On Wed, Jul 05, 2017 at 04:05:16PM +0300, Artturi Alm wrote:
> On Wed, Jul 05, 2017 at 11:27:06AM +0200, Mark Kettenis wrote:
> > > Date: Wed, 5 Jul 2017 09:34:59 +0300
> > > From: Artturi Alm <[email protected]>
> > >
> > > On Wed, Jul 05, 2017 at 02:27:46AM +0300, Artturi Alm wrote:
> > > > Hi,
> > > >
> > > > instead of messing w/bs_tags, use the fact pmap_kernel()->pm_refs is
> > > > going
> > > > to be 0 until pmap_bootstrap() has ran. tmp_bs_tag was unused, and
> > > > bootstrap_bs_map doesn't need/use the void *t-arg when being ran
> > > > indirectly
> > > > via armv7_bs_map().
> > > >
> > > > the whole existence of bootstrap_bs_map is another story, and the
> > > > comment in
> > > > /* Now, map the FDT area. */ is somewhat an stupid excuse, it's already
> > > > mapped
> > > > before initarm() w/VA=PA, and could well be _init()&_get_size()'d &
> > > > memcpy'ed
> > > > somewhere in reach within bootstrap KVA, guess diff might follow for
> > > > that,
> > > > too, if anyone has time for these simplifications.
> > > >
> > >
> > > Ok, i was wrong ^there, and the bootstrap code before initarm() didn't
> > > fill
> > > the L1 w/VA=PA anymore, for reasons i don't understand, so i 'fixed' it,
> > > with diff below. tested to boot and eeprom -p normally on cubie2 and
> > > wandb.
> > >
> > > i kept the diff minimal, to the point it does fdt_get_size() twice just
> > > like
> > > before, which i don't like, nor the name of size-variable and what not,
> > > but
> > > minimal it is. Would be the first step towards earlier physmem load :)
> > >
> > > -Artturi
> >
> > What are you trying to achieve heree?
> >
> > The current code quite deliberately does not create a cachable 1:1
> > mapping for the entire address space. Such a mapping is dangerous as
> > the CPU might speculatively load from any valid mapping and that is a
> > terrible idea for device mappings.
> >
>
> Point taken, and adapted the diff to map only 4mb at the expected fdt pa. So
> something like below, guess you read the one mail in this thread w/o diff
> in it, ofc. the aim is really higher, make arm/armv7 more consistent/
> readable/structured/cleaned/ all around, hoping it will make maintenance
> and future innovations easier or something, now stop worrying, i'm not
> NIH-patient about to design a new wheel or anything xD.
>
> Diff below is still rather raw, tested to boot and build a new kernel
> while running the diff correctly on sxi&imx, unfortunately the diff has a few
> unnecessary things in it, but the purpose of this is just to show the kind of
> things rather small reorganizing could bring.
>
> been up +24hrs, and might have had a few too long streches hacking w/o
> turning on the windows vm for a game or anything, so any stupid mistakes
> are because of that, i usually take a break atleast every 90mins or so:)
>
> And forgive the stupid ugly printf()s in _bs_valloc(), i forgot, and am
> already late from where i was supposed to be now, 'til later o/
> -Artturi
>
>
oh my, sorry to anyone if i wasted your time with the previous diff,
out of two branches, i chose the wrong one, and while mailing it,
i saw some extra cleanup, and the printfs, when i quickly visited
the end of the diff to make sure it was the 4mb one, for which i had
this diff on a separate non-cleanup branch, but thought this one had
just those that i saw and not much else, blind, even if my editor doesn't
color the diffs for me, was crazy to slip thru the renames.
won't happen again in the near future, i will try..
Correct diff below, the testing i said i had done, was with this
diff.
-Artturi
diff --git a/sys/arch/armv7/armv7/armv7_machdep.c
b/sys/arch/armv7/armv7/armv7_machdep.c
index aa1c549b29b..105fb3333f1 100644
--- a/sys/arch/armv7/armv7/armv7_machdep.c
+++ b/sys/arch/armv7/armv7/armv7_machdep.c
@@ -356,6 +356,30 @@ copy_io_area_map(pd_entry_t *new_pd)
}
}
+static inline paddr_t
+_bs_alloc(size_t sz)
+{
+ paddr_t addr, pa = 0;
+
+ for (sz = round_page(sz); sz > 0; sz -= PAGE_SIZE) {
+ if (uvm_page_physget(&addr) == FALSE)
+ panic("uvm_page_physget() failed");
+ memset((char *)addr, 0, PAGE_SIZE);
+ if (pa == 0)
+ pa = addr;
+ }
+ return pa;
+}
+
+/* RelativePA 2 KVA */
+#define _BS_RPA2KVA(x, y) (KERNEL_BASE + (x) - (y))
+static inline void
+_bs_valloc(pv_addr_t *pv, vsize_t sz, paddr_t off)
+{
+ pv->pv_pa = _bs_alloc(sz);
+ pv->pv_va = _BS_RPA2KVA(pv->pv_pa, off);
+}
+
/*
* u_int initarm(...)
*
@@ -379,7 +403,7 @@ initarm(void *arg0, void *arg1, void *arg2, paddr_t
loadaddr)
paddr_t memstart;
psize_t memsize;
paddr_t memend;
- void *config;
+ void *config = arg2;
size_t size;
void *node;
extern uint32_t esym; /* &_end if no symbols are loaded */
@@ -420,18 +444,8 @@ initarm(void *arg0, void *arg1, void *arg2, paddr_t
loadaddr)
tmp_bs_tag.bs_map = bootstrap_bs_map;
/*
- * Now, map the FDT area.
- *
- * As we don't know the size of a possible FDT, map the size of a
- * typical bootstrap bs map. The FDT might not be aligned, so this
- * might take up to two L1_S_SIZEd mappings.
- *
- * XXX: There's (currently) no way to unmap a bootstrap mapping, so
- * we might lose a bit of the bootstrap address space.
+ * Now, init the FDT @ PA, reloc and reinit to KVA later.
*/
- bootstrap_bs_map(NULL, (bus_addr_t)arg2, L1_S_SIZE, 0,
- (bus_space_handle_t *)&config);
-
if (!fdt_init(config) || fdt_get_size(config) == 0)
panic("initarm: no FDT");
@@ -477,6 +491,33 @@ initarm(void *arg0, void *arg1, void *arg2, paddr_t
loadaddr)
physmem = (physical_end - physical_start) / PAGE_SIZE;
+ /* Load memory into UVM. */
+#ifdef VERBOSE_INIT_ARM
+ printf("page ");
+#endif
+ uvm_setpagesize(); /* initialize PAGE_SIZE-dependent variables */
+ uvm_page_physload(atop(physical_freestart), atop(physical_freeend),
+ atop(physical_freestart), atop(physical_freeend), 0);
+
+ if (physical_start < loadaddr) {
+ uvm_page_physload(atop(physical_start), atop(loadaddr),
+ atop(physical_start), atop(loadaddr), 0);
+ physsegs--;
+ }
+
+ for (i = 1; i < physsegs; i++) {
+ if (fdt_get_reg(node, i, ®))
+ break;
+ if (reg.size == 0)
+ continue;
+
+ memstart = reg.addr;
+ memend = MIN(reg.addr + reg.size, (paddr_t)-PAGE_SIZE);
+ physmem += atop(memend - memstart);
+ uvm_page_physload(atop(memstart), atop(memend),
+ atop(memstart), atop(memend), 0);
+ }
+
#ifdef DEBUG
/* Tell the user about the memory */
printf("physmemory: %d pages at 0x%08lx -> 0x%08lx\n", physmem,
@@ -514,27 +555,27 @@ initarm(void *arg0, void *arg1, void *arg2, paddr_t
loadaddr)
/* Define a macro to simplify memory allocation */
#define valloc_pages(var, np) \
- alloc_pages((var).pv_pa, (np)); \
- (var).pv_va = KERNEL_BASE + (var).pv_pa - loadaddr;
+ _bs_valloc(&(var), ptoa((np)), loadaddr)
#define alloc_pages(var, np) \
- (var) = physical_freestart; \
- physical_freestart += ((np) * PAGE_SIZE); \
- if (physical_freeend < physical_freestart) \
- panic("initarm: out of memory"); \
- free_pages -= (np); \
- memset((char *)(var), 0, ((np) * PAGE_SIZE));
+ (var) = _bs_alloc(ptoa((np)))
loop1 = 0;
kernel_l1pt.pv_pa = 0;
+ physical_freestart = _bs_alloc(ptoa(NUM_KERNEL_PTS) + L1_TABLE_SIZE);
for (loop = 0; loop <= NUM_KERNEL_PTS; ++loop) {
/* Are we 16KB aligned for an L1 ? */
if (((physical_freestart) & (L1_TABLE_SIZE - 1)) == 0
&& kernel_l1pt.pv_pa == 0) {
- valloc_pages(kernel_l1pt, L1_TABLE_SIZE / PAGE_SIZE);
+ kernel_l1pt.pv_pa = physical_freestart;
+ kernel_l1pt.pv_va =
+ _BS_RPA2KVA(physical_freestart, loadaddr);
+ physical_freestart += L1_TABLE_SIZE;
} else {
- valloc_pages(kernel_pt_table[loop1],
- L2_TABLE_SIZE / PAGE_SIZE);
+ kernel_pt_table[loop1].pv_pa = physical_freestart;
+ kernel_pt_table[loop1].pv_va =
+ _BS_RPA2KVA(physical_freestart, loadaddr);
+ physical_freestart += L2_TABLE_SIZE;
++loop1;
}
}
@@ -572,11 +613,15 @@ initarm(void *arg0, void *arg1, void *arg2, paddr_t
loadaddr)
#endif
/*
- * Allocate pages for an FDT copy.
+ * Allocate pages for FDT, copy it there, and zero the original.
*/
size = fdt_get_size(config);
valloc_pages(fdt, round_page(size) / PAGE_SIZE);
memcpy((void *)fdt.pv_pa, config, size);
+ memset(config, 0, size);
+
+ /* Now we must reinit the FDT, using the virtual address. */
+ fdt_init((void *)fdt.pv_va);
/*
* XXX Defer this to later so that we can reclaim the memory
@@ -726,43 +771,12 @@ initarm(void *arg0, void *arg1, void *arg2, paddr_t
loadaddr)
prefetch_abort_handler_address = (u_int)prefetch_abort_handler;
undefined_handler_address = (u_int)undefinedinstruction_bounce;
- /* Now we can reinit the FDT, using the virtual address. */
- fdt_init((void *)fdt.pv_va);
-
/* Initialise the undefined instruction handlers */
#ifdef VERBOSE_INIT_ARM
printf("undefined ");
#endif
undefined_init();
- /* Load memory into UVM. */
-#ifdef VERBOSE_INIT_ARM
- printf("page ");
-#endif
- uvm_setpagesize(); /* initialize PAGE_SIZE-dependent variables */
- uvm_page_physload(atop(physical_freestart), atop(physical_freeend),
- atop(physical_freestart), atop(physical_freeend), 0);
-
- if (physical_start < loadaddr) {
- uvm_page_physload(atop(physical_start), atop(loadaddr),
- atop(physical_start), atop(loadaddr), 0);
- physsegs--;
- }
-
- node = fdt_find_node("/memory");
- for (i = 1; i < physsegs; i++) {
- if (fdt_get_reg(node, i, ®))
- break;
- if (reg.size == 0)
- continue;
-
- memstart = reg.addr;
- memend = MIN(reg.addr + reg.size, (paddr_t)-PAGE_SIZE);
- physmem += (memend - memstart) / PAGE_SIZE;
- uvm_page_physload(atop(memstart), atop(memend),
- atop(memstart), atop(memend), 0);
- }
-
/* Boot strap pmap telling it where the kernel page table is */
#ifdef VERBOSE_INIT_ARM
printf("pmap ");
diff --git a/sys/arch/armv7/armv7/locore0.S b/sys/arch/armv7/armv7/locore0.S
index 2a4e98cbe8c..f889dfeb204 100644
--- a/sys/arch/armv7/armv7/locore0.S
+++ b/sys/arch/armv7/armv7/locore0.S
@@ -136,6 +136,14 @@ _C_LABEL(bootstrap_start):
str r3, [r4, #8]
str r3, [r4, #0x14] // ram address for 0xc0000000
+ mov r1, r8, lsr #20
+ mov r2, r1, lsl #2
+ str r2, [r4, #(24 + 4)]
+ mov r1, r2, lsl #18
+ ldr r3, [r4, #(24 + 8)]
+ orr r3, r3, r1
+ str r3, [r4, #(24 + 8)]
+
/*
* the first entry has two fields that need to be updated for
* specific ram configuration of this board.
@@ -191,6 +199,9 @@ mmu_init_table:
/* map VA 0xc0000000..0xc3ffffff to PA */
MMU_INIT(0xc0000000, 0x00000000, 64,
L1_TYPE_S|L1_S_C|L1_S_V7_AP(AP_KRW)|L1_S_V7_AF)
+ /* map SDRAM VA==PA, WT cacheable @ pa = trunc_section(r8(=fdt)) */
+ MMU_INIT(0x00000000, 0x00000000, 4,
+ L1_TYPE_S|L1_S_C|L1_S_V7_AP(AP_KRW)|L1_S_V7_AF)
.word 0 /* end of table */