No one puts purposely wrong data in their DT. So > > therefore all data should be immutable? What happens when we need to > > fixup immutable data? > > There is for example the "status" property, which can take the values > "okay" or "disabled". This is used in other places for not so > unsimilar purposes. Actually it's even much simpler than that. If you look for example at PPC, you will see many memory nodes in the DTB that look like this one (this example is from arch/powerpc/boot/dts/canyonlands.dts): memory { device_type = "memory"; reg = <0x00000000 0x00000000 0x00000000>; /* Filled in by U-Boot */ }; I. e. the DT provides only dummy values. How about checking for these - if the DT contains such dummy entries, U-Boot will insert valid data; if the DT already contains real data, these take preference? [And if the latter differ from what U-Boot thinks is correct, then a big fat warning should be issued!] Best regards, Wolfgang Denk
On Tue, Feb 19, 2013 at 02:39:45PM -0600, Rob Herring wrote: > On 02/19/2013 01:26 PM, Wolfgang Denk wrote: [snip] > > I dislike this. If you want to make a property of your device tree > > "immutable", then you should mark it there as such. Instead of > > removing the code here, U-Boot should then check for such a property > > and leave the value unchanged. > > There is no such way defined to flag that and I'm not going to invent > one. It's not that the memory node is immutable, but the default should > not be "needs fixups". No one puts purposely wrong data in their DT. So > therefore all data should be immutable? What happens when we need to > fixup immutable data? It's not so simple. The first DTs I checked are the am335x ones and I see they hardcode to 256MB which isn't right, but no one has caught this since it's blindly fixed up. So as Wolfgang suggests next, we might need to go with replacing on 0x0 entries and warning when they don't match what U-Boot has detected.
>>>>> "Tom" == Tom Rini <trini@ti.com> writes:
Hi,
Tom> It's not so simple. The first DTs I checked are the am335x ones
Tom> and I see they hardcode to 256MB which isn't right, but no one has
Tom> caught this since it's blindly fixed up. So as Wolfgang suggests
Tom> next, we might need to go with replacing on 0x0 entries and
Tom> warning when they don't match what U-Boot has detected.
It is correct for the (classic) beaglebone. This also hits people using
appended dtb unless they have CONFIG_ARM_ATAG_DTB_COMPAT enabled (which
the probably do though).
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 On 02/20/2013 04:55 PM, Peter Korsgaard wrote: >>>>>> "Tom" == Tom Rini <trini@ti.com> writes: > > Hi, > > Tom> It's not so simple. The first DTs I checked are the am335x > ones Tom> and I see they hardcode to 256MB which isn't right, but > no one has Tom> caught this since it's blindly fixed up. So as > Wolfgang suggests Tom> next, we might need to go with replacing on > 0x0 entries and Tom> warning when they don't match what U-Boot has > detected. > > It is correct for the (classic) beaglebone. This also hits people > using appended dtb unless they have CONFIG_ARM_ATAG_DTB_COMPAT > enabled (which the probably do though). So yes, there are some cases where 256MB is the right value. But we always know dynamically how much we have, so we don't need to hard-code even in that case. And once we do change the behaviour, the appeneded tree hack shouldn't be relevant anyhow :) - -- Tom -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.11 (GNU/Linux) Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/ iQIcBAEBAgAGBQJRJUqnAAoJENk4IS6UOR1WAmgQALOHeh+p9aRLT+Hb0ZjpJk0c XGJzCO1EpdNpQBsgPxI2AzhPEHsVHgcdzXmzljHS4ukPWDbqihM4j7o9OHjuxAGY +8Yebw/E/1F7+wKXy013/3QFbl9mdd1OpTU95xarnaoe4+dP5pfa+Q/kH5mrc0U2 xiA4kDmInityA/DUuEP4d8mhnLQyn11CKS/LPKdT40HUGDk6ffzkmUCYEHBpwk11 QN7w6icIGE9JrMsQUOPir7CNDZsNCbeAtB8+gpStHxL63cHqxAh9R3yFyUc2uVHo 5NNkc71R0qLIrKMTbDjIfm1soTEWP94ob6VVPQafqSC7wQrK5r+2DNg9I7dN7xIr js5Y/BBbPuUZfMWiNWiyF+ONuoNNrz1+iLm9iLdaJajf5EuBQN/qUzNGKFTCsLne nR6cR0x/Ow+rN1TDmnSvmS845XUDK2qLY5nSthKieuFtoDY88j6VpR4ZO7+B6sSa 9MVx6LV8SO4USewl1FIpXoe6KC2AjdSK2rJN5i2SSQD6uxYzPCfBKezG7gi8Mfua v4M9W/jDU/avXo861XLeKgcZx0zA7PHxDhqgZDVbFs3QH3eobzZHGog001vM6mvr dEdUGxu0dVF9ocbF/2F6XxdMVxt2CK7ddbGCZy5V1R6ZCvOVZetCbQao8GB7ZqLU LMZf4b3iDGdrOfM57n5m =8tud -----END PGP SIGNATURE-----
diff --git a/arch/arm/lib/board.c b/arch/arm/lib/board.c index cfe32cc..235f953 100644 --- a/arch/arm/lib/board.c +++ b/arch/arm/lib/board.c @@ -212,8 +212,10 @@ int print_cpuinfo(void); void __dram_init_banksize(void) { +#if CONFIG_NR_DRAM_BANKS gd->bd->bi_dram[0].start = CONFIG_SYS_SDRAM_BASE; gd->bd->bi_dram[0].size = gd->ram_size; +#endif } void dram_init_banksize(void) __attribute__((weak, alias("__dram_init_banksize"))); diff --git a/arch/arm/lib/bootm.c b/arch/arm/lib/bootm.c index 1bd2730..1f54217 100644 --- a/arch/arm/lib/bootm.c +++ b/arch/arm/lib/bootm.c @@ -79,6 +79,7 @@ void arch_lmb_reserve(struct lmb *lmb) #ifdef CONFIG_OF_LIBFDT static int fixup_memory_node(void *blob) { +#if CONFIG_NR_DRAM_BANKS bd_t *bd = gd->bd; int bank; u64 start[CONFIG_NR_DRAM_BANKS]; @@ -90,6 +91,9 @@ static int fixup_memory_node(void *blob) } return fdt_fixup_memory_banks(blob, start, size, CONFIG_NR_DRAM_BANKS); +#else + return 0; +#endif } #endif