Cache: Fix a whole lot of things.

* Minor updates to the min & max cache sizes (16 & 64MB). Mostly to satisfy my power-of-two OCD.
  * Purge broken on-disk cache files
  * Optimize free RAM computations
  * Start dropping LRU items when running low on memory before pre-rendring (hinting) pages in non-reflowable documents.
  * Make serialize dump the most recently *displayed* page, as the actual MRU item is the most recently *hinted* page, not the current one.
  * Use more accurate item size estimations across the whole codebase.

TileCacheItem:

  * Drop lua-serialize in favor of Persist.

KoptInterface:

  * Drop lua-serialize in favor of Persist.
  * Make KOPTContext caching actually work by ensuring its hash is stable.

.github/ISSUE_TEMPLATE/bug_report.md

@@ -29,3 +29,12 @@ Android won't have a crash.log file because Google restricts what apps can log,
 Please try to include the relevant sections in your issue description.
 You can upload the whole `crash.log` file on GitHub by dragging and
 dropping it onto this textbox.
+
+If you instead opt to inline it, please do so behind a spoiler tag:
+<details>
+  <summary>crash.log</summary>
+  
+```
+<Paste crash.log content here>
+```
+</details>

defaults.lua

@@ -27,13 +27,13 @@ DHINTCOUNT = 1
 DRENDER_MODE = 0 -- 0 is COLOUR
 
 -- minimum cache size
-DGLOBAL_CACHE_SIZE_MINIMUM = 1024*1024*10
+DGLOBAL_CACHE_SIZE_MINIMUM = 1024*1024*16
 
 -- proportion of system free memory used as global cache
 DGLOBAL_CACHE_FREE_PROPORTION = 0.4
 
 -- maximum cache size
-DGLOBAL_CACHE_SIZE_MAXIMUM = 1024*1024*60
+DGLOBAL_CACHE_SIZE_MAXIMUM = 1024*1024*64
 
 -- background colour in non scroll mode: 8 = gray, 0 = white, 15 = black
 DBACKGROUND_COLOR = 0

frontend/apps/reader/modules/readerzooming.lua

@@ -458,9 +458,9 @@ function ReaderZooming:getZoom(pageno)
                          or self.zoom_factor
         zoom = zoom_w * zoom_factor
     end
-    if zoom and zoom > 10 and not Cache:willAccept(zoom * (self.dimen.w * self.dimen.h + 64)) then
+    if zoom and zoom > 10 and not Cache:willAccept(zoom * (self.dimen.w * self.dimen.h + 512)) then
         logger.dbg("zoom too large, adjusting")
-        while not Cache:willAccept(zoom * (self.dimen.w * self.dimen.h + 64)) do
+        while not Cache:willAccept(zoom * (self.dimen.w * self.dimen.h + 512)) do
             if zoom > 100 then
                 zoom = zoom - 50
             elseif zoom > 10 then

frontend/apps/reader/readerui.lua

@@ -733,13 +733,13 @@ function ReaderUI:onClose(full_refresh)
     if self.dialog ~= self then
         self:saveSettings()
     end
+    -- serialize last used items for later launch
+    Cache:serialize()
     if self.document ~= nil then
         logger.dbg("closing document")
         self:notifyCloseDocument()
     end
     UIManager:close(self.dialog, full_refresh and "full")
-    -- serialize last used items for later launch
-    Cache:serialize()
     if _running_instance == self then
         _running_instance = nil
     end

frontend/cache.lua

@@ -12,28 +12,99 @@ if CanvasContext.should_restrict_JIT then
     jit.off(true, true)
 end
 
+-- For documentation purposes, here's a battle-tested shell version of calcFreeMem
+--[[
+    if grep -q 'MemAvailable' /proc/meminfo ; then
+        # We'll settle for 85% of available memory to leave a bit of breathing room
+        tmpfs_size="$(awk '/MemAvailable/ {printf "%d", $2 * 0.85}' /proc/meminfo)"
+    elif grep -q 'Inactive(file)' /proc/meminfo ; then
+        # Basically try to emulate the kernel's computation, c.f., https://unix.stackexchange.com/q/261247
+        # Again, 85% of available memory
+        tmpfs_size="$(awk -v low=$(grep low /proc/zoneinfo | awk '{k+=$2}END{printf "%d", k}') \
+            '{a[$1]=$2}
+            END{
+                printf "%d", (a["MemFree:"]+a["Active(file):"]+a["Inactive(file):"]+a["SReclaimable:"]-(12*low))*0.85;
+            }' /proc/meminfo)"
+    else
+        # Ye olde crap workaround of Free + Buffers + Cache...
+        # Take it with a grain of salt, and settle for 80% of that...
+        tmpfs_size="$(awk \
+            '{a[$1]=$2}
+            END{
+                printf "%d", (a["MemFree:"]+a["Buffers:"]+a["Cached:"])*0.80;
+            }' /proc/meminfo)"
+    fi
+--]]
+
+-- And here's our simplified Lua version...
 local function calcFreeMem()
+    local memtotal, memfree, memavailable, buffers, cached
+
     local meminfo = io.open("/proc/meminfo", "r")
-    local freemem = 0
     if meminfo then
         for line in meminfo:lines() do
-            local free, buffer, cached, n
-            free, n = line:gsub("^MemFree:%s-(%d+) kB", "%1")
-            if n ~= 0 then freemem = freemem + tonumber(free)*1024 end
-            buffer, n = line:gsub("^Buffers:%s-(%d+) kB", "%1")
-            if n ~= 0 then freemem = freemem + tonumber(buffer)*1024 end
-            cached, n = line:gsub("^Cached:%s-(%d+) kB", "%1")
-            if n ~= 0 then freemem = freemem + tonumber(cached)*1024 end
+            if not memtotal then
+                memtotal = line:match("^MemTotal:%s-(%d+) kB")
+                if memtotal then
+                    -- Next!
+                    goto continue
+                end
+            end
+
+            if not memfree then
+                memfree = line:match("^MemFree:%s-(%d+) kB")
+                if memfree then
+                    -- Next!
+                    goto continue
+                end
+            end
+
+            if not memavailable then
+                memavailable = line:match("^MemAvailable:%s-(%d+) kB")
+                if memavailable then
+                    -- Best case scenario, we're done :)
+                    break
+                end
+            end
+
+            if not buffers then
+                buffers = line:match("^Buffers:%s-(%d+) kB")
+                if buffers then
+                    -- Next!
+                    goto continue
+                end
+            end
+
+            if not cached then
+                cached = line:match("^Cached:%s-(%d+) kB")
+                if cached then
+                    -- Ought to be the last entry we care about, we're done
+                    break
+                end
+            end
+
+            ::continue::
         end
         meminfo:close()
+    else
+        -- Not on Linux?
+        return 0, 0
+    end
+
+    if memavailable then
+        -- Leave a bit of margin, and report 85% of that...
+        return math.floor(memavailable * 0.85) * 1024, memtotal * 1024
+    else
+        -- Crappy Free + Buffers + Cache version, because the zoneinfo approach is a tad hairy...
+        -- So, leave an even larger margin, and only report 75% of that...
+        return math.floor((memfree + buffers + cached) * 0.75) * 1024, memtotal * 1024
     end
-    return freemem
 end
 
 local function calcCacheMemSize()
     local min = DGLOBAL_CACHE_SIZE_MINIMUM
     local max = DGLOBAL_CACHE_SIZE_MAXIMUM
-    local calc = calcFreeMem()*(DGLOBAL_CACHE_FREE_PROPORTION or 0)
+    local calc = calcFreeMem() * (DGLOBAL_CACHE_FREE_PROPORTION or 0)
     return math.min(max, math.max(min, calc))
 end
 
@@ -45,7 +116,7 @@ local cache_path = DataStorage:getDataDir() .. "/cache/"
 local function getDiskCache()
     local cached = {}
     for key_md5 in lfs.dir(cache_path) do
-        local file = cache_path..key_md5
+        local file = cache_path .. key_md5
         if lfs.attributes(file, "mode") == "file" then
             cached[key_md5] = file
         end
@@ -78,13 +149,13 @@ function Cache:_unref(key)
     for i = #self.cache_order, 1, -1 do
         if self.cache_order[i] == key then
             table.remove(self.cache_order, i)
+            break
         end
     end
 end
 
 -- internal: free cache item
 function Cache:_free(key)
-    if not self.cache[key] then return end
     self.current_memsize = self.current_memsize - self.cache[key].size
     self.cache[key]:onFree()
     self.cache[key] = nil
@@ -92,6 +163,8 @@ end
 
 -- drop an item named via key from the cache
 function Cache:drop(key)
+    if not self.cache[key] then return end
+
     self:_unref(key)
     self:_free(key)
 end
@@ -99,31 +172,37 @@ end
 function Cache:insert(key, object)
     -- make sure that one key only exists once: delete existing
     self:drop(key)
-    -- guarantee that we have enough memory in cache
-    if (object.size > self.max_memsize) then
-        logger.warn("too much memory claimed for", key)
+    -- If this object is single-handledly too large for the cache, we're done
+    if object.size > self.max_memsize then
+        logger.warn("Too much memory would be claimed by caching", key)
         return
     end
-    -- delete objects that least recently used
+    -- If inserting this obect would blow the cache's watermark,
+    -- start dropping least recently used items first.
     -- (they are at the end of the cache_order array)
     while self.current_memsize + object.size > self.max_memsize do
         local removed_key = table.remove(self.cache_order)
-        self:_free(removed_key)
+        if removed_key then
+            self:_free(removed_key)
+        else
+            logger.warn("Cache accounting is broken")
+            break
+        end
     end
-    -- insert new object in front of the LRU order
+    -- Insert new object in front of the LRU order
     table.insert(self.cache_order, 1, key)
     self.cache[key] = object
     self.current_memsize = self.current_memsize + object.size
 end
 
 --[[
---  check for cache item for key
+--  check for cache item by key
 --  if ItemClass is given, disk cache is also checked.
 --]]
 function Cache:check(key, ItemClass)
     if self.cache[key] then
         if self.cache_order[1] ~= key then
-            -- put key in front of the LRU list
+            -- Move key in front of the LRU list (i.e., MRU)
             self:_unref(key)
             table.insert(self.cache_order, 1, key)
         end
@@ -137,57 +216,77 @@ function Cache:check(key, ItemClass)
                 self:insert(key, item)
                 return item
             else
-                logger.warn("discard cache", msg)
+                logger.warn("Failed to load on-disk cache:", msg)
+                --- It's apparently unusable, purge it and refresh the snapshot.
+                os.remove(cached)
+                self:refreshSnapshot()
             end
         end
     end
 end
 
 function Cache:willAccept(size)
-    -- we only allow single objects to fill 75% of the cache
-    if size*4 < self.max_memsize*3 then
-        return true
-    end
+    -- We only allow single objects to fill 75% of the cache
+    return size*4 < self.max_memsize*3
 end
 
 function Cache:serialize()
-    -- calculate disk cache size
+    -- Calculate the current disk cache size
     local cached_size = 0
     local sorted_caches = {}
     for _, file in pairs(self.cached) do
         table.insert(sorted_caches, {file=file, time=lfs.attributes(file, "access")})
         cached_size = cached_size + (lfs.attributes(file, "size") or 0)
     end
-    table.sort(sorted_caches, function(v1,v2) return v1.time > v2.time end)
-    -- only serialize the most recently used cache
-    local cache_size = 0
+    table.sort(sorted_caches, function(v1, v2) return v1.time > v2.time end)
+
+    -- Only serialize the second most recently used cache item (as the MRU would be the *hinted* page).
+    local mru_key
+    local mru_found = 0
     for _, key in ipairs(self.cache_order) do
         local cache_item = self.cache[key]
 
-        -- only dump cache item that requests serialization explicitly
+        -- Only dump cache items that actually request persistence
         if cache_item.persistent and cache_item.dump then
-            local cache_full_path = cache_path..md5(key)
-            local cache_file_exists = lfs.attributes(cache_full_path)
-
-            if cache_file_exists then break end
+            mru_key = key
+            mru_found = mru_found + 1
+            if mru_found >= 2 then
+                -- We found the second MRU item, i.e., the *displayed* page
+                break
+            end
+        end
+    end
+    if mru_key then
+        local cache_full_path = cache_path .. md5(mru_key)
+        local cache_file_exists = lfs.attributes(cache_full_path)
 
-            logger.dbg("dump cache item", key)
-            cache_size = cache_item:dump(cache_full_path) or 0
-            if cache_size > 0 then break end
+        if not cache_file_exists then
+            logger.dbg("Dumping cache item", mru_key)
+            local cache_item = self.cache[mru_key]
+            local cache_size = cache_item:dump(cache_full_path)
+            if cache_size then
+                cached_size = cached_size + cache_size
+            end
         end
     end
-    -- set disk cache the same limit as memory cache
-    while cached_size + cache_size - self.max_memsize > 0 do
+
+    -- Allocate the same amount of storage to the disk cache than the memory cache
+    while cached_size > self.max_memsize do
         -- discard the least recently used cache
         local discarded = table.remove(sorted_caches)
-        cached_size = cached_size - lfs.attributes(discarded.file, "size")
-        os.remove(discarded.file)
+        if discarded then
+            cached_size = cached_size - lfs.attributes(discarded.file, "size")
+            os.remove(discarded.file)
+        else
+            logger.warn("Cache accounting is broken")
+            break
+        end
     end
-    -- disk cache may have changes so need to refresh disk cache snapshot
-    self.cached = getDiskCache()
+    -- We may have updated the disk cache's content, so refresh its state
+    self:refreshSnapshot()
 end
 
--- blank the cache
+-- Blank the cache
 function Cache:clear()
     for k, _ in pairs(self.cache) do
         self.cache[k]:onFree()
@@ -197,9 +296,41 @@ function Cache:clear()
     self.current_memsize = 0
 end
 
+-- Terribly crappy workaround: evict half the cache if we appear to be redlining on free RAM...
+function Cache:memoryPressureCheck()
+    local memfree, memtotal = calcFreeMem()
+
+    -- Nonsensical values? (!Linux), skip this.
+    if memtotal == 0 then
+        return
+    end
+
+    -- If less that 20% of the total RAM is free, drop half the Cache...
+    if memfree / memtotal < 0.20 then
+        logger.warn("Running low on memory, evicting half of the cache...")
+        for i = #self.cache_order / 2, 1, -1 do
+            local removed_key = table.remove(self.cache_order)
+            self:_free(removed_key)
+        end
+
+        -- And finish by forcing a GC sweep now...
+        collectgarbage()
+        collectgarbage()
+    end
+end
+
 -- Refresh the disk snapshot (mainly used by ui/data/onetime_migration)
 function Cache:refreshSnapshot()
     self.cached = getDiskCache()
 end
 
+-- Evict the disk cache (ditto)
+function Cache:clearDiskCache()
+    for _, file in pairs(self.cached) do
+        os.remove(file)
+    end
+
+    self:refreshSnapshot()
+end
+
 return Cache

frontend/cacheitem.lua

@@ -3,8 +3,11 @@ Inheritable abstraction for cache items
 --]]
 
 local CacheItem = {
-    size = 64, -- some reasonable default for simple Lua values / small tables
+    size = 128, -- some reasonable default for a small table.
 }
+--- NOTE: As far as size estimations go, the assumption is that a key, value pair should roughly take two words,
+---       and the most common items we cache are Geom-like tables (i.e., 4 key-value pairs).
+---       That's generally a low estimation, especially for larger tables, where memory allocation trickery may be happening.
 
 function CacheItem:new(o)
     o = o or {}

frontend/configurable.lua

@@ -1,3 +1,5 @@
+local ffiUtil = require("ffi/util")
+
 local Configurable = {}
 
 function Configurable:new(o)
@@ -8,7 +10,7 @@ function Configurable:new(o)
 end
 
 function Configurable:reset()
-    for key,value in pairs(self) do
+    for key, value in pairs(self) do
         local value_type = type(value)
         if value_type == "number" or value_type == "string" then
             self[key] = nil
@@ -18,7 +20,7 @@ end
 
 function Configurable:hash(sep)
     local hash = ""
-    for key,value in pairs(self) do
+    for key, value in ffiUtil.orderedPairs(self) do
         local value_type = type(value)
         if value_type == "number" or value_type == "string" then
             hash = hash..sep..value
@@ -31,7 +33,7 @@ function Configurable:loadDefaults(config_options)
     -- reset configurable before loading new options
     self:reset()
     local prefix = config_options.prefix.."_"
-    for i=1,#config_options do
+    for i=1, #config_options do
         local options = config_options[i].options
         for j=1,#options do
             local key = options[j].name
@@ -46,7 +48,7 @@ function Configurable:loadDefaults(config_options)
 end
 
 function Configurable:loadSettings(settings, prefix)
-    for key,value in pairs(self) do
+    for key, value in pairs(self) do
         local value_type = type(value)
         if value_type == "number" or value_type == "string"
             or value_type == "table" then
@@ -59,7 +61,7 @@ function Configurable:loadSettings(settings, prefix)
 end
 
 function Configurable:saveSettings(settings, prefix)
-    for key,value in pairs(self) do
+    for key, value in pairs(self) do
         local value_type = type(value)
         if value_type == "number" or value_type == "string"
             or value_type == "table" then

frontend/document/document.lua

@@ -93,9 +93,19 @@ end
 -- this might be overridden by a document implementation
 function Document:close()
     local DocumentRegistry = require("document/documentregistry")
-    if self.is_open and DocumentRegistry:closeDocument(self.file) == 0 then
-        self.is_open = false
-        self._document:close()
+    if self.is_open then
+        if DocumentRegistry:closeDocument(self.file) == 0 then
+            self.is_open = false
+            self._document:close()
+            self._document = nil
+
+            -- NOTE: DocumentRegistry:openDocument will force a GC sweep the next time we open a Document.
+            --       MµPDF will also do a bit of spring cleaning of its internal cache when opening a *different* document.
+        else
+            logger.warn("Tried to close a document with *multiple* remaining hot references")
+        end
+    else
+        logger.warn("Tried to close an already closed document")
     end
 end
 
@@ -375,7 +385,7 @@ function Document:renderPage(pageno, rect, zoom, rotation, gamma, render_mode)
     -- this will be the size we actually render
     local size = page_size
     -- we prefer to render the full page, if it fits into cache
-    if not Cache:willAccept(size.w * size.h + 64) then
+    if not Cache:willAccept(size.w * size.h * (self.render_color and 4 or 1) + 512) then
         -- whole page won't fit into cache
         logger.dbg("rendering only part of the page")
         --- @todo figure out how to better segment the page
@@ -392,11 +402,11 @@ function Document:renderPage(pageno, rect, zoom, rotation, gamma, render_mode)
     -- prepare cache item with contained blitbuffer
     tile = TileCacheItem:new{
         persistent = true,
-        size = size.w * size.h + 64, -- estimation
         excerpt = size,
         pageno = pageno,
         bb = Blitbuffer.new(size.w, size.h, self.render_color and self.color_bb_type or nil)
     }
+    tile.size = tonumber(tile.bb.stride) * tile.bb.h + 512 -- estimation
 
     -- create a draw context
     local dc = DrawContext.new()
@@ -429,6 +439,9 @@ end
 -- a hint for the cache engine to paint a full page to the cache
 --- @todo this should trigger a background operation
 function Document:hintPage(pageno, zoom, rotation, gamma, render_mode)
+    --- @note: Crappy safeguard around memory issues like in #7627: if we're eating too much RAM, drop half the cache...
+    Cache:memoryPressureCheck()
+
     logger.dbg("hinting page", pageno)
     self:renderPage(pageno, nil, zoom, rotation, gamma, render_mode)
 end

frontend/document/koptinterface.lua

@@ -10,9 +10,9 @@ local DEBUG = require("dbg")
 local Document = require("document/document")
 local Geom = require("ui/geometry")
 local KOPTContext = require("ffi/koptcontext")
+local Persist = require("persist")
 local TileCacheItem = require("document/tilecacheitem")
 local logger = require("logger")
-local serial = require("serialize")
 local util = require("ffi/util")
 
 local KoptInterface = {
@@ -36,16 +36,41 @@ end
 
 function ContextCacheItem:dump(filename)
     if self.kctx:isPreCache() == 0 then
-        logger.dbg("dumping koptcontext to", filename)
-        return serial.dump(self.size, KOPTContext.totable(self.kctx), filename)
+        logger.dbg("Dumping KOPTContext to", filename)
+
+        local cache_file = Persist:new{
+            path = filename,
+            codec = "zstd",
+        }
+
+        local t = KOPTContext.totable(self.kctx)
+        t.cache_size = self.size
+
+        local ok, size = cache_file:save(t)
+        if ok then
+            return size
+        else
+            logger.warn("Failed to dump KOPTContext")
+            return nil
+        end
     end
 end
 
 function ContextCacheItem:load(filename)
-    logger.dbg("loading koptcontext from", filename)
-    local size, kc_table = serial.load(filename)
-    self.size = size
-    self.kctx = KOPTContext.fromtable(kc_table)
+    logger.dbg("Loading KOPTContext from", filename)
+
+    local cache_file = Persist:new{
+        path = filename,
+        codec = "zstd",
+    }
+
+    local t = cache_file:load(filename)
+    if t then
+        self.size = t.cache_size
+        self.kctx = KOPTContext.fromtable(t)
+    else
+        logger.warn("Failed to load KOPTContext")
+    end
 end
 
 local OCREngine = CacheItem:new{}
@@ -154,7 +179,8 @@ Auto detect bbox.
 function KoptInterface:getAutoBBox(doc, pageno)
     local native_size = Document.getNativePageDimensions(doc, pageno)
     local bbox = {
-        x0 = 0, y0 = 0,
+        x0 = 0,
+        y0 = 0,
         x1 = native_size.w,
         y1 = native_size.h,
     }
@@ -172,7 +198,7 @@ function KoptInterface:getAutoBBox(doc, pageno)
         else
             bbox = Document.getPageBBox(doc, pageno)
         end
-        Cache:insert(hash, CacheItem:new{ autobbox = bbox })
+        Cache:insert(hash, CacheItem:new{ autobbox = bbox, size = 160 })
         page:close()
         kc:free()
         return bbox
@@ -207,7 +233,7 @@ function KoptInterface:getSemiAutoBBox(doc, pageno)
             auto_bbox = bbox
         end
         page:close()
-        Cache:insert(hash, CacheItem:new{ semiautobbox = auto_bbox })
+        Cache:insert(hash, CacheItem:new{ semiautobbox = auto_bbox, size = 160 })
         kc:free()
         return auto_bbox
     else
@@ -240,7 +266,7 @@ function KoptInterface:getCachedContext(doc, pageno)
         --self:logReflowDuration(pageno, dur)
         local fullwidth, fullheight = kc:getPageDim()
         logger.dbg("reflowed page", pageno, "fullwidth:", fullwidth, "fullheight:", fullheight)
-        self.last_context_size = fullwidth * fullheight + 128 -- estimation
+        self.last_context_size = fullwidth * fullheight + 3072 -- estimation
         Cache:insert(kctx_hash, ContextCacheItem:new{
             persistent = true,
             size = self.last_context_size,
@@ -251,7 +277,7 @@ function KoptInterface:getCachedContext(doc, pageno)
         -- wait for background thread
         local kc = self:waitForContext(cached.kctx)
         local fullwidth, fullheight = kc:getPageDim()
-        self.last_context_size = fullwidth * fullheight + 128 -- estimation
+        self.last_context_size = fullwidth * fullheight + 3072 -- estimation
         return kc
     end
 end
@@ -312,20 +338,20 @@ function KoptInterface:renderReflowedPage(doc, pageno, rect, zoom, rotation, ren
 
     local cached = Cache:check(renderpg_hash)
     if not cached then
-        -- do the real reflowing if kctx is not been cached yet
+        -- do the real reflowing if kctx has not been cached yet
         local kc = self:getCachedContext(doc, pageno)
         local fullwidth, fullheight = kc:getPageDim()
-        if not Cache:willAccept(fullwidth * fullheight / 2) then
+        if not Cache:willAccept(fullwidth * fullheight) then
             -- whole page won't fit into cache
             error("aborting, since we don't have enough cache for this page")
         end
         -- prepare cache item with contained blitbuffer
         local tile = TileCacheItem:new{
-            size = fullwidth * fullheight + 64, -- estimation
             excerpt = Geom:new{ w = fullwidth, h = fullheight },
             pageno = pageno,
         }
         tile.bb = kc:dstToBlitBuffer()
+        tile.size = tonumber(tile.bb.stride) * tile.bb.h + 512 -- estimation
         Cache:insert(renderpg_hash, tile)
         return tile
     else
@@ -363,7 +389,6 @@ function KoptInterface:renderOptimizedPage(doc, pageno, rect, zoom, rotation, re
         -- prepare cache item with contained blitbuffer
         local tile = TileCacheItem:new{
             persistent = true,
-            size = fullwidth * fullheight + 64, -- estimation
             excerpt = Geom:new{
                 x = 0, y = 0,
                 w = fullwidth,
@@ -372,6 +397,7 @@ function KoptInterface:renderOptimizedPage(doc, pageno, rect, zoom, rotation, re
             pageno = pageno,
         }
         tile.bb = kc:dstToBlitBuffer()
+        tile.size = tonumber(tile.bb.stride) * tile.bb.h + 512 -- estimation
         kc:free()
         Cache:insert(renderpg_hash, tile)
         return tile
@@ -478,8 +504,8 @@ function KoptInterface:getReflowedTextBoxes(doc, pageno)
             local kc = self:waitForContext(cached.kctx)
             --kc:setDebug()
             local fullwidth, fullheight = kc:getPageDim()
-            local boxes = kc:getReflowedWordBoxes("dst", 0, 0, fullwidth, fullheight)
-            Cache:insert(hash, CacheItem:new{ rfpgboxes = boxes })
+            local boxes, nr_word = kc:getReflowedWordBoxes("dst", 0, 0, fullwidth, fullheight)
+            Cache:insert(hash, CacheItem:new{ rfpgboxes = boxes, size = 192 * nr_word }) -- estimation
             return boxes
         end
     else
@@ -502,8 +528,8 @@ function KoptInterface:getNativeTextBoxes(doc, pageno)
             local kc = self:waitForContext(cached.kctx)
             --kc:setDebug()
             local fullwidth, fullheight = kc:getPageDim()
-            local boxes = kc:getNativeWordBoxes("dst", 0, 0, fullwidth, fullheight)
-            Cache:insert(hash, CacheItem:new{ nativepgboxes = boxes })
+            local boxes, nr_word = kc:getNativeWordBoxes("dst", 0, 0, fullwidth, fullheight)
+            Cache:insert(hash, CacheItem:new{ nativepgboxes = boxes, size = 192 * nr_word }) -- estimation
             return boxes
         end
     else
@@ -529,8 +555,8 @@ function KoptInterface:getReflowedTextBoxesFromScratch(doc, pageno)
             local fullwidth, fullheight = reflowed_kc:getPageDim()
             local kc = self:createContext(doc, pageno)
             kc:copyDestBMP(reflowed_kc)
-            local boxes = kc:getNativeWordBoxes("dst", 0, 0, fullwidth, fullheight)
-            Cache:insert(hash, CacheItem:new{ scratchrfpgboxes = boxes })
+            local boxes, nr_word = kc:getNativeWordBoxes("dst", 0, 0, fullwidth, fullheight)
+            Cache:insert(hash, CacheItem:new{ scratchrfpgboxes = boxes, size = 192 * nr_word }) -- estimation
             kc:free()
             return boxes
         end
@@ -575,8 +601,8 @@ function KoptInterface:getNativeTextBoxesFromScratch(doc, pageno)
         kc:setZoom(1.0)
         local page = doc._document:openPage(pageno)
         page:getPagePix(kc)
-        local boxes = kc:getNativeWordBoxes("src", 0, 0, page_size.w, page_size.h)
-        Cache:insert(hash, CacheItem:new{ scratchnativepgboxes = boxes })
+        local boxes, nr_word = kc:getNativeWordBoxes("src", 0, 0, page_size.w, page_size.h)
+        Cache:insert(hash, CacheItem:new{ scratchnativepgboxes = boxes, size = 192 * nr_word }) -- estimation
         page:close()
         kc:free()
         return boxes
@@ -607,7 +633,7 @@ function KoptInterface:getPageBlock(doc, pageno, x, y)
         local page = doc._document:openPage(pageno)
         page:getPagePix(kc)
         kc:findPageBlocks()
-        Cache:insert(hash, CacheItem:new{ kctx = kc })
+        Cache:insert(hash, CacheItem:new{ kctx = kc, size = 3072 }) -- estimation
         page:close()
         kctx = kc
     else
@@ -621,7 +647,7 @@ Get word from OCR providing selected word box.
 --]]
 function KoptInterface:getOCRWord(doc, pageno, wbox)
     if not Cache:check(self.ocrengine) then
-        Cache:insert(self.ocrengine, OCREngine:new{ ocrengine = KOPTContext.new() })
+        Cache:insert(self.ocrengine, OCREngine:new{ ocrengine = KOPTContext.new(), size = 3072 }) -- estimation
     end
     if doc.configurable.text_wrap == 1 then
         return self:getReflewOCRWord(doc, pageno, wbox.sbox)
@@ -648,7 +674,7 @@ function KoptInterface:getReflewOCRWord(doc, pageno, rect)
                 kc.getTOCRWord, kc, "dst",
                 rect.x, rect.y, rect.w, rect.h,
                 self.tessocr_data, self.ocr_lang, self.ocr_type, 0, 1)
-            Cache:insert(hash, CacheItem:new{ rfocrword = word })
+            Cache:insert(hash, CacheItem:new{ rfocrword = word, size = #word + 64 }) -- estimation
             return word
         end
     else
@@ -681,7 +707,7 @@ function KoptInterface:getNativeOCRWord(doc, pageno, rect)
             kc.getTOCRWord, kc, "src",
             0, 0, word_w, word_h,
             self.tessocr_data, self.ocr_lang, self.ocr_type, 0, 1)
-        Cache:insert(hash, CacheItem:new{ ocrword = word })
+        Cache:insert(hash, CacheItem:new{ ocrword = word, size = #word + 64 }) -- estimation
         logger.dbg("word", word)
         page:close()
         kc:free()
@@ -696,7 +722,7 @@ Get text from OCR providing selected text boxes.
 --]]
 function KoptInterface:getOCRText(doc, pageno, tboxes)
     if not Cache:check(self.ocrengine) then
-        Cache:insert(self.ocrengine, OCREngine:new{ ocrengine = KOPTContext.new() })
+        Cache:insert(self.ocrengine, OCREngine:new{ ocrengine = KOPTContext.new(), size = 3072 }) -- estimation
     end
     logger.info("Not implemented yet")
 end

frontend/document/pdfdocument.lua

@@ -152,9 +152,9 @@ function PdfDocument:getUsedBBox(pageno)
     if used.x1 > pwidth then used.x1 = pwidth end
     if used.y0 < 0 then used.y0 = 0 end
     if used.y1 > pheight then used.y1 = pheight end
-    --- @todo Give size for cacheitem?  02.12 2012 (houqp)
     Cache:insert(hash, CacheItem:new{
         ubbox = used,
+        size = 256, -- might be closer to 160
     })
     page:close()
     return used
@@ -170,6 +170,7 @@ function PdfDocument:getPageLinks(pageno)
     local links = page:getPageLinks()
     Cache:insert(hash, CacheItem:new{
         links = links,
+        size = 64 + (8 * 32 * #links),
     })
     page:close()
     return links

frontend/document/tilecacheitem.lua

@@ -1,6 +1,6 @@
 local Blitbuffer = require("ffi/blitbuffer")
 local CacheItem = require("cacheitem")
-local serial = require("serialize")
+local Persist = require("persist")
 local logger = require("logger")
 
 local TileCacheItem = CacheItem:new{}
@@ -12,19 +12,65 @@ function TileCacheItem:onFree()
     end
 end
 
+--- @note: Perhaps one day we'll be able to teach bitser or string.buffer about custom structs with pointers to buffers,
+---        so we won't have to do the BB tostring/fromstring dance anymore...
+function TileCacheItem:totable()
+    local t = {
+        size = self.size,
+        pageno = self.pageno,
+        excerpt = self.excerpt,
+        persistent = self.persistent,
+        bb = {
+            w = self.bb.w,
+            h = self.bb.h,
+            stride = tonumber(self.bb.stride),
+            fmt = self.bb:getType(),
+            data = Blitbuffer.tostring(self.bb),
+        },
+    }
+
+    return t
+end
+
 function TileCacheItem:dump(filename)
-    logger.dbg("dumping tile cache to", filename, self.excerpt)
-    return serial.dump(self.size, self.excerpt, self.pageno,
-            self.bb.w, self.bb.h, tonumber(self.bb.stride), self.bb:getType(),
-            Blitbuffer.tostring(self.bb), filename)
+    logger.dbg("Dumping tile cache to", filename, self.excerpt)
+
+    local cache_file = Persist:new{
+        path = filename,
+        codec = "zstd",
+    }
+
+    local ok, size = cache_file:save(self:totable())
+    if ok then
+        return size
+    else
+        logger.warn("Failed to dump tile cache")
+        return nil
+    end
+end
+
+function TileCacheItem:fromtable(t)
+    self.size = t.size
+    self.pageno = t.pageno
+    self.excerpt = t.excerpt
+    self.persistent = t.persistent
+    self.bb = Blitbuffer.fromstring(t.bb.w, t.bb.h, t.bb.fmt, t.bb.data, t.bb.stride)
 end
 
 function TileCacheItem:load(filename)
-    local w, h, stride, bb_type, bb_data
-    self.size, self.excerpt, self.pageno,
-            w, h, stride, bb_type, bb_data = serial.load(filename)
-    self.bb = Blitbuffer.fromstring(w, h, bb_type, bb_data, stride)
-    logger.dbg("loading tile cache from", filename, self)
+    local cache_file = Persist:new{
+        path = filename,
+        codec = "zstd",
+    }
+
+    local t = cache_file:load(filename)
+    if t then
+        self:fromtable(t)
+
+        logger.dbg("Loaded tile cache from", filename, self)
+    else
+        logger.warn("Failed to load tile cache from", filename)
+    end
 end
 
 return TileCacheItem

frontend/persist.lua

@@ -94,7 +94,8 @@ local codecs = {
             C.fclose(f)
             C.free(cbuff)
 
-            return true
+            --- @note: Slight API extension for TileCacheItem, which needs to know the on-disk size, and saves us a :size() call
+            return true, clen
         end,
 
         deserialize = function(path)
@@ -216,6 +217,9 @@ function Persist:save(t, as_bytecode)
         if not ok then
             return nil, err
         end
+
+        -- c.f., note above, err is the on-disk size
+        return true, err
     else
         local str, err = codecs[self.codec].serialize(t, as_bytecode)
         if not str then

frontend/ui/data/onetime_migration.lua

@@ -7,7 +7,7 @@ local lfs = require("libs/libkoreader-lfs")
 local logger = require("logger")
 
 -- Date at which the last migration snippet was added
-local CURRENT_MIGRATION_DATE = 20210414
+local CURRENT_MIGRATION_DATE = 20210503
 
 -- Retrieve the date of the previous migration, if any
 local last_migration_date = G_reader_settings:readSetting("last_migration_date", 0)
@@ -209,5 +209,13 @@ if last_migration_date < 20210414 then
     end
 end
 
+-- Cache, migration to Persist, https://github.com/koreader/koreader/pull/7624
+if last_migration_date < 20210503 then
+    logger.info("Performing one-time migration for 20210503")
+
+    local Cache = require("cache")
+    Cache:clearDiskCache()
+end
+
 -- We're done, store the current migration date
 G_reader_settings:saveSetting("last_migration_date", CURRENT_MIGRATION_DATE)

frontend/ui/rendertext.lua

@@ -113,7 +113,7 @@ function RenderText:getGlyph(face, charcode, bold)
         return
     end
     glyph = CacheItem:new{rendered_glyph}
-    glyph.size = glyph[1].bb:getWidth() * glyph[1].bb:getHeight() / 2 + 32
+    glyph.size = tonumber(glyph[1].bb.stride) * glyph[1].bb.h + 320
     GlyphCache:insert(hash, glyph)
     return rendered_glyph
 end
@@ -314,7 +314,7 @@ function RenderText:getGlyphByIndex(face, glyphindex, bold)
         return
     end
     glyph = CacheItem:new{rendered_glyph}
-    glyph.size = glyph[1].bb:getWidth() * glyph[1].bb:getHeight() / 2 + 32
+    glyph.size = tonumber(glyph[1].bb.stride) * glyph[1].bb.h + 320
     GlyphCache:insert(hash, glyph)
     return rendered_glyph
 end

frontend/ui/widget/imagewidget.lua

@@ -39,7 +39,7 @@ end
 local DPI_SCALE = get_dpi_scale()
 
 local ImageCache = Cache:new{
-    max_memsize = 5*1024*1024, -- 5M of image cache
+    max_memsize = 8*1024*1024, -- 8M of image cache
     current_memsize = 0,
     cache = {},
     -- this will hold the LRU order of the cache

plugins/opds.koplugin/opdsbrowser.lua

@@ -30,7 +30,7 @@ local CatalogCacheItem = CacheItem:new{
 
 -- cache catalog parsed from feed xml
 local CatalogCache = Cache:new{
-    max_memsize = 20*1024, -- keep only 20 cache items
+    max_memsize = 20*1024, -- keep only 20 items
     current_memsize = 0,
     cache = {},
     cache_order = {},