On Thu, Feb 01, 2024 at 01:10:18PM -0800, Oreoluwa Babatunde wrote: > > On 2/1/2024 11:46 AM, Rob Herring wrote: > > On Thu, Feb 01, 2024 at 09:08:06AM -0800, Oreoluwa Babatunde wrote: > >> On 1/30/2024 4:07 PM, Rob Herring wrote: > >>> On Fri, Jan 26, 2024 at 03:53:39PM -0800, Oreoluwa Babatunde wrote: > >>>> The reserved_mem array is used to store data for the different > >>>> reserved memory regions defined in the DT of a device. The array > >>>> stores information such as region name, node, start-address, and size > >>>> of the reserved memory regions. > >>>> > >>>> The array is currently statically allocated with a size of > >>>> MAX_RESERVED_REGIONS(64). This means that any system that specifies a > >>>> number of reserved memory regions greater than MAX_RESERVED_REGIONS(64) > >>>> will not have enough space to store the information for all the regions. > >>>> > >>>> Therefore, this series extends the use of the static array for > >>>> reserved_mem, and introduces a dynamically allocated array using > >>>> memblock_alloc() based on the number of reserved memory regions > >>>> specified in the DT. > >>>> > >>>> Some architectures such as arm64 require the page tables to be setup > >>>> before memblock allocated memory is writable. Therefore, the dynamic > >>>> allocation of the reserved_mem array will need to be done after the > >>>> page tables have been setup on these architectures. In most cases that > >>>> will be after paging_init(). > >>>> > >>>> Reserved memory regions can be divided into 2 groups. > >>>> i) Statically-placed reserved memory regions > >>>> i.e. regions defined in the DT using the @reg property. > >>>> ii) Dynamically-placed reserved memory regions. > >>>> i.e. regions specified in the DT using the @alloc_ranges > >>>> and @size properties. > >>>> > >>>> It is possible to call memblock_reserve() and memblock_mark_nomap() on > >>>> the statically-placed reserved memory regions and not need to save them > >>>> to the reserved_mem array until memory is allocated for it using > >>>> memblock, which will be after the page tables have been setup. > >>>> For the dynamically-placed reserved memory regions, it is not possible > >>>> to wait to store its information because the starting address is > >>>> allocated only at run time, and hence they need to be stored somewhere > >>>> after they are allocated. > >>>> Waiting until after the page tables have been setup to allocate memory > >>>> for the dynamically-placed regions is also not an option because the > >>>> allocations will come from memory that have already been added to the > >>>> page tables, which is not good for memory that is supposed to be > >>>> reserved and/or marked as nomap. > >>>> > >>>> Therefore, this series splits up the processing of the reserved memory > >>>> regions into two stages, of which the first stage is carried out by > >>>> early_init_fdt_scan_reserved_mem() and the second is carried out by > >>>> fdt_init_reserved_mem(). > >>>> > >>>> The early_init_fdt_scan_reserved_mem(), which is called before the page > >>>> tables are setup is used to: > >>>> 1. Call memblock_reserve() and memblock_mark_nomap() on all the > >>>> statically-placed reserved memory regions as needed. > >>>> 2. Allocate memory from memblock for the dynamically-placed reserved > >>>> memory regions and store them in the static array for reserved_mem. > >>>> memblock_reserve() and memblock_mark_nomap() are also called as > >>>> needed on all the memory allocated for the dynamically-placed > >>>> regions. > >>>> 3. Count the total number of reserved memory regions found in the DT. > >>>> > >>>> fdt_init_reserved_mem(), which should be called after the page tables > >>>> have been setup, is used to carry out the following: > >>>> 1. Allocate memory for the reserved_mem array based on the number of > >>>> reserved memory regions counted as mentioned above. > >>>> 2. Copy all the information for the dynamically-placed reserved memory > >>>> regions from the static array into the new allocated memory for the > >>>> reserved_mem array. > >>>> 3. Add the information for the statically-placed reserved memory into > >>>> reserved_mem array. > >>>> 4. Run the region specific init functions for each of the reserve memory > >>>> regions saved in the reserved_mem array. > >>> I don't see the need for fdt_init_reserved_mem() to be explicitly called > >>> by arch code. I said this already, but that can be done at the same time > >>> as unflattening the DT. The same conditions are needed for both: we need > >>> to be able to allocate memory from memblock. > >>> > >>> To put it another way, if fdt_init_reserved_mem() can be called "early", > >>> then unflattening could be moved earlier as well. Though I don't think > >>> we should optimize that. I'd rather see all arches call the DT functions > >>> at the same stages. > >> Hi Rob, > >> > >> The reason we moved fdt_init_reserved_mem() back into the arch specific code > >> was because we realized that there was no apparently obvious way to call > >> early_init_fdt_scan_reserved_mem() and fdt_init_reserved_mem() in the correct > >> order that will work for all archs if we placed fdt_init_reserved_mem() inside the > >> unflatten_devicetree() function. > >> > >> early_init_fdt_scan_reserved_mem() needs to be > >> called first before fdt_init_reserved_mem(). But on some archs, > >> unflatten_devicetree() is called before early_init_fdt_scan_reserved_mem(), which > >> means that if we have fdt_init_reserved_mem() inside the unflatten_devicetree() > >> function, it will be called before early_init_fdt_scan_reserved_mem(). > >> > >> This is connected to your other comments on Patch 7 & Patch 14. > >> I agree, unflatten_devicetree() should NOT be getting called before we reserve > >> memory for the reserved memory regions because that could cause memory to be > >> allocated from regions that should be reserved. > >> > >> Hence, resolving this issue should allow us to call fdt_init_reserved_mem() from > >> the unflatten_devicetree() function without it changing the order that we are > >> trying to have. > > There's one issue I've found which is unflatten_device_tree() isn't > > called for ACPI case on arm64. Turns out we need /reserved-memory > > handled in that case too. However, I think we're going to change > > calling unflatten_device_tree() unconditionally for another reason[1]. > > > > [1] https://lore.kernel.org/all/efe6a7886c3491cc9c225a903efa2b1e.sboyd@xxxxxxxxxx/ > > > >> I will work on implementing this and send another revision. > > I think we should go with a simpler route that's just copy the an > > initial array in initdata to a properly sized, allocated array like the > > patch below. Of course it will need some arch fixes and a follow-on > > patch to increase the initial array size. > > > > 8<-------------------------------------------------------------------- > > From: Rob Herring <robh@xxxxxxxxxx> > > Date: Wed, 31 Jan 2024 16:26:23 -0600 > > Subject: [PATCH] of: reserved-mem: Re-allocate reserved_mem array to actual > > size > > > > In preparation to increase the static reserved_mem array size yet again, > > copy the initial array to an allocated array sized based on the actual > > size needed. Now increasing the the size of the static reserved_mem > > array only eats up the initdata space. For platforms with reasonable > > number of reserved regions, we have a net gain in free memory. > > > > In order to do memblock allocations, fdt_init_reserved_mem() is moved a > > bit later to unflatten_device_tree(). On some arches this is effectively > > a nop. [...] > Hi Rob, > > One thing that could come up with this is that memory > for the dynamically-placed reserved memory regions > won't be allocated until we call fdt_init_reserved_mem(). > (i.e. reserved memory regions defined using @alloc-ranges > and @size properties) > > Since fdt_init_reserved_mem() is now being called from > unflatten_device_tree(), the page tables would have been > setup on most architectures, which means we will be > allocating from memory that have already been mapped. > > Could this be an issue for memory that is supposed to be > reserved? I suppose if the alloc-ranges region is not much bigger than the size and the kernel already made some allocation that landed in the region, then the allocation could fail. Not much we can do other than alloc the reserved regions as soon as possible. Are there cases where that's not happening? I suppose the kernel could try and avoid all alloc-ranges until they've been allocated, but that would have to be best effort. I've seen optimizations where it's desired to spread buffers across DRAM banks, so you could have N alloc-ranges for N banks that covers all of memory. There's also the issue that if you have more fixed regions than memblock can handle (128) before it can reallocate its arrays, then the page tables themselves could be allocated in reserved regions. > Especially for the regions that are specified as > no-map? 'no-map' is a hint, not a guarantee. Arm32 ignores it for regions within the kernel's linear map (at least it used to). I don't think anything changes here with it. Rob
